freebsd-skq/sys/vm/vm_pageout.c

1950 lines
56 KiB
C
Raw Normal View History

/*-
* SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
*
* Copyright (c) 1991 Regents of the University of California.
* All rights reserved.
* Copyright (c) 1994 John S. Dyson
* All rights reserved.
* Copyright (c) 1994 David Greenman
* All rights reserved.
* Copyright (c) 2005 Yahoo! Technologies Norway AS
* All rights reserved.
1994-05-24 10:09:53 +00:00
*
* This code is derived from software contributed to Berkeley by
* The Mach Operating System project at Carnegie-Mellon University.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. All advertising materials mentioning features or use of this software
* must display the following acknowledgement:
1994-05-24 10:09:53 +00:00
* This product includes software developed by the University of
* California, Berkeley and its contributors.
* 4. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
1994-08-02 07:55:43 +00:00
* from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
1994-05-24 10:09:53 +00:00
*
*
* Copyright (c) 1987, 1990 Carnegie-Mellon University.
* All rights reserved.
*
* Authors: Avadis Tevanian, Jr., Michael Wayne Young
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
*
1994-05-24 10:09:53 +00:00
* Permission to use, copy, modify and distribute this software and
* its documentation is hereby granted, provided that both the copyright
* notice and this permission notice appear in all copies of the
* software, derivative works or modified versions, and any portions
* thereof, and that both notices appear in supporting documentation.
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
*
* CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
* CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
1994-05-24 10:09:53 +00:00
* FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
*
1994-05-24 10:09:53 +00:00
* Carnegie Mellon requests users of this software to return to
*
* Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
* School of Computer Science
* Carnegie Mellon University
* Pittsburgh PA 15213-3890
*
* any improvements or extensions that they make and grant Carnegie the
* rights to redistribute these changes.
*/
/*
* The proverbial page-out daemon.
*/
2003-06-11 23:50:51 +00:00
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_vm.h"
1994-05-24 10:09:53 +00:00
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/eventhandler.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/kthread.h>
#include <sys/ktr.h>
#include <sys/mount.h>
#include <sys/racct.h>
#include <sys/resourcevar.h>
#include <sys/sched.h>
#include <sys/sdt.h>
1995-02-14 06:14:28 +00:00
#include <sys/signalvar.h>
#include <sys/smp.h>
#include <sys/time.h>
#include <sys/vnode.h>
#include <sys/vmmeter.h>
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
#include <sys/rwlock.h>
#include <sys/sx.h>
#include <sys/sysctl.h>
1994-05-24 10:09:53 +00:00
#include <vm/vm.h>
#include <vm/vm_param.h>
#include <vm/vm_object.h>
1994-05-24 10:09:53 +00:00
#include <vm/vm_page.h>
#include <vm/vm_map.h>
1994-05-24 10:09:53 +00:00
#include <vm/vm_pageout.h>
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct proc or any VM system structure will have to be rebuilt!!! Much needed overhaul of the VM system. Included in this first round of changes: 1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages, haspage, and sync operations are supported. The haspage interface now provides information about clusterability. All pager routines now take struct vm_object's instead of "pagers". 2) Improved data structures. In the previous paradigm, there is constant confusion caused by pagers being both a data structure ("allocate a pager") and a collection of routines. The idea of a pager structure has escentially been eliminated. Objects now have types, and this type is used to index the appropriate pager. In most cases, items in the pager structure were duplicated in the object data structure and thus were unnecessary. In the few cases that remained, a un_pager structure union was created in the object to contain these items. 3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now be removed. For instance, vm_object_enter(), vm_object_lookup(), vm_object_remove(), and the associated object hash list were some of the things that were removed. 4) simple_lock's removed. Discussion with several people reveals that the SMP locking primitives used in the VM system aren't likely the mechanism that we'll be adopting. Even if it were, the locking that was in the code was very inadequate and would have to be mostly re-done anyway. The locking in a uni-processor kernel was a no-op but went a long way toward making the code difficult to read and debug. 5) Places that attempted to kludge-up the fact that we don't have kernel thread support have been fixed to reflect the reality that we are really dealing with processes, not threads. The VM system didn't have complete thread support, so the comments and mis-named routines were just wrong. We now use tsleep and wakeup directly in the lock routines, for instance. 6) Where appropriate, the pagers have been improved, especially in the pager_alloc routines. Most of the pager_allocs have been rewritten and are now faster and easier to maintain. 7) The pagedaemon pageout clustering algorithm has been rewritten and now tries harder to output an even number of pages before and after the requested page. This is sort of the reverse of the ideal pagein algorithm and should provide better overall performance. 8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup have been removed. Some other unnecessary casts have also been removed. 9) Some almost useless debugging code removed. 10) Terminology of shadow objects vs. backing objects straightened out. The fact that the vm_object data structure escentially had this backwards really confused things. The use of "shadow" and "backing object" throughout the code is now internally consistent and correct in the Mach terminology. 11) Several minor bug fixes, including one in the vm daemon that caused 0 RSS objects to not get purged as intended. 12) A "default pager" has now been created which cleans up the transition of objects to the "swap" type. The previous checks throughout the code for swp->pg_data != NULL were really ugly. This change also provides the rudiments for future backing of "anonymous" memory by something other than the swap pager (via the vnode pager, for example), and it allows the decision about which of these pagers to use to be made dynamically (although will need some additional decision code to do this, of course). 13) (dyson) MAP_COPY has been deprecated and the corresponding "copy object" code has been removed. MAP_COPY was undocumented and non- standard. It was furthermore broken in several ways which caused its behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will continue to work correctly, but via the slightly different semantics of MAP_PRIVATE. 14) (dyson) Sharing maps have been removed. It's marginal usefulness in a threads design can be worked around in other ways. Both #12 and #13 were done to simplify the code and improve readability and maintain- ability. (As were most all of these changes) TODO: 1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing this will reduce the vnode pager to a mere fraction of its current size. 2) Rewrite vm_fault and the swap/vnode pagers to use the clustering information provided by the new haspage pager interface. This will substantially reduce the overhead by eliminating a large number of VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be improved to provide both a "behind" and "ahead" indication of contiguousness. 3) Implement the extended features of pager_haspage in swap_pager_haspage(). It currently just says 0 pages ahead/behind. 4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps via a much more general mechanism that could also be used for disk striping of regular filesystems. 5) Do something to improve the architecture of vm_object_collapse(). The fact that it makes calls into the swap pager and knows too much about how the swap pager operates really bothers me. It also doesn't allow for collapsing of non-swap pager objects ("unnamed" objects backed by other pagers).
1995-07-13 08:48:48 +00:00
#include <vm/vm_pager.h>
#include <vm/vm_phys.h>
#include <vm/swap_pager.h>
#include <vm/vm_extern.h>
#include <vm/uma.h>
1994-05-24 10:09:53 +00:00
/*
* System initialization
*/
/* the kernel process "vm_pageout"*/
2002-03-19 22:20:14 +00:00
static void vm_pageout(void);
static void vm_pageout_init(void);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
static int vm_pageout_clean(vm_page_t m, int *numpagedout);
static int vm_pageout_cluster(vm_page_t m);
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
static bool vm_pageout_scan(struct vm_domain *vmd, int pass);
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
int starting_page_shortage);
SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
NULL);
struct proc *pageproc;
static struct kproc_desc page_kp = {
"pagedaemon",
vm_pageout,
&pageproc
};
SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
&page_kp);
SDT_PROVIDER_DEFINE(vm);
SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/* Pagedaemon activity rates, in subdivisions of one second. */
#define VM_LAUNDER_RATE 10
#define VM_INACT_SCAN_RATE 2
int vm_pageout_deficit; /* Estimated number of pages deficit */
u_int vm_pageout_wakeup_thresh;
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
static int vm_pageout_oom_seq = 12;
bool vm_pageout_wanted; /* Event on which pageout daemon sleeps */
bool vm_pages_needed; /* Are threads waiting for free pages? */
1994-05-24 10:09:53 +00:00
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/* Pending request for dirty page laundering. */
static enum {
VM_LAUNDRY_IDLE,
VM_LAUNDRY_BACKGROUND,
VM_LAUNDRY_SHORTFALL
} vm_laundry_request = VM_LAUNDRY_IDLE;
static int vm_pageout_update_period;
2013-01-28 12:08:29 +00:00
static int disable_swap_pageouts;
static int lowmem_period = 10;
static time_t lowmem_uptime;
static int swapdev_enabled;
static int vm_panic_on_oom = 0;
SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
"panic on out of memory instead of killing the largest process");
SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
CTLFLAG_RWTUN, &vm_pageout_wakeup_thresh, 0,
"free page threshold for waking up the pageout daemon");
SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
"Maximum active LRU update period");
SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
"Low memory callback period");
SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
static int pageout_lock_miss;
SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
"back-to-back calls to oom detector to start OOM");
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
static int act_scan_laundry_weight = 3;
SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
&act_scan_laundry_weight, 0,
"weight given to clean vs. dirty pages in active queue scans");
static u_int vm_background_launder_target;
SYSCTL_UINT(_vm, OID_AUTO, background_launder_target, CTLFLAG_RWTUN,
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
&vm_background_launder_target, 0,
"background laundering target, in pages");
static u_int vm_background_launder_rate = 4096;
SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
&vm_background_launder_rate, 0,
"background laundering rate, in kilobytes per second");
static u_int vm_background_launder_max = 20 * 1024;
SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
&vm_background_launder_max, 0, "background laundering cap, in kilobytes");
int vm_pageout_page_count = 32;
1994-05-24 10:09:53 +00:00
int vm_page_max_wired; /* XXX max # of wired pages system-wide */
SYSCTL_INT(_vm, OID_AUTO, max_wired,
CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
1994-05-24 10:09:53 +00:00
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
static u_int isqrt(u_int num);
static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
static int vm_pageout_launder(struct vm_domain *vmd, int launder,
bool in_shortfall);
static void vm_pageout_laundry_worker(void *arg);
static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
/*
* Initialize a dummy page for marking the caller's place in the specified
* paging queue. In principle, this function only needs to set the flag
2016-08-05 16:44:11 +00:00
* PG_MARKER. Nonetheless, it write busies and initializes the hold count
* to one as safety precautions.
*/
static void
vm_pageout_init_marker(vm_page_t marker, u_short queue)
{
bzero(marker, sizeof(*marker));
marker->flags = PG_MARKER;
marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
marker->queue = queue;
marker->hold_count = 1;
}
/*
* vm_pageout_fallback_object_lock:
*
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
* Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
* known to have failed and page queue must be either PQ_ACTIVE or
* PQ_INACTIVE. To avoid lock order violation, unlock the page queue
* while locking the vm object. Use marker page to detect page queue
* changes and maintain notion of next page on page queue. Return
* TRUE if no changes were detected, FALSE otherwise. vm object is
* locked on return.
*
* This function depends on both the lock portion of struct vm_object
* and normal struct vm_page being type stable.
*/
static boolean_t
vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
{
struct vm_page marker;
struct vm_pagequeue *pq;
boolean_t unchanged;
u_short queue;
vm_object_t object;
queue = m->queue;
vm_pageout_init_marker(&marker, queue);
pq = vm_page_pagequeue(m);
object = m->object;
TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
vm_pagequeue_unlock(pq);
vm_page_unlock(m);
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
VM_OBJECT_WLOCK(object);
vm_page_lock(m);
vm_pagequeue_lock(pq);
/*
* The page's object might have changed, and/or the page might
* have moved from its original position in the queue. If the
* page's object has changed, then the caller should abandon
* processing the page because the wrong object lock was
* acquired. Use the marker's plinks.q, not the page's, to
* determine if the page has been moved. The state of the
* page's plinks.q can be indeterminate; whereas, the marker's
* plinks.q must be valid.
*/
*next = TAILQ_NEXT(&marker, plinks.q);
unchanged = m->object == object &&
m == TAILQ_PREV(&marker, pglist, plinks.q);
KASSERT(!unchanged || m->queue == queue,
("page %p queue %d %d", m, queue, m->queue));
TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
return (unchanged);
}
/*
* Lock the page while holding the page queue lock. Use marker page
* to detect page queue changes and maintain notion of next page on
* page queue. Return TRUE if no changes were detected, FALSE
* otherwise. The page is locked on return. The page queue lock might
* be dropped and reacquired.
*
* This function depends on normal struct vm_page being type stable.
*/
static boolean_t
vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
{
struct vm_page marker;
struct vm_pagequeue *pq;
boolean_t unchanged;
u_short queue;
vm_page_lock_assert(m, MA_NOTOWNED);
if (vm_page_trylock(m))
return (TRUE);
queue = m->queue;
vm_pageout_init_marker(&marker, queue);
pq = vm_page_pagequeue(m);
TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
vm_pagequeue_unlock(pq);
vm_page_lock(m);
vm_pagequeue_lock(pq);
/* Page queue might have changed. */
*next = TAILQ_NEXT(&marker, plinks.q);
unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
KASSERT(!unchanged || m->queue == queue,
("page %p queue %d %d", m, queue, m->queue));
TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
return (unchanged);
}
1994-05-24 10:09:53 +00:00
/*
* Scan for pages at adjacent offsets within the given page's object that are
* eligible for laundering, form a cluster of these pages and the given page,
* and launder that cluster.
1994-05-24 10:09:53 +00:00
*/
static int
vm_pageout_cluster(vm_page_t m)
1994-05-24 10:09:53 +00:00
{
2001-07-04 19:00:13 +00:00
vm_object_t object;
vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
vm_pindex_t pindex;
int ib, is, page_base, pageout_count;
vm_page_assert_locked(m);
object = m->object;
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
VM_OBJECT_ASSERT_WLOCKED(object);
pindex = m->pindex;
/*
* We can't clean the page if it is busy or held.
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct proc or any VM system structure will have to be rebuilt!!! Much needed overhaul of the VM system. Included in this first round of changes: 1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages, haspage, and sync operations are supported. The haspage interface now provides information about clusterability. All pager routines now take struct vm_object's instead of "pagers". 2) Improved data structures. In the previous paradigm, there is constant confusion caused by pagers being both a data structure ("allocate a pager") and a collection of routines. The idea of a pager structure has escentially been eliminated. Objects now have types, and this type is used to index the appropriate pager. In most cases, items in the pager structure were duplicated in the object data structure and thus were unnecessary. In the few cases that remained, a un_pager structure union was created in the object to contain these items. 3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now be removed. For instance, vm_object_enter(), vm_object_lookup(), vm_object_remove(), and the associated object hash list were some of the things that were removed. 4) simple_lock's removed. Discussion with several people reveals that the SMP locking primitives used in the VM system aren't likely the mechanism that we'll be adopting. Even if it were, the locking that was in the code was very inadequate and would have to be mostly re-done anyway. The locking in a uni-processor kernel was a no-op but went a long way toward making the code difficult to read and debug. 5) Places that attempted to kludge-up the fact that we don't have kernel thread support have been fixed to reflect the reality that we are really dealing with processes, not threads. The VM system didn't have complete thread support, so the comments and mis-named routines were just wrong. We now use tsleep and wakeup directly in the lock routines, for instance. 6) Where appropriate, the pagers have been improved, especially in the pager_alloc routines. Most of the pager_allocs have been rewritten and are now faster and easier to maintain. 7) The pagedaemon pageout clustering algorithm has been rewritten and now tries harder to output an even number of pages before and after the requested page. This is sort of the reverse of the ideal pagein algorithm and should provide better overall performance. 8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup have been removed. Some other unnecessary casts have also been removed. 9) Some almost useless debugging code removed. 10) Terminology of shadow objects vs. backing objects straightened out. The fact that the vm_object data structure escentially had this backwards really confused things. The use of "shadow" and "backing object" throughout the code is now internally consistent and correct in the Mach terminology. 11) Several minor bug fixes, including one in the vm daemon that caused 0 RSS objects to not get purged as intended. 12) A "default pager" has now been created which cleans up the transition of objects to the "swap" type. The previous checks throughout the code for swp->pg_data != NULL were really ugly. This change also provides the rudiments for future backing of "anonymous" memory by something other than the swap pager (via the vnode pager, for example), and it allows the decision about which of these pagers to use to be made dynamically (although will need some additional decision code to do this, of course). 13) (dyson) MAP_COPY has been deprecated and the corresponding "copy object" code has been removed. MAP_COPY was undocumented and non- standard. It was furthermore broken in several ways which caused its behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will continue to work correctly, but via the slightly different semantics of MAP_PRIVATE. 14) (dyson) Sharing maps have been removed. It's marginal usefulness in a threads design can be worked around in other ways. Both #12 and #13 were done to simplify the code and improve readability and maintain- ability. (As were most all of these changes) TODO: 1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing this will reduce the vnode pager to a mere fraction of its current size. 2) Rewrite vm_fault and the swap/vnode pagers to use the clustering information provided by the new haspage pager interface. This will substantially reduce the overhead by eliminating a large number of VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be improved to provide both a "behind" and "ahead" indication of contiguousness. 3) Implement the extended features of pager_haspage in swap_pager_haspage(). It currently just says 0 pages ahead/behind. 4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps via a much more general mechanism that could also be used for disk striping of regular filesystems. 5) Do something to improve the architecture of vm_object_collapse(). The fact that it makes calls into the swap pager and knows too much about how the swap pager operates really bothers me. It also doesn't allow for collapsing of non-swap pager objects ("unnamed" objects backed by other pagers).
1995-07-13 08:48:48 +00:00
*/
vm_page_assert_unbusied(m);
KASSERT(m->hold_count == 0, ("page %p is held", m));
pmap_remove_write(m);
vm_page_unlock(m);
1994-05-24 10:09:53 +00:00
mc[vm_pageout_page_count] = pb = ps = m;
pageout_count = 1;
page_base = vm_pageout_page_count;
ib = 1;
is = 1;
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct proc or any VM system structure will have to be rebuilt!!! Much needed overhaul of the VM system. Included in this first round of changes: 1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages, haspage, and sync operations are supported. The haspage interface now provides information about clusterability. All pager routines now take struct vm_object's instead of "pagers". 2) Improved data structures. In the previous paradigm, there is constant confusion caused by pagers being both a data structure ("allocate a pager") and a collection of routines. The idea of a pager structure has escentially been eliminated. Objects now have types, and this type is used to index the appropriate pager. In most cases, items in the pager structure were duplicated in the object data structure and thus were unnecessary. In the few cases that remained, a un_pager structure union was created in the object to contain these items. 3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now be removed. For instance, vm_object_enter(), vm_object_lookup(), vm_object_remove(), and the associated object hash list were some of the things that were removed. 4) simple_lock's removed. Discussion with several people reveals that the SMP locking primitives used in the VM system aren't likely the mechanism that we'll be adopting. Even if it were, the locking that was in the code was very inadequate and would have to be mostly re-done anyway. The locking in a uni-processor kernel was a no-op but went a long way toward making the code difficult to read and debug. 5) Places that attempted to kludge-up the fact that we don't have kernel thread support have been fixed to reflect the reality that we are really dealing with processes, not threads. The VM system didn't have complete thread support, so the comments and mis-named routines were just wrong. We now use tsleep and wakeup directly in the lock routines, for instance. 6) Where appropriate, the pagers have been improved, especially in the pager_alloc routines. Most of the pager_allocs have been rewritten and are now faster and easier to maintain. 7) The pagedaemon pageout clustering algorithm has been rewritten and now tries harder to output an even number of pages before and after the requested page. This is sort of the reverse of the ideal pagein algorithm and should provide better overall performance. 8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup have been removed. Some other unnecessary casts have also been removed. 9) Some almost useless debugging code removed. 10) Terminology of shadow objects vs. backing objects straightened out. The fact that the vm_object data structure escentially had this backwards really confused things. The use of "shadow" and "backing object" throughout the code is now internally consistent and correct in the Mach terminology. 11) Several minor bug fixes, including one in the vm daemon that caused 0 RSS objects to not get purged as intended. 12) A "default pager" has now been created which cleans up the transition of objects to the "swap" type. The previous checks throughout the code for swp->pg_data != NULL were really ugly. This change also provides the rudiments for future backing of "anonymous" memory by something other than the swap pager (via the vnode pager, for example), and it allows the decision about which of these pagers to use to be made dynamically (although will need some additional decision code to do this, of course). 13) (dyson) MAP_COPY has been deprecated and the corresponding "copy object" code has been removed. MAP_COPY was undocumented and non- standard. It was furthermore broken in several ways which caused its behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will continue to work correctly, but via the slightly different semantics of MAP_PRIVATE. 14) (dyson) Sharing maps have been removed. It's marginal usefulness in a threads design can be worked around in other ways. Both #12 and #13 were done to simplify the code and improve readability and maintain- ability. (As were most all of these changes) TODO: 1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing this will reduce the vnode pager to a mere fraction of its current size. 2) Rewrite vm_fault and the swap/vnode pagers to use the clustering information provided by the new haspage pager interface. This will substantially reduce the overhead by eliminating a large number of VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be improved to provide both a "behind" and "ahead" indication of contiguousness. 3) Implement the extended features of pager_haspage in swap_pager_haspage(). It currently just says 0 pages ahead/behind. 4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps via a much more general mechanism that could also be used for disk striping of regular filesystems. 5) Do something to improve the architecture of vm_object_collapse(). The fact that it makes calls into the swap pager and knows too much about how the swap pager operates really bothers me. It also doesn't allow for collapsing of non-swap pager objects ("unnamed" objects backed by other pagers).
1995-07-13 08:48:48 +00:00
/*
* We can cluster only if the page is not clean, busy, or held, and
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* the page is in the laundry queue.
*
* During heavy mmap/modification loads the pageout
* daemon can really fragment the underlying file
* due to flushing pages out of order and not trying to
* align the clusters (which leaves sporadic out-of-order
* holes). To solve this problem we do the reverse scan
* first and attempt to align our cluster, then do a
* forward scan if room remains.
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct proc or any VM system structure will have to be rebuilt!!! Much needed overhaul of the VM system. Included in this first round of changes: 1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages, haspage, and sync operations are supported. The haspage interface now provides information about clusterability. All pager routines now take struct vm_object's instead of "pagers". 2) Improved data structures. In the previous paradigm, there is constant confusion caused by pagers being both a data structure ("allocate a pager") and a collection of routines. The idea of a pager structure has escentially been eliminated. Objects now have types, and this type is used to index the appropriate pager. In most cases, items in the pager structure were duplicated in the object data structure and thus were unnecessary. In the few cases that remained, a un_pager structure union was created in the object to contain these items. 3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now be removed. For instance, vm_object_enter(), vm_object_lookup(), vm_object_remove(), and the associated object hash list were some of the things that were removed. 4) simple_lock's removed. Discussion with several people reveals that the SMP locking primitives used in the VM system aren't likely the mechanism that we'll be adopting. Even if it were, the locking that was in the code was very inadequate and would have to be mostly re-done anyway. The locking in a uni-processor kernel was a no-op but went a long way toward making the code difficult to read and debug. 5) Places that attempted to kludge-up the fact that we don't have kernel thread support have been fixed to reflect the reality that we are really dealing with processes, not threads. The VM system didn't have complete thread support, so the comments and mis-named routines were just wrong. We now use tsleep and wakeup directly in the lock routines, for instance. 6) Where appropriate, the pagers have been improved, especially in the pager_alloc routines. Most of the pager_allocs have been rewritten and are now faster and easier to maintain. 7) The pagedaemon pageout clustering algorithm has been rewritten and now tries harder to output an even number of pages before and after the requested page. This is sort of the reverse of the ideal pagein algorithm and should provide better overall performance. 8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup have been removed. Some other unnecessary casts have also been removed. 9) Some almost useless debugging code removed. 10) Terminology of shadow objects vs. backing objects straightened out. The fact that the vm_object data structure escentially had this backwards really confused things. The use of "shadow" and "backing object" throughout the code is now internally consistent and correct in the Mach terminology. 11) Several minor bug fixes, including one in the vm daemon that caused 0 RSS objects to not get purged as intended. 12) A "default pager" has now been created which cleans up the transition of objects to the "swap" type. The previous checks throughout the code for swp->pg_data != NULL were really ugly. This change also provides the rudiments for future backing of "anonymous" memory by something other than the swap pager (via the vnode pager, for example), and it allows the decision about which of these pagers to use to be made dynamically (although will need some additional decision code to do this, of course). 13) (dyson) MAP_COPY has been deprecated and the corresponding "copy object" code has been removed. MAP_COPY was undocumented and non- standard. It was furthermore broken in several ways which caused its behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will continue to work correctly, but via the slightly different semantics of MAP_PRIVATE. 14) (dyson) Sharing maps have been removed. It's marginal usefulness in a threads design can be worked around in other ways. Both #12 and #13 were done to simplify the code and improve readability and maintain- ability. (As were most all of these changes) TODO: 1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing this will reduce the vnode pager to a mere fraction of its current size. 2) Rewrite vm_fault and the swap/vnode pagers to use the clustering information provided by the new haspage pager interface. This will substantially reduce the overhead by eliminating a large number of VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be improved to provide both a "behind" and "ahead" indication of contiguousness. 3) Implement the extended features of pager_haspage in swap_pager_haspage(). It currently just says 0 pages ahead/behind. 4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps via a much more general mechanism that could also be used for disk striping of regular filesystems. 5) Do something to improve the architecture of vm_object_collapse(). The fact that it makes calls into the swap pager and knows too much about how the swap pager operates really bothers me. It also doesn't allow for collapsing of non-swap pager objects ("unnamed" objects backed by other pagers).
1995-07-13 08:48:48 +00:00
*/
more:
while (ib != 0 && pageout_count < vm_pageout_page_count) {
if (ib > pindex) {
ib = 0;
break;
}
if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
ib = 0;
break;
}
vm_page_test_dirty(p);
if (p->dirty == 0) {
ib = 0;
break;
}
vm_page_lock(p);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
if (!vm_page_in_laundry(p) ||
p->hold_count != 0) { /* may be undergoing I/O */
vm_page_unlock(p);
ib = 0;
break;
}
pmap_remove_write(p);
vm_page_unlock(p);
mc[--page_base] = pb = p;
++pageout_count;
++ib;
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
/*
* We are at an alignment boundary. Stop here, and switch
* directions. Do not clear ib.
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
*/
if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
break;
}
while (pageout_count < vm_pageout_page_count &&
pindex + is < object->size) {
if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
break;
vm_page_test_dirty(p);
if (p->dirty == 0)
break;
vm_page_lock(p);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
if (!vm_page_in_laundry(p) ||
p->hold_count != 0) { /* may be undergoing I/O */
vm_page_unlock(p);
break;
}
pmap_remove_write(p);
vm_page_unlock(p);
mc[page_base + pageout_count] = ps = p;
++pageout_count;
++is;
1994-05-24 10:09:53 +00:00
}
/*
* If we exhausted our forward scan, continue with the reverse scan
* when possible, even past an alignment boundary. This catches
* boundary conditions.
*/
if (ib != 0 && pageout_count < vm_pageout_page_count)
goto more;
return (vm_pageout_flush(&mc[page_base], pageout_count,
VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
}
/*
* vm_pageout_flush() - launder the given pages
*
* The given pages are laundered. Note that we setup for the start of
* I/O ( i.e. busy the page ), mark it read-only, and bump the object
* reference count all in here rather then in the parent. If we want
* the parent to do more sophisticated things we may have to change
* the ordering.
*
* Returned runlen is the count of pages between mreq and first
* page after mreq with status VM_PAGER_AGAIN.
* *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
* for any page in runlen set.
*/
int
vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
boolean_t *eio)
{
vm_object_t object = mc[0]->object;
int pageout_status[count];
int numpagedout = 0;
int i, runlen;
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
VM_OBJECT_ASSERT_WLOCKED(object);
/*
* Initiate I/O. Mark the pages busy and verify that they're valid
* and read-only.
*
* We do not have to fixup the clean/dirty bits here... we can
* allow the pager to do it after the I/O completes.
*
* NOTE! mc[i]->dirty may be partial or fragmented due to an
* edge case with file fragments.
*/
This mega-commit is meant to fix numerous interrelated problems. There has been some bitrot and incorrect assumptions in the vfs_bio code. These problems have manifest themselves worse on NFS type filesystems, but can still affect local filesystems under certain circumstances. Most of the problems have involved mmap consistancy, and as a side-effect broke the vfs.ioopt code. This code might have been committed seperately, but almost everything is interrelated. 1) Allow (pmap_object_init_pt) prefaulting of buffer-busy pages that are fully valid. 2) Rather than deactivating erroneously read initial (header) pages in kern_exec, we now free them. 3) Fix the rundown of non-VMIO buffers that are in an inconsistent (missing vp) state. 4) Fix the disassociation of pages from buffers in brelse. The previous code had rotted and was faulty in a couple of important circumstances. 5) Remove a gratuitious buffer wakeup in vfs_vmio_release. 6) Remove a crufty and currently unused cluster mechanism for VBLK files in vfs_bio_awrite. When the code is functional, I'll add back a cleaner version. 7) The page busy count wakeups assocated with the buffer cache usage were incorrectly cleaned up in a previous commit by me. Revert to the original, correct version, but with a cleaner implementation. 8) The cluster read code now tries to keep data associated with buffers more aggressively (without breaking the heuristics) when it is presumed that the read data (buffers) will be soon needed. 9) Change to filesystem lockmgr locks so that they use LK_NOPAUSE. The delay loop waiting is not useful for filesystem locks, due to the length of the time intervals. 10) Correct and clean-up spec_getpages. 11) Implement a fully functional nfs_getpages, nfs_putpages. 12) Fix nfs_write so that modifications are coherent with the NFS data on the server disk (at least as well as NFS seems to allow.) 13) Properly support MS_INVALIDATE on NFS. 14) Properly pass down MS_INVALIDATE to lower levels of the VM code from vm_map_clean. 15) Better support the notion of pages being busy but valid, so that fewer in-transit waits occur. (use p->busy more for pageouts instead of PG_BUSY.) Since the page is fully valid, it is still usable for reads. 16) It is possible (in error) for cached pages to be busy. Make the page allocation code handle that case correctly. (It should probably be a printf or panic, but I want the system to handle coding errors robustly. I'll probably add a printf.) 17) Correct the design and usage of vm_page_sleep. It didn't handle consistancy problems very well, so make the design a little less lofty. After vm_page_sleep, if it ever blocked, it is still important to relookup the page (if the object generation count changed), and verify it's status (always.) 18) In vm_pageout.c, vm_pageout_clean had rotted, so clean that up. 19) Push the page busy for writes and VM_PROT_READ into vm_pageout_flush. 20) Fix vm_pager_put_pages and it's descendents to support an int flag instead of a boolean, so that we can pass down the invalidate bit.
1998-03-07 21:37:31 +00:00
for (i = 0; i < count; i++) {
KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
("vm_pageout_flush: partially invalid page %p index %d/%d",
mc[i], i, count));
KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
("vm_pageout_flush: writeable page %p", mc[i]));
vm_page_sbusy(mc[i]);
This mega-commit is meant to fix numerous interrelated problems. There has been some bitrot and incorrect assumptions in the vfs_bio code. These problems have manifest themselves worse on NFS type filesystems, but can still affect local filesystems under certain circumstances. Most of the problems have involved mmap consistancy, and as a side-effect broke the vfs.ioopt code. This code might have been committed seperately, but almost everything is interrelated. 1) Allow (pmap_object_init_pt) prefaulting of buffer-busy pages that are fully valid. 2) Rather than deactivating erroneously read initial (header) pages in kern_exec, we now free them. 3) Fix the rundown of non-VMIO buffers that are in an inconsistent (missing vp) state. 4) Fix the disassociation of pages from buffers in brelse. The previous code had rotted and was faulty in a couple of important circumstances. 5) Remove a gratuitious buffer wakeup in vfs_vmio_release. 6) Remove a crufty and currently unused cluster mechanism for VBLK files in vfs_bio_awrite. When the code is functional, I'll add back a cleaner version. 7) The page busy count wakeups assocated with the buffer cache usage were incorrectly cleaned up in a previous commit by me. Revert to the original, correct version, but with a cleaner implementation. 8) The cluster read code now tries to keep data associated with buffers more aggressively (without breaking the heuristics) when it is presumed that the read data (buffers) will be soon needed. 9) Change to filesystem lockmgr locks so that they use LK_NOPAUSE. The delay loop waiting is not useful for filesystem locks, due to the length of the time intervals. 10) Correct and clean-up spec_getpages. 11) Implement a fully functional nfs_getpages, nfs_putpages. 12) Fix nfs_write so that modifications are coherent with the NFS data on the server disk (at least as well as NFS seems to allow.) 13) Properly support MS_INVALIDATE on NFS. 14) Properly pass down MS_INVALIDATE to lower levels of the VM code from vm_map_clean. 15) Better support the notion of pages being busy but valid, so that fewer in-transit waits occur. (use p->busy more for pageouts instead of PG_BUSY.) Since the page is fully valid, it is still usable for reads. 16) It is possible (in error) for cached pages to be busy. Make the page allocation code handle that case correctly. (It should probably be a printf or panic, but I want the system to handle coding errors robustly. I'll probably add a printf.) 17) Correct the design and usage of vm_page_sleep. It didn't handle consistancy problems very well, so make the design a little less lofty. After vm_page_sleep, if it ever blocked, it is still important to relookup the page (if the object generation count changed), and verify it's status (always.) 18) In vm_pageout.c, vm_pageout_clean had rotted, so clean that up. 19) Push the page busy for writes and VM_PROT_READ into vm_pageout_flush. 20) Fix vm_pager_put_pages and it's descendents to support an int flag instead of a boolean, so that we can pass down the invalidate bit.
1998-03-07 21:37:31 +00:00
}
vm_object_pip_add(object, count);
vm_pager_put_pages(object, mc, count, flags, pageout_status);
1994-05-24 10:09:53 +00:00
runlen = count - mreq;
if (eio != NULL)
*eio = FALSE;
for (i = 0; i < count; i++) {
vm_page_t mt = mc[i];
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct proc or any VM system structure will have to be rebuilt!!! Much needed overhaul of the VM system. Included in this first round of changes: 1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages, haspage, and sync operations are supported. The haspage interface now provides information about clusterability. All pager routines now take struct vm_object's instead of "pagers". 2) Improved data structures. In the previous paradigm, there is constant confusion caused by pagers being both a data structure ("allocate a pager") and a collection of routines. The idea of a pager structure has escentially been eliminated. Objects now have types, and this type is used to index the appropriate pager. In most cases, items in the pager structure were duplicated in the object data structure and thus were unnecessary. In the few cases that remained, a un_pager structure union was created in the object to contain these items. 3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now be removed. For instance, vm_object_enter(), vm_object_lookup(), vm_object_remove(), and the associated object hash list were some of the things that were removed. 4) simple_lock's removed. Discussion with several people reveals that the SMP locking primitives used in the VM system aren't likely the mechanism that we'll be adopting. Even if it were, the locking that was in the code was very inadequate and would have to be mostly re-done anyway. The locking in a uni-processor kernel was a no-op but went a long way toward making the code difficult to read and debug. 5) Places that attempted to kludge-up the fact that we don't have kernel thread support have been fixed to reflect the reality that we are really dealing with processes, not threads. The VM system didn't have complete thread support, so the comments and mis-named routines were just wrong. We now use tsleep and wakeup directly in the lock routines, for instance. 6) Where appropriate, the pagers have been improved, especially in the pager_alloc routines. Most of the pager_allocs have been rewritten and are now faster and easier to maintain. 7) The pagedaemon pageout clustering algorithm has been rewritten and now tries harder to output an even number of pages before and after the requested page. This is sort of the reverse of the ideal pagein algorithm and should provide better overall performance. 8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup have been removed. Some other unnecessary casts have also been removed. 9) Some almost useless debugging code removed. 10) Terminology of shadow objects vs. backing objects straightened out. The fact that the vm_object data structure escentially had this backwards really confused things. The use of "shadow" and "backing object" throughout the code is now internally consistent and correct in the Mach terminology. 11) Several minor bug fixes, including one in the vm daemon that caused 0 RSS objects to not get purged as intended. 12) A "default pager" has now been created which cleans up the transition of objects to the "swap" type. The previous checks throughout the code for swp->pg_data != NULL were really ugly. This change also provides the rudiments for future backing of "anonymous" memory by something other than the swap pager (via the vnode pager, for example), and it allows the decision about which of these pagers to use to be made dynamically (although will need some additional decision code to do this, of course). 13) (dyson) MAP_COPY has been deprecated and the corresponding "copy object" code has been removed. MAP_COPY was undocumented and non- standard. It was furthermore broken in several ways which caused its behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will continue to work correctly, but via the slightly different semantics of MAP_PRIVATE. 14) (dyson) Sharing maps have been removed. It's marginal usefulness in a threads design can be worked around in other ways. Both #12 and #13 were done to simplify the code and improve readability and maintain- ability. (As were most all of these changes) TODO: 1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing this will reduce the vnode pager to a mere fraction of its current size. 2) Rewrite vm_fault and the swap/vnode pagers to use the clustering information provided by the new haspage pager interface. This will substantially reduce the overhead by eliminating a large number of VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be improved to provide both a "behind" and "ahead" indication of contiguousness. 3) Implement the extended features of pager_haspage in swap_pager_haspage(). It currently just says 0 pages ahead/behind. 4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps via a much more general mechanism that could also be used for disk striping of regular filesystems. 5) Do something to improve the architecture of vm_object_collapse(). The fact that it makes calls into the swap pager and knows too much about how the swap pager operates really bothers me. It also doesn't allow for collapsing of non-swap pager objects ("unnamed" objects backed by other pagers).
1995-07-13 08:48:48 +00:00
KASSERT(pageout_status[i] == VM_PAGER_PEND ||
!pmap_page_is_write_mapped(mt),
("vm_pageout_flush: page %p is not write protected", mt));
switch (pageout_status[i]) {
case VM_PAGER_OK:
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_page_lock(mt);
if (vm_page_in_laundry(mt))
vm_page_deactivate_noreuse(mt);
vm_page_unlock(mt);
/* FALLTHROUGH */
case VM_PAGER_PEND:
numpagedout++;
break;
case VM_PAGER_BAD:
/*
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* The page is outside the object's range. We pretend
* that the page out worked and clean the page, so the
* changes will be lost if the page is reclaimed by
* the page daemon.
*/
vm_page_undirty(mt);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_page_lock(mt);
if (vm_page_in_laundry(mt))
vm_page_deactivate_noreuse(mt);
vm_page_unlock(mt);
break;
case VM_PAGER_ERROR:
case VM_PAGER_FAIL:
/*
* If the page couldn't be paged out to swap because the
* pager wasn't able to find space, place the page in
* the PQ_UNSWAPPABLE holding queue. This is an
* optimization that prevents the page daemon from
* wasting CPU cycles on pages that cannot be reclaimed
* becase no swap device is configured.
*
* Otherwise, reactivate the page so that it doesn't
* clog the laundry and inactive queues. (We will try
* paging it out again later.)
*/
vm_page_lock(mt);
if (object->type == OBJT_SWAP &&
pageout_status[i] == VM_PAGER_FAIL) {
vm_page_unswappable(mt);
numpagedout++;
} else
vm_page_activate(mt);
vm_page_unlock(mt);
if (eio != NULL && i >= mreq && i - mreq < runlen)
*eio = TRUE;
break;
case VM_PAGER_AGAIN:
if (i >= mreq && i - mreq < runlen)
runlen = i - mreq;
1994-05-24 10:09:53 +00:00
break;
}
1994-05-24 10:09:53 +00:00
/*
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
* If the operation is still going, leave the page busy to
* block all other accesses. Also, leave the paging in
* progress indicator set so that we don't attempt an object
* collapse.
1994-05-24 10:09:53 +00:00
*/
if (pageout_status[i] != VM_PAGER_PEND) {
vm_object_pip_wakeup(object);
vm_page_sunbusy(mt);
1994-05-24 10:09:53 +00:00
}
}
if (prunlen != NULL)
*prunlen = runlen;
return (numpagedout);
}
1994-05-24 10:09:53 +00:00
static void
vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
{
atomic_store_rel_int(&swapdev_enabled, 1);
}
static void
vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
{
if (swap_pager_nswapdev() == 1)
atomic_store_rel_int(&swapdev_enabled, 0);
}
/*
* Attempt to acquire all of the necessary locks to launder a page and
* then call through the clustering layer to PUTPAGES. Wait a short
* time for a vnode lock.
*
* Requires the page and object lock on entry, releases both before return.
* Returns 0 on success and an errno otherwise.
*/
static int
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_pageout_clean(vm_page_t m, int *numpagedout)
{
struct vnode *vp;
struct mount *mp;
vm_object_t object;
vm_pindex_t pindex;
int error, lockmode;
vm_page_assert_locked(m);
object = m->object;
VM_OBJECT_ASSERT_WLOCKED(object);
error = 0;
vp = NULL;
mp = NULL;
/*
* The object is already known NOT to be dead. It
* is possible for the vget() to block the whole
* pageout daemon, but the new low-memory handling
* code should prevent it.
*
* We can't wait forever for the vnode lock, we might
* deadlock due to a vn_read() getting stuck in
* vm_wait while holding this vnode. We skip the
* vnode if we can't get it in a reasonable amount
* of time.
*/
if (object->type == OBJT_VNODE) {
vm_page_unlock(m);
vp = object->handle;
if (vp->v_type == VREG &&
vn_start_write(vp, &mp, V_NOWAIT) != 0) {
mp = NULL;
error = EDEADLK;
goto unlock_all;
}
KASSERT(mp != NULL,
("vp %p with NULL v_mount", vp));
vm_object_reference_locked(object);
pindex = m->pindex;
VM_OBJECT_WUNLOCK(object);
lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
LK_SHARED : LK_EXCLUSIVE;
if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
vp = NULL;
error = EDEADLK;
goto unlock_mp;
}
VM_OBJECT_WLOCK(object);
/*
* Ensure that the object and vnode were not disassociated
* while locks were dropped.
*/
if (vp->v_object != object) {
error = ENOENT;
goto unlock_all;
}
vm_page_lock(m);
/*
* While the object and page were unlocked, the page
* may have been:
* (1) moved to a different queue,
* (2) reallocated to a different object,
* (3) reallocated to a different offset, or
* (4) cleaned.
*/
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
if (!vm_page_in_laundry(m) || m->object != object ||
m->pindex != pindex || m->dirty == 0) {
vm_page_unlock(m);
error = ENXIO;
goto unlock_all;
}
/*
* The page may have been busied or held while the object
* and page locks were released.
*/
if (vm_page_busied(m) || m->hold_count != 0) {
vm_page_unlock(m);
error = EBUSY;
goto unlock_all;
}
}
/*
* If a page is dirty, then it is either being washed
* (but not yet cleaned) or it is still in the
* laundry. If it is still in the laundry, then we
* start the cleaning operation.
*/
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
if ((*numpagedout = vm_pageout_cluster(m)) == 0)
error = EIO;
unlock_all:
VM_OBJECT_WUNLOCK(object);
unlock_mp:
vm_page_lock_assert(m, MA_NOTOWNED);
if (mp != NULL) {
if (vp != NULL)
vput(vp);
vm_object_deallocate(object);
vn_finished_write(mp);
}
return (error);
}
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* Attempt to launder the specified number of pages.
*
* Returns the number of pages successfully laundered.
*/
static int
vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
{
struct vm_pagequeue *pq;
vm_object_t object;
vm_page_t m, next;
int act_delta, error, maxscan, numpagedout, starting_target;
int vnodes_skipped;
bool pageout_ok, queue_locked;
starting_target = launder;
vnodes_skipped = 0;
/*
* Scan the laundry queues for pages eligible to be laundered. We stop
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* once the target number of dirty pages have been laundered, or once
* we've reached the end of the queue. A single iteration of this loop
* may cause more than one page to be laundered because of clustering.
*
* maxscan ensures that we don't re-examine requeued pages. Any
* additional pages written as part of a cluster are subtracted from
* maxscan since they must be taken from the laundry queue.
*
* As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
* swap devices are configured.
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
*/
if (atomic_load_acq_int(&swapdev_enabled))
pq = &vmd->vmd_pagequeues[PQ_UNSWAPPABLE];
else
pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
scan:
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_pagequeue_lock(pq);
maxscan = pq->pq_cnt;
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
queue_locked = true;
for (m = TAILQ_FIRST(&pq->pq_pl);
m != NULL && maxscan-- > 0 && launder > 0;
m = next) {
vm_pagequeue_assert_locked(pq);
KASSERT(queue_locked, ("unlocked laundry queue"));
KASSERT(vm_page_in_laundry(m),
("page %p has an inconsistent queue", m));
next = TAILQ_NEXT(m, plinks.q);
if ((m->flags & PG_MARKER) != 0)
continue;
KASSERT((m->flags & PG_FICTITIOUS) == 0,
("PG_FICTITIOUS page %p cannot be in laundry queue", m));
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("VPO_UNMANAGED page %p cannot be in laundry queue", m));
if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
vm_page_unlock(m);
continue;
}
object = m->object;
if ((!VM_OBJECT_TRYWLOCK(object) &&
(!vm_pageout_fallback_object_lock(m, &next) ||
m->hold_count != 0)) || vm_page_busied(m)) {
VM_OBJECT_WUNLOCK(object);
vm_page_unlock(m);
continue;
}
/*
* Unlock the laundry queue, invalidating the 'next' pointer.
* Use a marker to remember our place in the laundry queue.
*/
TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_laundry_marker,
plinks.q);
vm_pagequeue_unlock(pq);
queue_locked = false;
/*
* Invalid pages can be easily freed. They cannot be
* mapped; vm_page_free() asserts this.
*/
if (m->valid == 0)
goto free_page;
/*
* If the page has been referenced and the object is not dead,
* reactivate or requeue the page depending on whether the
* object is mapped.
*/
if ((m->aflags & PGA_REFERENCED) != 0) {
vm_page_aflag_clear(m, PGA_REFERENCED);
act_delta = 1;
} else
act_delta = 0;
if (object->ref_count != 0)
act_delta += pmap_ts_referenced(m);
else {
KASSERT(!pmap_page_is_mapped(m),
("page %p is mapped", m));
}
if (act_delta != 0) {
if (object->ref_count != 0) {
VM_CNT_INC(v_reactivated);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_page_activate(m);
/*
* Increase the activation count if the page
* was referenced while in the laundry queue.
* This makes it less likely that the page will
* be returned prematurely to the inactive
* queue.
*/
m->act_count += act_delta + ACT_ADVANCE;
/*
* If this was a background laundering, count
* activated pages towards our target. The
* purpose of background laundering is to ensure
* that pages are eventually cycled through the
* laundry queue, and an activation is a valid
* way out.
*/
if (!in_shortfall)
launder--;
goto drop_page;
} else if ((object->flags & OBJ_DEAD) == 0)
goto requeue_page;
}
/*
* If the page appears to be clean at the machine-independent
* layer, then remove all of its mappings from the pmap in
* anticipation of freeing it. If, however, any of the page's
* mappings allow write access, then the page may still be
* modified until the last of those mappings are removed.
*/
if (object->ref_count != 0) {
vm_page_test_dirty(m);
if (m->dirty == 0)
pmap_remove_all(m);
}
/*
* Clean pages are freed, and dirty pages are paged out unless
* they belong to a dead object. Requeueing dirty pages from
* dead objects is pointless, as they are being paged out and
* freed by the thread that destroyed the object.
*/
if (m->dirty == 0) {
free_page:
vm_page_free(m);
VM_CNT_INC(v_dfree);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
} else if ((object->flags & OBJ_DEAD) == 0) {
if (object->type != OBJT_SWAP &&
object->type != OBJT_DEFAULT)
pageout_ok = true;
else if (disable_swap_pageouts)
pageout_ok = false;
else
pageout_ok = true;
if (!pageout_ok) {
requeue_page:
vm_pagequeue_lock(pq);
queue_locked = true;
vm_page_requeue_locked(m);
goto drop_page;
}
/*
* Form a cluster with adjacent, dirty pages from the
* same object, and page out that entire cluster.
*
* The adjacent, dirty pages must also be in the
* laundry. However, their mappings are not checked
* for new references. Consequently, a recently
* referenced page may be paged out. However, that
* page will not be prematurely reclaimed. After page
* out, the page will be placed in the inactive queue,
* where any new references will be detected and the
* page reactivated.
*/
error = vm_pageout_clean(m, &numpagedout);
if (error == 0) {
launder -= numpagedout;
maxscan -= numpagedout - 1;
} else if (error == EDEADLK) {
pageout_lock_miss++;
vnodes_skipped++;
}
goto relock_queue;
}
drop_page:
vm_page_unlock(m);
VM_OBJECT_WUNLOCK(object);
relock_queue:
if (!queue_locked) {
vm_pagequeue_lock(pq);
queue_locked = true;
}
next = TAILQ_NEXT(&vmd->vmd_laundry_marker, plinks.q);
TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_laundry_marker, plinks.q);
}
vm_pagequeue_unlock(pq);
if (launder > 0 && pq == &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]) {
pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
goto scan;
}
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* Wakeup the sync daemon if we skipped a vnode in a writeable object
* and we didn't launder enough pages.
*/
if (vnodes_skipped > 0 && launder > 0)
(void)speedup_syncer();
return (starting_target - launder);
}
/*
* Compute the integer square root.
*/
static u_int
isqrt(u_int num)
{
u_int bit, root, tmp;
bit = 1u << ((NBBY * sizeof(u_int)) - 2);
while (bit > num)
bit >>= 2;
root = 0;
while (bit != 0) {
tmp = root + bit;
root >>= 1;
if (num >= tmp) {
num -= tmp;
root += bit;
}
bit >>= 2;
}
return (root);
}
/*
* Perform the work of the laundry thread: periodically wake up and determine
* whether any pages need to be laundered. If so, determine the number of pages
* that need to be laundered, and launder them.
*/
static void
vm_pageout_laundry_worker(void *arg)
{
struct vm_domain *domain;
struct vm_pagequeue *pq;
uint64_t nclean, ndirty;
u_int last_launder, wakeups;
int domidx, last_target, launder, shortfall, shortfall_cycle, target;
bool in_shortfall;
domidx = (uintptr_t)arg;
domain = &vm_dom[domidx];
pq = &domain->vmd_pagequeues[PQ_LAUNDRY];
KASSERT(domain->vmd_segs != 0, ("domain without segments"));
vm_pageout_init_marker(&domain->vmd_laundry_marker, PQ_LAUNDRY);
shortfall = 0;
in_shortfall = false;
shortfall_cycle = 0;
target = 0;
last_launder = 0;
/*
* Calls to these handlers are serialized by the swap syscall lock.
*/
(void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, domain,
EVENTHANDLER_PRI_ANY);
(void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, domain,
EVENTHANDLER_PRI_ANY);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* The pageout laundry worker is never done, so loop forever.
*/
for (;;) {
KASSERT(target >= 0, ("negative target %d", target));
KASSERT(shortfall_cycle >= 0,
("negative cycle %d", shortfall_cycle));
launder = 0;
wakeups = VM_CNT_FETCH(v_pdwakeups);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* First determine whether we need to launder pages to meet a
* shortage of free pages.
*/
if (shortfall > 0) {
in_shortfall = true;
shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
target = shortfall;
} else if (!in_shortfall)
goto trybackground;
else if (shortfall_cycle == 0 || vm_laundry_target() <= 0) {
/*
* We recently entered shortfall and began laundering
* pages. If we have completed that laundering run
* (and we are no longer in shortfall) or we have met
* our laundry target through other activity, then we
* can stop laundering pages.
*/
in_shortfall = false;
target = 0;
goto trybackground;
}
last_launder = wakeups;
launder = target / shortfall_cycle--;
goto dolaundry;
/*
* There's no immediate need to launder any pages; see if we
* meet the conditions to perform background laundering:
*
* 1. The ratio of dirty to clean inactive pages exceeds the
* background laundering threshold and the pagedaemon has
* been woken up to reclaim pages since our last
* laundering, or
* 2. we haven't yet reached the target of the current
* background laundering run.
*
* The background laundering threshold is not a constant.
* Instead, it is a slowly growing function of the number of
* page daemon wakeups since the last laundering. Thus, as the
* ratio of dirty to clean inactive pages grows, the amount of
* memory pressure required to trigger laundering decreases.
*/
trybackground:
nclean = vm_cnt.v_inactive_count + vm_cnt.v_free_count;
ndirty = vm_cnt.v_laundry_count;
if (target == 0 && wakeups != last_launder &&
ndirty * isqrt(wakeups - last_launder) >= nclean) {
target = vm_background_launder_target;
}
/*
* We have a non-zero background laundering target. If we've
* laundered up to our maximum without observing a page daemon
* wakeup, just stop. This is a safety belt that ensures we
* don't launder an excessive amount if memory pressure is low
* and the ratio of dirty to clean pages is large. Otherwise,
* proceed at the background laundering rate.
*/
if (target > 0) {
if (wakeups != last_launder) {
last_launder = wakeups;
last_target = target;
} else if (last_target - target >=
vm_background_launder_max * PAGE_SIZE / 1024) {
target = 0;
}
launder = vm_background_launder_rate * PAGE_SIZE / 1024;
launder /= VM_LAUNDER_RATE;
if (launder > target)
launder = target;
}
dolaundry:
if (launder > 0) {
/*
* Because of I/O clustering, the number of laundered
* pages could exceed "target" by the maximum size of
* a cluster minus one.
*/
target -= min(vm_pageout_launder(domain, launder,
in_shortfall), target);
pause("laundp", hz / VM_LAUNDER_RATE);
}
/*
* If we're not currently laundering pages and the page daemon
* hasn't posted a new request, sleep until the page daemon
* kicks us.
*/
vm_pagequeue_lock(pq);
if (target == 0 && vm_laundry_request == VM_LAUNDRY_IDLE)
(void)mtx_sleep(&vm_laundry_request,
vm_pagequeue_lockptr(pq), PVM, "launds", 0);
/*
* If the pagedaemon has indicated that it's in shortfall, start
* a shortfall laundering unless we're already in the middle of
* one. This may preempt a background laundering.
*/
if (vm_laundry_request == VM_LAUNDRY_SHORTFALL &&
(!in_shortfall || shortfall_cycle == 0)) {
shortfall = vm_laundry_target() + vm_pageout_deficit;
target = 0;
} else
shortfall = 0;
if (target == 0)
vm_laundry_request = VM_LAUNDRY_IDLE;
vm_pagequeue_unlock(pq);
}
}
/*
* vm_pageout_scan does the dirty work for the pageout daemon.
*
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* pass == 0: Update active LRU/deactivate pages
* pass >= 1: Free inactive pages
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
*
* Returns true if pass was zero or enough pages were freed by the inactive
* queue scan to meet the target.
*/
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
static bool
vm_pageout_scan(struct vm_domain *vmd, int pass)
{
vm_page_t m, next;
struct vm_pagequeue *pq;
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
vm_object_t object;
long min_scan;
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
int act_delta, addl_page_shortage, deficit, inactq_shortage, maxscan;
int page_shortage, scan_tick, scanned, starting_page_shortage;
boolean_t queue_locked;
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
/*
* If we need to reclaim memory ask kernel caches to return
* some. We rate limit to avoid thrashing.
*/
if (vmd == &vm_dom[0] && pass > 0 &&
(time_uptime - lowmem_uptime) >= lowmem_period) {
/*
* Decrease registered cache sizes.
*/
SDT_PROBE0(vm, , , vm__lowmem_scan);
EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
/*
* We do this explicitly after the caches have been
* drained above.
*/
uma_reclaim();
lowmem_uptime = time_uptime;
}
/*
* The addl_page_shortage is the number of temporarily
* stuck pages in the inactive queue. In other words, the
* number of pages from the inactive count that should be
* discounted in setting the target for the active queue scan.
*/
addl_page_shortage = 0;
/*
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
* Calculate the number of pages that we want to free. This number
* can be negative if many pages are freed between the wakeup call to
* the page daemon and this calculation.
*/
if (pass > 0) {
deficit = atomic_readandclear_int(&vm_pageout_deficit);
page_shortage = vm_paging_target() + deficit;
} else
page_shortage = deficit = 0;
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
starting_page_shortage = page_shortage;
/*
* Start scanning the inactive queue for pages that we can free. The
* scan will stop when we reach the target or we have scanned the
* entire queue. (Note that m->act_count is not used to make
* decisions for the inactive queue, only for the active queue.)
*/
pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
maxscan = pq->pq_cnt;
vm_pagequeue_lock(pq);
queue_locked = TRUE;
for (m = TAILQ_FIRST(&pq->pq_pl);
The buffer queue mechanism has been reformulated. Instead of having QUEUE_AGE, QUEUE_LRU, and QUEUE_EMPTY we instead have QUEUE_CLEAN, QUEUE_DIRTY, QUEUE_EMPTY, and QUEUE_EMPTYKVA. With this patch clean and dirty buffers have been separated. Empty buffers with KVM assignments have been separated from truely empty buffers. getnewbuf() has been rewritten and now operates in a 100% optimal fashion. That is, it is able to find precisely the right kind of buffer it needs to allocate a new buffer, defragment KVM, or to free-up an existing buffer when the buffer cache is full (which is a steady-state situation for the buffer cache). Buffer flushing has been reorganized. Previously buffers were flushed in the context of whatever process hit the conditions forcing buffer flushing to occur. This resulted in processes blocking on conditions unrelated to what they were doing. This also resulted in inappropriate VFS stacking chains due to multiple processes getting stuck trying to flush dirty buffers or due to a single process getting into a situation where it might attempt to flush buffers recursively - a situation that was only partially fixed in prior commits. We have added a new daemon called the buf_daemon which is responsible for flushing dirty buffers when the number of dirty buffers exceeds the vfs.hidirtybuffers limit. This daemon attempts to dynamically adjust the rate at which dirty buffers are flushed such that getnewbuf() calls (almost) never block. The number of nbufs and amount of buffer space is now scaled past the 8MB limit that was previously imposed for systems with over 64MB of memory, and the vfs.{lo,hi}dirtybuffers limits have been relaxed somewhat. The number of physical buffers has been increased with the intention that we will manage physical I/O differently in the future. reassignbuf previously attempted to keep the dirtyblkhd list sorted which could result in non-deterministic operation under certain conditions, such as when a large number of dirty buffers are being managed. This algorithm has been changed. reassignbuf now keeps buffers locally sorted if it can do so cheaply, and otherwise gives up and adds buffers to the head of the dirtyblkhd list. The new algorithm is deterministic but not perfect. The new algorithm greatly reduces problems that previously occured when write_behind was turned off in the system. The P_FLSINPROG proc->p_flag bit has been replaced by the more descriptive P_BUFEXHAUST bit. This bit allows processes working with filesystem buffers to use available emergency reserves. Normal processes do not set this bit and are not allowed to dig into emergency reserves. The purpose of this bit is to avoid low-memory deadlocks. A small race condition was fixed in getpbuf() in vm/vm_pager.c. Submitted by: Matthew Dillon <dillon@apollo.backplane.com> Reviewed by: Kirk McKusick <mckusick@mckusick.com>
1999-07-04 00:25:38 +00:00
m != NULL && maxscan-- > 0 && page_shortage > 0;
m = next) {
vm_pagequeue_assert_locked(pq);
KASSERT(queue_locked, ("unlocked inactive queue"));
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
KASSERT(vm_page_inactive(m), ("Inactive queue %p", m));
1994-05-24 10:09:53 +00:00
VM_CNT_INC(v_pdpages);
next = TAILQ_NEXT(m, plinks.q);
1994-05-24 10:09:53 +00:00
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
/*
* skip marker pages
*/
if (m->flags & PG_MARKER)
continue;
KASSERT((m->flags & PG_FICTITIOUS) == 0,
("Fictitious page %p cannot be in inactive queue", m));
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("Unmanaged page %p cannot be in inactive queue", m));
/*
* The page or object lock acquisitions fail if the
* page was removed from the queue or moved to a
* different position within the queue. In either
* case, addl_page_shortage should not be incremented.
*/
if (!vm_pageout_page_lock(m, &next))
goto unlock_page;
else if (m->hold_count != 0) {
/*
* Held pages are essentially stuck in the
* queue. So, they ought to be discounted
* from the inactive count. See the
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
* calculation of inactq_shortage before the
* loop over the active queue below.
*/
addl_page_shortage++;
goto unlock_page;
}
object = m->object;
if (!VM_OBJECT_TRYWLOCK(object)) {
if (!vm_pageout_fallback_object_lock(m, &next))
goto unlock_object;
else if (m->hold_count != 0) {
addl_page_shortage++;
goto unlock_object;
}
}
if (vm_page_busied(m)) {
/*
* Don't mess with busy pages. Leave them at
* the front of the queue. Most likely, they
* are being paged out and will leave the
* queue shortly after the scan finishes. So,
* they ought to be discounted from the
* inactive count.
*/
addl_page_shortage++;
unlock_object:
VM_OBJECT_WUNLOCK(object);
unlock_page:
vm_page_unlock(m);
continue;
}
KASSERT(m->hold_count == 0, ("Held page %p", m));
/*
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* Dequeue the inactive page and unlock the inactive page
* queue, invalidating the 'next' pointer. Dequeueing the
* page here avoids a later reacquisition (and release) of
* the inactive page queue lock when vm_page_activate(),
* vm_page_free(), or vm_page_launder() is called. Use a
* marker to remember our place in the inactive queue.
*/
TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_page_dequeue_locked(m);
vm_pagequeue_unlock(pq);
queue_locked = FALSE;
/*
* Invalid pages can be easily freed. They cannot be
* mapped, vm_page_free() asserts this.
*/
if (m->valid == 0)
goto free_page;
/*
* If the page has been referenced and the object is not dead,
* reactivate or requeue the page depending on whether the
* object is mapped.
*/
if ((m->aflags & PGA_REFERENCED) != 0) {
vm_page_aflag_clear(m, PGA_REFERENCED);
act_delta = 1;
} else
act_delta = 0;
if (object->ref_count != 0) {
act_delta += pmap_ts_referenced(m);
} else {
KASSERT(!pmap_page_is_mapped(m),
("vm_pageout_scan: page %p is mapped", m));
}
if (act_delta != 0) {
if (object->ref_count != 0) {
VM_CNT_INC(v_reactivated);
vm_page_activate(m);
/*
* Increase the activation count if the page
* was referenced while in the inactive queue.
* This makes it less likely that the page will
* be returned prematurely to the inactive
* queue.
*/
m->act_count += act_delta + ACT_ADVANCE;
goto drop_page;
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
} else if ((object->flags & OBJ_DEAD) == 0) {
vm_pagequeue_lock(pq);
queue_locked = TRUE;
m->queue = PQ_INACTIVE;
TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
vm_pagequeue_cnt_inc(pq);
goto drop_page;
}
}
/*
* If the page appears to be clean at the machine-independent
* layer, then remove all of its mappings from the pmap in
* anticipation of freeing it. If, however, any of the page's
* mappings allow write access, then the page may still be
* modified until the last of those mappings are removed.
*/
if (object->ref_count != 0) {
vm_page_test_dirty(m);
if (m->dirty == 0)
pmap_remove_all(m);
}
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* Clean pages can be freed, but dirty pages must be sent back
* to the laundry, unless they belong to a dead object.
* Requeueing dirty pages from dead objects is pointless, as
* they are being paged out and freed by the thread that
* destroyed the object.
*/
if (m->dirty == 0) {
free_page:
vm_page_free(m);
VM_CNT_INC(v_dfree);
--page_shortage;
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
} else if ((object->flags & OBJ_DEAD) == 0)
vm_page_launder(m);
drop_page:
vm_page_unlock(m);
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
VM_OBJECT_WUNLOCK(object);
if (!queue_locked) {
vm_pagequeue_lock(pq);
queue_locked = TRUE;
}
next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
}
vm_pagequeue_unlock(pq);
1994-05-24 10:09:53 +00:00
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* Wake up the laundry thread so that it can perform any needed
* laundering. If we didn't meet our target, we're in shortfall and
* need to launder more aggressively. If PQ_LAUNDRY is empty and no
* swap devices are configured, the laundry thread has no work to do, so
* don't bother waking it up.
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
*/
if (vm_laundry_request == VM_LAUNDRY_IDLE &&
starting_page_shortage > 0) {
pq = &vm_dom[0].vmd_pagequeues[PQ_LAUNDRY];
vm_pagequeue_lock(pq);
if (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled)) {
if (page_shortage > 0) {
vm_laundry_request = VM_LAUNDRY_SHORTFALL;
VM_CNT_INC(v_pdshortfalls);
} else if (vm_laundry_request != VM_LAUNDRY_SHORTFALL)
vm_laundry_request = VM_LAUNDRY_BACKGROUND;
wakeup(&vm_laundry_request);
}
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
vm_pagequeue_unlock(pq);
}
/*
* Wakeup the swapout daemon if we didn't free the targeted number of
* pages.
*/
if (page_shortage > 0)
vm_swapout_run();
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
/*
* If the inactive queue scan fails repeatedly to meet its
* target, kill the largest process.
*/
vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
/*
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
* Compute the number of pages we want to try to move from the
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* active queue to either the inactive or laundry queue.
*
* When scanning active pages, we make clean pages count more heavily
* towards the page shortage than dirty pages. This is because dirty
* pages must be laundered before they can be reused and thus have less
* utility when attempting to quickly alleviate a shortage. However,
* this weighting also causes the scan to deactivate dirty pages more
* more aggressively, improving the effectiveness of clustering and
* ensuring that they can eventually be reused.
1994-05-24 10:09:53 +00:00
*/
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
inactq_shortage = vm_cnt.v_inactive_target - (vm_cnt.v_inactive_count +
vm_cnt.v_laundry_count / act_scan_laundry_weight) +
vm_paging_target() + deficit + addl_page_shortage;
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
page_shortage *= act_scan_laundry_weight;
pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
vm_pagequeue_lock(pq);
maxscan = pq->pq_cnt;
/*
* If we're just idle polling attempt to visit every
* active page within 'update_period' seconds.
*/
scan_tick = ticks;
if (vm_pageout_update_period != 0) {
min_scan = pq->pq_cnt;
min_scan *= scan_tick - vmd->vmd_last_active_scan;
min_scan /= hz * vm_pageout_update_period;
} else
min_scan = 0;
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0))
vmd->vmd_last_active_scan = scan_tick;
/*
* Scan the active queue for pages that can be deactivated. Update
* the per-page activity counter and use it to identify deactivation
* candidates. Held pages may be deactivated.
*/
for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next,
scanned++) {
KASSERT(m->queue == PQ_ACTIVE,
("vm_pageout_scan: page %p isn't active", m));
next = TAILQ_NEXT(m, plinks.q);
if ((m->flags & PG_MARKER) != 0)
continue;
KASSERT((m->flags & PG_FICTITIOUS) == 0,
("Fictitious page %p cannot be in active queue", m));
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("Unmanaged page %p cannot be in active queue", m));
Eliminate checks for a page having a NULL object in vm_pageout_scan() and vm_pageout_page_stats(). These checks were recently introduced by the first page locking commit, r207410, but they are not needed. At the same time, eliminate some redundant accesses to the page's object field. (These accesses should have neen eliminated by r207410.) Make the assertion in vm_page_flag_set() stricter. Specifically, only managed pages should have PG_WRITEABLE set. Add a comment documenting an assertion to vm_page_flag_clear(). It has long been the case that fictitious pages have their wire count permanently set to one. Add comments to vm_page_wire() and vm_page_unwire() documenting this. Add assertions to these functions as well. Update the comment describing vm_page_unwire(). Much of the old comment had little to do with vm_page_unwire(), but a lot to do with _vm_page_deactivate(). Move relevant parts of the old comment to _vm_page_deactivate(). Only pages that belong to an object can be paged out. Therefore, it is pointless for vm_page_unwire() to acquire the page queues lock and enqueue such pages in one of the paging queues. Generally speaking, such pages are immediately freed after the call to vm_page_unwire(). Previously, it was the call to vm_page_free() that reacquired the page queues lock and removed these pages from the paging queues. Now, we will never acquire the page queues lock for this case. (It is also worth noting that since both vm_page_unwire() and vm_page_free() occurred with the page locked, the page daemon never saw the page with its object field set to NULL.) Change the panic with vm_page_unwire() to provide a more precise message. Reviewed by: kib@
2010-06-14 19:54:19 +00:00
if (!vm_pageout_page_lock(m, &next)) {
vm_page_unlock(m);
continue;
}
This set of commits to the VM system does the following, and contain contributions or ideas from Stephen McKay <syssgm@devetir.qld.gov.au>, Alan Cox <alc@cs.rice.edu>, David Greenman <davidg@freebsd.org> and me: More usage of the TAILQ macros. Additional minor fix to queue.h. Performance enhancements to the pageout daemon. Addition of a wait in the case that the pageout daemon has to run immediately. Slightly modify the pageout algorithm. Significant revamp of the pmap/fork code: 1) PTE's and UPAGES's are NO LONGER in the process's map. 2) PTE's and UPAGES's reside in their own objects. 3) TOTAL elimination of recursive page table pagefaults. 4) The page directory now resides in the PTE object. 5) Implemented pmap_copy, thereby speeding up fork time. 6) Changed the pv entries so that the head is a pointer and not an entire entry. 7) Significant cleanup of pmap_protect, and pmap_remove. 8) Removed significant amounts of machine dependent fork code from vm_glue. Pushed much of that code into the machine dependent pmap module. 9) Support more completely the reuse of already zeroed pages (Page table pages and page directories) as being already zeroed. Performance and code cleanups in vm_map: 1) Improved and simplified allocation of map entries. 2) Improved vm_map_copy code. 3) Corrected some minor problems in the simplify code. Implemented splvm (combo of splbio and splimp.) The VM code now seldom uses splhigh. Improved the speed of and simplified kmem_malloc. Minor mod to vm_fault to avoid using pre-zeroed pages in the case of objects with backing objects along with the already existant condition of having a vnode. (If there is a backing object, there will likely be a COW... With a COW, it isn't necessary to start with a pre-zeroed page.) Minor reorg of source to perhaps improve locality of ref.
1996-05-18 03:38:05 +00:00
/*
* The count for page daemon pages is updated after checking
* the page for eligibility.
This set of commits to the VM system does the following, and contain contributions or ideas from Stephen McKay <syssgm@devetir.qld.gov.au>, Alan Cox <alc@cs.rice.edu>, David Greenman <davidg@freebsd.org> and me: More usage of the TAILQ macros. Additional minor fix to queue.h. Performance enhancements to the pageout daemon. Addition of a wait in the case that the pageout daemon has to run immediately. Slightly modify the pageout algorithm. Significant revamp of the pmap/fork code: 1) PTE's and UPAGES's are NO LONGER in the process's map. 2) PTE's and UPAGES's reside in their own objects. 3) TOTAL elimination of recursive page table pagefaults. 4) The page directory now resides in the PTE object. 5) Implemented pmap_copy, thereby speeding up fork time. 6) Changed the pv entries so that the head is a pointer and not an entire entry. 7) Significant cleanup of pmap_protect, and pmap_remove. 8) Removed significant amounts of machine dependent fork code from vm_glue. Pushed much of that code into the machine dependent pmap module. 9) Support more completely the reuse of already zeroed pages (Page table pages and page directories) as being already zeroed. Performance and code cleanups in vm_map: 1) Improved and simplified allocation of map entries. 2) Improved vm_map_copy code. 3) Corrected some minor problems in the simplify code. Implemented splvm (combo of splbio and splimp.) The VM code now seldom uses splhigh. Improved the speed of and simplified kmem_malloc. Minor mod to vm_fault to avoid using pre-zeroed pages in the case of objects with backing objects along with the already existant condition of having a vnode. (If there is a backing object, there will likely be a COW... With a COW, it isn't necessary to start with a pre-zeroed page.) Minor reorg of source to perhaps improve locality of ref.
1996-05-18 03:38:05 +00:00
*/
VM_CNT_INC(v_pdpages);
/*
* Check to see "how much" the page has been used.
*/
if ((m->aflags & PGA_REFERENCED) != 0) {
vm_page_aflag_clear(m, PGA_REFERENCED);
act_delta = 1;
} else
act_delta = 0;
/*
* Perform an unsynchronized object ref count check. While
* the page lock ensures that the page is not reallocated to
* another object, in particular, one with unmanaged mappings
* that cannot support pmap_ts_referenced(), two races are,
* nonetheless, possible:
* 1) The count was transitioning to zero, but we saw a non-
* zero value. pmap_ts_referenced() will return zero
* because the page is not mapped.
* 2) The count was transitioning to one, but we saw zero.
* This race delays the detection of a new reference. At
* worst, we will deactivate and reactivate the page.
*/
if (m->object->ref_count != 0)
act_delta += pmap_ts_referenced(m);
/*
* Advance or decay the act_count based on recent usage.
*/
if (act_delta != 0) {
m->act_count += ACT_ADVANCE + act_delta;
if (m->act_count > ACT_MAX)
m->act_count = ACT_MAX;
} else
m->act_count -= min(m->act_count, ACT_DECLINE);
/*
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* Move this page to the tail of the active, inactive or laundry
* queue depending on usage.
*/
if (m->act_count == 0) {
/* Dequeue to avoid later lock recursion. */
vm_page_dequeue_locked(m);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* When not short for inactive pages, let dirty pages go
* through the inactive queue before moving to the
* laundry queues. This gives them some extra time to
* be reactivated, potentially avoiding an expensive
* pageout. During a page shortage, the inactive queue
* is necessarily small, so we may move dirty pages
* directly to the laundry queue.
*/
if (inactq_shortage <= 0)
vm_page_deactivate(m);
else {
/*
* Calling vm_page_test_dirty() here would
* require acquisition of the object's write
* lock. However, during a page shortage,
* directing dirty pages into the laundry
* queue is only an optimization and not a
* requirement. Therefore, we simply rely on
* the opportunistic updates to the page's
* dirty field by the pmap.
*/
if (m->dirty == 0) {
vm_page_deactivate(m);
inactq_shortage -=
act_scan_laundry_weight;
} else {
vm_page_launder(m);
inactq_shortage--;
}
}
} else
vm_page_requeue_locked(m);
vm_page_unlock(m);
1994-05-24 10:09:53 +00:00
}
vm_pagequeue_unlock(pq);
if (pass > 0)
vm_swapout_run_idle();
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
return (page_shortage <= 0);
}
static int vm_pageout_oom_vote;
/*
* The pagedaemon threads randlomly select one to perform the
* OOM. Trying to kill processes before all pagedaemons
* failed to reach free target is premature.
*/
static void
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
int starting_page_shortage)
{
int old_vote;
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
if (starting_page_shortage <= 0 || starting_page_shortage !=
page_shortage)
vmd->vmd_oom_seq = 0;
else
vmd->vmd_oom_seq++;
if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
if (vmd->vmd_oom) {
vmd->vmd_oom = FALSE;
atomic_subtract_int(&vm_pageout_oom_vote, 1);
}
return;
}
Rework the test which raises OOM condition. Right now, the code checks for the swap space consumption plus checks that the amount of the free pages exceeds some limit, in case pagedeamon did not coped with the page shortage in one of the late passes. This is wrong because it does not account for the presence of the reclamaible pages in the queues which are not selectable for reclaim immediately. E.g., on the swap-less systems, large active queue easily triggered OOM. Instead, only raise OOM when pagedaemon is unable to produce a free page in several back-to-back passes. Track the failed passes per pagedaemon thread. The number of passes to trigger OOM was selected empirically and tested both on small (32M-64M i386 VM) and large (32G amd64) configurations. If the specifics of the load require tuning, sysctl vm.pageout_oom_seq sets the number of back-to-back passes which must fail before OOM is raised. Each pass takes 1/2 of seconds. Less the value, more sensible the pagedaemon is to the page shortage. In future, some heuristic to calculate the value of the tunable might be designed based on the system configuration and load. But before it can be done, the i/o system must be fixed to reliably time-out pagedaemon writes, even if waiting for the memory to proceed. Then, code can account for the in-flight page-outs and postpone OOM until all of them finished, which should reduce the need in tuning. Right now, ignoring the in-flight writes and the counter allows to break deadlocks due to write path doing sleepable memory allocations. Reported by: Dmitry Sivachenko, bde, many others Tested by: pho, bde, tuexen (arm) Reviewed by: alc Discussed with: bde, imp Sponsored by: The FreeBSD Foundation MFC after: 3 weeks
2015-11-16 06:26:26 +00:00
/*
* Do not follow the call sequence until OOM condition is
* cleared.
*/
vmd->vmd_oom_seq = 0;
if (vmd->vmd_oom)
return;
vmd->vmd_oom = TRUE;
old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
if (old_vote != vm_ndomains - 1)
return;
/*
* The current pagedaemon thread is the last in the quorum to
* start OOM. Initiate the selection and signaling of the
* victim.
*/
vm_pageout_oom(VM_OOM_MEM);
/*
* After one round of OOM terror, recall our vote. On the
* next pass, current pagedaemon would vote again if the low
* memory condition is still there, due to vmd_oom being
* false.
*/
vmd->vmd_oom = FALSE;
atomic_subtract_int(&vm_pageout_oom_vote, 1);
}
/*
* The OOM killer is the page daemon's action of last resort when
* memory allocation requests have been stalled for a prolonged period
* of time because it cannot reclaim memory. This function computes
* the approximate number of physical pages that could be reclaimed if
* the specified address space is destroyed.
*
* Private, anonymous memory owned by the address space is the
* principal resource that we expect to recover after an OOM kill.
* Since the physical pages mapped by the address space's COW entries
* are typically shared pages, they are unlikely to be released and so
* they are not counted.
*
* To get to the point where the page daemon runs the OOM killer, its
* efforts to write-back vnode-backed pages may have stalled. This
* could be caused by a memory allocation deadlock in the write path
* that might be resolved by an OOM kill. Therefore, physical pages
* belonging to vnode-backed objects are counted, because they might
* be freed without being written out first if the address space holds
* the last reference to an unlinked vnode.
*
* Similarly, physical pages belonging to OBJT_PHYS objects are
* counted because the address space might hold the last reference to
* the object.
*/
static long
vm_pageout_oom_pagecount(struct vmspace *vmspace)
{
vm_map_t map;
vm_map_entry_t entry;
vm_object_t obj;
long res;
map = &vmspace->vm_map;
KASSERT(!map->system_map, ("system map"));
sx_assert(&map->lock, SA_LOCKED);
res = 0;
for (entry = map->header.next; entry != &map->header;
entry = entry->next) {
if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
continue;
obj = entry->object.vm_object;
if (obj == NULL)
continue;
if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
obj->ref_count != 1)
continue;
switch (obj->type) {
case OBJT_DEFAULT:
case OBJT_SWAP:
case OBJT_PHYS:
case OBJT_VNODE:
res += obj->resident_page_count;
break;
}
}
return (res);
}
void
vm_pageout_oom(int shortage)
{
struct proc *p, *bigproc;
vm_offset_t size, bigsize;
struct thread *td;
struct vmspace *vm;
bool breakout;
/*
* We keep the process bigproc locked once we find it to keep anyone
* from messing with it; however, there is a possibility of
* deadlock if process B is bigproc and one of its child processes
* attempts to propagate a signal to B while we are waiting for A's
* lock while walking this list. To avoid this, we don't block on
* the process lock but just skip a process if it is already locked.
*/
bigproc = NULL;
bigsize = 0;
sx_slock(&allproc_lock);
FOREACH_PROC_IN_SYSTEM(p) {
PROC_LOCK(p);
/*
* If this is a system, protected or killed process, skip it.
*/
if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
p->p_pid == 1 || P_KILLED(p) ||
(p->p_pid < 48 && swap_pager_avail != 0)) {
PROC_UNLOCK(p);
continue;
}
/*
* If the process is in a non-running type state,
* don't touch it. Check all the threads individually.
*/
breakout = false;
FOREACH_THREAD_IN_PROC(p, td) {
thread_lock(td);
if (!TD_ON_RUNQ(td) &&
!TD_IS_RUNNING(td) &&
!TD_IS_SLEEPING(td) &&
!TD_IS_SUSPENDED(td) &&
!TD_IS_SWAPPED(td)) {
thread_unlock(td);
breakout = true;
break;
}
thread_unlock(td);
}
if (breakout) {
PROC_UNLOCK(p);
continue;
}
/*
* get the process size
*/
vm = vmspace_acquire_ref(p);
if (vm == NULL) {
PROC_UNLOCK(p);
continue;
}
_PHOLD_LITE(p);
PROC_UNLOCK(p);
sx_sunlock(&allproc_lock);
if (!vm_map_trylock_read(&vm->vm_map)) {
vmspace_free(vm);
sx_slock(&allproc_lock);
PRELE(p);
continue;
}
size = vmspace_swap_count(vm);
if (shortage == VM_OOM_MEM)
size += vm_pageout_oom_pagecount(vm);
vm_map_unlock_read(&vm->vm_map);
vmspace_free(vm);
sx_slock(&allproc_lock);
/*
* If this process is bigger than the biggest one,
* remember it.
*/
if (size > bigsize) {
if (bigproc != NULL)
PRELE(bigproc);
bigproc = p;
bigsize = size;
} else {
PRELE(p);
}
}
sx_sunlock(&allproc_lock);
if (bigproc != NULL) {
if (vm_panic_on_oom != 0)
panic("out of swap space");
PROC_LOCK(bigproc);
killproc(bigproc, "out of swap space");
sched_nice(bigproc, PRIO_MIN);
_PRELE(bigproc);
PROC_UNLOCK(bigproc);
wakeup(&vm_cnt.v_free_count);
}
}
1994-05-24 10:09:53 +00:00
static void
vm_pageout_worker(void *arg)
{
struct vm_domain *domain;
int domidx, pass;
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
bool target_met;
domidx = (uintptr_t)arg;
domain = &vm_dom[domidx];
pass = 0;
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
target_met = true;
/*
* XXXKIB It could be useful to bind pageout daemon threads to
* the cores belonging to the domain, from which vm_page_array
* is allocated.
*/
KASSERT(domain->vmd_segs != 0, ("domain without segments"));
domain->vmd_last_active_scan = ticks;
vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
&domain->vmd_inacthead, plinks.q);
/*
* The pageout daemon worker is never done, so loop forever.
*/
while (TRUE) {
mtx_lock(&vm_page_queue_free_mtx);
/*
* Generally, after a level >= 1 scan, if there are enough
* free pages to wakeup the waiters, then they are already
* awake. A call to vm_page_free() during the scan awakened
* them. However, in the following case, this wakeup serves
* to bound the amount of time that a thread might wait.
* Suppose a thread's call to vm_page_alloc() fails, but
* before that thread calls VM_WAIT, enough pages are freed by
* other threads to alleviate the free page shortage. The
* thread will, nonetheless, wait until another page is freed
* or this wakeup is performed.
*/
if (vm_pages_needed && !vm_page_count_min()) {
vm_pages_needed = false;
wakeup(&vm_cnt.v_free_count);
}
/*
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
* Do not clear vm_pageout_wanted until we reach our free page
* target. Otherwise, we may be awakened over and over again,
* wasting CPU time.
*/
Change vm_pageout_scan() to return a value indicating whether the free page target was met. Previously, vm_pageout_worker() itself checked the length of the free page queues to determine whether vm_pageout_scan(pass >= 1)'s inactive queue scan freed enough pages to meet the free page target. Specifically, vm_pageout_worker() used vm_paging_needed(). The trouble with vm_paging_needed() is that it compares the length of the free page queues to the wakeup threshold for the page daemon, which is much lower than the free page target. Consequently, vm_pageout_worker() could conclude that the inactive queue scan succeeded in meeting its free page target when in fact it did not; and rather than immediately triggering an all-out laundering pass over the inactive queue, vm_pageout_worker() would go back to sleep waiting for the free page count to fall below the page daemon wakeup threshold again, at which point it will perform another limited (pass == 1) scan over the inactive queue. Changing vm_pageout_worker() to use vm_page_count_target() instead of vm_paging_needed() won't work because any page allocations that happen concurrently with the inactive queue scan will result in the free page count being below the target at the end of a successful scan. Instead, having vm_pageout_scan() return a value indicating success or failure is the most straightforward fix. Reviewed by: kib, markj MFC after: 3 weeks Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8111
2016-10-05 16:15:26 +00:00
if (vm_pageout_wanted && target_met)
vm_pageout_wanted = false;
/*
* Might the page daemon receive a wakeup call?
*/
if (vm_pageout_wanted) {
/*
* No. Either vm_pageout_wanted was set by another
* thread during the previous scan, which must have
* been a level 0 scan, or vm_pageout_wanted was
* already set and the scan failed to free enough
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
* pages. If we haven't yet performed a level >= 1
* (page reclamation) scan, then increase the level
* and scan again now. Otherwise, sleep a bit and
* try again later.
*/
mtx_unlock(&vm_page_queue_free_mtx);
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
if (pass >= 1)
pause("psleep", hz / VM_INACT_SCAN_RATE);
pass++;
} else {
/*
* Yes. Sleep until pages need to be reclaimed or
* have their reference stats updated.
*/
if (mtx_sleep(&vm_pageout_wanted,
&vm_page_queue_free_mtx, PDROP | PVM, "psleep",
hz) == 0) {
VM_CNT_INC(v_pdwakeups);
pass = 1;
} else
pass = 0;
}
target_met = vm_pageout_scan(domain, pass);
}
}
1994-05-24 10:09:53 +00:00
/*
* vm_pageout_init initialises basic pageout daemon settings.
1994-05-24 10:09:53 +00:00
*/
static void
vm_pageout_init(void)
1994-05-24 10:09:53 +00:00
{
/*
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
* Initialize some paging parameters.
1994-05-24 10:09:53 +00:00
*/
vm_cnt.v_interrupt_free_min = 2;
if (vm_cnt.v_page_count < 2000)
vm_pageout_page_count = 8;
/*
* v_free_reserved needs to include enough for the largest
* swap pager structures plus enough for any pv_entry structs
* when paging.
*/
if (vm_cnt.v_page_count > 1024)
vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
else
vm_cnt.v_free_min = 4;
vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
vm_cnt.v_interrupt_free_min;
vm_cnt.v_free_reserved = vm_pageout_page_count +
vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
vm_cnt.v_free_min += vm_cnt.v_free_reserved;
vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
/*
* Set the default wakeup threshold to be 10% above the minimum
* page limit. This keeps the steady state out of shortfall.
*/
vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
/*
* Set interval in seconds for active scan. We want to visit each
* page at least once every ten minutes. This is to prevent worst
* case paging behaviors with stale active LRU.
*/
if (vm_pageout_update_period == 0)
vm_pageout_update_period = 600;
These changes embody the support of the fully coherent merged VM buffer cache, much higher filesystem I/O performance, and much better paging performance. It represents the culmination of over 6 months of R&D. The majority of the merged VM/cache work is by John Dyson. The following highlights the most significant changes. Additionally, there are (mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to support the new VM/buffer scheme. vfs_bio.c: Significant rewrite of most of vfs_bio to support the merged VM buffer cache scheme. The scheme is almost fully compatible with the old filesystem interface. Significant improvement in the number of opportunities for write clustering. vfs_cluster.c, vfs_subr.c Upgrade and performance enhancements in vfs layer code to support merged VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff. vm_object.c: Yet more improvements in the collapse code. Elimination of some windows that can cause list corruption. vm_pageout.c: Fixed it, it really works better now. Somehow in 2.0, some "enhancements" broke the code. This code has been reworked from the ground-up. vm_fault.c, vm_page.c, pmap.c, vm_object.c Support for small-block filesystems with merged VM/buffer cache scheme. pmap.c vm_map.c Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of kernel PTs. vm_glue.c Much simpler and more effective swapping code. No more gratuitous swapping. proc.h Fixed the problem that the p_lock flag was not being cleared on a fork. swap_pager.c, vnode_pager.c Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the code doesn't need it anymore. machdep.c Changes to better support the parameter values for the merged VM/buffer cache scheme. machdep.c, kern_exec.c, vm_glue.c Implemented a seperate submap for temporary exec string space and another one to contain process upages. This eliminates all map fragmentation problems that previously existed. ffs_inode.c, ufs_inode.c, ufs_readwrite.c Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on busy buffers. Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
/* XXX does not really belong here */
1994-05-24 10:09:53 +00:00
if (vm_page_max_wired == 0)
vm_page_max_wired = vm_cnt.v_free_count / 3;
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
/*
* Target amount of memory to move out of the laundry queue during a
* background laundering. This is proportional to the amount of system
* memory.
*/
vm_background_launder_target = (vm_cnt.v_free_target -
vm_cnt.v_free_min) / 10;
}
/*
* vm_pageout is the high level pageout daemon.
*/
static void
vm_pageout(void)
{
int error;
#ifdef VM_NUMA_ALLOC
int i;
#endif
1994-05-24 10:09:53 +00:00
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct proc or any VM system structure will have to be rebuilt!!! Much needed overhaul of the VM system. Included in this first round of changes: 1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages, haspage, and sync operations are supported. The haspage interface now provides information about clusterability. All pager routines now take struct vm_object's instead of "pagers". 2) Improved data structures. In the previous paradigm, there is constant confusion caused by pagers being both a data structure ("allocate a pager") and a collection of routines. The idea of a pager structure has escentially been eliminated. Objects now have types, and this type is used to index the appropriate pager. In most cases, items in the pager structure were duplicated in the object data structure and thus were unnecessary. In the few cases that remained, a un_pager structure union was created in the object to contain these items. 3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now be removed. For instance, vm_object_enter(), vm_object_lookup(), vm_object_remove(), and the associated object hash list were some of the things that were removed. 4) simple_lock's removed. Discussion with several people reveals that the SMP locking primitives used in the VM system aren't likely the mechanism that we'll be adopting. Even if it were, the locking that was in the code was very inadequate and would have to be mostly re-done anyway. The locking in a uni-processor kernel was a no-op but went a long way toward making the code difficult to read and debug. 5) Places that attempted to kludge-up the fact that we don't have kernel thread support have been fixed to reflect the reality that we are really dealing with processes, not threads. The VM system didn't have complete thread support, so the comments and mis-named routines were just wrong. We now use tsleep and wakeup directly in the lock routines, for instance. 6) Where appropriate, the pagers have been improved, especially in the pager_alloc routines. Most of the pager_allocs have been rewritten and are now faster and easier to maintain. 7) The pagedaemon pageout clustering algorithm has been rewritten and now tries harder to output an even number of pages before and after the requested page. This is sort of the reverse of the ideal pagein algorithm and should provide better overall performance. 8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup have been removed. Some other unnecessary casts have also been removed. 9) Some almost useless debugging code removed. 10) Terminology of shadow objects vs. backing objects straightened out. The fact that the vm_object data structure escentially had this backwards really confused things. The use of "shadow" and "backing object" throughout the code is now internally consistent and correct in the Mach terminology. 11) Several minor bug fixes, including one in the vm daemon that caused 0 RSS objects to not get purged as intended. 12) A "default pager" has now been created which cleans up the transition of objects to the "swap" type. The previous checks throughout the code for swp->pg_data != NULL were really ugly. This change also provides the rudiments for future backing of "anonymous" memory by something other than the swap pager (via the vnode pager, for example), and it allows the decision about which of these pagers to use to be made dynamically (although will need some additional decision code to do this, of course). 13) (dyson) MAP_COPY has been deprecated and the corresponding "copy object" code has been removed. MAP_COPY was undocumented and non- standard. It was furthermore broken in several ways which caused its behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will continue to work correctly, but via the slightly different semantics of MAP_PRIVATE. 14) (dyson) Sharing maps have been removed. It's marginal usefulness in a threads design can be worked around in other ways. Both #12 and #13 were done to simplify the code and improve readability and maintain- ability. (As were most all of these changes) TODO: 1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing this will reduce the vnode pager to a mere fraction of its current size. 2) Rewrite vm_fault and the swap/vnode pagers to use the clustering information provided by the new haspage pager interface. This will substantially reduce the overhead by eliminating a large number of VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be improved to provide both a "behind" and "ahead" indication of contiguousness. 3) Implement the extended features of pager_haspage in swap_pager_haspage(). It currently just says 0 pages ahead/behind. 4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps via a much more general mechanism that could also be used for disk striping of regular filesystems. 5) Do something to improve the architecture of vm_object_collapse(). The fact that it makes calls into the swap pager and knows too much about how the swap pager operates really bothers me. It also doesn't allow for collapsing of non-swap pager objects ("unnamed" objects backed by other pagers).
1995-07-13 08:48:48 +00:00
swap_pager_swap_init();
Introduce a new page queue, PQ_LAUNDRY, for storing unreferenced, dirty pages, specificially, dirty pages that have passed once through the inactive queue. A new, dedicated thread is responsible for both deciding when to launder pages and actually laundering them. The new policy uses the relative sizes of the inactive and laundry queues to determine whether to launder pages at a given point in time. In general, this leads to more intelligent swapping behavior, since the laundry thread will avoid pageouts when the marginal benefit of doing so is low. Previously, without a dedicated queue for dirty pages, the page daemon didn't have the information to determine whether pageout provides any benefit to the system. Thus, the previous policy often resulted in small but steadily increasing amounts of swap usage when the system is under memory pressure, even when the inactive queue consisted mostly of clean pages. This change addresses that issue, and also paves the way for some future virtual memory system improvements by removing the last source of object-cached clean pages, i.e., PG_CACHE pages. The new laundry thread sleeps while waiting for a request from the page daemon thread(s). A request is raised by setting the variable vm_laundry_request and waking the laundry thread. We request launderings for two reasons: to try and balance the inactive and laundry queue sizes ("background laundering"), and to quickly make up for a shortage of free pages and clean inactive pages ("shortfall laundering"). When background laundering is requested, the laundry thread computes the number of page daemon wakeups that have taken place since the last laundering. If this number is large enough relative to the ratio of the laundry and (global) inactive queue sizes, we will launder vm_background_launder_target pages at vm_background_launder_rate KB/s. Otherwise, the laundry thread goes back to sleep without doing any work. When scanning the laundry queue during background laundering, reactivated pages are counted towards the laundry thread's target. In contrast, shortfall laundering is requested when an inactive queue scan fails to meet its target. In this case, the laundry thread attempts to launder enough pages to meet v_free_target within 0.5s, which is the inactive queue scan period. A laundry request can be latched while another is currently being serviced. In particular, a shortfall request will immediately preempt a background laundering. This change also redefines the meaning of vm_cnt.v_reactivated and removes the functions vm_page_cache() and vm_page_try_to_cache(). The new meaning of vm_cnt.v_reactivated now better reflects its name. It represents the number of inactive or laundry pages that are returned to the active queue on account of a reference. In collaboration with: markj Reviewed by: kib Tested by: pho Sponsored by: Dell EMC Isilon Differential Revision: https://reviews.freebsd.org/D8302
2016-11-09 18:48:37 +00:00
error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
0, 0, "laundry: dom0");
if (error != 0)
panic("starting laundry for domain 0, error %d", error);
#ifdef VM_NUMA_ALLOC
for (i = 1; i < vm_ndomains; i++) {
error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
curproc, NULL, 0, 0, "dom%d", i);
if (error != 0) {
panic("starting pageout for domain %d, error %d\n",
i, error);
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
}
1994-05-24 10:09:53 +00:00
}
#endif
error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
0, 0, "uma");
if (error != 0)
panic("starting uma_reclaim helper, error %d\n", error);
vm_pageout_worker((void *)(uintptr_t)0);
1994-05-24 10:09:53 +00:00
}
/*
* Unless the free page queue lock is held by the caller, this function
* should be regarded as advisory. Specifically, the caller should
* not msleep() on &vm_cnt.v_free_count following this function unless
* the free page queue lock is held until the msleep() is performed.
*/
void
2013-01-28 12:08:29 +00:00
pagedaemon_wakeup(void)
{
if (!vm_pageout_wanted && curthread->td_proc != pageproc) {
vm_pageout_wanted = true;
wakeup(&vm_pageout_wanted);
}
}