freebsd-dev/sys/vm/vm_reserv.c

1057 lines
31 KiB
C
Raw Normal View History

/*-
* Copyright (c) 2002-2006 Rice University
* Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
* All rights reserved.
*
* This software was developed for the FreeBSD Project by Alan L. Cox,
* Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
*
* 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.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS 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 COPYRIGHT
* HOLDERS 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.
*/
/*
* Superpage reservation management module
*
* Any external functions defined by this module are only to be used by the
* virtual memory system.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_vm.h"
#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mutex.h>
#include <sys/queue.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/sbuf.h>
#include <sys/sysctl.h>
#include <sys/systm.h>
#include <vm/vm.h>
#include <vm/vm_param.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <vm/vm_phys.h>
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
#include <vm/vm_radix.h>
#include <vm/vm_reserv.h>
/*
* The reservation system supports the speculative allocation of large physical
* pages ("superpages"). Speculative allocation enables the fully-automatic
* utilization of superpages by the virtual memory system. In other words, no
* programmatic directives are required to use superpages.
*/
#if VM_NRESERVLEVEL > 0
/*
* The number of small pages that are contained in a level 0 reservation
*/
#define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER)
/*
* The number of bits by which a physical address is shifted to obtain the
* reservation number
*/
#define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
/*
* The size of a level 0 reservation in bytes
*/
#define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT)
/*
* Computes the index of the small page underlying the given (object, pindex)
* within the reservation's array of small pages.
*/
#define VM_RESERV_INDEX(object, pindex) \
(((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
/*
* The size of a population map entry
*/
typedef u_long popmap_t;
/*
* The number of bits in a population map entry
*/
#define NBPOPMAP (NBBY * sizeof(popmap_t))
/*
* The number of population map entries in a reservation
*/
#define NPOPMAP howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
/*
* Clear a bit in the population map.
*/
static __inline void
popmap_clear(popmap_t popmap[], int i)
{
popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
}
/*
* Set a bit in the population map.
*/
static __inline void
popmap_set(popmap_t popmap[], int i)
{
popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
}
/*
* Is a bit in the population map clear?
*/
static __inline boolean_t
popmap_is_clear(popmap_t popmap[], int i)
{
return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
}
/*
* Is a bit in the population map set?
*/
static __inline boolean_t
popmap_is_set(popmap_t popmap[], int i)
{
return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
}
/*
* The reservation structure
*
* A reservation structure is constructed whenever a large physical page is
* speculatively allocated to an object. The reservation provides the small
* physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
* within that object. The reservation's "popcnt" tracks the number of these
* small physical pages that are in use at any given time. When and if the
* reservation is not fully utilized, it appears in the queue of partially-
* populated reservations. The reservation always appears on the containing
* object's list of reservations.
*
* A partially-populated reservation can be broken and reclaimed at any time.
*/
struct vm_reserv {
TAILQ_ENTRY(vm_reserv) partpopq;
LIST_ENTRY(vm_reserv) objq;
vm_object_t object; /* containing object */
vm_pindex_t pindex; /* offset within object */
vm_page_t pages; /* first page of a superpage */
int popcnt; /* # of pages in use */
char inpartpopq;
popmap_t popmap[NPOPMAP]; /* bit vector of used pages */
};
/*
* The reservation array
*
* This array is analoguous in function to vm_page_array. It differs in the
* respect that it may contain a greater number of useful reservation
* structures than there are (physical) superpages. These "invalid"
* reservation structures exist to trade-off space for time in the
* implementation of vm_reserv_from_page(). Invalid reservation structures are
* distinguishable from "valid" reservation structures by inspecting the
* reservation's "pages" field. Invalid reservation structures have a NULL
* "pages" field.
*
* vm_reserv_from_page() maps a small (physical) page to an element of this
* array by computing a physical reservation number from the page's physical
* address. The physical reservation number is used as the array index.
*
* An "active" reservation is a valid reservation structure that has a non-NULL
* "object" field and a non-zero "popcnt" field. In other words, every active
* reservation belongs to a particular object. Moreover, every active
* reservation has an entry in the containing object's list of reservations.
*/
static vm_reserv_t vm_reserv_array;
/*
* The partially-populated reservation queue
*
* This queue enables the fast recovery of an unused cached or free small page
* from a partially-populated reservation. The reservation at the head of
* this queue is the least-recently-changed, partially-populated reservation.
*
* Access to this queue is synchronized by the free page queue lock.
*/
static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
static long vm_reserv_broken;
SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
&vm_reserv_broken, 0, "Cumulative number of broken reservations");
static long vm_reserv_freed;
SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
&vm_reserv_freed, 0, "Cumulative number of freed reservations");
static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
static long vm_reserv_reclaimed;
SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
&vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
static void vm_reserv_break(vm_reserv_t rv, vm_page_t m);
static void vm_reserv_depopulate(vm_reserv_t rv, int index);
static vm_reserv_t vm_reserv_from_page(vm_page_t m);
static boolean_t vm_reserv_has_pindex(vm_reserv_t rv,
vm_pindex_t pindex);
static void vm_reserv_populate(vm_reserv_t rv, int index);
static void vm_reserv_reclaim(vm_reserv_t rv);
/*
* Describes the current state of the partially-populated reservation queue.
*/
static int
sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
{
struct sbuf sbuf;
vm_reserv_t rv;
int counter, error, level, unused_pages;
error = sysctl_wire_old_buffer(req, 0);
if (error != 0)
return (error);
sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n");
for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
counter = 0;
unused_pages = 0;
mtx_lock(&vm_page_queue_free_mtx);
TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
counter++;
unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
}
mtx_unlock(&vm_page_queue_free_mtx);
sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
unused_pages * ((int)PAGE_SIZE / 1024), counter);
}
error = sbuf_finish(&sbuf);
sbuf_delete(&sbuf);
return (error);
}
/*
* Reduces the given reservation's population count. If the population count
* becomes zero, the reservation is destroyed. Additionally, moves the
* reservation to the tail of the partially-populated reservation queue if the
* population count is non-zero.
*
* The free page queue lock must be held.
*/
static void
vm_reserv_depopulate(vm_reserv_t rv, int index)
{
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
KASSERT(rv->object != NULL,
("vm_reserv_depopulate: reserv %p is free", rv));
KASSERT(popmap_is_set(rv->popmap, index),
("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
index));
KASSERT(rv->popcnt > 0,
("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
if (rv->inpartpopq) {
TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
rv->inpartpopq = FALSE;
} else {
KASSERT(rv->pages->psind == 1,
("vm_reserv_depopulate: reserv %p is already demoted",
rv));
rv->pages->psind = 0;
}
popmap_clear(rv->popmap, index);
rv->popcnt--;
if (rv->popcnt == 0) {
LIST_REMOVE(rv, objq);
rv->object = NULL;
vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
vm_reserv_freed++;
} else {
rv->inpartpopq = TRUE;
TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
}
}
/*
* Returns the reservation to which the given page might belong.
*/
static __inline vm_reserv_t
vm_reserv_from_page(vm_page_t m)
{
return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
}
/*
* Returns TRUE if the given reservation contains the given page index and
* FALSE otherwise.
*/
static __inline boolean_t
vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
{
return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
}
/*
* Increases the given reservation's population count. Moves the reservation
* to the tail of the partially-populated reservation queue.
*
* The free page queue must be locked.
*/
static void
vm_reserv_populate(vm_reserv_t rv, int index)
{
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
KASSERT(rv->object != NULL,
("vm_reserv_populate: reserv %p is free", rv));
KASSERT(popmap_is_clear(rv->popmap, index),
("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
index));
KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
("vm_reserv_populate: reserv %p is already full", rv));
KASSERT(rv->pages->psind == 0,
("vm_reserv_populate: reserv %p is already promoted", rv));
if (rv->inpartpopq) {
TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
rv->inpartpopq = FALSE;
}
popmap_set(rv->popmap, index);
rv->popcnt++;
if (rv->popcnt < VM_LEVEL_0_NPAGES) {
rv->inpartpopq = TRUE;
TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
} else
rv->pages->psind = 1;
}
/*
* Allocates a contiguous set of physical pages of the given size "npages"
* from an existing or newly-created reservation. All of the physical pages
* must be at or above the given physical address "low" and below the given
* physical address "high". The given value "alignment" determines the
* alignment of the first physical page in the set. If the given value
* "boundary" is non-zero, then the set of physical pages cannot cross any
* physical address boundary that is a multiple of that value. Both
* "alignment" and "boundary" must be a power of two.
*
* The object and free page queue must be locked.
*/
vm_page_t
vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
{
vm_paddr_t pa, size;
vm_page_t m, m_ret, mpred, msucc;
vm_pindex_t first, leftcap, rightcap;
vm_reserv_t rv;
u_long allocpages, maxpages, minpages;
int i, index, n;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
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);
KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
/*
* Is a reservation fundamentally impossible?
*/
if (pindex < VM_RESERV_INDEX(object, pindex) ||
pindex + npages > object->size)
return (NULL);
/*
* All reservations of a particular size have the same alignment.
* Assuming that the first page is allocated from a reservation, the
* least significant bits of its physical address can be determined
* from its offset from the beginning of the reservation and the size
* of the reservation.
*
* Could the specified index within a reservation of the smallest
* possible size satisfy the alignment and boundary requirements?
*/
pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
if ((pa & (alignment - 1)) != 0)
return (NULL);
size = npages << PAGE_SHIFT;
if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
return (NULL);
/*
* Look for an existing reservation.
*/
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
mpred = vm_radix_lookup_le(&object->rtree, pindex);
if (mpred != NULL) {
KASSERT(mpred->pindex < pindex,
("vm_reserv_alloc_contig: pindex already allocated"));
rv = vm_reserv_from_page(mpred);
if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
goto found;
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
msucc = TAILQ_NEXT(mpred, listq);
} else
msucc = TAILQ_FIRST(&object->memq);
if (msucc != NULL) {
KASSERT(msucc->pindex > pindex,
("vm_reserv_alloc_page: pindex already allocated"));
rv = vm_reserv_from_page(msucc);
if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
goto found;
}
/*
* Could at least one reservation fit between the first index to the
* left that can be used and the first index to the right that cannot
* be used?
*/
first = pindex - VM_RESERV_INDEX(object, pindex);
if (mpred != NULL) {
if ((rv = vm_reserv_from_page(mpred))->object != object)
leftcap = mpred->pindex + 1;
else
leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
if (leftcap > first)
return (NULL);
}
minpages = VM_RESERV_INDEX(object, pindex) + npages;
maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
allocpages = maxpages;
if (msucc != NULL) {
if ((rv = vm_reserv_from_page(msucc))->object != object)
rightcap = msucc->pindex;
else
rightcap = rv->pindex;
if (first + maxpages > rightcap) {
if (maxpages == VM_LEVEL_0_NPAGES)
return (NULL);
allocpages = minpages;
}
}
/*
* Would the last new reservation extend past the end of the object?
*/
if (first + maxpages > object->size) {
/*
* Don't allocate the last new reservation if the object is a
* vnode or backed by another object that is a vnode.
*/
if (object->type == OBJT_VNODE ||
(object->backing_object != NULL &&
object->backing_object->type == OBJT_VNODE)) {
if (maxpages == VM_LEVEL_0_NPAGES)
return (NULL);
allocpages = minpages;
}
/* Speculate that the object may grow. */
}
/*
* Allocate and populate the new reservations. The alignment and
* boundary specified for this allocation may be different from the
* alignment and boundary specified for the requested pages. For
* instance, the specified index may not be the first page within the
* first new reservation.
*/
m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
if (m == NULL)
return (NULL);
m_ret = NULL;
index = VM_RESERV_INDEX(object, pindex);
do {
rv = vm_reserv_from_page(m);
KASSERT(rv->pages == m,
("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
rv));
KASSERT(rv->object == NULL,
("vm_reserv_alloc_contig: reserv %p isn't free", rv));
LIST_INSERT_HEAD(&object->rvq, rv, objq);
rv->object = object;
rv->pindex = first;
KASSERT(rv->popcnt == 0,
("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
rv));
KASSERT(!rv->inpartpopq,
("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
rv));
for (i = 0; i < NPOPMAP; i++)
KASSERT(rv->popmap[i] == 0,
("vm_reserv_alloc_contig: reserv %p's popmap is corrupted",
rv));
n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
for (i = 0; i < n; i++)
vm_reserv_populate(rv, index + i);
npages -= n;
if (m_ret == NULL) {
m_ret = &rv->pages[index];
index = 0;
}
m += VM_LEVEL_0_NPAGES;
first += VM_LEVEL_0_NPAGES;
allocpages -= VM_LEVEL_0_NPAGES;
} while (allocpages > 0);
return (m_ret);
/*
* Found a matching reservation.
*/
found:
index = VM_RESERV_INDEX(object, pindex);
/* Does the allocation fit within the reservation? */
if (index + npages > VM_LEVEL_0_NPAGES)
return (NULL);
m = &rv->pages[index];
pa = VM_PAGE_TO_PHYS(m);
if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
return (NULL);
/* Handle vm_page_rename(m, new_object, ...). */
for (i = 0; i < npages; i++)
if (popmap_is_set(rv->popmap, index + i))
return (NULL);
for (i = 0; i < npages; i++)
vm_reserv_populate(rv, index + i);
return (m);
}
/*
* Allocates a page from an existing or newly-created reservation.
*
* The page "mpred" must immediately precede the offset "pindex" within the
* specified object.
*
* The object and free page queue must be locked.
*/
vm_page_t
vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
{
vm_page_t m, msucc;
vm_pindex_t first, leftcap, rightcap;
vm_reserv_t rv;
int i, index;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
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);
/*
* Is a reservation fundamentally impossible?
*/
if (pindex < VM_RESERV_INDEX(object, pindex) ||
pindex >= object->size)
return (NULL);
/*
* Look for an existing reservation.
*/
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
if (mpred != NULL) {
KASSERT(mpred->object == object,
("vm_reserv_alloc_page: object doesn't contain mpred"));
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
KASSERT(mpred->pindex < pindex,
("vm_reserv_alloc_page: mpred doesn't precede pindex"));
rv = vm_reserv_from_page(mpred);
if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
goto found;
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
msucc = TAILQ_NEXT(mpred, listq);
} else
msucc = TAILQ_FIRST(&object->memq);
if (msucc != NULL) {
KASSERT(msucc->pindex > pindex,
("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
Sync back vmcontention branch into HEAD: Replace the per-object resident and cached pages splay tree with a path-compressed multi-digit radix trie. Along with this, switch also the x86-specific handling of idle page tables to using the radix trie. This change is supposed to do the following: - Allowing the acquisition of read locking for lookup operations of the resident/cached pages collections as the per-vm_page_t splay iterators are now removed. - Increase the scalability of the operations on the page collections. The radix trie does rely on the consumers locking to ensure atomicity of its operations. In order to avoid deadlocks the bisection nodes are pre-allocated in the UMA zone. This can be done safely because the algorithm needs at maximum one new node per insert which means the maximum number of the desired nodes is the number of available physical frames themselves. However, not all the times a new bisection node is really needed. The radix trie implements path-compression because UFS indirect blocks can lead to several objects with a very sparse trie, increasing the number of levels to usually scan. It also helps in the nodes pre-fetching by introducing the single node per-insert property. This code is not generalized (yet) because of the possible loss of performance by having much of the sizes in play configurable. However, efforts to make this code more general and then reusable in further different consumers might be really done. The only KPI change is the removal of the function vm_page_splay() which is now reaped. The only KBI change, instead, is the removal of the left/right iterators from struct vm_page, which are now reaped. Further technical notes broken into mealpieces can be retrieved from the svn branch: http://svn.freebsd.org/base/user/attilio/vmcontention/ Sponsored by: EMC / Isilon storage division In collaboration with: alc, jeff Tested by: flo, pho, jhb, davide Tested by: ian (arm) Tested by: andreast (powerpc)
2013-03-18 00:25:02 +00:00
rv = vm_reserv_from_page(msucc);
if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
goto found;
}
/*
* Could a reservation fit between the first index to the left that
* can be used and the first index to the right that cannot be used?
*/
first = pindex - VM_RESERV_INDEX(object, pindex);
if (mpred != NULL) {
if ((rv = vm_reserv_from_page(mpred))->object != object)
leftcap = mpred->pindex + 1;
else
leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
if (leftcap > first)
return (NULL);
}
if (msucc != NULL) {
if ((rv = vm_reserv_from_page(msucc))->object != object)
rightcap = msucc->pindex;
else
rightcap = rv->pindex;
if (first + VM_LEVEL_0_NPAGES > rightcap)
return (NULL);
}
/*
* Would a new reservation extend past the end of the object?
*/
if (first + VM_LEVEL_0_NPAGES > object->size) {
/*
* Don't allocate a new reservation if the object is a vnode or
* backed by another object that is a vnode.
*/
if (object->type == OBJT_VNODE ||
(object->backing_object != NULL &&
object->backing_object->type == OBJT_VNODE))
return (NULL);
/* Speculate that the object may grow. */
}
/*
* Allocate and populate the new reservation.
*/
m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
if (m == NULL)
return (NULL);
rv = vm_reserv_from_page(m);
KASSERT(rv->pages == m,
("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
KASSERT(rv->object == NULL,
("vm_reserv_alloc_page: reserv %p isn't free", rv));
LIST_INSERT_HEAD(&object->rvq, rv, objq);
rv->object = object;
rv->pindex = first;
KASSERT(rv->popcnt == 0,
("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
KASSERT(!rv->inpartpopq,
("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
for (i = 0; i < NPOPMAP; i++)
KASSERT(rv->popmap[i] == 0,
("vm_reserv_alloc_page: reserv %p's popmap is corrupted",
rv));
index = VM_RESERV_INDEX(object, pindex);
vm_reserv_populate(rv, index);
return (&rv->pages[index]);
/*
* Found a matching reservation.
*/
found:
index = VM_RESERV_INDEX(object, pindex);
m = &rv->pages[index];
/* Handle vm_page_rename(m, new_object, ...). */
if (popmap_is_set(rv->popmap, index))
return (NULL);
vm_reserv_populate(rv, index);
return (m);
}
/*
* Breaks the given reservation. Except for the specified cached or free
* page, all cached and free pages in the reservation are returned to the
* physical memory allocator. The reservation's population count and map are
* reset to their initial state.
*
* The given reservation must not be in the partially-populated reservation
* queue. The free page queue lock must be held.
*/
static void
vm_reserv_break(vm_reserv_t rv, vm_page_t m)
{
int begin_zeroes, hi, i, lo;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
KASSERT(rv->object != NULL,
("vm_reserv_break: reserv %p is free", rv));
KASSERT(!rv->inpartpopq,
("vm_reserv_break: reserv %p's inpartpopq is TRUE", rv));
LIST_REMOVE(rv, objq);
rv->object = NULL;
if (m != NULL) {
/*
* Since the reservation is being broken, there is no harm in
* abusing the population map to stop "m" from being returned
* to the physical memory allocator.
*/
i = m - rv->pages;
KASSERT(popmap_is_clear(rv->popmap, i),
("vm_reserv_break: reserv %p's popmap is corrupted", rv));
popmap_set(rv->popmap, i);
rv->popcnt++;
}
i = hi = 0;
do {
/* Find the next 0 bit. Any previous 0 bits are < "hi". */
lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
if (lo == 0) {
/* Redundantly clears bits < "hi". */
rv->popmap[i] = 0;
rv->popcnt -= NBPOPMAP - hi;
while (++i < NPOPMAP) {
lo = ffsl(~rv->popmap[i]);
if (lo == 0) {
rv->popmap[i] = 0;
rv->popcnt -= NBPOPMAP;
} else
break;
}
if (i == NPOPMAP)
break;
hi = 0;
}
KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
/* Convert from ffsl() to ordinary bit numbering. */
lo--;
if (lo > 0) {
/* Redundantly clears bits < "hi". */
rv->popmap[i] &= ~((1UL << lo) - 1);
rv->popcnt -= lo - hi;
}
begin_zeroes = NBPOPMAP * i + lo;
/* Find the next 1 bit. */
do
hi = ffsl(rv->popmap[i]);
while (hi == 0 && ++i < NPOPMAP);
if (i != NPOPMAP)
/* Convert from ffsl() to ordinary bit numbering. */
hi--;
vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
hi - begin_zeroes);
} while (i < NPOPMAP);
KASSERT(rv->popcnt == 0,
("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
vm_reserv_broken++;
}
/*
* Breaks all reservations belonging to the given object.
*/
void
vm_reserv_break_all(vm_object_t object)
{
vm_reserv_t rv;
mtx_lock(&vm_page_queue_free_mtx);
while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
KASSERT(rv->object == object,
("vm_reserv_break_all: reserv %p is corrupted", rv));
if (rv->inpartpopq) {
TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
rv->inpartpopq = FALSE;
}
vm_reserv_break(rv, NULL);
}
mtx_unlock(&vm_page_queue_free_mtx);
}
/*
* Frees the given page if it belongs to a reservation. Returns TRUE if the
* page is freed and FALSE otherwise.
*
* The free page queue lock must be held.
*/
boolean_t
vm_reserv_free_page(vm_page_t m)
{
vm_reserv_t rv;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
rv = vm_reserv_from_page(m);
if (rv->object == NULL)
return (FALSE);
if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE)
vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages,
VM_LEVEL_0_ORDER);
vm_reserv_depopulate(rv, m - rv->pages);
return (TRUE);
}
/*
* Initializes the reservation management system. Specifically, initializes
* the reservation array.
*
* Requires that vm_page_array and first_page are initialized!
*/
void
vm_reserv_init(void)
{
vm_paddr_t paddr;
int i;
/*
* Initialize the reservation array. Specifically, initialize the
* "pages" field for every element that has an underlying superpage.
*/
for (i = 0; phys_avail[i + 1] != 0; i += 2) {
paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
PHYS_TO_VM_PAGE(paddr);
paddr += VM_LEVEL_0_SIZE;
}
}
}
/*
* Returns a reservation level if the given page belongs to a fully-populated
* reservation and -1 otherwise.
*/
int
vm_reserv_level_iffullpop(vm_page_t m)
{
vm_reserv_t rv;
rv = vm_reserv_from_page(m);
return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
}
/*
* Prepare for the reactivation of a cached page.
*
* First, suppose that the given page "m" was allocated individually, i.e., not
* as part of a reservation, and cached. Then, suppose a reservation
* containing "m" is allocated by the same object. Although "m" and the
* reservation belong to the same object, "m"'s pindex may not match the
* reservation's.
*
* The free page queue must be locked.
*/
boolean_t
vm_reserv_reactivate_page(vm_page_t m)
{
vm_reserv_t rv;
int index;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
rv = vm_reserv_from_page(m);
if (rv->object == NULL)
return (FALSE);
KASSERT((m->flags & PG_CACHED) != 0,
("vm_reserv_reactivate_page: page %p is not cached", m));
if (m->object == rv->object &&
m->pindex - rv->pindex == (index = VM_RESERV_INDEX(m->object,
m->pindex)))
vm_reserv_populate(rv, index);
else {
KASSERT(rv->inpartpopq,
("vm_reserv_reactivate_page: reserv %p's inpartpopq is FALSE",
rv));
TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
rv->inpartpopq = FALSE;
/* Don't release "m" to the physical memory allocator. */
vm_reserv_break(rv, m);
}
return (TRUE);
}
/*
* Breaks the given partially-populated reservation, releasing its cached and
* free pages to the physical memory allocator.
*
* The free page queue lock must be held.
*/
static void
vm_reserv_reclaim(vm_reserv_t rv)
{
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
KASSERT(rv->inpartpopq,
("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
rv->inpartpopq = FALSE;
vm_reserv_break(rv, NULL);
vm_reserv_reclaimed++;
}
/*
* Breaks the reservation at the head of the partially-populated reservation
* queue, releasing its cached and free pages to the physical memory
* allocator. Returns TRUE if a reservation is broken and FALSE otherwise.
*
* The free page queue lock must be held.
*/
boolean_t
vm_reserv_reclaim_inactive(void)
{
vm_reserv_t rv;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
vm_reserv_reclaim(rv);
return (TRUE);
}
return (FALSE);
}
/*
* Searches the partially-populated reservation queue for the least recently
* active reservation with unused pages, i.e., cached or free, that satisfy the
* given request for contiguous physical memory. If a satisfactory reservation
* is found, it is broken. Returns TRUE if a reservation is broken and FALSE
* otherwise.
*
* The free page queue lock must be held.
*/
boolean_t
vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
u_long alignment, vm_paddr_t boundary)
{
vm_paddr_t pa, size;
vm_reserv_t rv;
int hi, i, lo, next_free;
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
if (npages > VM_LEVEL_0_NPAGES - 1)
return (FALSE);
size = npages << PAGE_SHIFT;
TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
if (pa + PAGE_SIZE - size < low) {
/* This entire reservation is too low; go to next. */
continue;
}
pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
if (pa + size > high) {
/* This entire reservation is too high; go to next. */
continue;
}
if (pa < low) {
/* Start the search for free pages at "low". */
i = (low - pa) / NBPOPMAP;
hi = (low - pa) % NBPOPMAP;
} else
i = hi = 0;
do {
/* Find the next free page. */
lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
while (lo == 0 && ++i < NPOPMAP)
lo = ffsl(~rv->popmap[i]);
if (i == NPOPMAP)
break;
/* Convert from ffsl() to ordinary bit numbering. */
lo--;
next_free = NBPOPMAP * i + lo;
pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
KASSERT(pa >= low,
("vm_reserv_reclaim_contig: pa is too low"));
if (pa + size > high) {
/* The rest of this reservation is too high. */
break;
} else if ((pa & (alignment - 1)) != 0 ||
((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
/* Continue with this reservation. */
hi = lo;
continue;
}
/* Find the next used page. */
hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
while (hi == 0 && ++i < NPOPMAP) {
if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
size) {
vm_reserv_reclaim(rv);
return (TRUE);
}
hi = ffsl(rv->popmap[i]);
}
/* Convert from ffsl() to ordinary bit numbering. */
if (i != NPOPMAP)
hi--;
if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
size) {
vm_reserv_reclaim(rv);
return (TRUE);
}
} while (i < NPOPMAP);
}
return (FALSE);
}
/*
* Transfers the reservation underlying the given page to a new object.
*
* The object must be locked.
*/
void
vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
vm_pindex_t old_object_offset)
{
vm_reserv_t rv;
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(new_object);
rv = vm_reserv_from_page(m);
if (rv->object == old_object) {
mtx_lock(&vm_page_queue_free_mtx);
if (rv->object == old_object) {
LIST_REMOVE(rv, objq);
LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
rv->object = new_object;
rv->pindex -= old_object_offset;
}
mtx_unlock(&vm_page_queue_free_mtx);
}
}
/*
* Allocates the virtual and physical memory required by the reservation
* management system's data structures, in particular, the reservation array.
*/
vm_paddr_t
vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
{
vm_paddr_t new_end;
size_t size;
/*
* Calculate the size (in bytes) of the reservation array. Round up
* from "high_water" because every small page is mapped to an element
* in the reservation array based on its physical address. Thus, the
* number of elements in the reservation array can be greater than the
* number of superpages.
*/
size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
/*
* Allocate and map the physical memory for the reservation array. The
* next available virtual address is returned by reference.
*/
new_end = end - round_page(size);
vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
VM_PROT_READ | VM_PROT_WRITE);
bzero(vm_reserv_array, size);
/*
* Return the next available physical address.
*/
return (new_end);
}
#endif /* VM_NRESERVLEVEL > 0 */