Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
/*-
|
|
|
|
|
* Copyright (c) 2002-2006 Rice University
|
|
|
|
|
* Copyright (c) 2007 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.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include <sys/cdefs.h>
|
|
|
|
|
__FBSDID("$FreeBSD$");
|
|
|
|
|
|
|
|
|
|
#include "opt_ddb.h"
|
|
|
|
|
|
|
|
|
|
#include <sys/param.h>
|
|
|
|
|
#include <sys/systm.h>
|
|
|
|
|
#include <sys/lock.h>
|
|
|
|
|
#include <sys/kernel.h>
|
|
|
|
|
#include <sys/malloc.h>
|
|
|
|
|
#include <sys/mutex.h>
|
|
|
|
|
#include <sys/queue.h>
|
|
|
|
|
#include <sys/sbuf.h>
|
|
|
|
|
#include <sys/sysctl.h>
|
|
|
|
|
#include <sys/vmmeter.h>
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
#include <sys/vnode.h>
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
|
|
|
|
|
#include <ddb/ddb.h>
|
|
|
|
|
|
|
|
|
|
#include <vm/vm.h>
|
|
|
|
|
#include <vm/vm_param.h>
|
|
|
|
|
#include <vm/vm_kern.h>
|
|
|
|
|
#include <vm/vm_object.h>
|
|
|
|
|
#include <vm/vm_page.h>
|
|
|
|
|
#include <vm/vm_phys.h>
|
2008-04-06 18:09:28 +00:00
|
|
|
#include <vm/vm_reserv.h>
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
|
2010-07-27 20:33:50 +00:00
|
|
|
/*
|
|
|
|
|
* VM_FREELIST_DEFAULT is split into VM_NDOMAIN lists, one for each
|
|
|
|
|
* domain. These extra lists are stored at the end of the regular
|
|
|
|
|
* free lists starting with VM_NFREELIST.
|
|
|
|
|
*/
|
|
|
|
|
#define VM_RAW_NFREELIST (VM_NFREELIST + VM_NDOMAIN - 1)
|
|
|
|
|
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
struct vm_freelist {
|
|
|
|
|
struct pglist pl;
|
|
|
|
|
int lcnt;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
struct vm_phys_seg {
|
|
|
|
|
vm_paddr_t start;
|
|
|
|
|
vm_paddr_t end;
|
|
|
|
|
vm_page_t first_page;
|
2010-07-27 20:33:50 +00:00
|
|
|
int domain;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
struct vm_freelist (*free_queues)[VM_NFREEPOOL][VM_NFREEORDER];
|
|
|
|
|
};
|
|
|
|
|
|
2010-07-27 20:33:50 +00:00
|
|
|
struct mem_affinity *mem_affinity;
|
|
|
|
|
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
static struct vm_phys_seg vm_phys_segs[VM_PHYSSEG_MAX];
|
|
|
|
|
|
|
|
|
|
static int vm_phys_nsegs;
|
|
|
|
|
|
|
|
|
|
static struct vm_freelist
|
2010-07-27 20:33:50 +00:00
|
|
|
vm_phys_free_queues[VM_RAW_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
|
|
|
|
|
static struct vm_freelist
|
|
|
|
|
(*vm_phys_lookup_lists[VM_NDOMAIN][VM_RAW_NFREELIST])[VM_NFREEPOOL][VM_NFREEORDER];
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
|
|
|
|
|
static int vm_nfreelists = VM_FREELIST_DEFAULT + 1;
|
|
|
|
|
|
|
|
|
|
static int cnt_prezero;
|
|
|
|
|
SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
|
|
|
|
|
&cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
|
|
|
|
|
|
|
|
|
|
static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
|
|
|
|
|
SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
|
|
|
|
|
NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
|
|
|
|
|
|
|
|
|
|
static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
|
|
|
|
|
SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
|
|
|
|
|
NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
|
|
|
|
|
|
2010-07-27 20:33:50 +00:00
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
static int sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS);
|
|
|
|
|
SYSCTL_OID(_vm, OID_AUTO, phys_lookup_lists, CTLTYPE_STRING | CTLFLAG_RD,
|
|
|
|
|
NULL, 0, sysctl_vm_phys_lookup_lists, "A", "Phys Lookup Lists");
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind,
|
|
|
|
|
int domain);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind);
|
|
|
|
|
static int vm_phys_paddr_to_segind(vm_paddr_t pa);
|
|
|
|
|
static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
|
|
|
|
|
int order);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Outputs the state of the physical memory allocator, specifically,
|
|
|
|
|
* the amount of physical memory in each free list.
|
|
|
|
|
*/
|
|
|
|
|
static int
|
|
|
|
|
sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
struct sbuf sbuf;
|
|
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
int error, flind, oind, pind;
|
|
|
|
|
|
2011-01-27 00:34:12 +00:00
|
|
|
error = sysctl_wire_old_buffer(req, 0);
|
|
|
|
|
if (error != 0)
|
|
|
|
|
return (error);
|
2010-09-16 16:13:12 +00:00
|
|
|
sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++) {
|
|
|
|
|
sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
|
|
|
|
|
"\n ORDER (SIZE) | NUMBER"
|
|
|
|
|
"\n ", flind);
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++)
|
|
|
|
|
sbuf_printf(&sbuf, " | POOL %d", pind);
|
|
|
|
|
sbuf_printf(&sbuf, "\n-- ");
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++)
|
|
|
|
|
sbuf_printf(&sbuf, "-- -- ");
|
|
|
|
|
sbuf_printf(&sbuf, "--\n");
|
|
|
|
|
for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
|
2010-10-30 18:00:53 +00:00
|
|
|
sbuf_printf(&sbuf, " %2d (%6dK)", oind,
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
1 << (PAGE_SHIFT - 10 + oind));
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++) {
|
|
|
|
|
fl = vm_phys_free_queues[flind][pind];
|
2010-10-30 18:00:53 +00:00
|
|
|
sbuf_printf(&sbuf, " | %6d", fl[oind].lcnt);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
sbuf_printf(&sbuf, "\n");
|
|
|
|
|
}
|
|
|
|
|
}
|
2010-09-16 16:13:12 +00:00
|
|
|
error = sbuf_finish(&sbuf);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
sbuf_delete(&sbuf);
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Outputs the set of physical memory segments.
|
|
|
|
|
*/
|
|
|
|
|
static int
|
|
|
|
|
sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
struct sbuf sbuf;
|
|
|
|
|
struct vm_phys_seg *seg;
|
|
|
|
|
int error, segind;
|
|
|
|
|
|
2011-01-27 00:34:12 +00:00
|
|
|
error = sysctl_wire_old_buffer(req, 0);
|
|
|
|
|
if (error != 0)
|
|
|
|
|
return (error);
|
2010-09-16 16:13:12 +00:00
|
|
|
sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
for (segind = 0; segind < vm_phys_nsegs; segind++) {
|
|
|
|
|
sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
|
|
|
|
|
seg = &vm_phys_segs[segind];
|
|
|
|
|
sbuf_printf(&sbuf, "start: %#jx\n",
|
|
|
|
|
(uintmax_t)seg->start);
|
|
|
|
|
sbuf_printf(&sbuf, "end: %#jx\n",
|
|
|
|
|
(uintmax_t)seg->end);
|
2010-07-27 20:33:50 +00:00
|
|
|
sbuf_printf(&sbuf, "domain: %d\n", seg->domain);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
|
|
|
|
|
}
|
2010-09-16 16:13:12 +00:00
|
|
|
error = sbuf_finish(&sbuf);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
sbuf_delete(&sbuf);
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
|
2010-07-27 20:33:50 +00:00
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
/*
|
|
|
|
|
* Outputs the set of free list lookup lists.
|
|
|
|
|
*/
|
|
|
|
|
static int
|
|
|
|
|
sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
struct sbuf sbuf;
|
|
|
|
|
int domain, error, flind, ndomains;
|
|
|
|
|
|
2011-01-27 00:34:12 +00:00
|
|
|
error = sysctl_wire_old_buffer(req, 0);
|
|
|
|
|
if (error != 0)
|
|
|
|
|
return (error);
|
2010-09-16 16:13:12 +00:00
|
|
|
sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
|
2011-01-27 00:34:12 +00:00
|
|
|
ndomains = vm_nfreelists - VM_NFREELIST + 1;
|
2010-07-27 20:33:50 +00:00
|
|
|
for (domain = 0; domain < ndomains; domain++) {
|
|
|
|
|
sbuf_printf(&sbuf, "\nDOMAIN %d:\n\n", domain);
|
|
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++)
|
|
|
|
|
sbuf_printf(&sbuf, " [%d]:\t%p\n", flind,
|
|
|
|
|
vm_phys_lookup_lists[domain][flind]);
|
|
|
|
|
}
|
2010-09-16 16:13:12 +00:00
|
|
|
error = sbuf_finish(&sbuf);
|
2010-07-27 20:33:50 +00:00
|
|
|
sbuf_delete(&sbuf);
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
/*
|
|
|
|
|
* Create a physical memory segment.
|
|
|
|
|
*/
|
|
|
|
|
static void
|
2010-07-27 20:33:50 +00:00
|
|
|
_vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind, int domain)
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
{
|
|
|
|
|
struct vm_phys_seg *seg;
|
|
|
|
|
#ifdef VM_PHYSSEG_SPARSE
|
|
|
|
|
long pages;
|
|
|
|
|
int segind;
|
|
|
|
|
|
|
|
|
|
pages = 0;
|
|
|
|
|
for (segind = 0; segind < vm_phys_nsegs; segind++) {
|
|
|
|
|
seg = &vm_phys_segs[segind];
|
|
|
|
|
pages += atop(seg->end - seg->start);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
|
|
|
|
|
("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
|
|
|
|
|
seg = &vm_phys_segs[vm_phys_nsegs++];
|
|
|
|
|
seg->start = start;
|
|
|
|
|
seg->end = end;
|
2010-07-27 20:33:50 +00:00
|
|
|
seg->domain = domain;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
#ifdef VM_PHYSSEG_SPARSE
|
|
|
|
|
seg->first_page = &vm_page_array[pages];
|
|
|
|
|
#else
|
|
|
|
|
seg->first_page = PHYS_TO_VM_PAGE(start);
|
2010-07-27 20:33:50 +00:00
|
|
|
#endif
|
|
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
if (flind == VM_FREELIST_DEFAULT && domain != 0) {
|
|
|
|
|
flind = VM_NFREELIST + (domain - 1);
|
|
|
|
|
if (flind >= vm_nfreelists)
|
|
|
|
|
vm_nfreelists = flind + 1;
|
|
|
|
|
}
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
#endif
|
|
|
|
|
seg->free_queues = &vm_phys_free_queues[flind];
|
|
|
|
|
}
|
|
|
|
|
|
2010-07-27 20:33:50 +00:00
|
|
|
static void
|
|
|
|
|
vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind)
|
|
|
|
|
{
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
if (mem_affinity == NULL) {
|
|
|
|
|
_vm_phys_create_seg(start, end, flind, 0);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for (i = 0;; i++) {
|
|
|
|
|
if (mem_affinity[i].end == 0)
|
|
|
|
|
panic("Reached end of affinity info");
|
|
|
|
|
if (mem_affinity[i].end <= start)
|
|
|
|
|
continue;
|
|
|
|
|
if (mem_affinity[i].start > start)
|
|
|
|
|
panic("No affinity info for start %jx",
|
|
|
|
|
(uintmax_t)start);
|
|
|
|
|
if (mem_affinity[i].end >= end) {
|
|
|
|
|
_vm_phys_create_seg(start, end, flind,
|
|
|
|
|
mem_affinity[i].domain);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
_vm_phys_create_seg(start, mem_affinity[i].end, flind,
|
|
|
|
|
mem_affinity[i].domain);
|
|
|
|
|
start = mem_affinity[i].end;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
/*
|
|
|
|
|
* Initialize the physical memory allocator.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
vm_phys_init(void)
|
|
|
|
|
{
|
|
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
int flind, i, oind, pind;
|
2010-07-27 20:33:50 +00:00
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
int ndomains, j;
|
|
|
|
|
#endif
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
|
|
|
|
|
for (i = 0; phys_avail[i + 1] != 0; i += 2) {
|
|
|
|
|
#ifdef VM_FREELIST_ISADMA
|
|
|
|
|
if (phys_avail[i] < 16777216) {
|
|
|
|
|
if (phys_avail[i + 1] > 16777216) {
|
|
|
|
|
vm_phys_create_seg(phys_avail[i], 16777216,
|
|
|
|
|
VM_FREELIST_ISADMA);
|
|
|
|
|
vm_phys_create_seg(16777216, phys_avail[i + 1],
|
|
|
|
|
VM_FREELIST_DEFAULT);
|
|
|
|
|
} else {
|
|
|
|
|
vm_phys_create_seg(phys_avail[i],
|
|
|
|
|
phys_avail[i + 1], VM_FREELIST_ISADMA);
|
|
|
|
|
}
|
|
|
|
|
if (VM_FREELIST_ISADMA >= vm_nfreelists)
|
|
|
|
|
vm_nfreelists = VM_FREELIST_ISADMA + 1;
|
|
|
|
|
} else
|
|
|
|
|
#endif
|
|
|
|
|
#ifdef VM_FREELIST_HIGHMEM
|
|
|
|
|
if (phys_avail[i + 1] > VM_HIGHMEM_ADDRESS) {
|
|
|
|
|
if (phys_avail[i] < VM_HIGHMEM_ADDRESS) {
|
|
|
|
|
vm_phys_create_seg(phys_avail[i],
|
|
|
|
|
VM_HIGHMEM_ADDRESS, VM_FREELIST_DEFAULT);
|
|
|
|
|
vm_phys_create_seg(VM_HIGHMEM_ADDRESS,
|
|
|
|
|
phys_avail[i + 1], VM_FREELIST_HIGHMEM);
|
|
|
|
|
} else {
|
|
|
|
|
vm_phys_create_seg(phys_avail[i],
|
|
|
|
|
phys_avail[i + 1], VM_FREELIST_HIGHMEM);
|
|
|
|
|
}
|
|
|
|
|
if (VM_FREELIST_HIGHMEM >= vm_nfreelists)
|
|
|
|
|
vm_nfreelists = VM_FREELIST_HIGHMEM + 1;
|
|
|
|
|
} else
|
|
|
|
|
#endif
|
|
|
|
|
vm_phys_create_seg(phys_avail[i], phys_avail[i + 1],
|
|
|
|
|
VM_FREELIST_DEFAULT);
|
|
|
|
|
}
|
|
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++) {
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++) {
|
|
|
|
|
fl = vm_phys_free_queues[flind][pind];
|
|
|
|
|
for (oind = 0; oind < VM_NFREEORDER; oind++)
|
|
|
|
|
TAILQ_INIT(&fl[oind].pl);
|
|
|
|
|
}
|
|
|
|
|
}
|
2010-07-27 20:33:50 +00:00
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
/*
|
|
|
|
|
* Build a free list lookup list for each domain. All of the
|
|
|
|
|
* memory domain lists are inserted at the VM_FREELIST_DEFAULT
|
|
|
|
|
* index in a round-robin order starting with the current
|
|
|
|
|
* domain.
|
|
|
|
|
*/
|
|
|
|
|
ndomains = vm_nfreelists - VM_NFREELIST + 1;
|
|
|
|
|
for (flind = 0; flind < VM_FREELIST_DEFAULT; flind++)
|
|
|
|
|
for (i = 0; i < ndomains; i++)
|
|
|
|
|
vm_phys_lookup_lists[i][flind] =
|
|
|
|
|
&vm_phys_free_queues[flind];
|
|
|
|
|
for (i = 0; i < ndomains; i++)
|
|
|
|
|
for (j = 0; j < ndomains; j++) {
|
|
|
|
|
flind = (i + j) % ndomains;
|
|
|
|
|
if (flind == 0)
|
|
|
|
|
flind = VM_FREELIST_DEFAULT;
|
|
|
|
|
else
|
|
|
|
|
flind += VM_NFREELIST - 1;
|
|
|
|
|
vm_phys_lookup_lists[i][VM_FREELIST_DEFAULT + j] =
|
|
|
|
|
&vm_phys_free_queues[flind];
|
|
|
|
|
}
|
|
|
|
|
for (flind = VM_FREELIST_DEFAULT + 1; flind < VM_NFREELIST;
|
|
|
|
|
flind++)
|
|
|
|
|
for (i = 0; i < ndomains; i++)
|
|
|
|
|
vm_phys_lookup_lists[i][flind + ndomains - 1] =
|
|
|
|
|
&vm_phys_free_queues[flind];
|
|
|
|
|
#else
|
|
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++)
|
|
|
|
|
vm_phys_lookup_lists[0][flind] = &vm_phys_free_queues[flind];
|
|
|
|
|
#endif
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Split a contiguous, power of two-sized set of physical pages.
|
|
|
|
|
*/
|
|
|
|
|
static __inline void
|
|
|
|
|
vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
|
|
|
|
|
{
|
|
|
|
|
vm_page_t m_buddy;
|
|
|
|
|
|
|
|
|
|
while (oind > order) {
|
|
|
|
|
oind--;
|
|
|
|
|
m_buddy = &m[1 << oind];
|
|
|
|
|
KASSERT(m_buddy->order == VM_NFREEORDER,
|
|
|
|
|
("vm_phys_split_pages: page %p has unexpected order %d",
|
|
|
|
|
m_buddy, m_buddy->order));
|
|
|
|
|
m_buddy->order = oind;
|
|
|
|
|
TAILQ_INSERT_HEAD(&fl[oind].pl, m_buddy, pageq);
|
|
|
|
|
fl[oind].lcnt++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialize a physical page and add it to the free lists.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
vm_phys_add_page(vm_paddr_t pa)
|
|
|
|
|
{
|
|
|
|
|
vm_page_t m;
|
|
|
|
|
|
|
|
|
|
cnt.v_page_count++;
|
|
|
|
|
m = vm_phys_paddr_to_vm_page(pa);
|
|
|
|
|
m->phys_addr = pa;
|
2011-01-17 19:17:26 +00:00
|
|
|
m->queue = PQ_NONE;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
m->segind = vm_phys_paddr_to_segind(pa);
|
|
|
|
|
m->flags = PG_FREE;
|
|
|
|
|
KASSERT(m->order == VM_NFREEORDER,
|
|
|
|
|
("vm_phys_add_page: page %p has unexpected order %d",
|
|
|
|
|
m, m->order));
|
|
|
|
|
m->pool = VM_FREEPOOL_DEFAULT;
|
|
|
|
|
pmap_page_init(m);
|
|
|
|
|
mtx_lock(&vm_page_queue_free_mtx);
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
cnt.v_free_count++;
|
2007-07-14 21:21:17 +00:00
|
|
|
vm_phys_free_pages(m, 0);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
mtx_unlock(&vm_page_queue_free_mtx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Allocate a contiguous, power of two-sized set of physical pages
|
|
|
|
|
* from the free lists.
|
2007-07-14 21:21:17 +00:00
|
|
|
*
|
|
|
|
|
* The free page queues must be locked.
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
*/
|
|
|
|
|
vm_page_t
|
2007-07-14 21:21:17 +00:00
|
|
|
vm_phys_alloc_pages(int pool, int order)
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
{
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
vm_page_t m;
|
|
|
|
|
int flind;
|
|
|
|
|
|
|
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++) {
|
|
|
|
|
m = vm_phys_alloc_freelist_pages(flind, pool, order);
|
|
|
|
|
if (m != NULL)
|
|
|
|
|
return (m);
|
|
|
|
|
}
|
|
|
|
|
return (NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Find and dequeue a free page on the given free list, with the
|
|
|
|
|
* specified pool and order
|
|
|
|
|
*/
|
|
|
|
|
vm_page_t
|
|
|
|
|
vm_phys_alloc_freelist_pages(int flind, int pool, int order)
|
|
|
|
|
{
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
struct vm_freelist *alt;
|
2010-07-27 20:33:50 +00:00
|
|
|
int domain, oind, pind;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
vm_page_t m;
|
|
|
|
|
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
KASSERT(flind < VM_NFREELIST,
|
|
|
|
|
("vm_phys_alloc_freelist_pages: freelist %d is out of range", flind));
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
KASSERT(pool < VM_NFREEPOOL,
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
KASSERT(order < VM_NFREEORDER,
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
("vm_phys_alloc_freelist_pages: order %d is out of range", order));
|
2010-07-27 20:33:50 +00:00
|
|
|
|
|
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
domain = PCPU_GET(domain);
|
|
|
|
|
#else
|
|
|
|
|
domain = 0;
|
|
|
|
|
#endif
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
|
2010-07-27 20:33:50 +00:00
|
|
|
fl = (*vm_phys_lookup_lists[domain][flind])[pool];
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
for (oind = order; oind < VM_NFREEORDER; oind++) {
|
|
|
|
|
m = TAILQ_FIRST(&fl[oind].pl);
|
|
|
|
|
if (m != NULL) {
|
|
|
|
|
TAILQ_REMOVE(&fl[oind].pl, m, pageq);
|
|
|
|
|
fl[oind].lcnt--;
|
|
|
|
|
m->order = VM_NFREEORDER;
|
|
|
|
|
vm_phys_split_pages(m, oind, fl, order);
|
|
|
|
|
return (m);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The given pool was empty. Find the largest
|
|
|
|
|
* contiguous, power-of-two-sized set of pages in any
|
|
|
|
|
* pool. Transfer these pages to the given pool, and
|
|
|
|
|
* use them to satisfy the allocation.
|
|
|
|
|
*/
|
|
|
|
|
for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++) {
|
2010-07-27 20:33:50 +00:00
|
|
|
alt = (*vm_phys_lookup_lists[domain][flind])[pind];
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
m = TAILQ_FIRST(&alt[oind].pl);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
if (m != NULL) {
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
TAILQ_REMOVE(&alt[oind].pl, m, pageq);
|
|
|
|
|
alt[oind].lcnt--;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
m->order = VM_NFREEORDER;
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
vm_phys_set_pool(pool, m, oind);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
vm_phys_split_pages(m, oind, fl, order);
|
|
|
|
|
return (m);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return (NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Allocate physical memory from phys_avail[].
|
|
|
|
|
*/
|
|
|
|
|
vm_paddr_t
|
|
|
|
|
vm_phys_bootstrap_alloc(vm_size_t size, unsigned long alignment)
|
|
|
|
|
{
|
|
|
|
|
vm_paddr_t pa;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
size = round_page(size);
|
|
|
|
|
for (i = 0; phys_avail[i + 1] != 0; i += 2) {
|
|
|
|
|
if (phys_avail[i + 1] - phys_avail[i] < size)
|
|
|
|
|
continue;
|
|
|
|
|
pa = phys_avail[i];
|
|
|
|
|
phys_avail[i] += size;
|
|
|
|
|
return (pa);
|
|
|
|
|
}
|
|
|
|
|
panic("vm_phys_bootstrap_alloc");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Find the vm_page corresponding to the given physical address.
|
|
|
|
|
*/
|
|
|
|
|
vm_page_t
|
|
|
|
|
vm_phys_paddr_to_vm_page(vm_paddr_t pa)
|
|
|
|
|
{
|
|
|
|
|
struct vm_phys_seg *seg;
|
|
|
|
|
int segind;
|
|
|
|
|
|
|
|
|
|
for (segind = 0; segind < vm_phys_nsegs; segind++) {
|
|
|
|
|
seg = &vm_phys_segs[segind];
|
|
|
|
|
if (pa >= seg->start && pa < seg->end)
|
|
|
|
|
return (&seg->first_page[atop(pa - seg->start)]);
|
|
|
|
|
}
|
2009-06-18 20:42:37 +00:00
|
|
|
return (NULL);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Find the segment containing the given physical address.
|
|
|
|
|
*/
|
|
|
|
|
static int
|
|
|
|
|
vm_phys_paddr_to_segind(vm_paddr_t pa)
|
|
|
|
|
{
|
|
|
|
|
struct vm_phys_seg *seg;
|
|
|
|
|
int segind;
|
|
|
|
|
|
|
|
|
|
for (segind = 0; segind < vm_phys_nsegs; segind++) {
|
|
|
|
|
seg = &vm_phys_segs[segind];
|
|
|
|
|
if (pa >= seg->start && pa < seg->end)
|
|
|
|
|
return (segind);
|
|
|
|
|
}
|
|
|
|
|
panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
|
|
|
|
|
(uintmax_t)pa);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Free a contiguous, power of two-sized set of physical pages.
|
2007-07-14 21:21:17 +00:00
|
|
|
*
|
|
|
|
|
* The free page queues must be locked.
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
vm_phys_free_pages(vm_page_t m, int order)
|
|
|
|
|
{
|
|
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
struct vm_phys_seg *seg;
|
|
|
|
|
vm_paddr_t pa, pa_buddy;
|
|
|
|
|
vm_page_t m_buddy;
|
|
|
|
|
|
|
|
|
|
KASSERT(m->order == VM_NFREEORDER,
|
2007-07-14 21:21:17 +00:00
|
|
|
("vm_phys_free_pages: page %p has unexpected order %d",
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
m, m->order));
|
|
|
|
|
KASSERT(m->pool < VM_NFREEPOOL,
|
2007-07-14 21:21:17 +00:00
|
|
|
("vm_phys_free_pages: page %p has unexpected pool %d",
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
m, m->pool));
|
|
|
|
|
KASSERT(order < VM_NFREEORDER,
|
2007-07-14 21:21:17 +00:00
|
|
|
("vm_phys_free_pages: order %d is out of range", order));
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
|
|
|
|
|
pa = VM_PAGE_TO_PHYS(m);
|
|
|
|
|
seg = &vm_phys_segs[m->segind];
|
|
|
|
|
while (order < VM_NFREEORDER - 1) {
|
|
|
|
|
pa_buddy = pa ^ (1 << (PAGE_SHIFT + order));
|
|
|
|
|
if (pa_buddy < seg->start ||
|
|
|
|
|
pa_buddy >= seg->end)
|
|
|
|
|
break;
|
|
|
|
|
m_buddy = &seg->first_page[atop(pa_buddy - seg->start)];
|
|
|
|
|
if (m_buddy->order != order)
|
|
|
|
|
break;
|
|
|
|
|
fl = (*seg->free_queues)[m_buddy->pool];
|
|
|
|
|
TAILQ_REMOVE(&fl[m_buddy->order].pl, m_buddy, pageq);
|
|
|
|
|
fl[m_buddy->order].lcnt--;
|
|
|
|
|
m_buddy->order = VM_NFREEORDER;
|
|
|
|
|
if (m_buddy->pool != m->pool)
|
|
|
|
|
vm_phys_set_pool(m->pool, m_buddy, order);
|
|
|
|
|
order++;
|
|
|
|
|
pa &= ~((1 << (PAGE_SHIFT + order)) - 1);
|
|
|
|
|
m = &seg->first_page[atop(pa - seg->start)];
|
|
|
|
|
}
|
|
|
|
|
m->order = order;
|
|
|
|
|
fl = (*seg->free_queues)[m->pool];
|
|
|
|
|
TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
|
|
|
|
|
fl[order].lcnt++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Set the pool for a contiguous, power of two-sized set of physical pages.
|
|
|
|
|
*/
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
void
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
vm_phys_set_pool(int pool, vm_page_t m, int order)
|
|
|
|
|
{
|
|
|
|
|
vm_page_t m_tmp;
|
|
|
|
|
|
|
|
|
|
for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
|
|
|
|
|
m_tmp->pool = pool;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2007-12-21 02:44:31 +00:00
|
|
|
* Search for the given physical page "m" in the free lists. If the search
|
|
|
|
|
* succeeds, remove "m" from the free lists and return TRUE. Otherwise, return
|
|
|
|
|
* FALSE, indicating that "m" is not in the free lists.
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
*
|
|
|
|
|
* The free page queues must be locked.
|
|
|
|
|
*/
|
2007-12-20 22:45:54 +00:00
|
|
|
boolean_t
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
vm_phys_unfree_page(vm_page_t m)
|
|
|
|
|
{
|
|
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
struct vm_phys_seg *seg;
|
|
|
|
|
vm_paddr_t pa, pa_half;
|
|
|
|
|
vm_page_t m_set, m_tmp;
|
|
|
|
|
int order;
|
|
|
|
|
|
|
|
|
|
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* First, find the contiguous, power of two-sized set of free
|
|
|
|
|
* physical pages containing the given physical page "m" and
|
|
|
|
|
* assign it to "m_set".
|
|
|
|
|
*/
|
|
|
|
|
seg = &vm_phys_segs[m->segind];
|
|
|
|
|
for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
|
2007-12-19 23:09:45 +00:00
|
|
|
order < VM_NFREEORDER - 1; ) {
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
order++;
|
|
|
|
|
pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
|
2008-04-05 05:02:53 +00:00
|
|
|
if (pa >= seg->start)
|
2007-12-20 22:45:54 +00:00
|
|
|
m_set = &seg->first_page[atop(pa - seg->start)];
|
|
|
|
|
else
|
|
|
|
|
return (FALSE);
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
}
|
2007-12-20 22:45:54 +00:00
|
|
|
if (m_set->order < order)
|
|
|
|
|
return (FALSE);
|
|
|
|
|
if (m_set->order == VM_NFREEORDER)
|
|
|
|
|
return (FALSE);
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
KASSERT(m_set->order < VM_NFREEORDER,
|
|
|
|
|
("vm_phys_unfree_page: page %p has unexpected order %d",
|
|
|
|
|
m_set, m_set->order));
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Next, remove "m_set" from the free lists. Finally, extract
|
|
|
|
|
* "m" from "m_set" using an iterative algorithm: While "m_set"
|
|
|
|
|
* is larger than a page, shrink "m_set" by returning the half
|
|
|
|
|
* of "m_set" that does not contain "m" to the free lists.
|
|
|
|
|
*/
|
|
|
|
|
fl = (*seg->free_queues)[m_set->pool];
|
|
|
|
|
order = m_set->order;
|
|
|
|
|
TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
|
|
|
|
|
fl[order].lcnt--;
|
|
|
|
|
m_set->order = VM_NFREEORDER;
|
|
|
|
|
while (order > 0) {
|
|
|
|
|
order--;
|
|
|
|
|
pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
|
|
|
|
|
if (m->phys_addr < pa_half)
|
|
|
|
|
m_tmp = &seg->first_page[atop(pa_half - seg->start)];
|
|
|
|
|
else {
|
|
|
|
|
m_tmp = m_set;
|
|
|
|
|
m_set = &seg->first_page[atop(pa_half - seg->start)];
|
|
|
|
|
}
|
|
|
|
|
m_tmp->order = order;
|
|
|
|
|
TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
|
|
|
|
|
fl[order].lcnt++;
|
|
|
|
|
}
|
|
|
|
|
KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
|
2007-12-20 22:45:54 +00:00
|
|
|
return (TRUE);
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Try to zero one physical page. Used by an idle priority thread.
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
*/
|
|
|
|
|
boolean_t
|
|
|
|
|
vm_phys_zero_pages_idle(void)
|
|
|
|
|
{
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
static struct vm_freelist *fl = vm_phys_free_queues[0][0];
|
|
|
|
|
static int flind, oind, pind;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
vm_page_t m, m_tmp;
|
|
|
|
|
|
|
|
|
|
mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
for (;;) {
|
|
|
|
|
TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
|
|
|
|
|
for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
|
|
|
|
|
if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
|
|
|
|
|
vm_phys_unfree_page(m_tmp);
|
|
|
|
|
cnt.v_free_count--;
|
|
|
|
|
mtx_unlock(&vm_page_queue_free_mtx);
|
|
|
|
|
pmap_zero_page_idle(m_tmp);
|
|
|
|
|
m_tmp->flags |= PG_ZERO;
|
|
|
|
|
mtx_lock(&vm_page_queue_free_mtx);
|
|
|
|
|
cnt.v_free_count++;
|
|
|
|
|
vm_phys_free_pages(m_tmp, 0);
|
|
|
|
|
vm_page_zero_count++;
|
|
|
|
|
cnt_prezero++;
|
|
|
|
|
return (TRUE);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
oind++;
|
|
|
|
|
if (oind == VM_NFREEORDER) {
|
|
|
|
|
oind = 0;
|
|
|
|
|
pind++;
|
|
|
|
|
if (pind == VM_NFREEPOOL) {
|
|
|
|
|
pind = 0;
|
|
|
|
|
flind++;
|
|
|
|
|
if (flind == vm_nfreelists)
|
|
|
|
|
flind = 0;
|
|
|
|
|
}
|
|
|
|
|
fl = vm_phys_free_queues[flind][pind];
|
|
|
|
|
}
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2007-06-16 05:25:53 +00:00
|
|
|
* Allocate a contiguous set of physical pages of the given size
|
|
|
|
|
* "npages" from the free lists. 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
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
* "alignment" and "boundary" must be a power of two.
|
|
|
|
|
*/
|
|
|
|
|
vm_page_t
|
|
|
|
|
vm_phys_alloc_contig(unsigned long npages, vm_paddr_t low, vm_paddr_t high,
|
2009-07-12 23:31:20 +00:00
|
|
|
unsigned long alignment, unsigned long boundary)
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
{
|
|
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
struct vm_phys_seg *seg;
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
struct vnode *vp;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
vm_paddr_t pa, pa_last, size;
|
2009-06-21 20:29:14 +00:00
|
|
|
vm_page_t deferred_vdrop_list, m, m_ret;
|
2010-07-27 20:33:50 +00:00
|
|
|
int domain, flind, i, oind, order, pind;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
|
2010-07-27 20:33:50 +00:00
|
|
|
#if VM_NDOMAIN > 1
|
|
|
|
|
domain = PCPU_GET(domain);
|
|
|
|
|
#else
|
|
|
|
|
domain = 0;
|
|
|
|
|
#endif
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
size = npages << PAGE_SHIFT;
|
|
|
|
|
KASSERT(size != 0,
|
|
|
|
|
("vm_phys_alloc_contig: size must not be 0"));
|
|
|
|
|
KASSERT((alignment & (alignment - 1)) == 0,
|
|
|
|
|
("vm_phys_alloc_contig: alignment must be a power of 2"));
|
|
|
|
|
KASSERT((boundary & (boundary - 1)) == 0,
|
|
|
|
|
("vm_phys_alloc_contig: boundary must be a power of 2"));
|
2009-06-21 20:29:14 +00:00
|
|
|
deferred_vdrop_list = NULL;
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
/* Compute the queue that is the best fit for npages. */
|
|
|
|
|
for (order = 0; (1 << order) < npages; order++);
|
|
|
|
|
mtx_lock(&vm_page_queue_free_mtx);
|
2008-04-06 18:09:28 +00:00
|
|
|
#if VM_NRESERVLEVEL > 0
|
|
|
|
|
retry:
|
|
|
|
|
#endif
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++) {
|
|
|
|
|
for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++) {
|
2010-07-27 20:33:50 +00:00
|
|
|
fl = (*vm_phys_lookup_lists[domain][flind])
|
|
|
|
|
[pind];
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
|
|
|
|
|
/*
|
|
|
|
|
* A free list may contain physical pages
|
|
|
|
|
* from one or more segments.
|
|
|
|
|
*/
|
|
|
|
|
seg = &vm_phys_segs[m_ret->segind];
|
|
|
|
|
if (seg->start > high ||
|
|
|
|
|
low >= seg->end)
|
|
|
|
|
continue;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Is the size of this allocation request
|
|
|
|
|
* larger than the largest block size?
|
|
|
|
|
*/
|
|
|
|
|
if (order >= VM_NFREEORDER) {
|
|
|
|
|
/*
|
|
|
|
|
* Determine if a sufficient number
|
|
|
|
|
* of subsequent blocks to satisfy
|
|
|
|
|
* the allocation request are free.
|
|
|
|
|
*/
|
|
|
|
|
pa = VM_PAGE_TO_PHYS(m_ret);
|
|
|
|
|
pa_last = pa + size;
|
|
|
|
|
for (;;) {
|
|
|
|
|
pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
|
|
|
|
|
if (pa >= pa_last)
|
|
|
|
|
break;
|
|
|
|
|
if (pa < seg->start ||
|
|
|
|
|
pa >= seg->end)
|
|
|
|
|
break;
|
|
|
|
|
m = &seg->first_page[atop(pa - seg->start)];
|
|
|
|
|
if (m->order != VM_NFREEORDER - 1)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
/* If not, continue to the next block. */
|
|
|
|
|
if (pa < pa_last)
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Determine if the blocks are within the given range,
|
|
|
|
|
* satisfy the given alignment, and do not cross the
|
|
|
|
|
* given boundary.
|
|
|
|
|
*/
|
|
|
|
|
pa = VM_PAGE_TO_PHYS(m_ret);
|
|
|
|
|
if (pa >= low &&
|
|
|
|
|
pa + size <= high &&
|
|
|
|
|
(pa & (alignment - 1)) == 0 &&
|
|
|
|
|
((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
|
|
|
|
|
goto done;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
2008-04-06 18:09:28 +00:00
|
|
|
#if VM_NRESERVLEVEL > 0
|
|
|
|
|
if (vm_reserv_reclaim_contig(size, low, high, alignment, boundary))
|
|
|
|
|
goto retry;
|
|
|
|
|
#endif
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
mtx_unlock(&vm_page_queue_free_mtx);
|
|
|
|
|
return (NULL);
|
|
|
|
|
done:
|
|
|
|
|
for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
|
|
|
|
|
fl = (*seg->free_queues)[m->pool];
|
|
|
|
|
TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
|
|
|
|
|
fl[m->order].lcnt--;
|
|
|
|
|
m->order = VM_NFREEORDER;
|
|
|
|
|
}
|
|
|
|
|
if (m_ret->pool != VM_FREEPOOL_DEFAULT)
|
|
|
|
|
vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
|
|
|
|
|
fl = (*seg->free_queues)[m_ret->pool];
|
|
|
|
|
vm_phys_split_pages(m_ret, oind, fl, order);
|
|
|
|
|
for (i = 0; i < npages; i++) {
|
|
|
|
|
m = &m_ret[i];
|
Redo the page table page allocation on MIPS, as suggested by
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
2010-07-21 09:27:00 +00:00
|
|
|
vp = vm_page_alloc_init(m);
|
|
|
|
|
if (vp != NULL) {
|
|
|
|
|
/*
|
|
|
|
|
* Enqueue the vnode for deferred vdrop().
|
|
|
|
|
*
|
|
|
|
|
* Unmanaged pages don't use "pageq", so it
|
|
|
|
|
* can be safely abused to construct a short-
|
|
|
|
|
* lived queue of vnodes.
|
|
|
|
|
*/
|
|
|
|
|
m->pageq.tqe_prev = (void *)vp;
|
|
|
|
|
m->pageq.tqe_next = deferred_vdrop_list;
|
|
|
|
|
deferred_vdrop_list = m;
|
Change the management of cached pages (PQ_CACHE) in two fundamental
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
2007-09-25 06:25:06 +00:00
|
|
|
}
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
for (; i < roundup2(npages, 1 << imin(oind, order)); i++) {
|
|
|
|
|
m = &m_ret[i];
|
|
|
|
|
KASSERT(m->order == VM_NFREEORDER,
|
|
|
|
|
("vm_phys_alloc_contig: page %p has unexpected order %d",
|
|
|
|
|
m, m->order));
|
2007-07-14 21:21:17 +00:00
|
|
|
vm_phys_free_pages(m, 0);
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
}
|
|
|
|
|
mtx_unlock(&vm_page_queue_free_mtx);
|
2009-06-21 20:29:14 +00:00
|
|
|
while (deferred_vdrop_list != NULL) {
|
|
|
|
|
vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
|
|
|
|
|
deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
|
|
|
|
|
}
|
Add a new physical memory allocator. However, do not yet connect it
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
2007-06-10 00:49:16 +00:00
|
|
|
return (m_ret);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#ifdef DDB
|
|
|
|
|
/*
|
|
|
|
|
* Show the number of physical pages in each of the free lists.
|
|
|
|
|
*/
|
|
|
|
|
DB_SHOW_COMMAND(freepages, db_show_freepages)
|
|
|
|
|
{
|
|
|
|
|
struct vm_freelist *fl;
|
|
|
|
|
int flind, oind, pind;
|
|
|
|
|
|
|
|
|
|
for (flind = 0; flind < vm_nfreelists; flind++) {
|
|
|
|
|
db_printf("FREE LIST %d:\n"
|
|
|
|
|
"\n ORDER (SIZE) | NUMBER"
|
|
|
|
|
"\n ", flind);
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++)
|
|
|
|
|
db_printf(" | POOL %d", pind);
|
|
|
|
|
db_printf("\n-- ");
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++)
|
|
|
|
|
db_printf("-- -- ");
|
|
|
|
|
db_printf("--\n");
|
|
|
|
|
for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
|
|
|
|
|
db_printf(" %2.2d (%6.6dK)", oind,
|
|
|
|
|
1 << (PAGE_SHIFT - 10 + oind));
|
|
|
|
|
for (pind = 0; pind < VM_NFREEPOOL; pind++) {
|
|
|
|
|
fl = vm_phys_free_queues[flind][pind];
|
|
|
|
|
db_printf(" | %6.6d", fl[oind].lcnt);
|
|
|
|
|
}
|
|
|
|
|
db_printf("\n");
|
|
|
|
|
}
|
|
|
|
|
db_printf("\n");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
#endif
|
