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)
vm_phys_free_pages(). Rename vm_phys_alloc_pages_locked() to
vm_phys_alloc_pages() and vm_phys_free_pages_locked() to
vm_phys_free_pages(). Add comments regarding the need for the free page
queues lock to be held by callers to these functions. No functional
changes.
Approved by: re (hrs)
vm_page_cowfault(). Initially, if vm_page_cowfault() sleeps, the given
page is wired, preventing it from being recycled. However, when
transmission of the page completes, the page is unwired and returned to
the page queues. At that point, the page is not in any special state
that prevents it from being recycled. Consequently, vm_page_cowfault()
should verify that the page is still held by the same vm object before
retrying the replacement of the page. Note: The containing object is,
however, safe from being recycled by virtue of having a non-zero
paging-in-progress count.
While I'm here, add some assertions and comments.
Approved by: re (rwatson)
MFC After: 3 weeks
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
In particular:
- Add an explicative table for locking of struct vmmeter members
- Apply new rules for some of those members
- Remove some unuseful comments
Heavily reviewed by: alc, bde, jeff
Approved by: jeff (mentor)
Now, we assume no more sched_lock protection for some of them and use the
distribuited loads method for vmmeter (distribuited through CPUs).
Reviewed by: alc, bde
Approved by: jeff (mentor)
Probabilly, a general approach is not the better solution here, so we should
solve the sched_lock protection problems separately.
Requested by: alc
Approved by: jeff (mentor)
vmcnts. This can be used to abstract away pcpu details but also changes
to use atomics for all counters now. This means sched lock is no longer
responsible for protecting counts in the switch routines.
Contributed by: Attilio Rao <attilio@FreeBSD.org>
VM_PHYSSEG_SPARSE depending on whether the physical address space is
densely or sparsely populated with memory. The effect of this
definition is to determine which of two implementations of
vm_page_array and PHYS_TO_VM_PAGE() is used. The legacy
implementation is obtained by defining VM_PHYSSEG_DENSE, and a new
implementation that trades off time for space is obtained by defining
VM_PHYSSEG_SPARSE. For now, all architectures except for ia64 and
sparc64 define VM_PHYSSEG_DENSE. Defining VM_PHYSSEG_SPARSE on ia64
allows the entirety of my Itanium 2's memory to be used. Previously,
only the first 1 GB could be used. Defining VM_PHYSSEG_SPARSE on
sparc64 allows USIIIi-based systems to boot without crashing.
This change is a combination of Nathan Whitehorn's patch and my own
work in perforce.
Discussed with: kmacy, marius, Nathan Whitehorn
PR: 112194
immediately flag any page that is allocated to a OBJT_PHYS object as
unmanaged in vm_page_alloc() rather than waiting for a later call to
vm_page_unmanage(). This allows for the elimination of some uses of
the page queues lock.
Change the type of the kernel and kmem objects from OBJT_DEFAULT to
OBJT_PHYS. This allows us to take advantage of the above change to
simplify the allocation of unmanaged pages in kmem_alloc() and
kmem_malloc().
Remove vm_page_unmanage(). It is no longer used.
vm_page_free_toq() to account for recent changes that allow
vm_page_free_toq() to be called on some pages without the page queues lock
being held, specifically, pages that are not contained in a vm object and
not a member of a page queue. (Examples of such pages include page table
pages, pv entry pages, and uma small alloc pages.)
is actually being added to the hold queue, not the free queue. At the same
time, avoid unnecessary tests to wake up threads waiting for free memory
and the idle thread that zeroes free pages. (These tests will be performed
later when the page finally moves from the hold queue to the free queue.)
inlined and a procedure call is made in the rare case, i.e., when it is
necessary to sleep. In this case, inlining the test actually makes the
kernel smaller.
page queues-synchronized flag. Reduce the scope of the page queues lock in
vm_fault() accordingly.
Move vm_fault()'s call to vm_object_set_writeable_dirty() outside of the
scope of the page queues lock. Reviewed by: tegge
Additionally, eliminate an unnecessary dereference in computing the
argument that is passed to vm_object_set_writeable_dirty().
synchronized by the lock on the object containing the page.
Transition PG_WANTED and PG_SWAPINPROG to use the new field,
eliminating the need for holding the page queues lock when setting
or clearing these flags. Rename PG_WANTED and PG_SWAPINPROG to
VPO_WANTED and VPO_SWAPINPROG, respectively.
Eliminate the assertion that the page queues lock is held in
vm_page_io_finish().
Eliminate the acquisition and release of the page queues lock
around calls to vm_page_io_finish() in kern_sendfile() and
vfs_unbusy_pages().
Originally, I had adopted sparc64's name, pmap_clear_write(), for the
function that is now pmap_remove_write(). However, this function is more
like pmap_remove_all() than like pmap_clear_modify() or
pmap_clear_reference(), hence, the name change.
The higher-level rationale behind this change is described in
src/sys/amd64/amd64/pmap.c revision 1.567. The short version is that I'm
trying to clean up and fix our support for execute access.
Reviewed by: marcel@ (ia64)
vm_page_startup(). As a result, we now only lookup the tunable once
instead of looking it up once for every physical page of memory in the
system. This cuts out about a 1 second or so delay in boot on x86
systems. The delay is much larger and more noticable on sun4v apparently.
Reported by: kmacy
MFC after: 1 week
These pages are allocated from the direct map, and were not previous
tracked. This included the vm_page_array and the early UMA bootstrap
pages.
Reviewed by: peter
via the debug.minidump sysctl and tunable.
Traditional dumps store all physical memory. This was once a good thing
when machines had a maximum of 64M of ram and 1GB of kvm. These days,
machines often have many gigabytes of ram and a smaller amount of kvm.
libkvm+kgdb don't have a way to access physical ram that is not mapped
into kvm at the time of the crash dump, so the extra ram being dumped
is mostly wasted.
Minidumps invert the process. Instead of dumping physical memory in
in order to guarantee that all of kvm's backing is dumped, minidumps
instead dump only memory that is actively mapped into kvm.
amd64 has a direct map region that things like UMA use. Obviously we
cannot dump all of the direct map region because that is effectively
an old style all-physical-memory dump. Instead, introduce a bitmap
and two helper routines (dump_add_page(pa) and dump_drop_page(pa)) that
allow certain critical direct map pages to be included in the dump.
uma_machdep.c's allocator is the intended consumer.
Dumps are a custom format. At the very beginning of the file is a header,
then a copy of the message buffer, then the bitmap of pages present in
the dump, then the final level of the kvm page table trees (2MB mappings
are expanded into a 4K page mappings), then the sparse physical pages
according to the bitmap. libkvm can now conveniently access the kvm
page table entries.
Booting my test 8GB machine, forcing it into ddb and forcing a dump
leads to a 48MB minidump. While this is a best case, I expect minidumps
to be in the 100MB-500MB range. Obviously, never larger than physical
memory of course.
minidumps are on by default. It would want be necessary to turn them off
if it was necessary to debug corrupt kernel page table management as that
would mess up minidumps as well.
Both minidumps and regular dumps are supported on the same machine.
and it has not plenty of free pages it tries to free pages in the cache queue.
Unfortunately freeing a cached page requires the locking of the object that
owns the page. However in the context of allocating pages we may not be able
to lock the object and thus can only TRY to lock the object. If the locking try
fails the cache page can not be freed and is activated to move it out of the way
so that we may try to free other cache pages.
If all pages in the cache belong to objects that are currently locked the
cache queue can be emptied without freeing a single page. This scenario caused
two problems:
1) vm_page_alloc always failed allocation when it tried freeing pages from
the cache queue and failed to do so. However if there are more than
cnt.v_interrupt_free_min pages on the free list it should return pages
when requested with priority VM_ALLOC_SYSTEM. Failure to do so can cause
resource exhaustion deadlocks.
2) Threads than need to allocate pages spend a lot of time cleaning up the
page queue without really getting anything done while the pagedaemon
needs to work overtime to refill the cache.
This change fixes the first problem. (1)
Reviewed by: tegge@
- provide an interface (macros) to the page coloring part of the VM system,
this allows to try different coloring algorithms without the need to
touch every file [1]
- make the page queue tuning values readable: sysctl vm.stats.pagequeue
- autotuning of the page coloring values based upon the cache size instead
of options in the kernel config (disabling of the page coloring as a
kernel option is still possible)
MD changes:
- detection of the cache size: only IA32 and AMD64 (untested) contains
cache size detection code, every other arch just comes with a dummy
function (this results in the use of default values like it was the
case without the autotuning of the page coloring)
- print some more info on Intel CPU's (like we do on AMD and Transmeta
CPU's)
Note to AMD owners (IA32 and AMD64): please run "sysctl vm.stats.pagequeue"
and report if the cache* values are zero (= bug in the cache detection code)
or not.
Based upon work by: Chad David <davidc@acns.ab.ca> [1]
Reviewed by: alc, arch (in 2004)
Discussed with: alc, Chad David, arch (in 2004)
by the zero-copy sockets method, and written to before the transmission
completes, we need to destroy all of the existing mappings to the page,
not just the one that we fault on. Otherwise, the mappings will no longer
be to the same page and changes made through one of the mappings will not
be visible through the others.
Observed by: tegge
If a copy-on-write fault occurs on the page, the new copy should inherit
a part of the original page's wire count.
Submitted by: tegge
MFC after: 1 week
