page zeroing thread before it has been created. It was possible for
calls to free() very early in the boot process to panic here because
the sleep queues were not yet initialised. Specifically, sysinit_add()
running at SI_SUB_KLD would trigger this if the array of pointers
became big enough to require uma_large_alloc() allocations.
Submitted by: peter
position that is 64-bit aligned and makes sure that the valid and
dirty fields are also 64-bit aligned. This means that if PAGE_SIZE
is 32K, the size of the vm_page structure is only increased by 8
bytes instead of 16 bytes. More importantly, the vm_page structure
is either 120 or 128 bytes on ia64. These are "interesting" sizes.
manipulating a vnode, e.g., calling vput(). This reduces contention for
Giant during many copy-on-write faults, resulting in some additional
speedup on SMPs.
Note: debug_mpsafevm must be enabled for this optimization to take effect.
"debug.mpsafevm" results in (almost) Giant-free execution of zero-fill
page faults. (Giant is held only briefly, just long enough to determine
if there is a vnode backing the faulting address.)
Also, condition the acquisition and release of Giant around calls to
pmap_remove() on "debug.mpsafevm".
The effect on performance is significant. On my dual Opteron, I see a
3.6% reduction in "buildworld" time.
- Use atomic operations to update several counters in vm_fault().
wait for system wires to disappear, do so (much more trivially) by
instead only checking for system wires of user maps and not kernel maps.
Alternative by: tor
Reviewed by: alc
algorithm built into the map entry splay tree. This replaces the
first_free hint in struct vm_map with two fields in vm_map_entry:
adj_free, the amount of free space following a map entry, and
max_free, the maximum amount of free space in the entry's subtree.
These fields make it possible to find a first-fit free region of a
given size in one pass down the tree, so O(log n) amortized using
splay trees.
This significantly reduces the overhead in vm_map_findspace() for
applications that mmap() many hundreds or thousands of regions, and
has a negligible slowdown (0.1%) on buildworld. See, for example, the
discussion of a micro-benchmark titled "Some mmap observations
compared to Linux 2.6/OpenBSD" on -hackers in late October 2003.
OpenBSD adopted this approach in March 2002, and NetBSD added it in
November 2003, both with Red-Black trees.
Submitted by: Mark W. Krentel
to avoid later changes before pmap_enter() and vm_fault_prefault()
has completed.
Simplify deadlock avoidance by not blocking on vm map relookup.
In collaboration with: alc
* Allow no-fault wiring/unwiring to succeed for consistency;
however, the wired count remains at zero, so it's a special case.
* Fix issues inside vm_map_wire() and vm_map_unwire() where the
exact state of user wiring (one or zero) and system wiring
(zero or more) could be confused; for example, system unwiring
could succeed in removing a user wire, instead of being an
error.
* Require all mappings to be unwired before they are deleted.
When VM space is still wired upon deletion, it will be waited
upon for the following unwire. This makes vslock(9) work
rather than allowing kernel-locked memory to be deleted
out from underneath of its consumer as it would before.
1. Move a comment to its proper place, updating it. (Except for white-
space, this comment had been unchanged since revision 1.1!)
2. Remove spl calls.
should be set to VM_PAGE_BITS_ALL before returning, to ensure that
neither vm_pager_get_pages nor vm_fault calls vm_page_zero_invalid
after dev_pager_getpages has returned.
Submitted by: tegge
so that they know whether the allocation is supposed to be able to sleep
or not.
* Allow uma_zone constructors and initialation functions to return either
success or error. Almost all of the ones in the tree currently return
success unconditionally, but mbuf is a notable exception: the packet
zone constructor wants to be able to fail if it cannot suballocate an
mbuf cluster, and the mbuf allocators want to be able to fail in general
in a MAC kernel if the MAC mbuf initializer fails. This fixes the
panics people are seeing when they run out of memory for mbuf clusters.
* Allow debug.nosleepwithlocks on WITNESS to be disabled, without changing
the default.
Both bmilekic and jeff have reviewed the changes made to make failable
zone allocations work.
those architectures without pmap locking.
- Eliminate the acquisition and release of Giant from vm_map_protect().
(Translation: mprotect(2) runs to completion without touching Giant on
alpha, amd64, i386 and ia64.)
maps. We always acquire the sx lock exclusively here, but we can't
use a mutex because we want to be able to sleep while holding the
lock. This is completely equivalent to what we were doing with the
lockmgr(9) locks before.
Approved by: alc
- Enable recursion on the page queues lock. This allows calls to
vm_page_alloc(VM_ALLOC_NORMAL) and UMA's obj_alloc() with the page
queues lock held. Such calls are made to allocate page table pages
and pv entries.
- The previous change enables a partial reversion of vm/vm_page.c
revision 1.216, i.e., the call to vm_page_alloc() by vm_page_cowfault()
now specifies VM_ALLOC_NORMAL rather than VM_ALLOC_INTERRUPT.
- Add partial locking to pmap_copy(). (As a side-effect, pmap_copy()
should now be faster on i386 SMP because it no longer generates IPIs
for TLB shootdown on the other processors.)
- Complete the locking of pmap_enter() and pmap_enter_quick(). (As of now,
all changes to a user-level pmap on alpha, amd64, and i386 are performed
with appropriate locking.)
- zone_large_init() stays pretty much the same.
- zone_small_init() will try to stash the slab header in the slab page
being allocated if the amount of calculated wasted space is less
than UMA_MAX_WASTE (for both the UMA_ZONE_REFCNT case and regular
case). If the amount of wasted space is >= UMA_MAX_WASTE, then
UMA_ZONE_OFFPAGE will be set and the slab header will be allocated
separately for better use of space.
- uma_startup() calculates the maximum ipers required in offpage slabs
(so that the offpage slab header zone(s) can be sized accordingly).
The algorithm used to calculate this replaces the old calculation
(which only happened to work coincidentally). We now iterate over
possible object sizes, starting from the smallest one, until we
determine that wastedspace calculated in zone_small_init() might
end up being greater than UMA_MAX_WASTE, at which point we use the
found object size to compute the maximum possible ipers. The
reason this works is because:
- wastedspace versus objectsize is a see-saw function with
local minima all equal to zero and local maxima growing
directly proportioned to objectsize. This implies that
for objects up to or equal a certain objectsize, the see-saw
remains entirely below UMA_MAX_WASTE, so for those objectsizes
it is impossible to ever go OFFPAGE for slab headers.
- ipers (items-per-slab) versus objectsize is an inversely
proportional function which falls off very quickly (very large
for small objectsizes).
- To determine the maximum ipers we'll ever need from OFFPAGE
slab headers we first find the largest objectsize for which
we are guaranteed to not go offpage for and use it to compute
ipers (as though we were offpage). Since the only objectsizes
allowed to go offpage are bigger than the found objectsize,
and since ipers vs objectsize is inversely proportional (and
monotonically decreasing), then we are guaranteed that the
ipers computed is always >= what we will ever need in offpage
slab headers.
- Define UMA_FRITM_SZ and UMA_FRITMREF_SZ to be the actual (possibly
padded) size of each freelist index so that offset calculations are
fixed.
This might fix weird data corruption problems and certainly allows
ARM to now boot to at least single-user (via simulator).
Tested on i386 UP by me.
Tested on sparc64 SMP by fenner.
Tested on ARM simulator to single-user by cognet.
vm/vm_object.c revision 1.88) and vm_object_sync() (originating in
vm/vm_map.c revision 1.36): When descending a chain of backing objects,
both use the wrong object's backing offset. Consequently, both may
operate on the wrong pages.
Quoting Matt, "This could be responsible for all of the sporatic madvise
oddness that has been reported over the years."
Reviewed by: Matt Dillon
- Push down Giant into shmexit(). (Giant is acquired only if the vmspace
contains shm segments.)
- Eliminate the acquisition of Giant from proc_rwmem().
- Reduce the scope of Giant in exit1(), uncovering the destruction of the
address space.
vmspace to the new vmspace in vmspace_exec() is mostly wasted effort. With
one exception, vm_swrss, the copied fields are immediately overwritten.
Instead, initialize these fields to zero in vmspace_alloc(), eliminating a
bcopy() from vmspace_exec() and a bzero() from vmspace_fork().
kmem_alloc_pageable(). The difference between these is that an errant
memory access to the zone will be detected sooner with
kmem_alloc_nofault().
The following changes serve to eliminate the following lock-order
reversal reported by witness:
1st 0xc1a3c084 vm object (vm object) @ vm/swap_pager.c:1311
2nd 0xc07acb00 swap_pager swhash (swap_pager swhash) @ vm/swap_pager.c:1797
3rd 0xc1804bdc vm object (vm object) @ vm/uma_core.c:931
There is no potential deadlock in this case. However, witness is unable
to recognize this because vm objects used by UMA have the same type as
ordinary vm objects. To remedy this, we make the following changes:
- Add a mutex type argument to VM_OBJECT_LOCK_INIT().
- Use the mutex type argument to assign distinct types to special
vm objects such as the kernel object, kmem object, and UMA objects.
- Define a static swap zone object for use by UMA. (Only static
objects are assigned a special mutex type.)
being incomplete, it currently has to know how to drop and pick back
up the vm_object's mutex if it has to sleep and drop the page queue
mutex. The problem with this is that if the page is busy, while we
are sleeping, the page can be freed and object disappear. When trying
to lock m->object, we'd get a stale or NULL pointer and crash.
The object is now cached, but this makes the assumption that
the object is referenced in some manner and will not itself
disappear while it is unlocked. Since this only happens if
the object is locked, I had to remove an assumption earlier in
contigmalloc() that reversed the order of locking the object and
doing vm_page_sleep_if_busy(), not the normal order.
init and fini handlers. Our vm system removes all userland mappings at
exit prior to calling pmap_release. It just so happens that we might
as well reuse the pmap for the next process since the userland slate
has already been wiped clean.
However. There is a functional benefit to this as well. For platforms
that share userland and kernel context in the same pmap, it means that
the kernel portion of a pmap remains valid after the vmspace has been
freed (process exit) and while it is in uma's cache. This is significant
for i386 SMP systems with kernel context borrowing because it avoids
a LOT of IPIs from the pmap_lazyfix() cleanup in the usual case.
Tested on: amd64, i386, sparc64, alpha
Glanced at by: alc
allocated as "no object" pages. Similar changes were made to the amd64
and i386 pmap last year. The primary reason being that maintaining
a pte object leads to lock order violations. A secondary reason being
that the pte object is redundant, i.e., the page table itself can be
used to lookup page table pages. (Historical note: The pte object
predates our ability to allocate "no object" pages. Thus, the pte
object was a necessary evil.)
- Unconditionally check the vm object lock's status in vm_page_remove().
Previously, this assertion could not be made on Alpha due to its use
of a pte object.
improved chance of working despite pressure from running programs.
Instead of trying to throw a bunch of pages out to swap and hope for
the best, only a range that can potentially fulfill contigmalloc(9)'s
request will have its contents paged out (potentially, not forcibly)
at a time.
The new contigmalloc operation still operates in three passes, but it
could potentially be tuned to more or less. The first pass only looks
at pages in the cache and free pages, so they would be thrown out
without having to block. If this is not enough, the subsequent passes
page out any unwired memory. To combat memory pressure refragmenting
the section of memory being laundered, each page is removed from the
systems' free memory queue once it has been freed so that blocking
later doesn't cause the memory laundered so far to get reallocated.
The page-out operations are now blocking, as it would make little sense
to try to push out a page, then get its status immediately afterward
to remove it from the available free pages queue, if it's unlikely to
have been freed. Another change is that if KVA allocation fails, the
allocated memory segment will be freed and not leaked.
There is a sysctl/tunable, defaulting to on, which causes the old
contigmalloc() algorithm to be used. Nonetheless, I have been using
vm.old_contigmalloc=0 for over a month. It is safe to switch at
run-time to see the difference it makes.
A new interface has been used which does not require mapping the
allocated pages into KVA: vm_page.h functions vm_page_alloc_contig()
and vm_page_release_contig(). These are what vm.old_contigmalloc=0
uses internally, so the sysctl/tunable does not affect their operation.
When using the contigmalloc(9) and contigfree(9) interfaces, memory
is now tracked with malloc(9) stats. Several functions have been
exported from kern_malloc.c to allow other subsystems to use these
statistics, as well. This invalidates the BUGS section of the
contigmalloc(9) manpage.
