recursion from the VM is handled (and the calling code that allocates
buckets knows how to deal with it), we do not want to prevent allocation
from the slab header zones (slabzone and slabrefzone) if uk_recurse is
not zero for them. The reason is that it could lead to NULL being
returned for the slab header allocations even in the M_WAITOK
case, and the caller can't handle that (this is also explained in a
comment with this commit).
The problem analysis is documented in our mailing lists:
http://docs.freebsd.org/cgi/getmsg.cgi?fetch=153445+0+archive/2004/freebsd-current/20041231.freebsd-current
(see entire thread for proper context).
Crash dump data provided by: Peter Holm <peter@holm.cc>
number of entries into bucket_zone_lookup(), which helps make more
clear the logic of consumers of bucket zones.
Annotate the behavior of bucket_init() with a comment indicating
how the various data structures, including the bucket lookup tables,
are initialized.
IF INVARIANTS is defined, and in the rare case that we have
allocated some objects from the slab and at least one initializer
on at least one of those objects failed, and we need to fail the
allocation and push the uninitialized items back into the slab
caches -- in that scenario, we would fail to [re]set the
bucket cache's ub_bucket item references to NULL, which would
eventually trigger a KASSERT.
state management corruption, mbuf leaks, general mbuf corruption,
and at least on i386 a first level splash damage radius that
encompasses up to about half a megabyte of the memory after
an mbuf cluster's allocation slab. In short, this has caused
instability nightmares anywhere the right kind of network traffic
is present.
When the polymorphic refcount slabs were added to UMA, the new types
were not used pervasively. In particular, the slab management
structure was turned into one for refcounts, and one for non-refcounts
(supposed to be mostly like the old slab management structure),
but the latter was almost always used through out. In general, every
access to zones with UMA_ZONE_REFCNT turned on corrupted the
"next free" slab offset offset and the refcount with each other and
with other allocations (on i386, 2 mbuf clusters per 4096 byte slab).
Fix things so that the right type is used to access refcounted zones
where it was not before. There are additional errors in gross
overestimation of padding, it seems, that would cause a large kegs
(nee zones) to be allocated when small ones would do. Unless I have
analyzed this incorrectly, it is not directly harmful.
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.
- 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.
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.)
and WITNESS is not built, then force all M_WAITOK allocations to
M_NOWAIT behavior (transparently). This is to be used temporarily
if wierd deadlocks are reported because we still have code paths
that perform M_WAITOK allocations with lock(s) held, which can
lead to deadlock. If WITNESS is compiled, then the sysctl is ignored
and we ask witness to tell us wether we have locks held, converting
to M_NOWAIT behavior only if it tells us that we do.
Note this removes the previous mbuf.h inclusion as well (only needed
by last revision), and cleans up unneeded [artificial] comparisons
to just the mbuf zones. The problem described above has nothing to
do with previous mbuf wait behavior; it is a general problem.
zones, and do it by direct comparison of uma_zone_t instead of strcmp.
The mbuf subsystem used to provide M_TRYWAIT/M_DONTWAIT semantics, but
this is mostly no longer the case. M_WAITOK has taken over the spot
M_TRYWAIT used to have, and for mbuf things, still may return NULL if
the code path is incorrectly holding a mutex going into mbuf allocation
functions.
The M_WAITOK/M_NOWAIT semantics are absolute; though it may deadlock
the system to try to malloc or uma_zalloc something with a mutex held
and M_WAITOK specified, it is absolutely required to not return NULL
and will result in instability and/or security breaches otherwise.
There is still room to add the WITNESS_WARN() to all cases so that
we are notified of the possibility of deadlocks, but it cannot change
the value of the "badness" variable and allow allocation to actually
fail except for the specialized cases which used to be M_TRYWAIT.
to failing -- that is, allocations via malloc(M_WAITOK) that are required
to never fail -- if WITNESS is not defined. While everyone should be
running WITNESS, in any case, zone "Mbuf" allocations are really the only
ones that should be screwed with by this hack.
This hack is crashing people, and would continue to do so with or without
WITNESS. Things shouldn't be allocating with M_WAITOK with locks held,
but it's not okay just to always remove M_WAITOK when !WITNESS.
Reported by: Bernd Walter <ticso@cicely5.cicely.de>
The general UMA lock is a recursion-allowed lock because
there is a code path where, while we're still configured
to use startup_alloc() for backend page allocations, we
may end up in uma_reclaim() which calls zone_foreach(zone_drain),
which grabs uma_mtx, only to later call into startup_alloc()
because while freeing we needed to allocate a bucket. Since
startup_alloc() also takes uma_mtx, we need to be able to
recurse on it.
This exact explanation also added as comment above mtx_init().
Trace showing recursion reported by: Peter Holm <peter-at-holm.cc>
internal reference counters, UMA_ZONE_NOFREE. This way, those slabs
(with their ref counts) will be effectively type-stable, then using
a trick like this on the refcount is no longer dangerous:
MEXT_REM_REF(m);
if (atomic_cmpset_int(m->m_ext.ref_cnt, 0, 1)) {
if (m->m_ext.ext_type == EXT_PACKET) {
uma_zfree(zone_pack, m);
return;
} else if (m->m_ext.ext_type == EXT_CLUSTER) {
uma_zfree(zone_clust, m->m_ext.ext_buf);
m->m_ext.ext_buf = NULL;
} else {
(*(m->m_ext.ext_free))(m->m_ext.ext_buf,
m->m_ext.ext_args);
if (m->m_ext.ext_type != EXT_EXTREF)
free(m->m_ext.ref_cnt, M_MBUF);
}
}
uma_zfree(zone_mbuf, m);
Previously, a second thread hitting the above cmpset might
actually read the refcnt AFTER it has already been freed. A very
rare occurance. Now we'll know that it won't be freed, though.
Spotted by: julian, pjd
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
pmap. For the kernel pmap, Giant is not required. In general, for
other pmaps, Giant is required by i386's pmap_pte() implementation.
Specifically, the use of PMAP2/PADDR2 is synchronized by Giant.
Note: In principle, updates to the kernel pmap's wired count could be
lost without Giant. However, in practice, we never use the kernel
pmap's wired count. This will be resolved when pmap locking appears.
- With the above change, cpu_thread_clean() and uma_large_free() need
not acquire Giant. (The first case is simply the revival of
i386/i386/vm_machdep.c's revision 1.226 by peter.)
when uma_reclaim() was called. This was introduced when the zone
working-set algorithm was removed in favor of using the per cpu caches
as the working set.
1) mp_maxid is a valid FreeBSD CPU ID in the range 0 .. MAXCPU - 1.
2) For all active CPUs in the system, PCPU_GET(cpuid) <= mp_maxid.
Approved by: re (scottl)
Tested on: i386, amd64, alpha
was equal to MAXCPU, we would overrun the pcpu_mtx array because maxcpu
was calculated incorrectly.
- Add some more debugging code so that memory leaks at the time of
uma_zdestroy() are more easily diagnosed.
Approved by: re (rwatson)
function, startup_alloc(), that is used for single page allocations prior
to the VM starting up. If it is used after the VM startups up, it
replaces the zone's allocf pointer with either page_alloc() or
uma_small_alloc() where appropriate.
Pointy hat to: me
Tested by: phk/amd64, me/x86
Temporarily disable the UMA_MD_SMALL_ALLOC stuff since recent commits
break sparc64, amd64, ia64 and alpha. It appears only i386 and maybe
powerpc were not broken.
working set cache. This has several advantages. Firstly, we never touch
the per cpu queues now in the timeout handler. This removes one more
reason for having per cpu locks. Secondly, it reduces the size of the zone
by 8 bytes, bringing it under 200 bytes for a single proc x86 box. This
tidies up other logic as well.
- The 'destroy' flag no longer needs to be passed to zone_drain() since it
always frees everything in the zone's slabs.
- cache_drain() is now only called from zone_dtor() and so it destroys by
default. It also does not need the destroy parameter now.
broken consumers of the malloc interface who assume that the allocated
address will be an even multiple of the size.
- Remove disabled time delay code on uma_reclaim(). The comment there said
it all. It was not an effective strategy and it should not be left in
#if 0'd for all eternity.
page_alloc() function from the slab_zalloc() function. This allows us
to unconditionally call uz_allocf().
- In page_alloc() cleanup the boot_pages logic some. Previously memory from
this cache that was not used by the time the system started was left in
the cache and never used. Typically this wasn't more than a few pages,
but now we will use this cache so long as memory is available.
by accepting the user supplied flags directly. Previously this was not
done so that flags for the same field would not be defined in two
different files. Add comments in each header instructing future
developers on how now to shoot their feet.
- Fix a test for !OFFPAGE which should have been a test for HASH. This would
have caused a panic if we had ever destructed a malloc zone. This also
opens up the possibility that other zones could use the vsetobj() method
rather than a hash.
don't cache as many items.
- Introduce the bucket_alloc(), bucket_free() functions to wrap bucket
allocation. These functions select the appropriate bucket zone to
allocate from or free to.
- Rename ub_ptr to ub_cnt to reflect a change in its use. ub_cnt now reflects
the count of free items in the bucket. This gets rid of many unnatural
subtractions by 1 throughout the code.
- Add ub_entries which reflects the number of entries possibly held in a
bucket.
UMA_ZFLAG_INTERNAL zones at all. Apparently, Wilko's alpha
was crashing while entering multi-user because, I think, we
were calculating the garbage cachefree for pcpu caches that
essentially don't exist for at least the 'zones' zone and it so
happened that we were reading from an unmapped location.
Confirmed to fix crash: wilko
Helped debug: wilko, gallatin
compare the zone element size (+1 for the byte of linkage) against
UMA_SLAB_SIZE - sizeof(struct uma_slab), and not just UMA_SLAB_SIZE.
Add a KASSERT in zone_small_init to make sure that the computed
ipers (items per slab) for the zone is not zero, despite the addition
of the check, just to be sure (this part submitted by: silby)
- UMA_ZONE_VM used to imply BUCKETCACHE. Now it implies
CACHEONLY instead. CACHEONLY is like BUCKETCACHE in the
case of bucket allocations, but in addition to that also ensures that
we don't setup the zone with OFFPAGE slab headers allocated from the
slabzone. This means that we're not allowed to have a UMA_ZONE_VM
zone initialized for large items (zone_large_init) because it would
require the slab headers to be allocated from slabzone, and hence
kmem_map. Some of the zones init'd with UMA_ZONE_VM are so init'd
before kmem_map is suballoc'd from kernel_map, which is why this
change is necessary.
concurrent invocations from acquiring the same address(es). Also, in case
of an incomplete allocation, free any allocated pages.
In collaboration with: tegge
sure that uma_dbg_free() is called if we're about to call
uma_zfree_internal() but we're asking it to skip the dtor and
uma_dbg_free() call itself. So, if we're about to call
uma_zfree_internal() from uma_zfree_arg() and skip == 1, call
uma_dbg_free() ourselves.
1) The race has to do with zone destruction. From the zone destructor we
would lock the zone, set the working set size to 0, then unlock the zone,
drain it, and then free the structure. Within the window following the
working-set-size set to 0 and unlocking of the zone and the point where
in zone_drain we re-acquire the zone lock, the uma timer routine could
have fired off and changed the working set size to something non-zero,
thereby potentially preventing us from completely freeing slabs before
destroying the zone (and thus leaking them).
2) The leak has to do with zone destruction as well. When destroying a
zone we would take care to free all the buckets cached in the zone, but
although we would drain the pcpu cache buckets, we would not free them.
This resulted in leaking a couple of bucket structures (512 bytes each)
per cpu on SMP during zone destruction.
While I'm here, also silence GCC warnings by turning uma_slab_alloc()
from inline to real function. It's too big to be an inline.
Reviewed by: JeffR
("UMA Zone") carefully, because it does not have pcpu caches allocated
at all. In the UP case, we did not catch this because one pcpu cache
is always allocated with the zone, but for the MP case, we were getting
bogus stats for this zone.
Tested by: Lukas Ertl <le@univie.ac.at>
- In sysctl_vm_zone use the per cpu locks to read the current cache
statistics this makes them more accurate while under heavy load.
Submitted by: tegge
so not only wastes memory but it can also cause a leak in zones that
will be destroyed later. The problem is that the slab allocation code
places newly created slabs on the partially allocated list because it
assumes that the caller will actually allocate some memory from it.
Failure to do so places an otherwise free slab on the partial slab list
where we wont find it later in zone_drain().
Continuously prodded to fix by: phk (Thanks)
