Originally I believed that these interfaces would be needed.
However, in practice it turned out that it was more straight
forward and maintainable to use the native Linux interfaces.
As such, this is all dead code and can be safely removed.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#109
To minimize the chance of triggering an OOM during direct reclaim.
The kmem caches have been improved to make a best effort to reclaim
at least one slab when a reclaim function is registered. This helps
avoid the case where objects are released but they are spread over
multiple slabs so no memory gets reclaimed.
Care has been taken to avoid deadlocking if the reclaim function
is unable to make forward progress. Additionally, the reclaim
function may be skipped entirely if there are already free slabs
which can be safely reaped.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#107
The Linux direct reclaim path uses this out of band value to
determine if forward progress is being made. Normally this is
incremented by kmem_freepages() which is part of the various
Linux slab implementations. However, since we are using none
of that infrastructure we're responsible for incrementing this
count.
If no forward progress is detected and a subsequent allocation
fails the OOM killer will be invoked. If there was forward
progress additional reclaim will be attempted via the page
cache and registerd shrinker until the allocation succeeds.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#107
When memory pressure triggers direct memory reclaim, a slabs age
and delay should not prevent it from being freed. This patch ensures
these values are ignored, allowing an empty slab to be freed in this
code path no matter the value of its age and delay.
This prevents needless scanning of the partial slabs and has been
observed to significantly reduce the total cpu usage. In addition,
it should allow for snappier reclaim under memory pressure.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#102
Previously, the SPL tried to maintain Solaris semantics by freeing
all available (empty) slabs from its slab caches when the shrinker
was called. This is not desirable when running on Linux. To make
the SPL shrinker more Linux friendly, the actual number of freed
slabs from each of the slab caches is now derived from nr_to_scan
and skc_slab_objs.
Additionally, an accounting bug was fixed in spl_slab_reclaim()
which could cause us to reclaim one more slab than requested.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#101
Leverage the existing generic 64-bit division operations which
were originally implemented for x86 to support ARM. All that is
required is to make the symbols available to the linker with the
expected names.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
As of the removal of the taskq work list made in commit:
commit 2c02b71b14
Author: Prakash Surya <surya1@llnl.gov>
Date: Mon Dec 5 17:32:48 2011 -0800
Replace tq_work_list and tq_threads in taskq_t
To lay the ground work for introducing the taskq_dispatch_prealloc()
interface, the tq_work_list and tq_threads fields had to be replaced
with new alternatives in the taskq_t structure.
the comment above taskq_wait_check has been incorrect. This change is an
attempt at bringing that description more in line with the current
implementation. Essentially, references to the old task work list had to
be updated to reference the new taskq thread active list.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #65
Long ago I added support to the spl for condition variable names
because I thought they might be needed. It turns out they aren't.
In fact the official Solaris cv_init(9F) man page discourages
their use in the kernel.
cv_init(9F)
Parameters
name - Descriptive string. This is obsolete and should be
NULL. (Non-NULL strings are legal, but they're a
waste of kernel memory.)
Therefore, I'm removing them from the spl to reclaim this memory
and adding an ASSERT() to ensure no new consumers are added which
make use of the name.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Include the ZFS_META_RELEASE in the module load/unload messages
to more clearly indicate exactly what version of the SPL has
been loaded.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Add the bare minimum functionality to support dynamic kstats. A
complete kstat implementation should be done as part of issue #84.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #84
Until now the notion of an internal debug logging infrastructure
was conflated with enabling ASSERT()s. This patch clarifies things
by cleanly breaking the two subsystem apart. The result of this
is the following behavior.
--enable-debug - Enable/disable code wrapped in ASSERT()s.
--disable-debug ASSERT()s are used to check invariants and
are never required for correct operation.
They are disabled by default because they
may impact performance.
--enable-debug-log - Enable/disable the debug log infrastructure.
--disable-debug-log This infrastructure allows the spl code and
its consumer to log messages to an in-kernel
log. The granularity of the logging can be
controlled by a debug mask. By default the
mask disables most debug messages resulting
in a negligible performance impact. Because
of this the debug log is enabled by default.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Testing has shown that tq->tq_lock can be highly contended when a
large number of small work items are dispatched. The lock hold time
is reduced by the following changes:
1) Use exclusive threads in the work_waitq
When a single work item is dispatched we only need to wake a single
thread to service it. The current implementation uses non-exclusive
threads so all threads are woken when the dispatcher calls wake_up().
If a large number of threads are in the queue this overhead can become
non-negligible.
2) Conditionally add/remove threads from work waitq
Taskq threads need only add themselves to the work wait queue if
there are no pending work items.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #32
This reverts commit ec2b41049f.
A race condition was introduced by which a wake_up() call can be lost
after the taskq thread determines there is no pending work items,
leading to deadlock:
1. taksq thread enables interrupts
2. dispatcher thread runs, queues work item, call wake_up()
3. taskq thread runs, adds self to waitq, sleeps
This could easily happen if an interrupt for an IO completion was
outstanding at the point where the taskq thread reenables interrupts,
just before the call to add_wait_queue_exclusive(). The handler would
run immediately within the race window.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #32
Testing has shown that tq->tq_lock can be highly contended when a
large number of small work items are dispatched. The lock hold time
is reduced by the following changes:
1) Use exclusive threads in the work_waitq
When a single work item is dispatched we only need to wake a single
thread to service it. The current implementation uses non-exclusive
threads so all threads are woken when the dispatcher calls wake_up().
If a large number of threads are in the queue this overhead can become
non-negligible.
2) Conditionally add/remove threads from work waitq outside of tq_lock
Taskq threads need only add themselves to the work wait queue if there
are no pending work items. Furthermore, the add and remove function
calls can be made outside of the taskq lock since the wait queues are
protected from concurrent access by their own spinlocks.
3) Call wake_up() outside of tq->tq_lock
Again, the wait queues are protected by their own spinlock, so the
dispatcher functions can drop tq->tq_lock before calling wake_up().
A new splat test taskq:contention was added in a prior commit to measure
the impact of these changes. The following table summarizes the
results using data from the kernel lock profiler.
tq_lock time %diff Wall clock (s) %diff
original: 39117614.10 0 41.72 0
exclusive threads: 31871483.61 18.5 34.2 18.0
unlocked add/rm waitq: 13794303.90 64.7 16.17 61.2
unlocked wake_up(): 1589172.08 95.9 16.61 60.2
Each row reflects the average result over 5 test runs.
/proc/lock_stats was zeroed out before and collected after each run.
Column 1 is the cumulative hold time in microseconds for tq->tq_lock.
The tests are cumulative; each row reflects the code changes of the
previous rows. %diff is calculated with respect to "original" as
100*(orig-new)/orig.
Although calling wake_up() outside of the taskq lock dramatically
reduced the taskq lock hold time, the test actually took slightly more
wall clock time. This is because the point of contention shifts from
the taskq lock to the wait queue lock. But the change still seems
worthwhile since it removes our taskq implementation as a bottleneck,
assuming the small increase in wall clock time to be statistical
noise.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#32
The Proxmox VE kernel contains a patch which renames the function
invalidate_inodes() to invalidate_inodes_check(). In the process
it adds a 'check' argument and a '#define invalidate_inodes(x)'
compatibility wrapper for legacy callers. Therefore, if either
of these functions are exported invalidate_inodes() can be
safely used.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#58
A preallocated taskq_ent_t's tqent_flags must be checked prior to
servicing the taskq_ent_t. Once a preallocated taskq entry is serviced,
the ownership of the entry is handed back to the caller of
taskq_dispatch, thus the entry's contents can potentially be mangled.
In particular, this is a problem in the case where a preallocated taskq
entry is serviced, and the caller clears it's tqent_flags field. Thus,
when the function returns and task_done is called, it looks as though
the entry is **not** a preallocated task (when in fact it **is** a
preallocated task).
In this situation, task_done will place the preallocated taskq_ent_t
structure onto the taskq_t's free list. This is a **huge** mistake. If
the taskq_ent_t is then freed by the caller of taskq_dispatch, the
taskq_t's free list will hold a pointer to garbage data. Even worse, if
nothing has over written the freed memory before the pointer is
dereferenced, it may still look as though it points to a valid list_head
belonging to a taskq_ent_t structure.
Thus, the task entry's flags are now copied prior to servicing the task.
This copy is then checked to see if it is a preallocated task, and
determine if the entry needs to be passed down to the task_done
function.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#71
The taskq_t's active thread list is sorted based on its
tqt_ent->tqent_id field. The list is kept sorted solely by inserting
new taskq_thread_t's in their correct sorted location; no other
means is used. This means that once inserted, if a taskq_thread_t's
tqt_ent->tqent_id field changes, the list runs the risk of no
longer being sorted.
Prior to the introduction of the taskq_dispatch_prealloc() interface,
this was not a problem as a taskq_ent_t actively being serviced under
the old interface should always have a static tqent_id field. Thus,
once the taskq_thread_t is added to the taskq_t's active thread list,
the taskq_thread_t's tqt_ent->tqent_id field would remain constant.
Now, this is no longer the case. Currently, if using the
taskq_dispatch_prealloc() interface, any given taskq_ent_t actively
being serviced _may_ have its tqent_id value incremented. This happens
when the preallocated taskq_ent_t structure is recursively dispatched.
Thus, a taskq_thread_t could potentially have its tqt_ent->tqent_id
field silently modified from under its feet. If this were to happen
to a taskq_thread_t on a taskq_t's active thread list, this would
compromise the integrity of the order of the list (as the list
_may_ no longer be sorted).
To get around this, the taskq_thread_t's taskq_ent_t pointer was
replaced with its own static copy of the tqent_id. So, as a taskq_ent_t
is pulled off of the taskq_t's pending list, a static copy of its
tqent_id is made and this copy is used to sort the active thread
list. Using a static copy is key in ensuring the integrity of the
order of the active thread list. Even if the underlying taskq_ent_t
is recursively dispatched (as has its tqent_id modified), this
static copy stored inside the taskq_thread_t will remain constant.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #71
This patch implements the taskq_dispatch_prealloc() interface which
was introduced by the following illumos-gate commit. It allows for
a preallocated taskq_ent_t to be used when dispatching items to a
taskq. This eliminates a memory allocation which helps minimize
lock contention in the taskq when dispatching functions.
commit 5aeb94743e3be0c51e86f73096334611ae3a058e
Author: Garrett D'Amore <garrett@nexenta.com>
Date: Wed Jul 27 07:13:44 2011 -0700
734 taskq_dispatch_prealloc() desired
943 zio_interrupt ends up calling taskq_dispatch with TQ_SLEEP
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #65
To lay the ground work for introducing the taskq_dispatch_prealloc()
interface, the tq_work_list and tq_threads fields had to be replaced
with new alternatives in the taskq_t structure.
The tq_threads field was replaced with tq_thread_list. Rather than
storing the pointers to the taskq's kernel threads in an array, they are
now stored as a list. In addition to laying the ground work for the
taskq_dispatch_prealloc() interface, this change could also enable taskq
threads to be dynamically created and destroyed as threads can now be
added and removed to this list relatively easily.
The tq_work_list field was replaced with tq_active_list. Instead of
keeping a list of taskq_ent_t's which are currently being serviced, a
list of taskq_threads currently servicing a taskq_ent_t is kept. This
frees up the taskq_ent_t's tqent_list field when it is being serviced
(i.e. now when a taskq_ent_t is being serviced, it's tqent_list field
will be empty).
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #65
The spl_task structure was renamed to taskq_ent, and all of
its fields were renamed to have a prefix of 'tqent' rather
than 't'. This was to align with the naming convention which
the ZFS code assumes. Previously these fields were private
so the name never mattered.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #65
This is a bit of cleanup I'd been meaning to get to for a while
to reduce the chance of a type conflict. Well that conflict
finally occurred with the kstat_init() function which conflicts
with a function in the 2.6.32-6-pve kernel.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#56
As of Linux 3.1 the shrink_dcache_memory and shrink_icache_memory
functions have been removed. This same task is now accomplished
more cleanly with per super block shrinkers. This unfortunately
leaves us no easy way to support the dnlc_reduce_cache() function.
This support has always been entirely optional. So when no
reasonable interface is available allow the dnlc_reduce_cache()
function to effectively become a no-op.
The downside of this change is that it will prevent the zfs arc
meta data limts from being enforced. However, the current zfs
implementation in this regard is already flawed and needs to
be reworked. If the arc needs to enfore a meta data limit it
will need to be extended to coordinate directly with the zpl.
This will allow us to drop all this compatibility code and get
more fine grained control over the cache management.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #52
Prior to Linux 3.1 the kern_path_parent symbol was exported for
use by kernel modules. As of Linux 3.1 it is now longer easily
available. To handle this case the spl will now dynamically
look up address of the missing symbol at module load time.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #52
Be careful not to unconditionally clear the PF_MEMALLOC bit in
the task structure. It may have already been set when entering
kv_alloc() in which case it must remain set on exit. In
particular the kswapd thread will have PF_MEMALLOC set in
order to prevent it from entering direct reclaim. By clearing
it we allow the following NULL deref to potentially occur.
BUG: unable to handle kernel NULL pointer dereference at (null)
IP: [<ffffffff8109c7ab>] balance_pgdat+0x25b/0x4ff
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes ZFS issue #287
This would cause problems when using 'zfs send' with a file as the
target (rather than a pipe or a socket as is usually the case) as
for each write the destination offset in the file would be 0.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes ZFS issue #391
The typo did not have any effect (apart from a negligible performance
impact) because skc->skc_flags * KMC_OFFSLAB is always non-null when
at least one bit in skc->skc_flags is set.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
In a non-debug build the ASSERT() would be optimized away
which could cause pending work items to not be cancelled.
We must also use cancel_delayed_work_sync() rather than just
cancel_delayed_work() to actually wait until work items have
completed. Otherwise they might accidentally access free'd
memory.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes ZFS bugs #279, #62, #363, #418
File descriptors are a per-process resource. The same descriptor
in different processes can refer to different files. find_file()
incorrectly assumed that file descriptors are globally unique.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes ZFS issue #386
No longer print the following warning to the console when the
/etc/hostid file is missing. This is the expected default behavior.
Keeping the hostid in sync with the initramfs is now accomplished
by creating the /etc/hostid in the initramfs not on the system.
SPL: The /etc/hostid file is not found.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Deprecate the /usr/bin/hostid call by reading the /etc/hostid file
directly. Add the spl_hostid_path parameter to override the default
/etc/hostid path.
Rename the set_hostid() function to hostid_exec() to better reflect
actual behavior and complement the new hostid_read() function.
Use HW_INVALID_HOSTID as the spl_hostid sentinel value because
zero seems to be a valid gethostid() result on Linux.
Update the the wrapper macros for the memory shrinker to handle
this 4th API change. The callback function now takes a
shrink_control structure. This is certainly a step in the
right direction but it's annoying to have to accomidate yet
another version of the API.
It has become necessary to be able to optionally disable
direct memory reclaim for certain taskqs. To support
this the TASKQ_NORECLAIM flags has been added which sets
the PF_MEMALLOC bit for all threads in the taskq.
Change the SPL kernel messages for module loading and module
unloading so that they are similar to the ZFS kernel messages.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
This reverts commit 1814251453.
Demote the gawk call back to awk and ensure that stderr is attached. GNU gawk
tolerates a missing stderr handle, but many utilities do not, which could be
why a regular awk call was unexplainably failing on some systems.
Use argv[0] instead of sh_path for consistency internally and with other Linux
drivers.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Provide a call_usermodehelper() alternative by letting the hostid be passed as
a module parameter like this:
$ modprobe spl spl_hostid=0x12345678
Internally change the spl_hostid variable to unsigned long because that is the
type that the coreutils /usr/bin/hostid returns.
Move the hostid command into GET_HOSTID_CMD for consistency with the similar
GET_KALLSYMS_ADDR_CMD invocation.
Use argv[0] instead of sh_path for consistency internally and with other Linux
drivers.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The function zlib_deflate_workspacesize() now take 2 arguments.
This was done to avoid always having to allocate the maximum size
workspace (268K). The caller can now specific the windowBits and
memLevel compression parameters to get a smaller workspace.
For our purposes we introduce a spl_zlib_deflate_workspacesize()
wrapper which accepts both arguments. When the two argument
version of zlib_deflate_workspacesize() is available the arguments
are passed through. When it's not we assume the worst case and
a maximally sized workspace is used.
The path_lookup() function has been renamed to kern_path_parent()
and the flags argument has been removed. The only behavior now
offered is that of LOOKUP_PARENT. The spl already always passed
this flag so dropping the flag does not impact us.
This is a long over due compatibility change. Way, way, way back
in 2007 there was a push to remove all consumers of SPIN_LOCK_UNLOCKED.
Finally, in 2011 with 2.6.39 all the consumers have been updated
and SPIN_LOCK_UNLOCKED was removed. It's about time we use the
new API as well, this change does exactly that. DEFINE_SPINLOCK()
was available as far back as 2.6.12 so there doesn't need to be
any additional autoconf-foo for this change.
Flagged by the default -Wunused-but-set-variable gcc option when
running under Fedora 15. Since it's correct this variable is
entirely unused this commit removes it.
To resolve a potiential filesystem corruption issue a second
argument was added to invalidate_inodes(). This argument controls
whether dirty inodes are dropped or treated as busy when invalidating
a super block. When only the legacy API is available the second
argument will be dropped for compatibility.
Provide the dnlc_reduce_cache() function which attempts to prune
cached entries from the dcache and icache. After the entries are
pruned any slabs which they may have been using are reaped.
Note the API takes a reclaim percentage but we don't have easy
access to the total number of cache entries to calculate the
reclaim count. However, in practice this doesn't need to be
exactly correct. We simply need to reclaim some useful fraction
(but not all) of the cache. The caller can determine if more
needs to be done.
One of the most common things you want to know when looking at
the slab is how much memory is being used. This information was
available in /proc/spl/kmem/slab but only on a per-slab basis.
This commit adds the following /proc/sys/kernel/spl/kmem/slab*
entries to make total slab usage easily available at a glance.
slab_kmem_total - Total kmem slab size
slab_kmem_avail - Alloc'd kmem slab size
slab_kmem_max - Max observed kmem slab size
slab_vmem_total - Total vmem slab size
slab_vmem_avail - Alloc'd vmem slab size
slab_vmem_max - Max observed vmem slab size
NOTE: The slab_*_max values are expected to over report because
they show maximum values since boot, not current values.
The 'slab_fail', 'slab_create', and 'slab_destroy' columns in the slab
output have been removed because they are virtually always zero and
not very useful.
The much more useful 'size' and 'alloc' columns have been added which
show the total slab size and how much of the total size has been
allocated to objects.
Finally, the formatting has been updated to be much more human
readable while still being friendly for tool like awk to parse.
The Linux shrinker has gone through three API changes since 2.6.22.
Rather than force every caller to understand all three APIs this
change consolidates the compatibility code in to the mm-compat.h
header. The caller then can then use a single spl provided
shrinker API which does the right thing for your kernel.
SPL_SHRINKER_CALLBACK_PROTO(shrinker_callback, cb, nr_to_scan, gfp_mask);
SPL_SHRINKER_DECLARE(shrinker_struct, shrinker_callback, seeks);
spl_register_shrinker(&shrinker_struct);
spl_unregister_shrinker(&&shrinker_struct);
spl_exec_shrinker(&shrinker_struct, nr_to_scan, gfp_mask);
Solaris credentials don't have an fsuid/fsguid field but Linux
credentials do. To handle this case the Solaris API is being
modestly extended to include the crgetfsuid()/crgetfsgid()
helper functions.
Addititionally, because the crget*() helpers are implemented
identically regardless of HAVE_CRED_STRUCT they have been
moved outside the #ifdef to common code. This simplification
means we only have one version of the helper to keep to to date.
As part of vmalloc() a __pte_alloc_kernel() allocation may occur. This
internal allocation does not honor the gfp flags passed to vmalloc().
This means even when vmalloc(GFP_NOFS) is called it is possible that a
synchronous reclaim will occur. This reclaim can trigger file IO which
can result in a deadlock. This issue can be avoided by explicitly
setting PF_MEMALLOC on the process to subvert synchronous reclaim when
vmalloc() is called with !__GFP_FS.
An example stack of the deadlock can be found here (1), along with the
upstream kernel bug (2), and the original bug discussion on the
linux-mm mailing list (3). This code can be properly autoconf'ed
when the upstream bug is fixed.
1) http://github.com/behlendorf/zfs/issues/labels/Vmalloc#issue/133
2) http://bugzilla.kernel.org/show_bug.cgi?id=30702
3) http://marc.info/?l=linux-mm&m=128942194520631&w=4
The xvattr support in the spl has always simply consisted of
defining a couple structures and a few #defines. This was enough
to enable compilation of code which just passed xvattr types
around but not enough to effectively manipulate them.
This change removes even this minimal support leaving it up
to packages which leverage the spl to prove the full xvattr
support. By removing it from the spl we ensure not conflict
with the higher level packages.
This just leaves minimal vnode support for basical manipulation
of files. This code is does have the proper support functions
in the spl and a set of regression tests.
Additionally, this change removed the unused 'caller_context_t *'
type and replaces it with a 'void *'.
While portions of the code needed to support z_compress_level() and
z_uncompress() where in place. In reality the current implementation
was non-functional, it just was compilable.
The critical missing component was to setup a workspace for the
compress/uncompress stream structures to use. A kmem_cache was
added for the workspace area because we require a large chunk
of memory. This avoids to need to continually alloc/free this
memory and vmap() the pages which is very slow. Several objects
will reside in the per-cpu kmem_cache making them quick to acquire
and release. A further optimization would be to adjust the
implementation to additional ensure the memory is local to the cpu.
Currently that may not be the case.
