For earlier versions of the kernel with memalloc_noio_save, it only turns
off __GFP_IO but leaves __GFP_FS untouched during direct reclaim. This
would cause threads to direct reclaim into ZFS and cause deadlock.
Instead, we should stick to using spl_fstrans_mark. Since we would
explicitly turn off both __GFP_IO and __GFP_FS before allocation, it
will work on every version of the kernel.
This impacts kernel versions 3.9-3.17, see upstream kernel commit
torvalds/linux@934f307 for reference.
Signed-off-by: Chunwei Chen <david.chen@osnexus.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tim Chase <tim@chase2k.com>
Closes#515
Issue zfsonlinux/zfs#4111
This is needed for architectures that do not have a builtin prefetchw()
Signed-off-by: Dimitri John Ledkov <xnox@ubuntu.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#502
Linux slab will automatically free empty slab when number of partial slab is
over min_partial, so we don't need to explicitly shrink it. In fact, calling
kmem_cache_shrink from shrinker will cause heavy contention on
kmem_cache_node->list_lock, to the point that it might cause __slab_free to
livelock (see zfsonlinux/zfs#3936)
Signed-off-by: Chunwei Chen <david.chen@osnexus.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes zfsonlinux/zfs#3936
Closes#487
Allocate a kmem cache magazine for every possible CPU which might
be added to the system. This ensures that when one of these CPUs
is enabled it can be safely used immediately.
For many systems the number of online CPUs is identical to the
number of present CPUs so this does imply an increased memory
footprint. In fact, dynamically allocating the array of magazine
pointers instead of using the worst case NR_CPUS can end up
decreasing our memory footprint.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ned Bass <bass6@llnl.gov>
Closes#482
Support grsecurity/PaX kernel configurations where
CONFIG_PAX_USERCOPY_SLABS are enabled. When this kernel option
is enabled slabs which are used to copy between user and kernel
space must be created with SLAB_USERCOPY.
Stock Linux kernels do not have a SLAB_USERCOPY definition so
this causes no change in behavior for non-PAX-enabled kernels.
Verified-by: Wuffleton <null@wuffleton.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #2977
Issue #3796
On Linux the meaning of a processes priority is inverted with respect
to illumos. High values on Linux indicate a _low_ priority while high
value on illumos indicate a _high_ priority.
In order to preserve the logical meaning of the minclsyspri and
maxclsyspri macros when they are used by the illumos wrapper functions
their values have been inverted. This way when changes are merged
from upstream illumos we won't need to remember to invert the macro.
It could also lead to confusion.
Note this change also reverts some of the priorities changes in prior
commit 62aa81a. The rational is as follows:
spl_kmem_cache - High priority may result in blocked memory allocs
spl_system_taskq - May perform I/O for file backed VDEVs
spl_dynamic_taskq - New taskq threads should be spawned promptly
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ned Bass <bass6@llnl.gov>
Issue zfsonlinux/zfs#3607
As described in spl_kmem_cache_destroy() the ->skc_ref count was
added to address the case of a cache reap or grow racing with a
destroy. They are not strictly needed in the alloc/free paths
because consumers of the cache are responsible for not using it
while it's being destroyed.
Removing this code is desirable because there is some evidence that
contention on this atomic negative impacts performance on large-scale
NUMA systems.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tim Chase <tim@chase2k.com>
Issue #463
Add a new defclsyspri macro which can be used to request the default
Linux scheduler priority. Neither the minclsyspri or maxclsyspri map
to the default Linux kernel thread priority. This makes it awkward to
create taskqs which run with the same priority as the rest of the kernel
threads on the system which can lead to performance issues.
All SPL callers which previously used minclsyspri or maxclsyspri have
been changed to use defclsyspri. The vast majority of callers were
part of the test suite which won't have an external impact. The few
places where it could impact performance the change was from maxclsyspri
to defclsyspri. This makes it more likely the process will be scheduled
which may help performance.
To facilitate further performance analysis the spl_taskq_thread_priority
module option has been added. When disabled (0) all newly created kernel
threads will use the default kernel thread priority. When enabled (1)
the specified taskq priority will be used. By default this value is
enabled (1).
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Add the TASKQ_DYNAMIC flag to the kmem_cache and system taskqs
to reduce the number of idle threads on the system. Additional
threads will be created on demand up to the previous maximum
thread counts. This should have minimal, if any, impact on
performance.
This makes the system taskq consistent with illumos which is
always created as a dynamic taskq with up to 64 threads.
The task limits for the kmem_cache have been increased to avoid
any unnessisary throttling and to keep a larger reserve of
task_t structures on the free list.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tim Chase <tim@chase2k.com>
Closes#458
Avoid deadlocks when entering the shrinker from a PF_FSTRANS context.
This patch also reverts commit d0d5dd7 which added MUTEX_FSTRANS. Its
use has been deprecated within ZFS as it was an ineffective mechanism
to eliminate deadlocks. Among other things, it introduced the need for
strict ordering of mutex locking and unlocking in order that the
PF_FSTRANS flag wouldn't set incorrectly.
Signed-off-by: Tim Chase <tim@chase2k.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#446
The __get_free_pages() function must be used in place of kmalloc()
to ensure the __GFP_COMP is strictly honored. This is due to
kmalloc() being layered on the generic Linux slab caches. It
wasn't until recently that all caches were created using __GFP_COMP.
This means that it is possible for a kmalloc() which passed the
__GFP_COMP flag to be returned a non-compound allocation.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The kmem cache implementation always adds new slabs by dispatching a
task to the spl_kmem_cache taskq to perform the allocation. This is
done because large slabs must be allocated using vmalloc(). It is
possible these allocations will block on IO because the GFP_NOIO flag
is not honored. This can result in a deadlock.
Therefore, a deadlock detection strategy was implemented to deal with
this case. When it is determined, by timeout, that the spl_kmem_cache
thread has deadlocked attempting to add a new slab. Then all callers
attempting to allocate from the cache fall back to using kmalloc()
which does honor all passed flags.
This logic was correct but an optimization in the code allowed for a
deadlock. Because only slabs backed by vmalloc() can deadlock in the
way described above. An optimization was made to only invoke this
deadlock detection code for vmalloc() backed caches. This had the
advantage of making it easy to distinguish these objects when they
were freed.
But this isn't strictly safe. If all the spl_kmem_cache threads end
up deadlocked than we can't grow any of the other caches either. This
can once again result in a deadlock if memory needs to be allocated
from one of these other caches to ensure forward progress.
The fix here is to remove the optimization which limits this fall back
allocation stratagy to vmalloc() backed caches. Doing this means we
may need to take the cache lock in spl_kmem_cache_free() call path.
But this small cost can be mitigated by ignoring objects with virtual
addresses.
For good measure the default number of spl_kmem_cache threads has been
increased from 1 to 4, and made tunable. This alone wouldn't resolve
the original issue since it's still possible for all the threads to be
deadlocked. However, it does help responsiveness by ensuring that a
single deadlocked spl_kmem_cache thread doesn't block allocations from
other caches until the timeout is reached.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
This change is designed to improve the memory utilization of
slabs by more carefully setting their size. The way the code
currently works is problematic for slabs which contain large
objects (>1MB). This is due to slabs being unconditionally
rounded up to a power of two which may result in unused space
at the end of the slab.
The reason the existing code rounds up every slab is because it
assumes it will backed by the buddy allocator. Since the buddy
allocator can only performs power of two allocations this is
desirable because it avoids wasting any space. However, this
logic breaks down if slab is backed by vmalloc() which operates
at a page level granularity. In this case, the optimal thing to
do is calculate the minimum required slab size given certain
constraints (object size, alignment, objects/slab, etc).
Therefore, this patch reworks the spl_slab_size() function so
that it sizes KMC_KMEM slabs differently than KMC_VMEM slabs.
KMC_KMEM slabs are rounded up to the nearest power of two, and
KMC_VMEM slabs are allowed to be the minimum required size.
This change also reduces the default number of objects per slab.
This reduces how much memory a single cache object can pin, which
can result in significant memory saving for highly fragmented
caches. But depending on the workload it may result in slabs
being allocated and freed more frequently. In practice, this
has been shown to be a better default for most workloads.
Also the maximum slab size has been reduced to 4MB on 32-bit
systems. Due to the limited virtual address space it's critical
the we be as frugal as possible. A limit of 4M still lets us
reasonably comfortably allocate a limited number of 1MB objects.
Finally, the kmem:slab_small and kmem:slab_large SPLAT tests
were extended to provide better test coverage of various object
sizes and alignments. Caches are created with random parameters
and their basic functionality is verified by allocating several
slabs worth of objects.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reduce the threshold for detecting a kmem cache deadlock by 10x
from HZ to HZ/10. The reduced value is still several orders of
magnitude large enough to avoid being triggered incorrectly. By
reducing it we allow the system to resolve the issue more quickly.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Many people have noticed that the kmem cache implementation is slow
to release its memory. This patch makes the reclaim behavior more
aggressive by immediately freeing a slab once it is empty. Unused
objects which are cached in the magazines will still prevent a slab
from being freed.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The comment above the Linux 3.16 kernel's clear_bit() states:
/**
* clear_bit - Clears a bit in memory
* @nr: Bit to clear
* @addr: Address to start counting from
*
* clear_bit() is atomic and may not be reordered. However, it does
* not contain a memory barrier, so if it is used for locking purposes,
* you should call smp_mb__before_atomic() and/or smp_mb__after_atomic()
* in order to ensure changes are visible on other processors.
*/
This comment does not make sense in the context of x86 because x86 maps the
operations to barrier(), which is a compiler barrier. However, it does make
sense to me when I consider architectures that reorder around atomic
instructions. In such situations, a processor is allowed to execute the
wake_up_bit() before clear_bit() and we have a race. There are a few
architectures that suffer from this issue.
In such situations, the other processor would wake-up, see the bit is still
taken and go to sleep, while the one responsible for waking it up will
assume that it did its job and continue.
This patch implements a wrapper that maps smp_mb__{before,after}_atomic() to
smp_mb__{before,after}_clear_bit() on older kernels and changes our code to
leverage it in a manner consistent with the mainline kernel.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The port of XFS to Linux introduced a thread-specific PF_FSTRANS bit
that is used to mark contexts which are processing transactions. When
set, allocations in this context can dip into kernel memory reserves
to avoid deadlocks during writeback. Linux 3.9 provided the additional
PF_MEMALLOC_NOIO for disabling __GFP_IO in page allocations, which XFS
began using in 3.15.
This patch implements hooks for marking transactions via PF_FSTRANS.
When an allocation is performed in the context of PF_FSTRANS, any
KM_SLEEP allocation is transparently converted to a GFP_NOIO allocation.
Additionally, when using a Linux 3.9 or newer kernel, it will set
PF_MEMALLOC_NOIO to prevent direct reclaim from entering pageout() on
on any KM_PUSHPAGE or KM_NOSLEEP allocation. This effectively allows
the spl_vmalloc() helper function to be used safely in a thread which
is responsible for IO.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
This patch achieves the following goals:
1. It replaces the preprocessor kmem flag to gfp flag mapping with
proper translation logic. This eliminates the potential for
surprises that were previously possible where kmem flags were
mapped to gfp flags.
2. It maps vmem_alloc() allocations to kmem_alloc() for allocations
sized less than or equal to the newly-added spl_kmem_alloc_max
parameter. This ensures that small allocations will not contend
on a single global lock, large allocations can still be handled,
and potentially limited virtual address space will not be squandered.
This behavior is entirely different than under Illumos due to
different memory management strategies employed by the respective
kernels. However, this functionally provides the semantics required.
3. The --disable-debug-kmem, --enable-debug-kmem (default), and
--enable-debug-kmem-tracking allocators have been unified in to
a single spl_kmem_alloc_impl() allocation function. This was
done to simplify the code and make it more maintainable.
4. Improve portability by exposing an implementation of the memory
allocations functions that can be safely used in the same way
they are used on Illumos. Specifically, callers may safely
use KM_SLEEP in contexts which perform filesystem IO. This
allows us to eliminate an entire class of Linux specific changes
which were previously required to avoid deadlocking the system.
This change will be largely transparent to existing callers but there
are a few caveats:
1. Because the headers were refactored and extraneous includes removed
callers may find they need to explicitly add additional #includes.
In particular, kmem_cache.h must now be explicitly includes to
access the SPL's kmem cache implementation. This behavior is
different from Illumos but it was done to avoid always masking
the Linux slab functions when kmem.h is included.
2. Callers, like Lustre, which made assumptions about the definitions
of KM_SLEEP, KM_NOSLEEP, and KM_PUSHPAGE will need to be updated.
Other callers such as ZFS which did not will not require changes.
3. KM_PUSHPAGE is no longer overloaded to imply GFP_NOIO. It retains
its original meaning of allowing allocations to access reserved
memory. KM_PUSHPAGE callers can be converted back to KM_SLEEP.
4. The KM_NODEBUG flags has been retired and the default warning
threshold increased to 32k.
5. The kmem_virt() functions has been removed. For callers which
need to distinguish between a physical and virtual address use
is_vmalloc_addr().
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Address all cstyle issues in the kmem, vmem, and kmem_cache source
and headers. This will done to make it easier to review subsequent
changes which will rework the kmem/vmem implementation.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
This change introduces no functional changes to the memory management
interfaces. It only restructures the existing codes by separating the
kmem, vmem, and kmem cache implementations in the separate source and
header files.
Splitting this functionality in to separate files required the addition
of spl_vmem_{init,fini}() and spl_kmem_cache_{initi,fini}() functions.
Additionally, several minor changes to the #include's were required to
accommodate the removal of extraneous header from kmem.h.
But again, while large this patch introduces no functional changes.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
