For a generic explanation of why mutexs needed to be reimplemented
to work with the kernel lock profiling see commits:
e811949a57 and
d28db80fd0
The specific changes made to the mutex implemetation are as follows.
The Linux mutex structure is now directly embedded in the kmutex_t.
This allows a kmutex_t to be directly case to a mutex struct and
passed directly to the Linux primative.
Just like with the rwlocks it is critical that these functions be
implemented as '#defines to ensure the location information is
preserved. The preprocessor can then do a direct replacement of
the Solaris primative with the linux primative.
Just as with the rwlocks we need to track the lock owner. Here
things get a little more interesting because depending on your
kernel version, and how you've built your kernel Linux may already
do this for you. If your running a 2.6.29 or newer kernel on a
SMP system the lock owner will be tracked. This was added to Linux
to support adaptive mutexs, more on that shortly. Alternately, your
kernel might track the lock owner if you've set CONFIG_DEBUG_MUTEXES
in the kernel build. If neither of the above things is true for
your kernel the kmutex_t type will include and track the lock owner
to ensure correct behavior. This is all handled by a new autoconf
check called SPL_AC_MUTEX_OWNER.
Concerning adaptive mutexs these are a very recent development and
they did not make it in to either the latest FC11 of SLES11 kernels.
Ideally, I'd love to see this kernel change appear in one of these
distros because it does help performance. From Linux kernel commit:
0d66bf6d3514b35eb6897629059443132992dbd7
"Testing with Ingo's test-mutex application...
gave a 345% boost for VFS scalability on my testbox"
However, if you don't want to backport this change yourself you
can still simply export the task_curr() symbol. The kmutex_t
implementation will use this symbol when it's available to
provide it's own adaptive mutexs.
Finally, DEBUG_MUTEX support was removed including the proc handlers.
This was done because now that we are cleanly integrated with the
kernel profiling all this information and much much more is available
in debug kernel builds. This code was now redundant.
Update mutexs validated on:
- SLES10 (ppc64)
- SLES11 (x86_64)
- CHAOS4.2 (x86_64)
- RHEL5.3 (x86_64)
- RHEL6 (x86_64)
- FC11 (x86_64)
The behavior of RW_*_HELD was updated because it was not quite right.
It is not sufficient to return non-zero when the lock is help, we must
only do this when the current task in the holder.
This means we need to track the lock owner which is not something
tracked in a Linux semaphore. After some experimentation the
solution I settled on was to embed the Linux semaphore at the start
of a larger krwlock_t structure which includes the owner field.
This maintains good performance and allows us to cleanly intergrate
with the kernel lock analysis tools. My reasons:
1) By placing the Linux semaphore at the start of krwlock_t we can
then simply cast krwlock_t to a rw_semaphore and pass that on to
the linux kernel. This allows us to use '#defines so the preprocessor
can do direct replacement of the Solaris primative with the linux
equivilant. This is important because it then maintains the location
information for each rw_* call point.
2) Additionally, by adding the owner to krwlock_t we can keep this
needed extra information adjacent to the lock itself. This removes
the need for a fancy lookup to get the owner which is optimal for
performance. We can also leverage the existing spin lock in the
semaphore to ensure owner is updated correctly.
3) All helper functions which do not need to strictly be implemented
as a define to preserve location information can be done as a static
inline function.
4) Adding the owner to krwlock_t allows us to remove all memory
allocations done during lock initialization. This is good for all
the obvious reasons, we do give up the ability to specific the lock
name. The Linux profiling tools will stringify the lock name used
in the code via the preprocessor and use that.
Update rwlocks validated on:
- SLES10 (ppc64)
- SLES11 (x86_64)
- CHAOS4.2 (x86_64)
- RHEL5.3 (x86_64)
- RHEL6 (x86_64)
- FC11 (x86_64)
It turns out that the previous rwlock implementation worked well but
did not integrate properly with the upstream kernel lock profiling/
analysis tools. This is a major problem since it would be awfully
nice to be able to use the automatic lock checker and profiler.
The problem is that the upstream lock tools use the pre-processor
to create a lock class for each uniquely named locked. Since the
rwsem was embedded in a wrapper structure the name was always the
same. The effect was that we only ended up with one lock class for
the entire SPL which caused the lock dependency checker to flag
nearly everything as a possible deadlock.
The solution was to directly map a krwlock to a Linux rwsem using
a typedef there by eliminating the wrapper structure. This was not
done initially because the rwsem implementation is specific to the arch.
To fully implement the Solaris krwlock API using only the provided rwsem
API is not possible. It can only be done by directly accessing some of
the internal data member of the rwsem structure.
For example, the Linux API provides a different function for dropping
a reader vs writer lock. Whereas the Solaris API uses the same function
and the caller does not pass in what type of lock it is. This means to
properly drop the lock we need to determine if the lock is currently a
reader or writer lock. Then we need to call the proper Linux API function.
Unfortunately, there is no provided API for this so we must extracted this
information directly from arch specific lock implementation. This is
all do able, and what I did, but it does complicate things considerably.
The good news is that in addition to the profiling benefits of this
change. We may see performance improvements due to slightly reduced
overhead when creating rwlocks and manipulating them.
The only function I was forced to sacrafice was rw_owner() because this
information is simply not stored anywhere in the rwsem. Luckily this
appears not to be a commonly used function on Solaris, and it is my
understanding it is mainly used for debugging anyway.
In addition to the core rwlock changes, extensive updates were made to
the rwlock regression tests. Each class of test was extended to provide
more API coverage and to be more rigerous in checking for misbehavior.
This is a pretty significant change and with that in mind I have been
careful to validate it on several platforms before committing. The full
SPLAT regression test suite was run numberous times on all of the following
platforms. This includes various kernels ranging from 2.6.16 to 2.6.29.
- SLES10 (ppc64)
- SLES11 (x86_64)
- CHAOS4.2 (x86_64)
- RHEL5.3 (x86_64)
- RHEL6 (x86_64)
- FC11 (x86_64)
Basically everything we need to monitor the global memory state of
the system is now cleanly available via global_page_state(). The
problem is that this interface is still fairly recent, and there
has been one change in the page state enum which we need to handle.
These changes basically boil down to the following:
- If global_page_state() is available we should use it. Several
autoconf checks have been added to detect the correct enum names.
- If global_page_state() is not available check to see if
get_zone_counts() symbol is available and use that.
- If the get_zone_counts() symbol is not exported we have no choice
be to dynamically aquire it at load time. This is an absolute
last resort for old kernel which we don't want to patch to
cleanly export the symbol.
This interface is going away, and it's not as if most callers actually
use crhold/crfree when working with credentials. So it'll be okay
they we're not taking a reference on the task structure the odds of
it going away while working with a credential and pretty small.
The previous credential implementation simply provided the needed types and
a couple of dummy functions needed. This update correctly ties the basic
Solaris credential API in to one of two Linux kernel APIs.
Prior to 2.6.29 the linux kernel embeded all credentials in the task
structure. For these kernels, we pass around the entire task struct as if
it were the credential, then we use the helper functions to extract the
credential related bits.
As of 2.6.29 a new credential type was added which we can and do fairly
cleanly layer on top of. Once again the helper functions nicely hide
the implementation details from all callers.
Three tests were added to the splat test framework to verify basic
correctness. They should be extended as needed when need credential
functions are added.
Simply pass the ioctl on to the normal handler. If the ioctl
helper macros are used correctly this should be safe as they
will handle the packing/unpacking of the data encoded in the
ioctl command. And actually, if the caller does not use the
IO* macros at all, and just passes small values, it will probably
be OK as well. We only get in to trouble if they try and use
the upper 32-bits. Endianness is not really a concern here, we
we are pretty much assumed they user and kernel will match.
used to scale the number of threads based on the number of online
CPUs. As CPUs are added/removed we should rescale the thread
count appropriately, but currently this is only done at create.
- Initial SLES testing uncovered a long standing bug in the debug
tracing. The tcd_for_each() macro expected a NULL to terminate
the trace_data[i] array but this was only ever true due to luck.
All trace_data[] iterators are now properly capped by TCD_TYPE_MAX.
- SPLAT_MAJOR 229 conflicted with a 'hvc' device on my SLES system.
Since this was always an arbitrary choice I picked something else.
- The HAVE_PGDAT_LIST case should set pgdat_list_addr to the value stored
at the address of the memory location returned by kallsyms_lookup_name().
- Prior to 2.6.17 there were no *_pgdat helper functions in mm/mmzone.c.
Instead for_each_zone() operated directly on pgdat_list which may or
may not have been exported depending on how your kernel was compiled.
Now new configure checks determine if you have the helpers or not, and
if the needed symbols are exported. If they are not exported then they
are dynamically aquired at runtime by kallsyms_lookup_name().
- Configure check for SLES specific API change to vfs_unlink()
and vfs_rename() which added a 'struct vfsmount *' argument.
This was for something called the linux-security-module, but
it appears that it was never adopted upstream.
- Configure check, the div64_64() function was renamed to
div64_u64() as of 2.6.26.
- Configure check, the global_page_state() fuction was introduced
in 2.6.18 kernels. The earlier 2.6.16 based SLES10 must not try
and use it, thankfully get_zone_counts() is still available.
- To simplify debugging poison all symbols aquired dynamically
using spl_kallsyms_lookup_name() with SYMBOL_POISON.
- Add console messages when the user mode helpers fail.
- spl_kmem_init_globals() use bit shifts instead of division.
- When the monotonic clock is unavailable __gethrtime() must perform
the HZ division as an 'unsigned long long' because the SPL only
implements __udivdi3(), and not __divdi3() for 'long long' division
on 32-bit arches.
We need dependent packages to be able to include spl_config.h so they
can leverage the configure checks the SPL has done. This is important
because several of the spl headers need the results of these checks to
work properly. Unfortunately, the autoheader build product is always
private to a particular build and defined certain common things.
(PACKAGE, VERSION, etc). This prevents other packages which also use
autoheader from being include because the definitions conflict. To
avoid this problem the SPL build system leverage AH_BOTTOM to include
a spl_unconfig.h at the botton of the autoheader build product. This
custom include undefs all known shared symbols to prevent the confict.
This does however mean that those definition are also not availble
to the SPL package either. The SPL package therefore uses the
equivilant SPL_META_* definitions.
In the interests of portability I have added a FC10/i686 box to
my list of development platforms. The hope is this will allow me
to keep current with upstream kernel API changes, and at the same
time ensure I don't accidentally break x86 support. This patch
resolves all remaining issues observed under that environment.
1) SPL_AC_ZONE_STAT_ITEM_FIA autoconf check added. As of 2.6.21
the kernel added a clean API for modules to get the global count
for free, inactive, and active pages. The SPL attempts to detect
if this API is available and directly map spl_global_page_state()
to global_page_state(). If the full API is not available then
spl_global_page_state() is implemented as a thin layer to get
these values via get_zone_counts() if that symbol is available.
2) New kmem:vmem_size regression test added to validate correct
vmem_size() functionality. The test case acquires the current
global vmem state, allocates from the vmem region, then verifies
the allocation is correctly reflected in the vmem_size() stats.
3) Change splat_kmem_cache_thread_test() to always use KMC_KMEM
based memory. On x86 systems with limited virtual address space
failures resulted due to exhaustig the address space. The tests
really need to problem exhausting all memory on the system thus
we need to use the physical address space.
4) Change kmem:slab_lock to cap it's memory usage at availrmem
instead of using the native linux nr_free_pages(). This provides
additional test coverage of the SPL Linux VM integration.
5) Change kmem:slab_overcommit to perform allocation of 256K
instead of 1M. On x86 based systems it is not possible to create
a kmem backed slab with entires of that size. To compensate for
this the number of allocations performed in increased by 4x.
6) Additional autoconf documentation for proposed upstream API
changes to make additional symbols available to modules.
7) Console error messages added when spl_kallsyms_lookup_name()
fails to locate an expected symbol. This causes the module to fail
to load and we need to know exactly which symbol was not available.
I'm very surprised this has not surfaced until now. But the taskq_wait()
implementation work only wait successfully the first time it was called.
Subsequent usage of taskq_wait() on the taskq would not wait.
The issue was caused by tq->tq_lowest_id being set to MAX_INT after the
first wait completed. This caused subsequent waits which check that the
waiting id is less than the lowest taskq id to always succeed. The fix
is to ensure that tq->tq_lowest_id is never set larger than tq->tq_next.id.
Additional fixes which were added to this patch include:
1) Fix a race by placing the taskq_wait_check() in the tq->tq_lock spinlock.
2) taskq_wait() should wait for the largest outstanding id.
3) Multiple spelling corrections.
4) Added taskq wait regression test to validate correct behavior.
Fixes hostid mismatch which leads to assertion failure when the hostid/hw_serial is a 10-character decimal number:
$ zpool status
pool: lustre
state: ONLINE
lt-zpool: zpool_main.c:3176: status_callback: Assertion `reason == ZPOOL_STATUS_OK' failed.
zsh: 5262 abort zpool status
An update to the build system to properly support all commonly
used Makefile targets these include:
make all # Build everything
make install # Install everything
make clean # Clean up build products
make distclean # Clean up everything
make dist # Create package tarball
make srpm # Create package source RPM
make rpm # Create package binary RPMs
make tags # Create ctags and etags for everything
Extra care was taken to ensure that the source RPMs are fully
rebuildable against Fedora/RHEL/Chaos kernels. To build binary
RPMs from the source RPM for your system simply run:
rpmbuild --rebuild spl-x.y.z-1.src.rpm
This will produce two binary RPMs with correct 'requires'
dependencies for your kernel. One will contain all spl modules
and support utilities, the other is a devel package for compiling
additional kernel modules which are dependant on the spl.
spl-x.y.z-1_<kernel version>.x86_64.rpm
spl-devel-x.y.2-1_<kernel version>.x86_64.rpm
Remove all instances of functions being reimplemented in the SPL.
When the prototypes are available in the linux headers but the
function address itself is not exported use kallsyms_lookup_name()
to find the address. The function name itself can them become a
define which calls a function pointer. This is preferable to
reimplementing the function in the SPL because it ensures we get
the correct version of the function for the running kernel. This
is actually pretty safe because the prototype is defined in the
headers so we know we are calling the function properly.
This patch also includes a rhel5 kernel patch we exports the needed
symbols so we don't need to use kallsyms_lookup_name(). There are
autoconf checks to detect if the symbol is exported and if so to
use it directly. We should add patches for stock upstream kernels
as needed if for no other reason than so we can easily track which
additional symbols we needed exported. Those patches can also be
used by anyone willing to rebuild their kernel, but this should
not be a requirement. The rhel5 version of the export-symbols
patch has been applied to the chaos kernel.
Additional fixes:
1) Implement vmem_size() function using get_vmalloc_info()
2) SPL_CHECK_SYMBOL_EXPORT macro updated to use $LINUX_OBJ instead
of $LINUX because Module.symvers is a build product. When
$LINUX_OBJ != $LINUX we will not properly detect exported symbols.
3) SPL_LINUX_COMPILE_IFELSE macro updated to add include2 and
$LINUX/include search paths to allow proper compilation when
the kernel target build directory is not the source directory.
Minimal support added for the zone_get_hostid() function. Only
global zones are supported therefore this function must be called
with a NULL argumment. Additionally, I've added the HW_HOSTID_LEN
define and updated all instances where a hard coded magic value
of 11 was used; "A good riddance of bad rubbish!"
This check was originally added to detect double initializations
of mutex types (which it did find). Unfortunately, Coverity is
right that there is a very small chance we could trigger the
assertion by accident because an uninitialized stack variable
happens to contain the mutex magic. This is particularly unlikely
since we do poison the mutexs when destroyed but still possible.
Therefore I'm simply removing the assertion.
- The previous magazine ageing sceme relied on the on_each_cpu()
function to call spl_magazine_age() on each cpu. It turns out
this could deadlock with do_flush_tlb_all() which also relies
on the IPI based on_each_cpu(). To avoid this problem a per-
magazine delayed work item is created and indepentantly
scheduled to the correct cpu removing the need for on_each_cpu().
- Additionally two unused fields were removed from the type
spl_kmem_cache_t, they were hold overs from previous cleanup.
- struct work_struct work
- struct timer_list timer
- spl_slab_reclaim() 'continue' changed back to 'break' from commit
37db7d8cf9. The original was correct,
I have added a comment to ensure this does not happen again.
- spl_slab_reclaim() further optimized by moving the destructor call
in spl_slab_free() outside the skc->skc_lock. This minimizes the
length of time the spin lock is held, allows the destructors to
be invoked concurrently for different objects, and as a bonus makes
it safe (although unwise) to sleep in the destructors.
- Default SPL_KMEM_CACHE_DELAY changed to 15 to match Solaris.
- Aged out slab checking occurs every SPL_KMEM_CACHE_DELAY / 3.
- skc->skc_reap tunable added whichs allows callers of
spl_slab_reclaim() to cap the number of slabs reclaimed.
On Solaris all eligible slabs are always reclaimed, and this
is still the default behavior. However, I suspect that is
not always wise for reasons such as in the next comment.
- spl_slab_reclaim() added cond_resched() while walking the
slab/object free lists. Soft lockups were observed when
freeing large numbers of vmalloc'd slabs/objets.
- spl_slab_reclaim() 'sks->sks_ref > 0' check changes from
incorrect 'break' to 'continue' to ensure all slabs are
checked.
- spl_cache_age() reworked to avoid a deadlock with
do_flush_tlb_all() which occured because we slept waiting
for completion in spl_cache_age(). To waiting for magazine
reclamation to finish is not required so we no longer wait.
- spl_magazine_create() and spl_magazine_destroy() shifted
back to using for_each_online_cpu() instead of the
spl_on_each_cpu() approach which was of course a bad idea
due to memory allocations which Ricardo pointed out.
Added support for Solaris swapfs_minfree, and swapfs_reserve tunables.
In additional availrmem is now available and return a reasonable value
which is reasonably analogous to the Solaris meaning. On linux we
return the sun of free and inactive pages since these are all easily
reclaimable.
All tunables are available in /proc/sys/kernel/spl/vm/* and they may
need a little adjusting once we observe the real behavior. Some of
the defaults are mapped to similar linux counterparts, others are
straight from the OpenSolaris defaults.
Support added to provide reasonable values for the global Solaris
VM variables: minfree, desfree, lotsfree, needfree. These values
are set to the sum of their per-zone linux counterparts which
should be close enough for Solaris consumers.
When a non-GPL app links against the SPL we cannot use the udev
interfaces, which means non of the device special files are created.
Because of this I had added a poor mans udev which cause the SPL
to invoke an upcall and create the basic devices when a minor
is registered. When a minor is unregistered we use the vnode
interface to unlink the special file.
- Added SPL_AC_3ARGS_ON_EACH_CPU configure check to determine
if the older 4 argument version of on_each_cpu() should be
used or the new 3 argument version. The retry argument was
dropped in the new API which was never used anyway.
- Updated work queue compatibility wrappers. The old way this
worked was to pass a data point when initialized the workqueue.
The new API assumed the work item is embedding in a structure
and we us container_of() to find that data pointer.
- Updated skc->skc_flags to be an unsigned long which is now
type checked in the bit operations. This silences the warnings.
- Updated autogen products and splat tests accordingly
- Added slab work queue task which gradually ages and free's slabs
from the cache which have not been used recently.
- Optimized slab packing algorithm to ensure each slab contains the
maximum number of objects without create to large a slab.
- Fix deadlock, we can never call kv_free() under the skc_lock. We
now unlink the objects and slabs from the cache itself and attach
them to a private work list. The contents of the list are then
subsequently freed outside the spin lock.
- Move magazine create/destroy operation on to local cpu.
- Further performace optimizations by minimize the usage of the large
per-cache skc_lock. This includes the addition of KMC_BIT_REAPING
bit mask which is used to prevent concurrent reaping, and to defer
new slab creation when reaping is occuring.
- Add KMC_BIT_DESTROYING bit mask which is set when the cache is being
destroyed, this is used to catch any task accessing the cache while
it is being destroyed.
- Add comments to all the functions and additional comments to try
and make everything as clear as possible.
- Major cleanup and additions to the SPLAT kmem tests to more
rigerously stress the cache implementation and look for any problems.
This includes correctness and performance tests.
- Updated portable work queue interfaces
