I blame Mikey for this. He elevated my slightly dubious testcase:
to benchmark status. And naturally we need to be number 1 at creating
zeros. So lets improve __clear_user some more.
As Paul suggests we can use dcbz for large lengths. This patch gets
the destination cacheline aligned then uses dcbz on whole cachelines.
Before:
10485760000 bytes (10 GB) copied, 0.414744 s, 25.3 GB/s
After:
10485760000 bytes (10 GB) copied, 0.268597 s, 39.0 GB/s
39 GB/s, a new record.
Signed-off-by: Anton Blanchard <anton@samba.org>
Tested-by: Olof Johansson <olof@lixom.net>
Acked-by: Olof Johansson <olof@lixom.net>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
At the moment all queues in a multiqueue adapter will serialise
against the IOMMU table lock. This is proving to be a big issue,
especially with 10Gbit ethernet.
This patch creates 4 pools and tries to spread the load across
them. If the table is under 1GB in size we revert back to the
original behaviour of 1 pool and 1 largealloc pool.
We create a hash to map CPUs to pools. Since we prefer interrupts to
be affinitised to primary CPUs, without some form of hashing we are
very likely to end up using the same pool. As an example, POWER7
has 4 way SMT and with 4 pools all primary threads will map to the
same pool.
The largealloc pool is reduced from 1/2 to 1/4 of the space to
partially offset the overhead of breaking the table up into pools.
Some performance numbers were obtained with a Chelsio T3 adapter on
two POWER7 boxes, running a 100 session TCP round robin test.
Performance improved 69% with this patch applied.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
In preparation for IOMMU pools, push the spinlock into
iommu_range_alloc and __iommu_free.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch moves tce_free outside of the lock in iommu_free.
Some performance numbers were obtained with a Chelsio T3 adapter on
two POWER7 boxes, running a 100 session TCP round robin test.
Performance improved 25% with this patch applied.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We currently hold the IOMMU spinlock around tce_build and tce_flush.
This causes our spinlock hold times to be much higher than required
and can impact multiqueue adapters.
This patch moves tce_build and tce_flush outside of the lock in
iommu_alloc, and tce_flush outside of the lock in iommu_free.
Some performance numbers were obtained with a Chelsio T3 adapter on
two POWER7 boxes, running a 100 session TCP round robin test.
Performance improved 32% with this patch applied.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
tce_buildmulti_pSeriesLP uses a per cpu page to communicate with the
hypervisor. We currently rely on the IOMMU table spinlock but
subsequent patches will be removing that so disable interrupts
around all accesses of tce_page.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Implement a POWER7 optimised memcpy using VMX and enhanced prefetch
instructions.
This is a copy of the POWER7 optimised copy_to_user/copy_from_user
loop. Detailed implementation and performance details can be found in
commit a66086b819 (powerpc: POWER7 optimised
copy_to_user/copy_from_user using VMX).
I noticed memcpy issues when profiling a RAID6 workload:
.memcpy
.async_memcpy
.async_copy_data
.__raid_run_ops
.handle_stripe
.raid5d
.md_thread
I created a simplified testcase by building a RAID6 array with 4 1GB
ramdisks (booting with brd.rd_size=1048576):
# mdadm -CR -e 1.2 /dev/md0 --level=6 -n4 /dev/ram[0-3]
I then timed how long it took to write to the entire array:
# dd if=/dev/zero of=/dev/md0 bs=1M
Before: 892 MB/s
After: 999 MB/s
A 12% improvement.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Version 2.06 of the POWER ISA introduced enhanced touch instructions,
allowing us to specify a number of attributes including the length of
a stream.
This patch adds a software stream for both loads and stores in the
POWER7 copy_tofrom_user loop. Since the setup is quite complicated
and we have to use an eieio to ensure correct ordering of the "GO"
command we only do this for copies above 4kB.
To quantify any performance improvements we need a working set
bigger than the caches so we operate on a 1GB file:
# dd if=/dev/zero of=/tmp/foo bs=1M count=1024
And we compare how fast we can read the file:
# dd if=/tmp/foo of=/dev/null bs=1M
before: 7.7 GB/s
after: 9.6 GB/s
A 25% improvement.
The worst case for this patch will be a completely L1 cache contained
copy of just over 4kB. We can test this with the copy_to_user
testcase we used to tune copy_tofrom_user originally:
http://ozlabs.org/~anton/junkcode/copy_to_user.c
# time ./copy_to_user2 -l 4224 -i 10000000
before: 6.807 s
after: 6.946 s
A 2% slowdown, which seems reasonable considering our data is unlikely
to be completely L1 contained.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
While creating the PCI root bus through function pci_create_root_bus()
of PCI core, it should have assigned the secondary bus number for the
newly created PCI root bus. Thus we needn't do the explicit assignment
for the secondary bus number again in pcibios_scan_phb().
Signed-off-by: Gavin Shan <shangw@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The form affinity for NUMA is set to 1 if the firmware supports
OPAL. Otherwise, we have to retrieve that from OF node "/chosen".
For the latter case, OF node "/chosen" reference count was never
decreased.
Signed-off-by: Gavin Shan <shangw@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Implement a POWER7 optimised copy_page using VMX and enhanced
prefetch instructions. We use enhanced prefetch hints to prefetch
both the load and store side. We copy a cacheline at a time and
fall back to regular loads and stores if we are unable to use VMX
(eg we are in an interrupt).
The following microbenchmark was used to assess the impact of
the patch:
http://ozlabs.org/~anton/junkcode/page_fault_file.c
We test MAP_PRIVATE page faults across a 1GB file, 100 times:
# time ./page_fault_file -p -l 1G -i 100
Before: 22.25s
After: 18.89s
17% faster
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Subsequent patches will add more VMX library functions and it makes
sense to keep all the c-code helper functions in the one file.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
mtmsrd is an expensive instruction, we save a few cycles by
doing it once instead of twice.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Version 2.06 of the POWER ISA introduced enhanced touch instructions,
allowing us to specify a number of attributes including the length of
a stream.
This patch adds a software stream for both loads and stores in the
POWER7 copy_tofrom_user loop. Since the setup is quite complicated
and we have to use an eieio to ensure correct ordering of the "GO"
command we only do this for copies above 4kB.
To quantify any performance improvements we need a working set
bigger than the caches so we operate on a 1GB file:
# dd if=/dev/zero of=/tmp/foo bs=1M count=1024
And we compare how fast we can read the file:
# dd if=/tmp/foo of=/dev/null bs=1M
before: 7.7 GB/s
after: 9.6 GB/s
A 25% improvement.
The worst case for this patch will be a completely L1 cache contained
copy of just over 4kB. We can test this with the copy_to_user
testcase we used to tune copy_tofrom_user originally:
http://ozlabs.org/~anton/junkcode/copy_to_user.c
# time ./copy_to_user2 -l 4224 -i 10000000
before: 6.807 s
after: 6.946 s
A 2% slowdown, which seems reasonable considering our data is unlikely
to be completely L1 contained.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
1) call_function.lock used in smp_call_function_many() is just to protect
call_function.queue and &data->refs, cpu_online_mask is outside of the
lock. And it's not necessary to protect cpu_online_mask,
because data->cpumask is pre-calculate and even if a cpu is brougt up
when calling arch_send_call_function_ipi_mask(), it's harmless because
validation test in generic_smp_call_function_interrupt() will take care
of it.
2) For cpu down issue, stop_machine() will guarantee that no concurrent
smp_call_fuction() is processing.
Signed-off-by: Yong Zhang <yong.zhang0@gmail.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
I noticed __clear_user high up in a profile of one of my RAID stress
tests. The testcase was doing a dd from /dev/zero which ends up
calling __clear_user.
__clear_user is basically a loop with a single 4 byte store which
is horribly slow. We can do much better by aligning the desination
and doing 32 bytes of 8 byte stores in a loop.
The following testcase was used to verify the patch:
http://ozlabs.org/~anton/junkcode/stress_clear_user.c
To show the improvement in performance I ran a dd from /dev/zero
to /dev/null on a POWER7 box:
Before:
# dd if=/dev/zero of=/dev/null bs=1M count=10000
10485760000 bytes (10 GB) copied, 3.72379 s, 2.8 GB/s
After:
# time dd if=/dev/zero of=/dev/null bs=1M count=10000
10485760000 bytes (10 GB) copied, 0.728318 s, 14.4 GB/s
Over 5x faster.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
irq_entry, irq_exit, timer_interrupt_entry and timer_interrupt_exit
all do the same thing so use DECLARE_EVENT_CLASS to avoid duplicating
everything 4 times.
This saves quite a lot of space in both instruction text and data:
text data bss dec hex filename
9265 19622 16 28903 70e7 arch/powerpc/kernel/irq.o
6817 19019 16 25852 64fc arch/powerpc/kernel/irq.o
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When looking through some instruction traces I noticed our tracepoint
checks were inline. It turns out we don't have CONFIG_JUMP_LABEL
enabled.
By enabling CONFIG_JUMP_LABEL we replace a load/compare/branch with
a nop at every tracepoint call. For example in do_IRQ:
CONFIG_JUMP_LABEL disabled:
stdx 3,11,9
lwz 0,8(29)
cmpwi 7,0,0
bne- 7,.L124
bl .irq_enter
CONFIG_JUMP_LABEL enabled:
stdx 3,11,9
nop
bl .irq_enter
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The following patch is to remove the pseries_notify_add_cpu() call
and replace it by a hot plug notifier.
This would prevent cpuidle resources being released and allocated each
time cpu comes online on pseries.
The earlier design was causing a lockdep problem
in start_secondary as reported on this thread
-https://lkml.org/lkml/2012/5/17/2
This applies on 3.4-rc7
Signed-off-by: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
An upcoming release of firmware will add DDW extensions, in particular
an API to "reset" the DMA window to the original configuration (32-bit,
2GB in size). With that API available, we can safely remove the default
window, increasing the resources available to firmware for creation of
larger windows for the slot in question -- if we encounter an error, we
can use the new API to reset the state of the slot.
Further, this same release of firmware will make it a hard requirement
for OSes to release the existing window before any other windows will be
shown as available, to avoid conflicts in addressing between the two
windows.
In anticipation of these changes, always remove the default window
before we do any DDW manipulations.
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The patch_instruction() interface is made to modify kernel text. It is
safer to use that then the probe_kernel_write() when modifying kernel
code.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
For ftrace to use the patch_instruction code, it needs to check for
faults on write. Ftrace updates code all over the kernel, and we need to
know if code is updated or not due to protections that are placed on
some portions of the kernel. If ftrace does not detect a fault, it will
error later on, and it will be much more difficult to find the problem.
By changing patch_instruction() to detect faults, then ftrace will be
able to make use of it too.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
PowerPC does not have the synchronization issues that x86 has with
modifying code on one CPU while another CPU is executing it.
The other CPU will either see the old or new code without any
issues, unlike x86 which may issue a GPF.
Instead of calling the heavy stop_machine, just update the code.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Currently we build all board files regardless of the final zImage
target. This is sub-optimal (in terms on compilation) and leads to
problems in one platform needlessly causing failures for other
platforms.
Use the Kconfig variables to selectively construct this board files to
build.
Signed-off-by: Tony Breeds <tony@bakeyournoodle.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Pull a couple more powerpc fixes from Benjamin Herrenschmidt:
"Here are two more fixes that I "missed" when scrubbing patchwork last
week which are worth still having in 3.5."
* 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/benh/powerpc:
powerpc/kvm: sldi should be sld
powerpc/xmon: Use cpumask iterator to avoid warning
commit 43220aa0f2
md/raid5: fix a hang on device failure.
fixed a hang, but introduced a refcounting in-balance so
that if the presence of bad-blocks ever caused an rdev to
be 'blocked' we would increment the refcount on the rdev and
never decrement it.
So added the needed rdev_dec_pending when md_wait_for_blocked_rdev
is not called.
Reported-by: majianpeng <majianpeng@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
In ops_run_io(), the call to md_wait_for_blocked_rdev will decrement
nr_pending so we lose the reference we hold on the rdev.
So atomic_inc it first to maintain the reference.
This bug was introduced by commit 73e92e51b7
md/raid5. Don't write to known bad block on doubtful devices.
which appeared in 3.0, so patch is suitable for stable kernels since
then.
Cc: stable@vger.kernel.org
Signed-off-by: majianpeng <majianpeng@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
In chunk_aligned_read() we are adding data_offset before calling
is_badblock. But is_badblock also adds data_offset, so that is bad.
So move the addition of data_offset to after the call to
is_badblock.
This bug was introduced by commit 31c176ecdf
md/raid5: avoid reading from known bad blocks.
which first appeared in 3.0. So that patch is suitable for any
-stable kernel from 3.0.y onwards. However it will need minor
revision for most of those (as the comment didn't appear until
recently).
Cc: stable@vger.kernel.org
Signed-off-by: majianpeng <majianpeng@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
If a RAID5 has both a failed device and a device marked as
'WantReplacement', then we should preferentially replace the failed
device.
However the current code replaces whichever is found first.
So split into 2 loops, check fail failed/missing first, and only check
for WantReplacement if nothing is failed or missing.
Reported-by: majianpeng <majianpeng@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
If a RAID10 has an odd number of chunks - as might happen when there
are an odd number of devices - the last chunk has no pair and so is
not mirrored. We don't store data there, but when recovering the last
device in an array we retry to recover that last chunk from a
non-existent location. This results in an error, and the recovery
aborts.
When we get to that last chunk we should just stop - there is nothing
more to do anyway.
This bug has been present since the introduction of RAID10, so the
patch is appropriate for any -stable kernel.
Cc: stable@vger.kernel.org
Reported-by: Christian Balzer <chibi@gol.com>
Tested-by: Christian Balzer <chibi@gol.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Prune this down to just the struct kvm_irqfd so we can avoid
changing function definition for every flag or field we use.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Acked-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
While we are resolving directory modifications in the
tree log, we are triggering delayed metadata updates to
the filesystem btrees.
This commit forces the delayed updates to run so the
replay code can find any modifications done. It stops
us from crashing because the directory deleltion replay
expects items to be removed immediately from the tree.
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
cc: stable@kernel.org
We can race with unlink and not actually be able to do our igrab in
btrfs_add_ordered_extent. This will result in all sorts of problems.
Instead of doing the complicated work to try and handle returning an error
properly from btrfs_add_ordered_extent, just hold a ref to the inode during
writepages. If we cannot grab a ref we know we're freeing this inode anyway
and can just drop the dirty pages on the floor, because screw them we're
going to invalidate them anyway. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The tree log stuff can have allocated space that we end up having split
across a bitmap and a real extent. The free space code does not deal with
this, it assumes that if it finds an extent or bitmap entry that the entire
range must fall within the entry it finds. This isn't necessarily the case,
so rework the remove function so it can handle this case properly. This
fixed two panics the user hit, first in the case where the space was
initially in a bitmap and then in an extent entry, and then the reverse
case. Thanks,
Reported-and-tested-by: Shaun Reich <sreich@kde.org>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
When we're evicting an inode during log recovery, we need to ensure that the inode
is not in orphan state any more, which means inode's run_time flags has _no_
BTRFS_INODE_HAS_ORPHAN_ITEM. Thus, the BUG_ON was triggered because of a wrong
check for the flags.
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We used the wrong ioctl macro for the getflags ioctl before.
As we don't have the set/getflags ioctls in the user space ioctl.h
at the moment, it's safe to fix it now.
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Alexander Block <ablock84@googlemail.com>
This introduces btrfs_resume_balance_async(), which, given that
restriper state was recovered earlier by btrfs_recover_balance(),
resumes balance in btrfs-balance kthread.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Fix a bug that triggered asserts in btrfs_balance() in both normal and
resume modes -- restriper state was not properly restored on read-only
mounts. This factors out resuming code from btrfs_restore_balance(),
which is now also called earlier in the mount sequence to avoid the
problem of some early writes getting the old profile.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Miao pointed out there's a problem with mixing dio writes and buffered
reads. If the read happens between us invalidating the page range and
actually locking the extent we can bring in pages into page cache. Then
once the write finishes if somebody tries to read again it will just find
uptodate pages and we'll read stale data. So we need to lock the extent and
check for uptodate bits in the range. If there are uptodate bits we need to
unlock and invalidate again. This will keep this race from happening since
we will hold the extent locked until we create the ordered extent, and then
teh read side always waits for ordered extents. There was also a race in
how we updated i_size, previously we were relying on the generic DIO stuff
to adjust the i_size after the DIO had completed, but this happens outside
of the extent lock which means reads could come in and not see the updated
i_size. So instead move this work into where we create the extents, and
then this way the update ordered i_size stuff works properly in the endio
handlers. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
It is normal behaviour of the low level btrfs function btrfs_map_bio()
to complete a bio with -EIO if the device is missing, instead of just
preventing the bio creation in an earlier step.
This used to cause I/O statistic read error increments and annoying
printk_ratelimited messages. This commit fixes the issue.
Signed-off-by: Stefan Behrens <sbehrens@giantdisaster.de>
Reported-by: Carey Underwood <cwillu@cwillu.com>
The SYS_NIRQ1 pin is the interupt line for the PMIC part of the TWL6030
and interrupts from the PMIC are needed as wakeup sources.
Ensure this pin is mux'd as input and has wakeup enabled so PMIC
interupts (e.g. RTC) can be used as wakeup sources.
Tested on OMAP4430/Panda.
Signed-off-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
In order for suspend/resume dependencies to work correctly, I2C has to
be initialized (more specifically, registered with the driver core)
before MMC. Without this, the MMC driver fails to adjust the VMMC
regulator (using i2c writes) during the suspend path.
Problem found testing suspend/resume on 3730/OveroSTORM platform.
Signed-off-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
write_file_bool() modifies 32 bits of data, so "amd_iommu_unmap_flush"
needs to be 32 bits as well or we'll corrupt memory. Fortunately it
looks like the data is aligned with a gap after the declaration so this
is harmless in production.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Even though it has "bool" in the name, you have pass a u32 pointer to
debugfs_create_bool(). Otherwise you get memory corruption in
write_file_bool(). Fortunately in this case the corruption happens in
an alignment hole between variables so it doesn't cause any problems.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
allo_pdir() is called in smmu_iommu_domain_init() with spin_lock
held. memory allocations in it have to be atomic/unsleepable.
Signed-off-by: Hiroshi DOYU <hdoyu@nvidia.com>
Reported-by: Chris Wright <chrisw@sous-sol.org>
Acked-by: Chris Wright <chrisw@sous-sol.org>
Cc: stable@vger.kernel.org
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
