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>
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>
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>
Implement a POWER7 optimised copy_to_user/copy_from_user using VMX.
For large aligned copies this new loop is over 10% faster, and for
large unaligned copies it is over 200% faster.
If we take a fault we fall back to the old version, this keeps
things relatively simple and easy to verify.
On POWER7 unaligned stores rarely slow down - they only flush when
a store crosses a 4KB page boundary. Furthermore this flush is
handled completely in hardware and should be 20-30 cycles.
Unaligned loads on the other hand flush much more often - whenever
crossing a 128 byte cache line, or a 32 byte sector if either sector
is an L1 miss.
Considering this information we really want to get the loads aligned
and not worry about the alignment of the stores. Microbenchmarks
confirm that this approach is much faster than the current unaligned
copy loop that uses shifts and rotates to ensure both loads and
stores are aligned.
We also want to try and do the stores in cacheline aligned, cacheline
sized chunks. If the store queue is unable to merge an entire
cacheline of stores then the L2 cache will have to do a
read/modify/write. Even worse, we will serialise this with the stores
in the next iteration of the copy loop since both iterations hit
the same cacheline.
Based on this, the new loop does the following things:
1 - 127 bytes
Get the source 8 byte aligned and use 8 byte loads and stores. Pretty
boring and similar to how the current loop works.
128 - 4095 bytes
Get the source 8 byte aligned and use 8 byte loads and stores,
1 cacheline at a time. We aren't doing the stores in cacheline
aligned chunks so we will potentially serialise once per cacheline.
Even so it is much better than the loop we have today.
4096 - bytes
If both source and destination have the same alignment get them both
16 byte aligned, then get the destination cacheline aligned. Do
cacheline sized loads and stores using VMX.
If source and destination do not have the same alignment, we get the
destination cacheline aligned, and use permute to do aligned loads.
In both cases the VMX loop should be optimal - we always do aligned
loads and stores and are always doing stores in cacheline aligned,
cacheline sized chunks.
To be able to use VMX we must be careful about interrupts and
sleeping. We don't use the VMX loop when in an interrupt (which should
be rare anyway) and we wrap the VMX loop in disable/enable_pagefault
and fall back to the existing copy_tofrom_user loop if we do need to
sleep.
The VMX breakpoint of 4096 bytes was chosen using this microbenchmark:
http://ozlabs.org/~anton/junkcode/copy_to_user.c
Since we are using VMX and there is a cost to saving and restoring
the user VMX state there are two broad cases we need to benchmark:
- Best case - userspace never uses VMX
- Worst case - userspace always uses VMX
In reality a userspace process will sit somewhere between these two
extremes. Since we need to test both aligned and unaligned copies we
end up with 4 combinations. The point at which the VMX loop begins to
win is:
0% VMX
aligned 2048 bytes
unaligned 2048 bytes
100% VMX
aligned 16384 bytes
unaligned 8192 bytes
Considering this is a microbenchmark, the data is hot in cache and
the VMX loop has better store queue merging properties we set the
breakpoint to 4096 bytes, a little below the unaligned breakpoints.
Some future optimisations we can look at:
- Looking at the perf data, a significant part of the cost when a
task is always using VMX is the extra exception we take to restore
the VMX state. As such we should do something similar to the x86
optimisation that restores FPU state for heavy users. ie:
/*
* If the task has used fpu the last 5 timeslices, just do a full
* restore of the math state immediately to avoid the trap; the
* chances of needing FPU soon are obviously high now
*/
preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5;
and
/*
* fpu_counter contains the number of consecutive context switches
* that the FPU is used. If this is over a threshold, the lazy fpu
* saving becomes unlazy to save the trap. This is an unsigned char
* so that after 256 times the counter wraps and the behavior turns
* lazy again; this to deal with bursty apps that only use FPU for
* a short time
*/
- We could create a paca bit to mirror the VMX enabled MSR bit and check
that first, avoiding multiple calls to calling enable_kernel_altivec.
That should help with iovec based system calls like readv.
- We could have two VMX breakpoints, one for when we know the user VMX
state is loaded into the registers and one when it isn't. This could
be a second bit in the paca so we can calculate the break points quickly.
- One suggestion from Ben was to save and restore the VSX registers
we use inline instead of using enable_kernel_altivec.
[BenH: Fixed a problem with preempt and fixed build without CONFIG_ALTIVEC]
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Anton Blanchard