The implementation for memory copy functions on ARM had a (disabled)
provision for aligning the source pointer before loading registers with
data. Turns out that aligning the _destination_ pointer is much more
useful, as the read side is already sufficiently helped with the use of
preload.
So this changes the definition of the CALGN() macro to target the
destination pointer instead, and turns it on for Feroceon processors
where the gain is very noticeable.
Signed-off-by: Nicolas Pitre <nico@marvell.com>
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Fix a lot of typos. Eyeballed by jmc@ in OpenBSD.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Patch from Nicolas Pitre
This patch provides a new implementation for optimized memory copy
functions on ARM. It is made of two levels: a template that consists of
the core copy code and separate files that define macros to be used with
the core code depending on the type of copy needed. This allows for best
performances while sharing the same core for implementing memcpy(),
copy_from_user() and copy_to_user() for instance.
Two reasons for this work:
1) the current copy_to_user/copy_from_user implementation assumes no
task switch will ever occur in the middle of each copied page making
it completely unsafe with CONFIG_PREEMPT=y.
2) current copy implementations are measurably suboptimal and optimizing
different implementations separately is a pain and more opportunities
for bugs.
The reason for (1) is the fact that copy inside user pages are performed
with the ldm instruction which has no mean for testing user protections
and could possibly race with process preemption bypassing the COW mechanism
for example. This is a longstanding issue that we said ought to be fixed
for about two years now. The solution is to substitute those ldm insns
with a series of ldrt or strt insns to enforce user memory protection.
At least on StrongARM and XScale cores the ldm is not faster than the
equivalent ldr/str insns with a warm i-cache so there is no measurable
performance degradation with that change. The fact that the copy code is
a template makes it pretty easy to reuse the same core code as for memcpy
and benefit from the same performance optimizations.
Now (2) is best demonstrated with actual throughput measurements.
First, here is a summary of memcopy tests performed on a StrongARM core:
PTR alignment buffer size kernel version this version
------------------------------------------------------------
aligned 32 59.73 107.43
unaligned 32 61.31 74.72
aligned 100 132.47 136.15
unaligned 100 103.84 123.76
aligned 4096 130.67 130.80
unaligned 4096 130.68 130.64
aligned 1048576 68.03 68.18
unaligned 1048576 68.03 68.18
The buffer size is in bytes and the measured speed in MB/s. The copy
was performed repeatedly with given buffer and throughput averaged over
3 seconds.
Here we can see that the current kernel version has a higher entry cost
that shows up with small buffers. As buffer size grows both implementation
converge to the same throughput.
Now here's the exact same test performed on an XScale core (PXA255):
PTR alignment buffer size kernel version this version
------------------------------------------------------------
aligned 32 46.99 77.58
unaligned 32 53.61 59.59
aligned 100 107.19 136.59
unaligned 100 83.61 97.58
aligned 4096 129.13 129.98
unaligned 4096 128.36 128.53
aligned 1048576 53.76 59.41
unaligned 1048576 33.67 56.96
Again we can see the entry setup cost being higher for the current kernel
before getting to the main copy loop. Then throughput results converge
as long as the buffer remains in the cache. Then the 1MB case shows more
differences probably due to better pld placement and/or less instruction
interlocks in this proposed implementation.
Disclaimer: The PXA system was running with slower clocks than the
StrongARM system so trying to infer any conclusion by comparing those
separate sets of results side by side would be completely inappropriate.
So... What this patch does is to replace both memcpy and memmove with
an implementation based on the provided copy code template. The memmove
code is kept separate since it is used only if the memory areas involved
do overlap in which case the code is a transposition of the template but
with the copy occurring in the opposite direction (trying to fit that
mode into the template turned it into a mess not worth it for memmove
alone). And obviously both memcpy and memmove were tested with all kinds
of pointer alignments and buffer sizes to exercise all code paths for
correctness.
The next patch will provide the now trivial replacement implementation
copy_to_user and copy_from_user.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
