The memory pinning code in uaccess_with_memcpy.c does not check
for HugeTLB or THP pmds, and will enter an infinite loop should
a __copy_to_user or __clear_user occur against a huge page.
This patch adds detection code for huge pages to pin_page_for_write.
As this code can be executed in a fast path it refers to the actual
pmds rather than the vma. If a HugeTLB or THP is found (they have
the same pmd representation on ARM), the page table spinlock is
taken to prevent modification whilst the page is pinned.
On ARM, huge pages are only represented as pmds, thus no huge pud
checks are performed. (For huge puds one would lock the page table
in a similar manner as in the pmd case).
Two helper functions are introduced; pmd_thp_or_huge will check
whether or not a page is huge or transparent huge (which have the
same pmd layout on ARM), and pmd_hugewillfault will detect whether
or not a page fault will occur on write to the page.
Running the following test (with the chunking from read_zero
removed):
$ dd if=/dev/zero of=/dev/null bs=10M count=1024
Gave: 2.3 GB/s backed by normal pages,
2.9 GB/s backed by huge pages,
5.1 GB/s backed by huge pages, with page mask=HPAGE_MASK.
After some discussion, it was decided not to adopt the HPAGE_MASK,
as this would have a significant detrimental effect on the overall
system latency due to page_table_lock being held for too long.
This could be revisited if split huge page locks are adopted.
Signed-off-by: Steve Capper <steve.capper@linaro.org>
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Add pud_offset() et.al. between the pgd and pmd code in preparation of
using pgtable-nopud.h rather than 4level-fixup.h.
This incorporates a fix from Jamie Iles <jamie@jamieiles.com> for
uaccess_with_memcpy.c.
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Previous size thresholds were guessed from various user space benchmarks
using a kernel with and without the alternative uaccess option. This
is however not as precise as a kernel based test to measure the real
speed of each method.
This adds a simple test bench to show the time needed for each method.
With this, the optimal size treshold for the alternative implementation
can be determined with more confidence. It appears that the optimal
threshold for both copy_to_user and clear_user is around 64 bytes. This
is not a surprise knowing that the memcpy and memset implementations
need at least 64 bytes to achieve maximum throughput.
One might suggest that such test be used to determine the optimal
threshold at run time instead, but results are near enough to 64 on
tested targets concerned by this alternative copy_to_user implementation,
so adding some overhead associated with a variable threshold is probably
not worth it for now.
Signed-off-by: Nicolas Pitre <nico@marvell.com>
Because the alternate copy_to_user implementation has a higher setup cost
than the standard implementation, the size of the memory area to copy
is tested and the standard implementation invoked instead when that size
is too small. Still, that test is made after the processor has preserved
a bunch of registers on the stack which have to be reloaded right away
needlessly in that case, causing a measurable performance regression
compared to plain usage of the standard implementation only.
To make the size test overhead negligible, let's factorize it out of
the alternate copy_to_user function where it is clear to the compiler
that no stack frame is needed. Thanks to CONFIG_ARM_UNWIND allowing
for frame pointers to be disabled and tail call optimization to kick in,
the overhead in the small copy case becomes only 3 assembly instructions.
A similar trick is applied to clear_user as well.
Signed-off-by: Nicolas Pitre <nico@marvell.com>
This implements {copy_to,clear}_user() by faulting in the userland
pages and then using the regular kernel mem{cpy,set}() to copy the
data (while holding the page table lock). This is a win if the regular
mem{cpy,set}() implementations are faster than the user copy functions,
which is the case e.g. on Feroceon, where 8-word STMs (which memcpy()
uses under the right conditions) give significantly higher memory write
throughput than a sequence of individual 32bit stores.
Here are numbers for page sized buffers on some Feroceon cores:
- copy_to_user on Orion5x goes from 51 MB/s to 83 MB/s
- clear_user on Orion5x goes from 89MB/s to 314MB/s
- copy_to_user on Kirkwood goes from 240 MB/s to 356 MB/s
- clear_user on Kirkwood goes from 367 MB/s to 1108 MB/s
- copy_to_user on Disco-Duo goes from 248 MB/s to 398 MB/s
- clear_user on Disco-Duo goes from 328 MB/s to 1741 MB/s
Because the setup cost is non negligible, this is worthwhile only if
the amount of data to copy is large enough. The operation falls back
to the standard implementation when the amount of data is below a certain
threshold. This threshold was determined empirically, however some targets
could benefit from a lower runtime determined value for optimal results
eventually.
In the copy_from_user() case, this technique does not provide any
worthwhile performance gain due to the fact that any kind of read access
allocates the cache and subsequent 32bit loads are just as fast as the
equivalent 8-word LDM.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Signed-off-by: Nicolas Pitre <nico@marvell.com>
Tested-by: Martin Michlmayr <tbm@cyrius.com>
Steven Capper