diff options
author | Christian Borntraeger <borntraeger@de.ibm.com> | 2015-01-30 10:31:13 +0100 |
---|---|---|
committer | Martin Schwidefsky <schwidefsky@de.ibm.com> | 2015-02-26 09:24:49 +0100 |
commit | fb3d1c085c05e0e4b112d915dbd06b20b259e6c5 (patch) | |
tree | 650756a903449ccc32a9cc8c3c4cbadae347d929 /arch/s390/pci | |
parent | f0483044c1c96089256cda4cf182eea1ead77fe4 (diff) | |
download | linux-fb3d1c085c05e0e4b112d915dbd06b20b259e6c5.tar.gz linux-fb3d1c085c05e0e4b112d915dbd06b20b259e6c5.tar.bz2 linux-fb3d1c085c05e0e4b112d915dbd06b20b259e6c5.zip |
s390: let the compiler do page clearing
The hardware folks told me that for page clearing "when you exactly
know what to do, hand written xc+pfd is usally faster then mvcl for
page clearing, as it saves millicode overhead and parameter parsing
and checking" as long as you dont need the cache bypassing.
Turns out that gcc already does a proper xc,pfd loop.
A small test on z196 that does
buff = mmap(NULL, bufsize,PROT_EXEC|PROT_WRITE|PROT_READ,AP_PRIVATE| MAP_ANONYMOUS,0,0);
for ( i = 0; i < bufsize; i+= 256)
buff[i] = 0x5;
gets 20% faster (touches every cache line of a page)
and
buff = mmap(NULL, bufsize,PROT_EXEC|PROT_WRITE|PROT_READ,AP_PRIVATE| MAP_ANONYMOUS,0,0);
for ( i = 0; i < bufsize; i+= 4096)
buff[i] = 0x5;
is within noise ratio (touches one cache line of a page).
As the clear_page is usually called for first memory accesses
we can assume that at least one cache line is used afterwards,
so this change should be always better.
Another benchmark, a make -j 40 of my testsuite in tmpfs with
hot caches on a 32cpu system:
-- unpatched -- -- patched --
real 0m1.017s real 0m0.994s (~2% faster, but in noise)
user 0m5.339s user 0m5.016s (~6% faster)
sys 0m0.691s sys 0m0.632s (~8% faster)
Let use the same define to memset as the asm-generic variant
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Diffstat (limited to 'arch/s390/pci')
0 files changed, 0 insertions, 0 deletions