On 4/15/23 00:02, Matthew Wilcox wrote:
On Fri, Apr 14, 2023 at 10:47:39PM +0200, Bernd Schubert wrote:
Up to bs=512K it works fine, 1M (and for what it matters
already 768K) introduce the order=0 issue.
Hmm, I replaced memcpy with dumb version, that copies byte by byte - problem
gone. Is it possible that the optimized memcpy causes caused kind of random
memory access and confuses mm / readahead?
And somehow your memcpy or system is not doing that?
Oh, that would make sense! If the memcpy() works high-to-low, then
you'd see exactly the behaviour you're reporting. Whereas low-to-high
results in the behaviour I'm seeing.
In my case it is not exactly high-to-low, it is more low, high, then
from high to low. Issue goes away with a sufficiently large RA size. And
RA behaves better when POSIX_MADV_SEQUENTIAL is not done.
For sure memcpy implementation and cpu depending (I have tested on avx
and avx2 systems).
Hm. I don't know what to do about that. Maybe a "sufficiently large"
memcpy should call posix_madvise(src, n, POSIX_MADV_WILLNEED)
What would speak against ignoring POSIX_MADV_SEQUENTIAL for RA when it
would detect that something breaks the contract?
But then I don't know how much mmap/memcpy is used for large memcpy or
memcmp - maybe we should just document the issue right now in the mmap
man page? Especially that POSIX_MADV_SEQUENTIAL should be avoided?
Thanks,
Bernd