On 2020-10-14 8:28 a.m., Ingo Molnar wrote:
* Ankur Arora <ankur.a.arora@xxxxxxxxxx> wrote:Uncached writes are suitable for circumstances where the region written to is not expected to be read again soon, or the region written to is large enough that there's no expectation that we will find the writes in the cache. Accordingly switch to using clear_page_uncached() for gigantic pages. Signed-off-by: Ankur Arora <ankur.a.arora@xxxxxxxxxx> --- mm/memory.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/mm/memory.c b/mm/memory.c index eeae590e526a..4d2c58f83ab1 100644 --- a/mm/memory.c +++ b/mm/memory.c @@ -5092,7 +5092,7 @@ static void clear_gigantic_page(struct page *page, for (i = 0; i < pages_per_huge_page; i++, p = mem_map_next(p, page, i)) { cond_resched(); - clear_user_highpage(p, addr + i * PAGE_SIZE); + clear_user_highpage_uncached(p, addr + i * PAGE_SIZE); } }So this does the clearing in 4K chunks, and your measurements suggest that short memory clearing is not as efficient, right?
I did not measure that separately (though I should), but the performance numbers around that were somewhat puzzling. For MOVNTI, the performance via perf bench (single call to memset_movnti()) is pretty close (within margin of error) to what we see with the page-fault workload (4K chunks in clear_page_nt().) With 'REP;STOS' though, there's degradation (~30% Broadwell, ~5% Rome) between perf bench (single call to memset_erms()) compared to the page-fault workload (4K chunks in clear_page_erms()). In the second case, we are executing a lot more 'REP;STOS' loops while the number of instructions in the first case is pretty much the same, so maybe that's what accounts for it. But I checked and we are not frontend bound. Maybe there are high setup costs for 'REP;STOS' on Broadwell? It does advertise X86_FEATURE_ERMS though...
I'm wondering whether it would make sense to do 2MB chunked clearing on 64-bit CPUs, instead of 512x 4k clearing? Both 2MB and GB pages are continuous in memory, so accessible to these instructions in a single narrow loop.
Yeah, I think it makes sense to do and should be quite straight-forward as well. I'll try that out. I suspect it might help the X86_FEATURE_NT_BAD models more but there's no reason why for it to hurt anywhere. Ankur
Thanks,
Ingo
