Mac M1 fix proposal. #278
Mac M1 fix proposal. #278
Conversation
davidchisnall
left a comment
There was a problem hiding this comment.
Thanks for working on this, a couple of comments inline.
src/pal/pal_apple.h
Outdated
| template<bool page_aligned = false> | ||
| static void zero(void* p, size_t size) noexcept | ||
| { | ||
| ::bzero(p, size); |
There was a problem hiding this comment.
If this is required only on AArch64 (is it only AArch64 macOS, or iOS derivatives as well?) then can we guard it on a constexpr if check that the AAL is for AArch64, not x86?
There was a problem hiding this comment.
fair enough. yes Mac M1 is "only" a super iOs device anyway.
src/pal/pal_apple.h
Outdated
| static void notify_not_using(void* p, size_t size) noexcept | ||
| { | ||
| SNMALLOC_ASSERT(is_aligned_block<PALBSD::page_size>(p, size)); | ||
| while (madvise(p, size, MADV_FREE_REUSABLE) == -1 && errno == EAGAIN) |
There was a problem hiding this comment.
I don't believe this is correct. MADV_FREE_REUSABLE is completely undocumented (and therefore Apple reserves the right to remove it or modify its semantics in minor version updates) but the comments in Apple's open-source libmalloc indicate that you must madvise with MADV_FREE_REUSE before reuse. We would need a notify_using implementation that does this. I don't know what the performance differences are between MADV_FREE and the pair of MADV_FREE_REUSEABLE + MADV_FREE_REUSE, if it's significant then we might want to provide both options to allow users to tune for lower memory or higher performance.
There was a problem hiding this comment.
Yes this doesd even appear on the man page. however it s not related at all to performance (it s mostly the same) but memory accounting "faithfulness" but fair point I can give up this part.
There was a problem hiding this comment.
From what I can tell, it's not just about accounting, it has different semantics. The traditional MADV_FREE has to track the dirty bit to tell whether you've used the page and so can require an IPI and TLB shoot-down to make the page read-only in the map before it's actually removed. This version can reclaim the page immediately but requires a system call (which explicitly synchronises any thread that might be about to use the page) before it can be reused.
It looks like it's a good fit for snmalloc, but we do need the corresponding notify_using function in the PAL.
There was a problem hiding this comment.
Might be better in another PR as here it s mainly to fix some unit tests crashing whereas this is improvement.
There was a problem hiding this comment.
Sounds good, thanks!
The actual code works fine on the usual mac Intel however, some failures occurs with some unit tests on the ARM h/w when zero'ing page ranges.
e3c68ce to
4810079
Compare
The actual code works fine on the usual Mac Intel however, some failures
occurs with some unit tests on the ARM h/w when zero'ing page range.
While at it, using MADV_FREE_REUSABLE which is Darwin's specific but helps
command line tools to report better memory accounting.