Re: [PATCH v2 2/2] mm: speed up mremap by 500x on large regions
Not 32-bit ARM? On Fri, Oct 12, 2018 at 6:35 PM, Joel Fernandes wrote: > On Fri, Oct 12, 2018 at 11:18:36AM -0700, David Miller wrote: >> From: Joel Fernandes > [...] >> > Also, do we not flush the caches from any path when we munmap >> > address space? We do call do_munmap on the old mapping from mremap >> > after moving to the new one. >> >> Sparc makes sure that shared mapping have consistent colors. Therefore >> all that's left are private mappings and those will be initialized by >> block stores to clear the page out or similar. >> >> Also, when creating new mappings, we flush the D-cache when necessary >> in update_mmu_cache(). >> >> We also maintain a bit in the page struct to track when a page which >> was potentially written to on one cpu ends up mapped into another >> address space and flush as necessary. >> >> The cache is write-through, which simplifies the preconditions we have >> to maintain. > > Makes sense, thanks. For the moment I sent patches to enable this on arm64 > and x86. We can enable it on sparc as well at a later time as it sounds it > could be a safe optimization to apply to that architecture as well. > > thanks, > > - Joel > ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH v2 2/2] mm: speed up mremap by 500x on large regions
I wonder whether it makes sense to expose to userspace somehow whether mremap is "fast" for a particular architecture. If a feature relies on fast mremap, it might be better for some userland component to disable that feature entirely rather than blindly use mremap and end up performing very poorly. If we're disabling fast mremap when THP is enabled, the userland component can't just rely on an architecture switch and some kind of runtime feature detection becomes even more important. On Fri, Oct 12, 2018 at 6:44 PM, Joel Fernandes wrote: > On Fri, Oct 12, 2018 at 06:39:45PM -0700, Daniel Colascione wrote: >> Not 32-bit ARM? > > Well, I didn't want to enable every possible architecture we could in a > single go. Certainly arm32 can be a follow on enablement as can be other > architectures. The point of this series is to upstream this feature and > enable a hand-picked few architectures as a first step. > > thanks, > > - Joel > ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH v2 2/2] mm: speed up mremap by 500x on large regions
On Fri, Oct 12, 2018 at 7:10 PM, Joel Fernandes wrote: > On Fri, Oct 12, 2018 at 06:54:33PM -0700, Daniel Colascione wrote: >> I wonder whether it makes sense to expose to userspace somehow whether >> mremap is "fast" for a particular architecture. If a feature relies on >> fast mremap, it might be better for some userland component to disable >> that feature entirely rather than blindly use mremap and end up >> performing very poorly. If we're disabling fast mremap when THP is >> enabled, the userland component can't just rely on an architecture >> switch and some kind of runtime feature detection becomes even more >> important. > > I hate to point out that its forbidden to top post on LKML :-) > https://kernelnewbies.org/mailinglistguidelines > So don't that Mr. Dan! :D Guilty as charged. I really should switch back to Gnus. :-) > But anyway, I think this runtime detection thing is not needed. THP is > actually expected to be as fast as this anyway, so if that's available then > we should already be as fast. Ah, I think the commit message is confusing. (Or else I'm misreading the patch now.) It's not quite that we're disabling the feature when THP is enabled anywhere, but rather that we use the move_huge_pmd path for huge PMDs and use the new code only for non-huge PMDs. (Right?) If that's the case, the commit message shouldn't say "Incase THP is enabled, the optimization is skipped". Even if THP is enabled on a system generally, we might use the new PMD-moving code for mapping types that don't support THP-ization, right? > This is for non-THP where THP cannot be enabled > and there is still room for some improvement. Most/all architectures will be > just fine with this. This flag is more of a safety-net type of thing where in > the future if there is this one or two weird architectures that don't play > well, then they can turn it off at the architecture level by not selecting > the flag. See my latest patches for the per-architecture compile-time > controls. Ideally we'd like to blanket turn it on on all, but this is just > playing it extra safe as Kirill and me were discussing on other threads. Sure. I'm just pointing out that the 500x performance different turns the operation into a qualitatively different feature, so if we expect to actually ship a mainstream architecture without support for this thing, we should make it explicit. If we're not, we shouldn't. ___ linux-snps-arc mailing list linux-snps-arc@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-snps-arc
