On Fri, Feb 2, 2024 at 4:42 PM Alexandre Ghiti <alexghiti@xxxxxxxxxxxx> wrote: > > Hi Andrea, > > On Thu, Feb 1, 2024 at 4:03 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote: > > > > On Wed, Jan 31, 2024 at 04:59:29PM +0100, Alexandre Ghiti wrote: > > > The preventive sfence.vma were emitted because new mappings must be made > > > visible to the page table walker but Svvptc guarantees that xRET act as > > > a fence, so no need to sfence.vma for the uarchs that implement this > > > extension. > > > > AFAIU, your first submission shows that you don't need that xRET property. > > Similarly for other archs. What was rationale behind this Svvptc change? > > Actually, the ARC has just changed its mind and removed this new The wording was incorrect here, the ARC did not state anything, the author of Svvptc proposed an amended version of the spec that removes this behaviour and that's under discussion. > behaviour from the Svvptc extension, so we will take some gratuitous > page faults (but that should be outliners), which makes riscv similar > to x86 and arm64. > > > > > > > > This allows to drastically reduce the number of sfence.vma emitted: > > > > > > * Ubuntu boot to login: > > > Before: ~630k sfence.vma > > > After: ~200k sfence.vma > > > > > > * ltp - mmapstress01 > > > Before: ~45k > > > After: ~6.3k > > > > > > * lmbench - lat_pagefault > > > Before: ~665k > > > After: 832 (!) > > > > > > * lmbench - lat_mmap > > > Before: ~546k > > > After: 718 (!) > > > > This Svvptc seems to move/add the "burden" of the synchronization to xRET: > > Perhaps integrate the above counts w/ the perf gains in the cover letter? > > Yes, I'll copy that to the cover letter. > > Thanks for your interest! > > Alex > > > > > Andrea
