On 3/13/2023 11:00 PM, Richard Henderson wrote:
On 3/13/23 07:13, Wu, Fei2 wrote:
Hi Richard,
Sorry for disturbing you. I'm doing some perf profiling on qemu-riscv64,
I see 10%+ faster to build stress-ng without the following patch. I know
it's incorrect to just skip this patch, I'm wondering if we can do
something on intercepting mmap/mprotect (very rare), e.g. even
invalidating all the TBs, but keep the cross-page block chaining.
It also affects breakpoints.
I have no good ideas for how to keep cross-page block chaining without
breaking either of these use cases. If you come up with a good idea,
please post on qemu-devel for discussion.
Thank you for reply. I am new to qemu/tcg, lots of details and
backgrounds need to catch up.
If we only want to address user-mode qemu, and assume this cross-page
chain, first page -> second page:
* breakpoints. If a new bp is added to second page, the chain is hard to
maintain, but it looks acceptable to flush all TBs and fall back to
current non-cross-page implementation during debugging? I think It's
different from the full system situation here:
https://gitlab.com/qemu-project/qemu/-/issues/404
* mprotect. If the 2nd page remains 'X' permission after mprotect, the
chain is still valid, if it's changed to non-X, then the syscall
interceptor will change the permission of corresponding host page to
non-X, it will be segfault as expected?
* mmap. I cannot figure out the situation. Is there any unit test for
this, or could you please shed some light?