Message ID | 20200819182856.4893-1-robert.foley@linaro.org |
---|---|
Headers | show |
Series | accel/tcg: remove implied BQL from cpu_handle_interrupt/exception path | expand |
On Wed, 19 Aug 2020 14:28:49 -0400 Robert Foley <robert.foley@linaro.org> wrote: > The purpose of this change is to set the groundwork > so that an arch could move towards removing > the BQL from the cpu_handle_interrupt/exception paths. > > The BQL is a bottleneck in scaling to more cores. > And this cpu_handle_interrupt/exception path is one of > the key BQL users as measured by the QEMU sync profiling (qsp). > > As the first step in removing the BQL from this path, we will make > changes to the core/common functions of cpu_handle_interrupt/exception > to drop the holding of the BQL. The holding of the BQL is pushed down > to the per-arch implementation code. I have only skimmed the patches I was cc:ed on so far, but the series seems sane to me in general. > > This patch goes through several transitions of the code in order to > maintain correctness (bisectability). In order to maintain > bisectability across these steps some patches need to touch many > files across different arches, however most of the changes are trivial. > > The general order of the changes is below where each step > represents one patch. > > 1) rename all *_do_interrupt functions to *_do_interrupt_locked I'm wondering whether this renaming could be done in an automated way (e.g. via Coccinelle). Reviewing the method for the renaming is often easier than looking at a lot of similar code patterns. > > 2) add a new function *_do_interrupt that takes the BQL and calls > *_do_interrupt_locked, point ->do_interrupt to it, and remove > the BQL from cpu-exec.c's cpu_handle_exception. > > 3) modify the BQL critical sections around > ->cpu_exec_interrupt, so that the BQL critical section covers just the > call to ->cpu_exec_interrupt. > > 4/5) same as 1/2 for ->cpu_exec_interrupt. This removes the BQL > from cpu_handle_exception. The method of doing this in steps looks fine, although the patches produced are a bit unwieldy -- that's something we have to live with, I guess.
On Fri, 21 Aug 2020 at 06:56, Cornelia Huck <cohuck@redhat.com> wrote: > > > > As the first step in removing the BQL from this path, we will make > > changes to the core/common functions of cpu_handle_interrupt/exception > > to drop the holding of the BQL. The holding of the BQL is pushed down > > to the per-arch implementation code. > > I have only skimmed the patches I was cc:ed on so far, but the series > seems sane to me in general. Thanks for the feedback ! > > > > > This patch goes through several transitions of the code in order to > > maintain correctness (bisectability). In order to maintain > > bisectability across these steps some patches need to touch many > > files across different arches, however most of the changes are trivial. > > > > The general order of the changes is below where each step > > represents one patch. > > > > 1) rename all *_do_interrupt functions to *_do_interrupt_locked > > I'm wondering whether this renaming could be done in an automated way > (e.g. via Coccinelle). Reviewing the method for the renaming is often > easier than looking at a lot of similar code patterns. Good point. We will look into this. Thanks, -Rob Foley > > > > > 2) add a new function *_do_interrupt that takes the BQL and calls > > *_do_interrupt_locked, point ->do_interrupt to it, and remove > > the BQL from cpu-exec.c's cpu_handle_exception. > > > > 3) modify the BQL critical sections around > > ->cpu_exec_interrupt, so that the BQL critical section covers just the > > call to ->cpu_exec_interrupt. > > > > 4/5) same as 1/2 for ->cpu_exec_interrupt. This removes the BQL > > from cpu_handle_exception. > > The method of doing this in steps looks fine, although the patches > produced are a bit unwieldy -- that's something we have to live with, I > guess. >