[PATCH v4 0/4] Miscellaneous about sbi_tlb and sbi_ipi

[Xiang W]

在 2023-04-06星期四的 12:30 +0530，Anup Patel写道：
> On Thu, Apr 6, 2023 at 10:27 AM Anup Patel <anup at brainfault.org> wrote:
> > 
> > On Thu, Mar 16, 2023 at 1:52 PM Xiang W <wxjstz at 126.com> wrote:
> > > 
> > > 在 2023-02-02星期四的 16:51 +0800，Xiang W写道：
> > > > Removed the mutual waiting between process and sync in sbi_tlb
> > > > Update ipi first then wait together
> > > > 
> > > Please review!
> > > Optimization can improve tlb sync efficiency.
> > > 
> > > sequence diagram before optimization:
> > > ===================================================================
> > > send ipi hart       hart 0       hart 1     hart 2   ...
> > >       +               +           +           +
> > >       |               |           |           |
> > >       | update hart 0 |           |           |
> > >       +<------------->+--+        |           |
> > >       |               |  |process |           |
> > >       |               |  |        |           |
> > >       |               +<-+        |           |
> > >       | sync          |           |           |
> > >       +<------------->+           |           |
> > >       | update hart 1             |           |
> > >       +<------------------------->+--+        |
> > >       |                           |  |process |
> > >       |                           |  |        |
> > >       |                           +<-+        |
> > >       | sync                      |           |
> > >       +<------------------------->+           |
> > >       | update hart 2                         |
> > >       +<------------------------------------->+--+
> > >       |                                       |  |process
> > >       |                                       |  |
> > >       |                                       +<-+
> > >       | sync                                  |
> > >       +<------------------------------------->+
> > >       .
> > >       .
> > >       .
> > > 
> > > sequence diagram after optimization:
> > > ===================================================================
> > > send ipi hart       hart 0       hart 1     hart 2   ...
> > >       +               +           +           +
> > >       |               |           |           |
> > >       | update hart 0 |           |           |
> > >       +<------------->+--+        |           |
> > >       | update hart 1 |  |process |           |
> > >       +<--------------|--|------->+--+        |
> > >       | update hart 2 +<-+        |  |process |
> > >       +<--------------|-----------|--|------->+--+
> > >       |               |           +<-+        |  |process
> > >       |               |           |           |  |
> > >       | sync          |           |           +<-+
> > >       +<------------->+           |           |
> > >       | sync                      |           |
> > >       +<------------------------->+           |
> > >       | sync                                  |
> > >       +<------------------------------------->+
> > >       .
> > >       .
> > >       .
> > 
> > Sorry for being slow on this series. Overall, the approach looks
> > good but I need to review individual patches.
> > 
> > In the meantime, I suggest you try running stress-ng one
> > QEMU virt machine with this patch applied in OpenSBI.
> 
> To run stress-ng on QEMU virt machine, launch QEMU virt
> machine with at least 1G RAM and Fedora RootFS located
> at:
> https://dl.fedoraproject.org/pub/alt/risc-v/repo/virt-builder-images/images/
> https://dl.fedoraproject.org/pub/alt/risc-v/repo/virt-builder-images/images/Fedora-Developer-37-20221130.n.0-nvme.raw.img.xz
> 
> I typically run a cyclic stress-ng test as follows:
> stress-ng --verbose --seq 4 -t 20 -x cyclic

The following is the log of my test:

without patches: https://drive.google.com/file/d/1I1pfA39tJvcZqFI55FSPxD-6o7mBaD1x/view?usp=share_link
with patches: https://drive.google.com/file/d/1FRRa6HJilS87dTgFYdA4XJRHhPIOBZms/view?usp=share_link

Regards,
Xiang W

> 
> Regards,
> Anup
> 
> > 
> > Regards,
> > Anup
> > 
> > > 
> > > Regards,
> > > Xiang W
> > > 
> > > > V4 Changes:
> > > > PATCH1: Fixed not adding one to tlb_sync when SBI_FIFO_UNCHANGED
> > > > PATCH4: Add comments
> > > > 
> > > > V3 Changes:
> > > > PATCH2: Remove missing sync in update
> > > > add PATCH4: Optimize sbi_tlb queue waiting
> > > > 
> > > > V2 Changes:
> > > > PATCH2: Fix m changes after update IPIs
> > > > 
> > > > Xiang W (4):
> > > >   lib: sbi: Optimize sbi_tlb
> > > >   lib: sbi: Optimize sbi_ipi
> > > >   lib: sbi: rename function of sbi_ipi
> > > >   lib: sbi: Optimize sbi_tlb queue waiting
> > > > 
> > > >  include/sbi/sbi_ipi.h |  3 +++
> > > >  lib/sbi/sbi_ipi.c     | 43 ++++++++++++++++++++++++++++++-------
> > > >  lib/sbi/sbi_tlb.c     | 49 +++++++++++++++++++------------------------
> > > >  3 files changed, 61 insertions(+), 34 deletions(-)
> > > > 
> > > 
> > > 
> > > --
> > > opensbi mailing list
> > > opensbi at lists.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/opensbi