On Sun Feb 7, 2021 at 4:12 AM CST, Christophe Leroy wrote:
>
>
> Le 06/02/2021 à 18:39, Christopher M. Riedl a écrit :
> > On Sat Feb 6, 2021 at 10:32 AM CST, Christophe Leroy wrote:
> >>
> >>
> >> Le 20/10/2020 à 04:01, Christopher M. Riedl a écrit :
> >>> On Fri Oct 16, 2020 at 10:48 AM CDT, Christophe Leroy wrote:
> >>>>
> >>>>
> >>>> Le 15/10/2020 à 17:01, Christopher M. Riedl a écrit :
> >>>>> Reuse the "safe" implementation from signal.c except for calling
> >>>>> unsafe_copy_from_user() to copy into a local buffer. Unlike the
> >>>>> unsafe_copy_{vsx,fpr}_to_user() functions the "copy from" functions
> >>>>> cannot use unsafe_get_user() directly to bypass the local buffer since
> >>>>> doing so significantly reduces signal handling performance.
> >>>>
> >>>> Why can't the functions use unsafe_get_user(), why does it significantly
> >>>> reduces signal handling
> >>>> performance ? How much significant ? I would expect that not going
> >>>> through an intermediate memory
> >>>> area would be more efficient
> >>>>
> >>>
> >>> Here is a comparison, 'unsafe-signal64-regs' avoids the intermediate
> >>> buffer:
> >>>
> >>> | | hash | radix |
> >>> | -------------------- | ------ | ------ |
> >>> | linuxppc/next | 289014 | 158408 |
> >>> | unsafe-signal64 | 298506 | 253053 |
> >>> | unsafe-signal64-regs | 254898 | 220831 |
> >>>
> >>> I have not figured out the 'why' yet. As you mentioned in your series,
> >>> technically calling __copy_tofrom_user() is overkill for these
> >>> operations. The only obvious difference between unsafe_put_user() and
> >>> unsafe_get_user() is that we don't have asm-goto for the 'get' variant.
> >>> Instead we wrap with unsafe_op_wrap() which inserts a conditional and
> >>> then goto to the label.
> >>>
> >>> Implemenations:
> >>>
> >>> #define unsafe_copy_fpr_from_user(task, from, label) do { \
> >>> struct task_struct *__t = task; \
> >>> u64 __user *buf = (u64 __user *)from; \
> >>> int i; \
> >>> \
> >>> for (i = 0; i < ELF_NFPREG - 1; i++) \
> >>> unsafe_get_user(__t->thread.TS_FPR(i), &buf[i], label); \
> >>> unsafe_get_user(__t->thread.fp_state.fpscr, &buf[i], label); \
> >>> } while (0)
> >>>
> >>> #define unsafe_copy_vsx_from_user(task, from, label) do { \
> >>> struct task_struct *__t = task; \
> >>> u64 __user *buf = (u64 __user *)from; \
> >>> int i; \
> >>> \
> >>> for (i = 0; i < ELF_NVSRHALFREG ; i++) \
> >>>
> >>> unsafe_get_user(__t->thread.fp_state.fpr[i][TS_VSRLOWOFFSET], \
> >>> &buf[i], label); \
> >>> } while (0)
> >>>
> >>
> >> Do you have CONFIG_PROVE_LOCKING or CONFIG_DEBUG_ATOMIC_SLEEP enabled in
> >> your config ?
> >
> > I don't have these set in my config (ppc64le_defconfig). I think I
> > figured this out - the reason for the lower signal throughput is the
> > barrier_nospec() in __get_user_nocheck(). When looping we incur that
> > cost on every iteration. Commenting it out results in signal performance
> > of ~316K w/ hash on the unsafe-signal64-regs branch. Obviously the
> > barrier is there for a reason but it is quite costly.
>
> Interesting.
>
> Can you try with the patch I just sent out
> https://patchwork.ozlabs.org/project/linuxppc-dev/patch/c72f014730823b413528e90ab6c4d3bcb79f8497.1612692067.git.christophe.le...@csgroup.eu/
Yeah that patch solves the problem. Using unsafe_get_user() in a loop is
actually faster on radix than using the intermediary buffer step. A
summary of results below (unsafe-signal64-v6 uses unsafe_get_user() and
avoids the local buffer):
| | hash | radix |
| -------------------------------- | ------ | ------ |
| unsafe-signal64-v5 | 194533 | 230089 |
| unsafe-signal64-v6 | 176739 | 202840 |
| unsafe-signal64-v5+barrier patch | 203037 | 234936 |
| unsafe-signal64-v6+barrier patch | 205484 | 241030 |
I am still expecting some comments/feedback on my v5 before sending out
v6. Should I include your patch in my series as well?
>
> Thanks
> Christophe
