1. Patchew
  2. QEMU
  3. View series

[Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing

Emilio G. Cota posted 5 patches 7 years ago
Failed in applying to current master (apply log)
There is a newer version of this series
[Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing
Posted by Emilio G. Cota 7 years ago 
v1: https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01146.html

Changes since v1:

- Add tlb_index and tlb_entry helpers from Richard

- Introduce sizeof_tlb() and tlb_n_entries()

- Extract tlb_mask as its own array in CPUArchState, as
  suggested by Richard. For the associated helpers (tlb_index etc)
  I tried several approaches, and performance-wise they're all
  the same, so went for the simplest implementation.

- Use uintptr_t for tlb_mask, as done in Richard's patch
  + tcg/i386: use hrexw when reading tlb_mask
  + Define tlbtype and tlbrexw solely based on TARGET_PAGE_BITS

- Rename tlb_is_invalid to tlb_entry_is_empty, comparing all
  fields (except .addend) against -1.

- Rename CPUTLBDesc.used to .n_used_entries.

- Drop the MIN/MAX CPU_TLB_BITS patches, defining instead
  some values for MIN/MAX as well as a default.

- Use new_size and old_size consistently in the resizing function,
  as suggested by Richard.

- Add an additional chart to the last patch, where softmmu
  performance is compared against user-mode:
    https://imgur.com/a/eXkjMCE

You can fetch this series from:
  https://github.com/cota/qemu/tree/tlb-dyn-v2

Note that it applies on top of my tlb-lock-v4 series:
  https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01421.html

Thanks,

		Emilio
Re: [Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing
Posted by Alex Bennée 7 years ago 
Emilio G. Cota <cota@braap.org> writes:

> v1: https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01146.html
>
> Changes since v1:

Hmm I'm seeing some qtest failures, for example:

  $ make check-qtest-alpha V=1
  ...
  QTEST_QEMU_BINARY=alpha-softmmu/qemu-system-alpha QTEST_QEMU_IMG=qemu-img MALLOC_PERTURB_=${MALLOC_PERTURB_:-$(( ${RANDOM:-0} % 255 + 1))} gtester -k --verbose -m=quick test
  s/boot-serial-test tests/qmp-test tests/qmp-cmd-test tests/device-introspect-test tests/cdrom-test tests/machine-none-test tests/qom-test tests/test-hmp
  TEST: tests/boot-serial-test... (pid=31091)
    /alpha/boot-serial/clipper:                                          Broken pipe
  tests/libqtest.c:129: kill_qemu() detected QEMU death from signal 11 (Segmentation fault) (core dumped)
  FAIL
  GTester: last random seed: R02S948c4a5112fd7682934f4d96e1aff38e
  (pid=31099)
  FAIL: tests/boot-serial-test


>
> - Add tlb_index and tlb_entry helpers from Richard
>
> - Introduce sizeof_tlb() and tlb_n_entries()
>
> - Extract tlb_mask as its own array in CPUArchState, as
>   suggested by Richard. For the associated helpers (tlb_index etc)
>   I tried several approaches, and performance-wise they're all
>   the same, so went for the simplest implementation.
>
> - Use uintptr_t for tlb_mask, as done in Richard's patch
>   + tcg/i386: use hrexw when reading tlb_mask
>   + Define tlbtype and tlbrexw solely based on TARGET_PAGE_BITS
>
> - Rename tlb_is_invalid to tlb_entry_is_empty, comparing all
>   fields (except .addend) against -1.
>
> - Rename CPUTLBDesc.used to .n_used_entries.
>
> - Drop the MIN/MAX CPU_TLB_BITS patches, defining instead
>   some values for MIN/MAX as well as a default.
>
> - Use new_size and old_size consistently in the resizing function,
>   as suggested by Richard.
>
> - Add an additional chart to the last patch, where softmmu
>   performance is compared against user-mode:
>     https://imgur.com/a/eXkjMCE
>
> You can fetch this series from:
>   https://github.com/cota/qemu/tree/tlb-dyn-v2
>
> Note that it applies on top of my tlb-lock-v4 series:
>   https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01421.html
>
> Thanks,
>
> 		Emilio


--
Alex Bennée
Re: [Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing
Posted by Emilio G. Cota 7 years ago 
On Tue, Oct 09, 2018 at 13:34:40 +0100, Alex Bennée wrote:
> 
> Emilio G. Cota <cota@braap.org> writes:
> 
> > v1: https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01146.html
> >
> > Changes since v1:
> 
> Hmm I'm seeing some qtest failures, for example:
> 
>   $ make check-qtest-alpha V=1
>   ...
>   QTEST_QEMU_BINARY=alpha-softmmu/qemu-system-alpha QTEST_QEMU_IMG=qemu-img MALLOC_PERTURB_=${MALLOC_PERTURB_:-$(( ${RANDOM:-0} % 255 + 1))} gtester -k --verbose -m=quick test
>   s/boot-serial-test tests/qmp-test tests/qmp-cmd-test tests/device-introspect-test tests/cdrom-test tests/machine-none-test tests/qom-test tests/test-hmp
>   TEST: tests/boot-serial-test... (pid=31091)
>     /alpha/boot-serial/clipper:                                          Broken pipe
>   tests/libqtest.c:129: kill_qemu() detected QEMU death from signal 11 (Segmentation fault) (core dumped)
>   FAIL
>   GTester: last random seed: R02S948c4a5112fd7682934f4d96e1aff38e
>   (pid=31099)
>   FAIL: tests/boot-serial-test

I'm pretty sure that the problem is that tlb_init is not being
called at all. Note that this applies to the tlb-lock series
as well, although there we're just calling qemu_spin_init,
which is not really necessary because CPUArchState is 0-allocated.

I'll take a look.

Thanks,

		E.
Re: [Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing
Posted by Alex Bennée 7 years ago 
Emilio G. Cota <cota@braap.org> writes:

> On Tue, Oct 09, 2018 at 13:34:40 +0100, Alex Bennée wrote:
>>
>> Emilio G. Cota <cota@braap.org> writes:
>>
>> > v1: https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01146.html
>> >
>> > Changes since v1:
>>
>> Hmm I'm seeing some qtest failures, for example:
>>
>>   $ make check-qtest-alpha V=1
>>   ...
>>   QTEST_QEMU_BINARY=alpha-softmmu/qemu-system-alpha QTEST_QEMU_IMG=qemu-img MALLOC_PERTURB_=${MALLOC_PERTURB_:-$(( ${RANDOM:-0} % 255 + 1))} gtester -k --verbose -m=quick test
>>   s/boot-serial-test tests/qmp-test tests/qmp-cmd-test tests/device-introspect-test tests/cdrom-test tests/machine-none-test tests/qom-test tests/test-hmp
>>   TEST: tests/boot-serial-test... (pid=31091)
>>     /alpha/boot-serial/clipper:                                          Broken pipe
>>   tests/libqtest.c:129: kill_qemu() detected QEMU death from signal 11 (Segmentation fault) (core dumped)
>>   FAIL
>>   GTester: last random seed: R02S948c4a5112fd7682934f4d96e1aff38e
>>   (pid=31099)
>>   FAIL: tests/boot-serial-test
>
> I'm pretty sure that the problem is that tlb_init is not being
> called at all. Note that this applies to the tlb-lock series
> as well, although there we're just calling qemu_spin_init,
> which is not really necessary because CPUArchState is 0-allocated.
>
> I'll take a look.

Yeah I hadn't tried to bisect it, but I'm on top of tlb-lock-v4 as
requested.

--
Alex Bennée
Re: [Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing
Posted by Emilio G. Cota 7 years ago 
On Tue, Oct 09, 2018 at 15:45:36 +0100, Alex Bennée wrote:
> 
> Emilio G. Cota <cota@braap.org> writes:
> 
> > On Tue, Oct 09, 2018 at 13:34:40 +0100, Alex Bennée wrote:
> >>
> >> Emilio G. Cota <cota@braap.org> writes:
> >>
> >> > v1: https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01146.html
> >> >
> >> > Changes since v1:
> >>
> >> Hmm I'm seeing some qtest failures, for example:
> >>
> >>   $ make check-qtest-alpha V=1
> >>   ...
> >>   QTEST_QEMU_BINARY=alpha-softmmu/qemu-system-alpha QTEST_QEMU_IMG=qemu-img MALLOC_PERTURB_=${MALLOC_PERTURB_:-$(( ${RANDOM:-0} % 255 + 1))} gtester -k --verbose -m=quick test
> >>   s/boot-serial-test tests/qmp-test tests/qmp-cmd-test tests/device-introspect-test tests/cdrom-test tests/machine-none-test tests/qom-test tests/test-hmp
> >>   TEST: tests/boot-serial-test... (pid=31091)
> >>     /alpha/boot-serial/clipper:                                          Broken pipe
> >>   tests/libqtest.c:129: kill_qemu() detected QEMU death from signal 11 (Segmentation fault) (core dumped)
> >>   FAIL
> >>   GTester: last random seed: R02S948c4a5112fd7682934f4d96e1aff38e
> >>   (pid=31099)
> >>   FAIL: tests/boot-serial-test
> >
> > I'm pretty sure that the problem is that tlb_init is not being
> > called at all. Note that this applies to the tlb-lock series
> > as well, although there we're just calling qemu_spin_init,
> > which is not really necessary because CPUArchState is 0-allocated.
> >
> > I'll take a look.
> 
> Yeah I hadn't tried to bisect it, but I'm on top of tlb-lock-v4 as
> requested.

It's the tlb_flush in alpha_cpu_initfn:

static void alpha_cpu_initfn(Object *obj)
{
    CPUState *cs = CPU(obj);
    AlphaCPU *cpu = ALPHA_CPU(obj);
    CPUAlphaState *env = &cpu->env;

    cs->env_ptr = env;
    tlb_flush(cs);

We call tlb_init later on at realize time.

I think we can safely get rid of this tlb_flush. Unicore
also has it. I'll add patches for both to the tlb-lock series.

		Emilio
Re: [Qemu-devel] [RFC v2 0/5] Dynamic TLB sizing
Posted by Alex Bennée 7 years ago 
Emilio G. Cota <cota@braap.org> writes:

> On Tue, Oct 09, 2018 at 15:45:36 +0100, Alex Bennée wrote:
>>
>> Emilio G. Cota <cota@braap.org> writes:
>>
>> > On Tue, Oct 09, 2018 at 13:34:40 +0100, Alex Bennée wrote:
>> >>
>> >> Emilio G. Cota <cota@braap.org> writes:
>> >>
>> >> > v1: https://lists.gnu.org/archive/html/qemu-devel/2018-10/msg01146.html
>> >> >
>> >> > Changes since v1:
>> >>
>> >> Hmm I'm seeing some qtest failures, for example:
>> >>
>> >>   $ make check-qtest-alpha V=1
>> >>   ...
>> >>   QTEST_QEMU_BINARY=alpha-softmmu/qemu-system-alpha QTEST_QEMU_IMG=qemu-img MALLOC_PERTURB_=${MALLOC_PERTURB_:-$(( ${RANDOM:-0} % 255 + 1))} gtester -k --verbose -m=quick test
>> >>   s/boot-serial-test tests/qmp-test tests/qmp-cmd-test tests/device-introspect-test tests/cdrom-test tests/machine-none-test tests/qom-test tests/test-hmp
>> >>   TEST: tests/boot-serial-test... (pid=31091)
>> >>     /alpha/boot-serial/clipper:                                          Broken pipe
>> >>   tests/libqtest.c:129: kill_qemu() detected QEMU death from signal 11 (Segmentation fault) (core dumped)
>> >>   FAIL
>> >>   GTester: last random seed: R02S948c4a5112fd7682934f4d96e1aff38e
>> >>   (pid=31099)
>> >>   FAIL: tests/boot-serial-test
>> >
>> > I'm pretty sure that the problem is that tlb_init is not being
>> > called at all. Note that this applies to the tlb-lock series
>> > as well, although there we're just calling qemu_spin_init,
>> > which is not really necessary because CPUArchState is 0-allocated.
>> >
>> > I'll take a look.
>>
>> Yeah I hadn't tried to bisect it, but I'm on top of tlb-lock-v4 as
>> requested.
>
> It's the tlb_flush in alpha_cpu_initfn:
>
> static void alpha_cpu_initfn(Object *obj)
> {
>     CPUState *cs = CPU(obj);
>     AlphaCPU *cpu = ALPHA_CPU(obj);
>     CPUAlphaState *env = &cpu->env;
>
>     cs->env_ptr = env;
>     tlb_flush(cs);
>
> We call tlb_init later on at realize time.
>
> I think we can safely get rid of this tlb_flush.

Agreed.

> Unicore
> also has it. I'll add patches for both to the tlb-lock series.

Weirdly this didn't fail, but I agree it's superfluous.

>
> 		Emilio


--
Alex Bennée

Patchew 2.1-devel-b67e4c2

© 2016 - 2025 Red Hat, Inc.