Skip to contentSkip to conversation

drm/i915: check to see if SIMD registers are available before using SIMD

PATCH In review Jason A. Donenfeld 6 participants 14 messages Apr 30, 2020 – May 4, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915: check to see if SIMD registers are available before using SIMD

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

---
 drivers/gpu/drm/i915/i915_memcpy.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/gpu/drm/i915/i915_memcpy.c +12 -0
index fdd550405fd3..7c0e022586bc 100644
@@ -24,6 +24,7 @@
 
 #include <linux/kernel.h>
 #include <asm/fpu/api.h>
+#include <asm/simd.h>
 
 #include "i915_memcpy.h"
 
@@ -38,6 +39,12 @@ static DEFINE_STATIC_KEY_FALSE(has_movntdqa);
 #ifdef CONFIG_AS_MOVNTDQA
 static void __memcpy_ntdqa(void *dst, const void *src, unsigned long len)
 {
+	if (unlikely(!may_use_simd())) {
+		memcpy(dst, src, len);
+		return;
+	}
+
+
 	kernel_fpu_begin();
 
 	while (len >= 4) {
@@ -67,6 +74,11 @@ static void __memcpy_ntdqa(void *dst, const void *src, unsigned long len)
 
 static void __memcpy_ntdqu(void *dst, const void *src, unsigned long len)
 {
+	if (unlikely(!may_use_simd())) {
+		memcpy(dst, src, len);
+		return;
+	}
+
 	kernel_fpu_begin();
 
 	while (len >= 4) {
Signature · 1 line
2.26.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 6 lines · Jason A. Donenfeld

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

That would indicate that these functions are used from IRQ/softirq which break otherwise if the kernel is also using the registers. The crypto code uses it for that purpose.

So

May I ask how large the memcpy can be? I'm asking in case it is large and an explicit rescheduling point might be needed.

Quoted 2 lines

Sebastian

#2 lore raw

From: Sebastian Andrzej Siewior

Quoted 13 lines

Sent: 01 May 2020 11:42

Quoted 6 lines · Jason A. Donenfeld

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

That would indicate that these functions are used from IRQ/softirq which break otherwise if the kernel is also using the registers. The crypto code uses it for that purpose.

Quoted 5 lines

So

May I ask how large the memcpy can be? I'm asking in case it is large and an explicit rescheduling point might be needed.

It is also quite likely that a 'rep movs' copy will be at least just as fast on modern hardware.

Clearly if you are copying to/from PCIe memory you need the largest resisters possible - but I think the graphics buffers are mapped cached? (Otherwise I wouldn't see 3ms 'spins' while it invalidates the entire screen buffer cache.)

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#3 lore raw
In reply to Sebastian Andrzej Siewior
Quoted 1 line · Sebastian Andrzej Siewior

Thanks.

Quoted 3 lines

May I ask how large the memcpy can be? I'm asking in case it is large and an explicit rescheduling point might be needed.

Yea I was worried about that too. I'm not an i915 developer, but so far as I can tell:

- The path from intel_engine_cmd_parser is <= 256 KiB for "known users", so that's rather large.
- The path from perf_memcpy is either 4k, 64k, or 4M, depending on the type of object, so that seems gigantic, but I think that might be selftest code.
- The path from compress_page appears to be PAGE_SIZE, so 4k, which meshes with the limits we set agreed on few weeks ago for the crypto stuff.
- The path from guc_read_update_log_buffer appears to be 8k, 32k, 2M, or 8M, depending on the type of object, so that seems absurdly huge and doesn't appear to be selftest code either like the other case.

I have no doubt the i915 developers will jump in here waving their arms, but either way, it sure seems to me like you might have a point.

#4 lore raw

[Intel-gfx] ✗ Fi.CI.CHECKPATCH: warning for drm/i915: check to see if SIMD registers are available before using SIMD

In reply to Jason A. Donenfeld
== Series Details ==

== Series Details ==

Series: drm/i915: check to see if SIMD registers are available before using SIMD
URL : https://patchwork.freedesktop.org/series/76825/
State : warning

== Summary ==

$ dim checkpatch origin/drm-tip
962131944f1f drm/i915: check to see if SIMD registers are available before using SIMD
-:38: CHECK:LINE_SPACING: Please don't use multiple blank lines
#38: FILE: drivers/gpu/drm/i915/i915_memcpy.c:46:
+
+

total: 0 errors, 0 warnings, 1 checks, 30 lines checked

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw

[Intel-gfx] ✓ Fi.CI.BAT: success for drm/i915: check to see if SIMD registers are available before using SIMD

In reply to Jason A. Donenfeld
== Series Details ==

== Series Details ==

Series: drm/i915: check to see if SIMD registers are available before using SIMD
URL : https://patchwork.freedesktop.org/series/76825/
State : success

== Summary ==

CI Bug Log - changes from CI_DRM_8407 -> Patchwork_17544
====================================================

Summary
-------

**SUCCESS**

No regressions found.

External URL: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/index.html

Known issues
------------

Here are the changes found in Patchwork_17544 that come from known issues:

### IGT changes ###

#### Possible fixes ####

* igt@i915_selftest@live@execlists:
- fi-skl-guc: [INCOMPLETE][1] ([i915#1795]) -> [PASS][2]
[1]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/fi-skl-guc/igt@i915_selftest@[email protected]
[2]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/fi-skl-guc/igt@i915_selftest@[email protected]

[i915#1795]: https://gitlab.freedesktop.org/drm/intel/issues/1795

Participating hosts (50 -> 43)
------------------------------

Missing (7): fi-ilk-m540 fi-hsw-4200u fi-byt-squawks fi-bsw-cyan fi-ctg-p8600 fi-byt-clapper fi-bdw-samus

Build changes
-------------

* CI: CI-20190529 -> None
* Linux: CI_DRM_8407 -> Patchwork_17544

CI-20190529: 20190529
CI_DRM_8407: ecb91f743598d9d646fa046c7341058e48494e7c @ git://anongit.freedesktop.org/gfx-ci/linux
IGT_5626: f27fdfff026276ac75c69e487c929a843f66f6ca @ git://anongit.freedesktop.org/xorg/app/intel-gpu-tools
Patchwork_17544: 962131944f1f279983832e13cdf45517f2abfa06 @ git://anongit.freedesktop.org/gfx-ci/linux

== Linux commits ==

962131944f1f drm/i915: check to see if SIMD registers are available before using SIMD

== Logs ==

For more details see: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/index.html
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#6 lore raw
In reply to Jason A. Donenfeld
Quoted 6 lines · Jason A. Donenfeld

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

Err, why does i915 implements its own uncached memcpy instead of relying on core functionality to start with?

#7 lore raw
Quoted 11 lines · Christoph Hellwig
Quoted 6 lines · Jason A. Donenfeld

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

Err, why does i915 implements its own uncached memcpy instead of relying on core functionality to start with?

I was wondering the same. It sure does seem like this ought to be more generalized functionality, with a name that represents the type of transfer it's optimized for (wc or similar).

#8 lore raw
In reply to Christoph Hellwig

Quoting Christoph Hellwig (2020-05-01 19:07:31)

Quoted 10 lines · Jason A. Donenfeld
Quoted 6 lines

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

Err, why does i915 implements its own uncached memcpy instead of relying on core functionality to start with?

What is this core functionality that provides movntqda? -Chris

#9 lore raw
Quoted 4 lines · Chris Wilson
Quoted 2 lines

Err, why does i915 implements its own uncached memcpy instead of relying on core functionality to start with?

What is this core functionality that provides movntqda?

A sensible name might be memcpy_uncached or mempcy_nontemporal. But the important point is that this should be arch code with a common fallback rather than hacking it up in drivers.

#10 lore raw

From: Christoph Hellwig

Quoted 11 lines

Sent: 04 May 2020 17:03

Quoted 4 lines · Chris Wilson
Quoted 2 lines

Err, why does i915 implements its own uncached memcpy instead of relying on core functionality to start with?

What is this core functionality that provides movntqda?

A sensible name might be memcpy_uncached or mempcy_nontemporal. But the important point is that this should be arch code with a common fallback rather than hacking it up in drivers.

More the point, you are trying to do a copy where:
1) The kernel isn't expected to read the data - so can bypass the cache. and maybe:
2) The data needs flushing from the cache to actual memory. and maybe:
3) The cache lines need invalidating.

The fallbacks depend on the required behaviour.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#11 lore raw

[Intel-gfx] ✓ Fi.CI.IGT: success for drm/i915: check to see if SIMD registers are available before using SIMD

In reply to Jason A. Donenfeld
== Series Details ==

== Series Details ==

Series: drm/i915: check to see if SIMD registers are available before using SIMD
URL : https://patchwork.freedesktop.org/series/76825/
State : success

== Summary ==

CI Bug Log - changes from CI_DRM_8407_full -> Patchwork_17544_full
====================================================

Summary
-------

**SUCCESS**

No regressions found.

Known issues
------------

Here are the changes found in Patchwork_17544_full that come from known issues:

### IGT changes ###

#### Issues hit ####

* igt@i915_suspend@fence-restore-untiled:
- shard-kbl: [PASS][1] -> [DMESG-WARN][2] ([i915#180])
[1]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-kbl4/igt@[email protected]
[2]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-kbl6/igt@[email protected]

* igt@kms_cursor_legacy@pipe-d-torture-bo:
- shard-tglb: [PASS][3] -> [DMESG-WARN][4] ([i915#128])
[3]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-tglb6/igt@[email protected]
[4]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-tglb6/igt@[email protected]

* igt@kms_draw_crc@draw-method-rgb565-mmap-wc-untiled:
- shard-glk: [PASS][5] -> [FAIL][6] ([i915#52] / [i915#54])
[5]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-glk1/igt@[email protected]
[6]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-glk7/igt@[email protected]

* igt@kms_lease@lease_again:
- shard-snb: [PASS][7] -> [SKIP][8] ([fdo#109271]) +1 similar issue
[7]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-snb1/igt@kms_lease@lease_again.html
[8]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-snb5/igt@kms_lease@lease_again.html

* igt@kms_pipe_crc_basic@nonblocking-crc-pipe-b-frame-sequence:
- shard-skl: [PASS][9] -> [FAIL][10] ([i915#53])
[9]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-skl7/igt@[email protected]
[10]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-skl6/igt@[email protected]

* igt@kms_pipe_crc_basic@suspend-read-crc-pipe-c:
- shard-apl: [PASS][11] -> [DMESG-WARN][12] ([i915#180])
[11]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-apl2/igt@[email protected]
[12]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-apl4/igt@[email protected]

* igt@kms_plane_alpha_blend@pipe-b-coverage-7efc:
- shard-skl: [PASS][13] -> [FAIL][14] ([fdo#108145] / [i915#265]) +1 similar issue
[13]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-skl7/igt@[email protected]
[14]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-skl6/igt@[email protected]

* igt@kms_psr@psr2_cursor_mmap_cpu:
- shard-iclb: [PASS][15] -> [SKIP][16] ([fdo#109441]) +1 similar issue
[15]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-iclb2/igt@kms_psr@psr2_cursor_mmap_cpu.html
[16]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-iclb6/igt@kms_psr@psr2_cursor_mmap_cpu.html

#### Possible fixes ####

* igt@gem_exec_params@invalid-bsd-ring:
- shard-iclb: [SKIP][17] ([fdo#109276]) -> [PASS][18]
[17]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-iclb3/igt@[email protected]
[18]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-iclb4/igt@[email protected]

* igt@kms_cursor_crc@pipe-b-cursor-suspend:
- shard-skl: [INCOMPLETE][19] ([i915#300]) -> [PASS][20]
[19]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-skl10/igt@[email protected]
[20]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-skl9/igt@[email protected]

* igt@kms_cursor_legacy@pipe-b-torture-bo:
- shard-snb: [DMESG-WARN][21] ([i915#128]) -> [PASS][22]
[21]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-snb6/igt@[email protected]
[22]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-snb2/igt@[email protected]

* igt@kms_frontbuffer_tracking@fbc-1p-pri-indfb-multidraw:
- shard-snb: [SKIP][23] ([fdo#109271]) -> [PASS][24] +1 similar issue
[23]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-snb2/igt@[email protected]
[24]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-snb6/igt@[email protected]

* igt@kms_plane@plane-panning-bottom-right-suspend-pipe-a-planes:
- shard-apl: [DMESG-WARN][25] ([i915#180]) -> [PASS][26] +1 similar issue
[25]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-apl2/igt@[email protected]
[26]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-apl4/igt@[email protected]

* igt@kms_plane_alpha_blend@pipe-c-coverage-7efc:
- shard-skl: [FAIL][27] ([fdo#108145] / [i915#265]) -> [PASS][28]
[27]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-skl6/igt@[email protected]
[28]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-skl8/igt@[email protected]

* igt@kms_psr@psr2_cursor_render:
- shard-iclb: [SKIP][29] ([fdo#109441]) -> [PASS][30] +1 similar issue
[29]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-iclb8/igt@kms_psr@psr2_cursor_render.html
[30]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-iclb2/igt@kms_psr@psr2_cursor_render.html

* igt@kms_vblank@pipe-a-ts-continuation-suspend:
- shard-kbl: [DMESG-WARN][31] ([i915#180]) -> [PASS][32] +1 similar issue
[31]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-kbl4/igt@[email protected]
[32]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-kbl7/igt@[email protected]

#### Warnings ####

* igt@i915_pm_rpm@modeset-lpsp:
- shard-snb: [SKIP][33] ([fdo#109271]) -> [INCOMPLETE][34] ([i915#82])
[33]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-snb5/igt@[email protected]
[34]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-snb6/igt@[email protected]

* igt@kms_plane_alpha_blend@pipe-a-alpha-basic:
- shard-apl: [FAIL][35] ([fdo#108145] / [i915#265] / [i915#95]) -> [FAIL][36] ([fdo#108145] / [i915#265])
[35]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_8407/shard-apl3/igt@[email protected]
[36]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/shard-apl1/igt@[email protected]

{name}: This element is suppressed. This means it is ignored when computing
the status of the difference (SUCCESS, WARNING, or FAILURE).

[fdo#108145]: https://bugs.freedesktop.org/show_bug.cgi?id=108145
[fdo#109271]: https://bugs.freedesktop.org/show_bug.cgi?id=109271
[fdo#109276]: https://bugs.freedesktop.org/show_bug.cgi?id=109276
[fdo#109441]: https://bugs.freedesktop.org/show_bug.cgi?id=109441
[i915#128]: https://gitlab.freedesktop.org/drm/intel/issues/128
[i915#1542]: https://gitlab.freedesktop.org/drm/intel/issues/1542
[i915#180]: https://gitlab.freedesktop.org/drm/intel/issues/180
[i915#198]: https://gitlab.freedesktop.org/drm/intel/issues/198
[i915#265]: https://gitlab.freedesktop.org/drm/intel/issues/265
[i915#300]: https://gitlab.freedesktop.org/drm/intel/issues/300
[i915#52]: https://gitlab.freedesktop.org/drm/intel/issues/52
[i915#53]: https://gitlab.freedesktop.org/drm/intel/issues/53
[i915#54]: https://gitlab.freedesktop.org/drm/intel/issues/54
[i915#79]: https://gitlab.freedesktop.org/drm/intel/issues/79
[i915#82]: https://gitlab.freedesktop.org/drm/intel/issues/82
[i915#95]: https://gitlab.freedesktop.org/drm/intel/issues/95

Participating hosts (10 -> 10)
------------------------------

No changes in participating hosts

Build changes
-------------

* CI: CI-20190529 -> None
* Linux: CI_DRM_8407 -> Patchwork_17544

CI-20190529: 20190529
CI_DRM_8407: ecb91f743598d9d646fa046c7341058e48494e7c @ git://anongit.freedesktop.org/gfx-ci/linux
IGT_5626: f27fdfff026276ac75c69e487c929a843f66f6ca @ git://anongit.freedesktop.org/xorg/app/intel-gpu-tools
Patchwork_17544: 962131944f1f279983832e13cdf45517f2abfa06 @ git://anongit.freedesktop.org/gfx-ci/linux
piglit_4509: fdc5a4ca11124ab8413c7988896eec4c97336694 @ git://anongit.freedesktop.org/piglit

== Logs ==

For more details see: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_17544/index.html
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#12 lore raw
In reply to Jason A. Donenfeld

Quoting Jason A. Donenfeld (2020-04-30 23:10:16)

Quoted 8 lines

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

The same argument as on the previous submission is that we return to the caller if we can't use movntqda as their fallback path should be faster than uncached memcpy.
-Chris

#13 lore raw
Quoted 9 lines · Chris Wilson

Quoting Jason A. Donenfeld (2020-04-30 23:10:16)

Quoted 7 lines

Sometimes it's not okay to use SIMD registers, the conditions for which have changed subtly from kernel release to kernel release. Usually the pattern is to check for may_use_simd() and then fallback to using something slower in the unlikely case SIMD registers aren't available. So, this patch fixes up i915's accelerated memcpy routines to fallback to boring memcpy if may_use_simd() is false.

Quoted 5 lines
Quoted 1 line

The same argument as on the previous submission is that we return to the caller if we can't use movntqda as their fallback path should be faster than uncached memcpy.

Oh, THAT's what you meant before. Okay, will follow up.