From: "Christian König" <christian.koenig@amd.com>
To: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>,
dri-devel@lists.freedesktop.org
Cc: kernel-dev@igalia.com, amd-gfx@lists.freedesktop.org,
intel-xe@lists.freedesktop.org,
"Michel Dänzer" <michel.daenzer@mailbox.org>,
"Maíra Canal" <mcanal@igalia.com>
Subject: Re: [PATCH v5 6/6] drm/syncobj: Add a fast path to drm_syncobj_array_find
Date: Thu, 12 Jun 2025 09:21:43 +0200 [thread overview]
Message-ID: <c17f25ef-390a-43e3-a467-45e64048ed6c@amd.com> (raw)
In-Reply-To: <10e83252-e565-4cb4-9bc2-ae238528df92@igalia.com>
On 6/11/25 17:29, Tvrtko Ursulin wrote:
>
> On 11/06/2025 15:21, Christian König wrote:
>> On 6/11/25 16:00, Tvrtko Ursulin wrote:
>>> Running the Cyberpunk 2077 benchmark we can observe that the lookup helper
>>> is relatively hot, but the 97% of the calls are for a single object. (~3%
>>> for two points, and never more than three points. While a more trivial
>>> workload like vkmark under Plasma is even more skewed to single point
>>> lookups.)
>>>
>>> Therefore lets add a fast path to bypass the kmalloc_array/kfree and use a
>>> pre-allocated stack array for those cases.
>>
>> Have you considered using memdup_user()? That's using a separate bucket IIRC and might give similar performance.
>
> I haven't but I can try it. I would be surprised if it made a (positive) difference though.
Yeah, it's mostly for extra security I think.
>
> And I realised I need to repeat the benchmarks anyway, since in v4 I had to stop doing access_ok+__get_user, after kernel test robot let me know 64-bit get_user is a not a thing on all platforms. I thought the gains are from avoiding allocations but, as you say, now I need to see if copy_from_user doesn't nullify them..
>
>> If that is still not sufficient I'm really wondering if we shouldn't have a macro for doing this. It's a really common use case as far as I can see.
>
> Hmm macro for what exactly?
Like a macro which uses an array on the stack for small (<4) number of values and k(v)malloc() for large ones.
IIRC there is also a relatively new functionality which allows releasing the memory automatically when we leave the function.
Regards,
Christian.
>
> Regards,
>
> Tvrtko
>
>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
>>> Reviewed-by: Maíra Canal <mcanal@igalia.com>
>>> ---
>>> v2:
>>> * Added comments describing how the fast path arrays were sized.
>>> * Make container freeing criteria clearer by using a boolean.
>>> ---
>>> drivers/gpu/drm/drm_syncobj.c | 56 +++++++++++++++++++++++++++--------
>>> 1 file changed, 44 insertions(+), 12 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
>>> index be5905dca87f..65c301852f0d 100644
>>> --- a/drivers/gpu/drm/drm_syncobj.c
>>> +++ b/drivers/gpu/drm/drm_syncobj.c
>>> @@ -1259,6 +1259,8 @@ EXPORT_SYMBOL(drm_timeout_abs_to_jiffies);
>>> static int drm_syncobj_array_find(struct drm_file *file_private,
>>> u32 __user *handles,
>>> uint32_t count,
>>> + struct drm_syncobj **stack_syncobjs,
>>> + u32 stack_count,
>>> struct drm_syncobj ***syncobjs_out)
>>> {
>>> struct drm_syncobj **syncobjs;
>>> @@ -1268,9 +1270,13 @@ static int drm_syncobj_array_find(struct drm_file *file_private,
>>> if (!access_ok(handles, count * sizeof(*handles)))
>>> return -EFAULT;
>>> - syncobjs = kmalloc_array(count, sizeof(*syncobjs), GFP_KERNEL);
>>> - if (!syncobjs)
>>> - return -ENOMEM;
>>> + if (count > stack_count) {
>>> + syncobjs = kmalloc_array(count, sizeof(*syncobjs), GFP_KERNEL);
>>> + if (!syncobjs)
>>> + return -ENOMEM;
>>> + } else {
>>> + syncobjs = stack_syncobjs;
>>> + }
>>> for (i = 0; i < count; i++) {
>>> u32 handle;
>>> @@ -1292,25 +1298,31 @@ static int drm_syncobj_array_find(struct drm_file *file_private,
>>> err_put_syncobjs:
>>> while (i-- > 0)
>>> drm_syncobj_put(syncobjs[i]);
>>> - kfree(syncobjs);
>>> +
>>> + if (syncobjs != stack_syncobjs)
>>> + kfree(syncobjs);
>>> return ret;
>>> }
>>> static void drm_syncobj_array_free(struct drm_syncobj **syncobjs,
>>> - uint32_t count)
>>> + uint32_t count,
>>> + bool free_container)
>>> {
>>> uint32_t i;
>>> for (i = 0; i < count; i++)
>>> drm_syncobj_put(syncobjs[i]);
>>> - kfree(syncobjs);
>>> +
>>> + if (free_container)
>>> + kfree(syncobjs);
>>> }
>>> int
>>> drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
>>> struct drm_file *file_private)
>>> {
>>> + struct drm_syncobj *stack_syncobjs[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
>>> struct drm_syncobj_wait *args = data;
>>> ktime_t deadline, *pdeadline = NULL;
>>> u32 count = args->count_handles;
>>> @@ -1336,6 +1348,8 @@ drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
>>> ret = drm_syncobj_array_find(file_private,
>>> u64_to_user_ptr(args->handles),
>>> count,
>>> + stack_syncobjs,
>>> + ARRAY_SIZE(stack_syncobjs),
>>> &syncobjs);
>>> if (ret < 0)
>>> return ret;
>>> @@ -1354,7 +1368,7 @@ drm_syncobj_wait_ioctl(struct drm_device *dev, void *data,
>>> &first,
>>> pdeadline);
>>> - drm_syncobj_array_free(syncobjs, count);
>>> + drm_syncobj_array_free(syncobjs, count, syncobjs != stack_syncobjs);
>>> if (timeout < 0)
>>> return timeout;
>>> @@ -1368,6 +1382,7 @@ int
>>> drm_syncobj_timeline_wait_ioctl(struct drm_device *dev, void *data,
>>> struct drm_file *file_private)
>>> {
>>> + struct drm_syncobj *stack_syncobjs[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
>>> struct drm_syncobj_timeline_wait *args = data;
>>> ktime_t deadline, *pdeadline = NULL;
>>> u32 count = args->count_handles;
>>> @@ -1394,6 +1409,8 @@ drm_syncobj_timeline_wait_ioctl(struct drm_device *dev, void *data,
>>> ret = drm_syncobj_array_find(file_private,
>>> u64_to_user_ptr(args->handles),
>>> count,
>>> + stack_syncobjs,
>>> + ARRAY_SIZE(stack_syncobjs),
>>> &syncobjs);
>>> if (ret < 0)
>>> return ret;
>>> @@ -1412,7 +1429,7 @@ drm_syncobj_timeline_wait_ioctl(struct drm_device *dev, void *data,
>>> &first,
>>> pdeadline);
>>> - drm_syncobj_array_free(syncobjs, count);
>>> + drm_syncobj_array_free(syncobjs, count, syncobjs != stack_syncobjs);
>>> if (timeout < 0)
>>> return timeout;
>>> @@ -1529,6 +1546,7 @@ int
>>> drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
>>> struct drm_file *file_private)
>>> {
>>> + struct drm_syncobj *stack_syncobjs[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
>>> struct drm_syncobj_array *args = data;
>>> struct drm_syncobj **syncobjs;
>>> uint32_t i;
>>> @@ -1546,6 +1564,8 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
>>> ret = drm_syncobj_array_find(file_private,
>>> u64_to_user_ptr(args->handles),
>>> args->count_handles,
>>> + stack_syncobjs,
>>> + ARRAY_SIZE(stack_syncobjs),
>>> &syncobjs);
>>> if (ret < 0)
>>> return ret;
>>> @@ -1553,7 +1573,8 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
>>> for (i = 0; i < args->count_handles; i++)
>>> drm_syncobj_replace_fence(syncobjs[i], NULL);
>>> - drm_syncobj_array_free(syncobjs, args->count_handles);
>>> + drm_syncobj_array_free(syncobjs, args->count_handles,
>>> + syncobjs != stack_syncobjs);
>>> return 0;
>>> }
>>> @@ -1562,6 +1583,7 @@ int
>>> drm_syncobj_signal_ioctl(struct drm_device *dev, void *data,
>>> struct drm_file *file_private)
>>> {
>>> + struct drm_syncobj *stack_syncobjs[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
>>> struct drm_syncobj_array *args = data;
>>> struct drm_syncobj **syncobjs;
>>> uint32_t i;
>>> @@ -1579,6 +1601,8 @@ drm_syncobj_signal_ioctl(struct drm_device *dev, void *data,
>>> ret = drm_syncobj_array_find(file_private,
>>> u64_to_user_ptr(args->handles),
>>> args->count_handles,
>>> + stack_syncobjs,
>>> + ARRAY_SIZE(stack_syncobjs),
>>> &syncobjs);
>>> if (ret < 0)
>>> return ret;
>>> @@ -1589,7 +1613,8 @@ drm_syncobj_signal_ioctl(struct drm_device *dev, void *data,
>>> break;
>>> }
>>> - drm_syncobj_array_free(syncobjs, args->count_handles);
>>> + drm_syncobj_array_free(syncobjs, args->count_handles,
>>> + syncobjs != stack_syncobjs);
>>> return ret;
>>> }
>>> @@ -1598,6 +1623,7 @@ int
>>> drm_syncobj_timeline_signal_ioctl(struct drm_device *dev, void *data,
>>> struct drm_file *file_private)
>>> {
>>> + struct drm_syncobj *stack_syncobjs[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
>>> struct drm_syncobj_timeline_array *args = data;
>>> uint64_t __user *points = u64_to_user_ptr(args->points);
>>> uint32_t i, j, count = args->count_handles;
>>> @@ -1617,6 +1643,8 @@ drm_syncobj_timeline_signal_ioctl(struct drm_device *dev, void *data,
>>> ret = drm_syncobj_array_find(file_private,
>>> u64_to_user_ptr(args->handles),
>>> count,
>>> + stack_syncobjs,
>>> + ARRAY_SIZE(stack_syncobjs),
>>> &syncobjs);
>>> if (ret < 0)
>>> return ret;
>>> @@ -1653,7 +1681,7 @@ drm_syncobj_timeline_signal_ioctl(struct drm_device *dev, void *data,
>>> err_chains:
>>> kfree(chains);
>>> out:
>>> - drm_syncobj_array_free(syncobjs, count);
>>> + drm_syncobj_array_free(syncobjs, count, syncobjs != stack_syncobjs);
>>> return ret;
>>> }
>>> @@ -1661,6 +1689,7 @@ drm_syncobj_timeline_signal_ioctl(struct drm_device *dev, void *data,
>>> int drm_syncobj_query_ioctl(struct drm_device *dev, void *data,
>>> struct drm_file *file_private)
>>> {
>>> + struct drm_syncobj *stack_syncobjs[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
>>> struct drm_syncobj_timeline_array *args = data;
>>> struct drm_syncobj **syncobjs;
>>> uint64_t __user *points = u64_to_user_ptr(args->points);
>>> @@ -1679,6 +1708,8 @@ int drm_syncobj_query_ioctl(struct drm_device *dev, void *data,
>>> ret = drm_syncobj_array_find(file_private,
>>> u64_to_user_ptr(args->handles),
>>> args->count_handles,
>>> + stack_syncobjs,
>>> + ARRAY_SIZE(stack_syncobjs),
>>> &syncobjs);
>>> if (ret < 0)
>>> return ret;
>>> @@ -1722,7 +1753,8 @@ int drm_syncobj_query_ioctl(struct drm_device *dev, void *data,
>>> if (ret)
>>> break;
>>> }
>>> - drm_syncobj_array_free(syncobjs, args->count_handles);
>>> + drm_syncobj_array_free(syncobjs, args->count_handles,
>>> + syncobjs != stack_syncobjs);
>>> return ret;
>>> }
>>
>
next prev parent reply other threads:[~2025-06-12 7:21 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-06-11 14:00 [PATCH v5 0/6] A few drm_syncobj optimisations Tvrtko Ursulin
2025-06-11 14:00 ` [PATCH v5 1/6] drm/syncobj: Remove unhelpful helper Tvrtko Ursulin
2025-06-11 14:00 ` [PATCH v5 2/6] drm/syncobj: Do not allocate an array to store zeros when waiting Tvrtko Ursulin
2025-06-11 14:00 ` [PATCH v5 3/6] drm/syncobj: Avoid one temporary allocation in drm_syncobj_array_find Tvrtko Ursulin
2025-06-11 14:00 ` [PATCH v5 4/6] drm/syncobj: Avoid temporary allocation in drm_syncobj_timeline_signal_ioctl Tvrtko Ursulin
2025-06-11 14:15 ` Christian König
2025-06-11 14:00 ` [PATCH v5 5/6] drm/syncobj: Add a fast path to drm_syncobj_array_wait_timeout Tvrtko Ursulin
2025-06-11 14:00 ` [PATCH v5 6/6] drm/syncobj: Add a fast path to drm_syncobj_array_find Tvrtko Ursulin
2025-06-11 14:21 ` Christian König
2025-06-11 15:29 ` Tvrtko Ursulin
2025-06-12 7:21 ` Christian König [this message]
2025-06-12 10:58 ` Tvrtko Ursulin
2025-06-12 12:02 ` Christian König
2025-06-11 14:37 ` ✗ CI.checkpatch: warning for A few drm_syncobj optimisations (rev3) Patchwork
2025-06-11 14:38 ` ✓ CI.KUnit: success " Patchwork
2025-06-11 14:49 ` ✓ CI.Build: " Patchwork
2025-06-11 14:52 ` ✓ CI.Hooks: " Patchwork
2025-06-11 14:53 ` ✓ CI.checksparse: " Patchwork
2025-06-11 15:33 ` ✓ Xe.CI.BAT: " Patchwork
2025-06-11 18:10 ` ✗ Xe.CI.Full: failure " Patchwork
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=c17f25ef-390a-43e3-a467-45e64048ed6c@amd.com \
--to=christian.koenig@amd.com \
--cc=amd-gfx@lists.freedesktop.org \
--cc=dri-devel@lists.freedesktop.org \
--cc=intel-xe@lists.freedesktop.org \
--cc=kernel-dev@igalia.com \
--cc=mcanal@igalia.com \
--cc=michel.daenzer@mailbox.org \
--cc=tvrtko.ursulin@igalia.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox