On Sat, 14 Jun 2025 08:43:13 +0200, Christophe Leroy wrote: > > This series converts all variants of SNDRV_PCM_IOCTL_SYNC_PTR to > user_access_begin/user_access_end() in order to reduce the CPU load > measured in function snd_pcm_ioctl. > > With the current implementation, "perf top" reports a high load in > snd_pcm_iotcl(). Most calls to that function are SNDRV_PCM_IOCTL_SYNC_PTR. > > 14.20% test_perf [.] engine_main > ==> 12.86% [kernel] [k] snd_pcm_ioctl > 11.91% [kernel] [k] finish_task_switch.isra.0 > 4.15% [kernel] [k] snd_pcm_group_unlock_irq.part.0 > 4.07% libc.so.6 [.] __ioctl_time64 > 3.58% libasound.so.2.0.0 [.] __snd_pcm_mmap_begin_generic > 3.37% [kernel] [k] sys_ioctl > 2.96% libasound.so.2.0.0 [.] snd_pcm_hw_avail_update > 2.73% libasound.so.2.0.0 [.] __snd_pcm_mmap_begin > 2.58% [kernel] [k] system_call_exception > 1.93% libasound.so.2.0.0 [.] sync_ptr1 > 1.85% libasound.so.2.0.0 [.] snd_pcm_unlock > 1.84% libasound.so.2.0.0 [.] snd_pcm_mmap_begin > 1.83% libasound.so.2.0.0 [.] bad_pcm_state > 1.68% libasound.so.2.0.0 [.] snd_pcm_mmap_avail > 1.67% libasound.so.2.0.0 [.] snd_pcm_avail_update > > A tentative was done with going via intermediaire structs on stack to > replace the multiple get_user() and put_user() with copy_from_user() > and copy_to_user(). But copy_from_user() calls _copy_from_user() and > copy_to_user() calls _copy_to_user(). Both then call __copy_tofrom_user(). > In total it is 16.4% so it is worse than before. > > 14.47% test_perf [.] engine_main > 12.00% [kernel] [k] finish_task_switch.isra.0 > ==> 8.37% [kernel] [k] snd_pcm_ioctl > 5.44% libc.so.6 [.] __ioctl_time64 > 5.03% [kernel] [k] snd_pcm_group_unlock_irq.part.0 > ==> 4.86% [kernel] [k] __copy_tofrom_user > 4.62% [kernel] [k] sys_ioctl > 3.22% [kernel] [k] system_call_exception > 2.42% libasound.so.2.0.0 [.] snd_pcm_mmap_begin > 2.31% [kernel] [k] fdget > 2.23% libasound.so.2.0.0 [.] __snd_pcm_mmap_begin_generic > 2.19% [kernel] [k] syscall_exit_prepare > 1.92% libasound.so.2.0.0 [.] snd_pcm_mmap_avail > 1.86% libasound.so.2.0.0 [.] __snd_pcm_mmap_begin > 1.68% libasound.so.2.0.0 [.] snd_pcm_hw_avail_update > ==> 1.67% [kernel] [k] _copy_from_user > 1.66% libasound.so.2.0.0 [.] bad_pcm_state > ==> 1.53% [kernel] [k] _copy_to_user > 1.40% libasound.so.2.0.0 [.] sync_ptr1 > > With this series which uses unsafe_put_user() and unsafe_get_user(), > the load is significantly reduced: > > 17.46% test_perf [.] engine_main > 9.14% [kernel] [k] finish_task_switch.isra.0 > ==> 4.92% [kernel] [k] snd_pcm_ioctl > 3.99% [kernel] [k] snd_pcm_group_unlock_irq.part.0 > 3.71% libc.so.6 [.] __ioctl_time64 > 3.61% libasound.so.2.0.0 [.] __snd_pcm_mmap_begin_generic > 2.72% libasound.so.2.0.0 [.] sync_ptr1 > 2.65% [kernel] [k] system_call_exception > 2.46% [kernel] [k] sys_ioctl > 2.43% [kernel] [k] __rseq_handle_notify_resume > 2.34% [kernel] [k] do_epoll_wait > 2.30% libasound.so.2.0.0 [.] __snd_pcm_mmap_commit > 2.14% libasound.so.2.0.0 [.] __snd_pcm_avail > 2.04% libasound.so.2.0.0 [.] snd_pcm_hw_avail_update > 1.89% libasound.so.2.0.0 [.] snd_pcm_lock > 1.84% libasound.so.2.0.0 [.] snd_pcm_mmap_avail > 1.76% libasound.so.2.0.0 [.] __snd_pcm_avail_update > 1.61% libasound.so.2.0.0 [.] bad_pcm_state > 1.60% libasound.so.2.0.0 [.] __snd_pcm_mmap_begin > 1.49% libasound.so.2.0.0 [.] query_status_data > > Since v2: > - Fix macros to skip user_read_access_end() when user_read_access_begin() > failed > - Fix some tabulations for properly aligning backslashes > > Since RFC: > - Added a cover letter to summarize some of the measurements done on and > around the RFC > - Fixed relevant checkpatch feedback > - Split last patch in two > > Christophe Leroy (4): > ALSA: pcm: refactor copy from/to user in SNDRV_PCM_IOCTL_SYNC_PTR > ALSA: pcm: Convert SNDRV_PCM_IOCTL_SYNC_PTR to > user_access_begin/user_access_end() > ALSA: pcm: Replace [audio_]tstamp_[n]sec by struct __snd_timespec in > struct snd_pcm_mmap_status32 > ALSA: pcm: Convert snd_pcm_sync_ptr() to > user_access_begin/user_access_end()
Applied now all patches. Thanks! Takashi
