Message ID | 20190613050403.GA21502@ravnborg.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | Drop use of DRM_WAIT_ON() [Was: drm/drm_vblank: Change EINVAL by the correct errno] | expand |
Hi Rodrigo et al. On Thu, Jun 13, 2019 at 07:04:03AM +0200, Sam Ravnborg wrote: > Hi Rodrigo. > > On Wed, Jun 12, 2019 at 11:10:54PM -0300, Rodrigo Siqueira wrote: > > For historical reason, the function drm_wait_vblank_ioctl always return > > -EINVAL if something gets wrong. This scenario limits the flexibility > > for the userspace make detailed verification of the problem and take > > some action. In particular, the validation of “if (!dev->irq_enabled)” > > in the drm_wait_vblank_ioctl is responsible for checking if the driver > > support vblank or not. If the driver does not support VBlank, the > > function drm_wait_vblank_ioctl returns EINVAL which does not represent > > the real issue; this patch changes this behavior by return EOPNOTSUPP. > > Additionally, some operations are unsupported by this function, and > > returns EINVAL; this patch also changes the return value to EOPNOTSUPP > > in this case. Lastly, the function drm_wait_vblank_ioctl is invoked by > > libdrm, which is used by many compositors; because of this, it is > > important to check if this change breaks any compositor. In this sense, > > the following projects were examined: > > > > * Drm-hwcomposer > > * Kwin > > * Sway > > * Wlroots > > * Wayland-core > > * Weston > > * Xorg (67 different drivers) > > > > For each repository the verification happened in three steps: > > > > * Update the main branch > > * Look for any occurrence "drmWaitVBlank" with the command: > > git grep -n "drmWaitVBlank" > > * Look in the git history of the project with the command: > > git log -SdrmWaitVBlank > > > > Finally, none of the above projects validate the use of EINVAL which > > make safe, at least for these projects, to change the return values. > > > > Change since V2: > > Daniel Vetter and Chris Wilson > > - Replace ENOTTY by EOPNOTSUPP > > - Return EINVAL if the parameters are wrong > > > > Signed-off-by: Rodrigo Siqueira <rodrigosiqueiramelo@gmail.com> > > --- > > Update: > > Now IGT has a way to validate if a driver has vblank support or not. > > See: https://gitlab.freedesktop.org/drm/igt-gpu-tools/commit/2d244aed69165753f3adbbd6468db073dc1acf9A > > > > drivers/gpu/drm/drm_vblank.c | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c > > index 0d704bddb1a6..d76a783a7d4b 100644 > > --- a/drivers/gpu/drm/drm_vblank.c > > +++ b/drivers/gpu/drm/drm_vblank.c > > @@ -1578,10 +1578,10 @@ int drm_wait_vblank_ioctl(struct drm_device *dev, void *data, > > unsigned int flags, pipe, high_pipe; > > > > if (!dev->irq_enabled) > > - return -EINVAL; > > + return -EOPNOTSUPP; > > > > if (vblwait->request.type & _DRM_VBLANK_SIGNAL) > > - return -EINVAL; > > + return -EOPNOTSUPP; > > > > if (vblwait->request.type & > > ~(_DRM_VBLANK_TYPES_MASK | _DRM_VBLANK_FLAGS_MASK | > > When touching this function, could I ask you to take a look at > eliminating the use of DRM_WAIT_ON()? > It comes from the deprecated drm_os_linux.h header, and it is only of > the few remaining users of DRM_WAIT_ON(). > > Below you can find my untested first try - where I did an attempt not to > change behaviour. intel-gfx did not like the patch - so no need to spend time looking at the patch until I have that fixed. Sam
Hi Sam, I tested it by using VKMS and kms_flip, and tests related to “vblank” fails (e.g., basic-flip-vs-wf_vblank, blocking-absolute-wf_vblank, flip-vs-absolute-wf_vblank, etc). I tried to dig into this issue, and you can see my comments inline: On Thu, Jun 13, 2019 at 2:04 AM Sam Ravnborg <sam@ravnborg.org> wrote: > > Hi Rodrigo. > > On Wed, Jun 12, 2019 at 11:10:54PM -0300, Rodrigo Siqueira wrote: > > For historical reason, the function drm_wait_vblank_ioctl always return > > -EINVAL if something gets wrong. This scenario limits the flexibility > > for the userspace make detailed verification of the problem and take > > some action. In particular, the validation of “if (!dev->irq_enabled)” > > in the drm_wait_vblank_ioctl is responsible for checking if the driver > > support vblank or not. If the driver does not support VBlank, the > > function drm_wait_vblank_ioctl returns EINVAL which does not represent > > the real issue; this patch changes this behavior by return EOPNOTSUPP. > > Additionally, some operations are unsupported by this function, and > > returns EINVAL; this patch also changes the return value to EOPNOTSUPP > > in this case. Lastly, the function drm_wait_vblank_ioctl is invoked by > > libdrm, which is used by many compositors; because of this, it is > > important to check if this change breaks any compositor. In this sense, > > the following projects were examined: > > > > * Drm-hwcomposer > > * Kwin > > * Sway > > * Wlroots > > * Wayland-core > > * Weston > > * Xorg (67 different drivers) > > > > For each repository the verification happened in three steps: > > > > * Update the main branch > > * Look for any occurrence "drmWaitVBlank" with the command: > > git grep -n "drmWaitVBlank" > > * Look in the git history of the project with the command: > > git log -SdrmWaitVBlank > > > > Finally, none of the above projects validate the use of EINVAL which > > make safe, at least for these projects, to change the return values. > > > > Change since V2: > > Daniel Vetter and Chris Wilson > > - Replace ENOTTY by EOPNOTSUPP > > - Return EINVAL if the parameters are wrong > > > > Signed-off-by: Rodrigo Siqueira <rodrigosiqueiramelo@gmail.com> > > --- > > Update: > > Now IGT has a way to validate if a driver has vblank support or not. > > See: https://gitlab.freedesktop.org/drm/igt-gpu-tools/commit/2d244aed69165753f3adbbd6468db073dc1acf9A > > > > drivers/gpu/drm/drm_vblank.c | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c > > index 0d704bddb1a6..d76a783a7d4b 100644 > > --- a/drivers/gpu/drm/drm_vblank.c > > +++ b/drivers/gpu/drm/drm_vblank.c > > @@ -1578,10 +1578,10 @@ int drm_wait_vblank_ioctl(struct drm_device *dev, void *data, > > unsigned int flags, pipe, high_pipe; > > > > if (!dev->irq_enabled) > > - return -EINVAL; > > + return -EOPNOTSUPP; > > > > if (vblwait->request.type & _DRM_VBLANK_SIGNAL) > > - return -EINVAL; > > + return -EOPNOTSUPP; > > > > if (vblwait->request.type & > > ~(_DRM_VBLANK_TYPES_MASK | _DRM_VBLANK_FLAGS_MASK | > > When touching this function, could I ask you to take a look at > eliminating the use of DRM_WAIT_ON()? > It comes from the deprecated drm_os_linux.h header, and it is only of > the few remaining users of DRM_WAIT_ON(). > > Below you can find my untested first try - where I did an attempt not to > change behaviour. > > Sam > > commit 17b119b02467356198b57bca9949b146082bcaa1 > Author: Sam Ravnborg <sam@ravnborg.org> > Date: Thu May 30 09:38:47 2019 +0200 > > drm/vblank: drop use of DRM_WAIT_ON() > > DRM_WAIT_ON() is from the deprecated drm_os_linux header and > the modern replacement is the wait_event_*. > > The return values differ, so a conversion is needed to > keep the original interface towards userspace. > Introduced a switch/case to make code obvious and to allow > different debug prints depending on the result. > > The timeout value of 3 * HZ was translated to 30 msec > > Signed-off-by: Sam Ravnborg <sam@ravnborg.org> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > Cc: Maxime Ripard <maxime.ripard@bootlin.com> > Cc: Sean Paul <sean@poorly.run> > Cc: David Airlie <airlied@linux.ie> > Cc: Daniel Vetter <daniel@ffwll.ch> > > diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c > index 0d704bddb1a6..51fc6b106333 100644 > --- a/drivers/gpu/drm/drm_vblank.c > +++ b/drivers/gpu/drm/drm_vblank.c > @@ -31,7 +31,6 @@ > #include <drm/drm_drv.h> > #include <drm/drm_framebuffer.h> > #include <drm/drm_print.h> > -#include <drm/drm_os_linux.h> > #include <drm/drm_vblank.h> > > #include "drm_internal.h" > @@ -1668,18 +1667,27 @@ int drm_wait_vblank_ioctl(struct drm_device *dev, void *data, > if (req_seq != seq) { > DRM_DEBUG("waiting on vblank count %llu, crtc %u\n", > req_seq, pipe); > - DRM_WAIT_ON(ret, vblank->queue, 3 * HZ, > - vblank_passed(drm_vblank_count(dev, pipe), > - req_seq) || > - !READ_ONCE(vblank->enabled)); > + ret = wait_event_interruptible_timeout(vblank->queue, > + vblank_passed(drm_vblank_count(dev, pipe), req_seq) || > + !READ_ONCE(vblank->enabled), > + msecs_to_jiffies(30)); Maybe I’m wrong, but msecs_to_jiffies(30) is much smaller than 3 * HZ. Right? > } > > - if (ret != -EINTR) { > + switch (ret) { > + case 1: > + ret = 0; > drm_wait_vblank_reply(dev, pipe, &vblwait->reply); > - > DRM_DEBUG("crtc %d returning %u to client\n", > pipe, vblwait->reply.sequence); > - } else { > + break; > + case 0: > + ret = -EBUSY; After applying your patch, I noticed that drm_wait_vblank_ioctl() consistently returns EBUSY, which is the cause of the errors in the userspace. After that, I decided to take a look at DRM_WAIT_ON; See below the code and my comments: #define DRM_WAIT_ON( ret, queue, timeout, condition ) \ do { \ DECLARE_WAITQUEUE(entry, current); \ unsigned long end = jiffies + (timeout); \ add_wait_queue(&(queue), &entry); \ \ for (;;) { \ __set_current_state(TASK_INTERRUPTIBLE); \ if (condition) \ break; \ if (time_after_eq(jiffies, end)) { \ ret = -EBUSY; \ break; \ } \ I think that your code does not handle this condition for EBUSY in the same way of DRM_WAIT_ON(), or did I miss something? Best regards > + drm_wait_vblank_reply(dev, pipe, &vblwait->reply); > + DRM_DEBUG("timeout waiting for vblank. crtc %d returning %u to client\n", > + pipe, vblwait->reply.sequence); > + break; > + default: > + ret = -EINTR; > DRM_DEBUG("crtc %d vblank wait interrupted by signal\n", pipe); > } >
On Thu, Jun 13, 2019 at 07:04:03AM +0200, Sam Ravnborg wrote: > Hi Rodrigo. > > On Wed, Jun 12, 2019 at 11:10:54PM -0300, Rodrigo Siqueira wrote: > > For historical reason, the function drm_wait_vblank_ioctl always return > > -EINVAL if something gets wrong. This scenario limits the flexibility > > for the userspace make detailed verification of the problem and take > > some action. In particular, the validation of “if (!dev->irq_enabled)” > > in the drm_wait_vblank_ioctl is responsible for checking if the driver > > support vblank or not. If the driver does not support VBlank, the > > function drm_wait_vblank_ioctl returns EINVAL which does not represent > > the real issue; this patch changes this behavior by return EOPNOTSUPP. > > Additionally, some operations are unsupported by this function, and > > returns EINVAL; this patch also changes the return value to EOPNOTSUPP > > in this case. Lastly, the function drm_wait_vblank_ioctl is invoked by > > libdrm, which is used by many compositors; because of this, it is > > important to check if this change breaks any compositor. In this sense, > > the following projects were examined: > > > > * Drm-hwcomposer > > * Kwin > > * Sway > > * Wlroots > > * Wayland-core > > * Weston > > * Xorg (67 different drivers) > > > > For each repository the verification happened in three steps: > > > > * Update the main branch > > * Look for any occurrence "drmWaitVBlank" with the command: > > git grep -n "drmWaitVBlank" > > * Look in the git history of the project with the command: > > git log -SdrmWaitVBlank > > > > Finally, none of the above projects validate the use of EINVAL which > > make safe, at least for these projects, to change the return values. > > > > Change since V2: > > Daniel Vetter and Chris Wilson > > - Replace ENOTTY by EOPNOTSUPP > > - Return EINVAL if the parameters are wrong > > > > Signed-off-by: Rodrigo Siqueira <rodrigosiqueiramelo@gmail.com> > > --- > > Update: > > Now IGT has a way to validate if a driver has vblank support or not. > > See: https://gitlab.freedesktop.org/drm/igt-gpu-tools/commit/2d244aed69165753f3adbbd6468db073dc1acf9A > > > > drivers/gpu/drm/drm_vblank.c | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c > > index 0d704bddb1a6..d76a783a7d4b 100644 > > --- a/drivers/gpu/drm/drm_vblank.c > > +++ b/drivers/gpu/drm/drm_vblank.c > > @@ -1578,10 +1578,10 @@ int drm_wait_vblank_ioctl(struct drm_device *dev, void *data, > > unsigned int flags, pipe, high_pipe; > > > > if (!dev->irq_enabled) > > - return -EINVAL; > > + return -EOPNOTSUPP; > > > > if (vblwait->request.type & _DRM_VBLANK_SIGNAL) > > - return -EINVAL; > > + return -EOPNOTSUPP; > > > > if (vblwait->request.type & > > ~(_DRM_VBLANK_TYPES_MASK | _DRM_VBLANK_FLAGS_MASK | > > When touching this function, could I ask you to take a look at > eliminating the use of DRM_WAIT_ON()? > It comes from the deprecated drm_os_linux.h header, and it is only of > the few remaining users of DRM_WAIT_ON(). IIRC all previous attempts at removing that ended up with regressions. I think there are some dragons lurking inside that macro.
diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c index 0d704bddb1a6..51fc6b106333 100644 --- a/drivers/gpu/drm/drm_vblank.c +++ b/drivers/gpu/drm/drm_vblank.c @@ -31,7 +31,6 @@ #include <drm/drm_drv.h> #include <drm/drm_framebuffer.h> #include <drm/drm_print.h> -#include <drm/drm_os_linux.h> #include <drm/drm_vblank.h> #include "drm_internal.h" @@ -1668,18 +1667,27 @@ int drm_wait_vblank_ioctl(struct drm_device *dev, void *data, if (req_seq != seq) { DRM_DEBUG("waiting on vblank count %llu, crtc %u\n", req_seq, pipe); - DRM_WAIT_ON(ret, vblank->queue, 3 * HZ, - vblank_passed(drm_vblank_count(dev, pipe), - req_seq) || - !READ_ONCE(vblank->enabled)); + ret = wait_event_interruptible_timeout(vblank->queue, + vblank_passed(drm_vblank_count(dev, pipe), req_seq) || + !READ_ONCE(vblank->enabled), + msecs_to_jiffies(30)); } - if (ret != -EINTR) { + switch (ret) { + case 1: + ret = 0; drm_wait_vblank_reply(dev, pipe, &vblwait->reply); - DRM_DEBUG("crtc %d returning %u to client\n", pipe, vblwait->reply.sequence); - } else { + break; + case 0: + ret = -EBUSY; + drm_wait_vblank_reply(dev, pipe, &vblwait->reply); + DRM_DEBUG("timeout waiting for vblank. crtc %d returning %u to client\n", + pipe, vblwait->reply.sequence); + break; + default: + ret = -EINTR; DRM_DEBUG("crtc %d vblank wait interrupted by signal\n", pipe); }