Re: [RFC 3/3] drm: Update file owner during use
On 11/01/2023 22:19, Daniel Vetter wrote: On Tue, Jan 10, 2023 at 01:14:51PM +, Tvrtko Ursulin wrote: On 06/01/2023 18:00, Daniel Vetter wrote: On Fri, Jan 06, 2023 at 03:53:13PM +0100, Christian König wrote: Am 06.01.23 um 11:53 schrieb Daniel Vetter: On Fri, Jan 06, 2023 at 11:32:17AM +0100, Christian König wrote: Am 05.01.23 um 13:32 schrieb Daniel Vetter: [SNIP] For the case of an master fd I actually don't see the reason why we should limit that? And fd can become master if it either was master before or has CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid here? This is just info/debug printing, I don't see the connection to drm_auth/master stuff? Aside from the patch mixes up the master opener and the current user due to fd passing or something like that. That's exactly why my comment meant as well. The connect is that the drm_auth/master code currently the pid/tgid as indicator if the "owner" of the fd has changed and so if an access should be allowed or not. I find that approach a bit questionable. Note that we cannot do that (I think at least, after pondering this some more) because it would break the logind master fd passing scheme - there the receiving compositor is explicitly _not_ allowed to acquire master rights on its own. So the master priviledges must not move with the fd or things can go wrong. That could be the rational behind that, but why doesn't logind then just pass on a normal render node to the compositor? Because the compositor wants the kms node. We have 3 access levels in drm - render stuff - modeset stuff (needs a card* node and master rights for changing things) - set/drop master (needs root) logind wants to give the compositor modeset access, but not master drop/set access (because vt switching is controlled by logind). The pid check in drm_auth is for the use-case where you start your compositor on a root vt (or setuid-root), and then want to make sure that after cred dropping, set/drop master keeps working. Because in that case the vt switch dance is done by the compositor. Maybe we should document this stuff a bit better :-) Maybe add a friendly warning? E.g. like "Don't touch it, it works!" :) I think Tvrtko just volunteered for that :-) Maybe addition in the drm-uapi.rst section would be good that fills out the gaps we have. I can attempt to copy, paste and tidy what you wrote here, albeit with less than full degree of authority. Assuming into the existing comment above drm_master_check_perm? I'd put it into the DOC: section in drm_auth.c so it shows up in the drm-uapi.rst docs. Or do a new one if you want to split this out and then include it in the drm-uapi.rst. But in terms of where my series is going next I would need some clarification in the other sub-thread. Maybe I'm lost on what the leftover confusion is? This one seemed to be it? The question of whether you are now okay with my approach to track file_priv->pid if !was_master, or your counter-proposal to have file_priv->pid and file_priv->"render_user_pid" is still relevant. If the latter then what semantics have been settled at, one-shot transition or not? I had an issue with one-shot and even didn't fully understand what you did not like about my proposal. Regards, Tvrtko
Re: [RFC 3/3] drm: Update file owner during use
On Tue, Jan 10, 2023 at 01:14:51PM +, Tvrtko Ursulin wrote: > > On 06/01/2023 18:00, Daniel Vetter wrote: > > On Fri, Jan 06, 2023 at 03:53:13PM +0100, Christian König wrote: > > > Am 06.01.23 um 11:53 schrieb Daniel Vetter: > > > > On Fri, Jan 06, 2023 at 11:32:17AM +0100, Christian König wrote: > > > > > Am 05.01.23 um 13:32 schrieb Daniel Vetter: > > > > > > [SNIP] > > > > > > > For the case of an master fd I actually don't see the reason why > > > > > > > we should > > > > > > > limit that? And fd can become master if it either was master > > > > > > > before or has > > > > > > > CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid > > > > > > > here? > > > > > > This is just info/debug printing, I don't see the connection to > > > > > > drm_auth/master stuff? Aside from the patch mixes up the master > > > > > > opener and > > > > > > the current user due to fd passing or something like that. > > > > > That's exactly why my comment meant as well. > > > > > > > > > > The connect is that the drm_auth/master code currently the pid/tgid as > > > > > indicator if the "owner" of the fd has changed and so if an access > > > > > should be > > > > > allowed or not. I find that approach a bit questionable. > > > > > > > > > > > Note that we cannot do that (I think at least, after pondering this > > > > > > some > > > > > > more) because it would break the logind master fd passing scheme - > > > > > > there > > > > > > the receiving compositor is explicitly _not_ allowed to acquire > > > > > > master > > > > > > rights on its own. So the master priviledges must not move with the > > > > > > fd or > > > > > > things can go wrong. > > > > > That could be the rational behind that, but why doesn't logind then > > > > > just > > > > > pass on a normal render node to the compositor? > > > > Because the compositor wants the kms node. We have 3 access levels in > > > > drm > > > > > > > > - render stuff > > > > - modeset stuff (needs a card* node and master rights for changing > > > > things) > > > > - set/drop master (needs root) > > > > > > > > logind wants to give the compositor modeset access, but not master > > > > drop/set access (because vt switching is controlled by logind). > > > > > > > > The pid check in drm_auth is for the use-case where you start your > > > > compositor on a root vt (or setuid-root), and then want to make sure > > > > that after cred dropping, set/drop master keeps working. Because in that > > > > case the vt switch dance is done by the compositor. > > > > > > > > Maybe we should document this stuff a bit better :-) > > > > > > Maybe add a friendly warning? E.g. like "Don't touch it, it works!" :) > > > > I think Tvrtko just volunteered for that :-) Maybe addition in the > > drm-uapi.rst section would be good that fills out the gaps we have. > > I can attempt to copy, paste and tidy what you wrote here, albeit with less > than full degree of authority. Assuming into the existing comment above > drm_master_check_perm? I'd put it into the DOC: section in drm_auth.c so it shows up in the drm-uapi.rst docs. Or do a new one if you want to split this out and then include it in the drm-uapi.rst. > But in terms of where my series is going next I would need some > clarification in the other sub-thread. Maybe I'm lost on what the leftover confusion is? This one seemed to be it? -Daniel > > Regards, > > Tvrtko > > > > So basically this is a valid use case where logind set/get the master > > > status > > > of a fd while the compositor uses the master functionality? > > > > Yup, and the compositor is _not_ allowed to call these. Despite that it's > > the exact sime struct file - it has to be the same struct file in both > > loging and compositor, otherwise logind cannot orchestratet the vt switch > > dance for the compositors. Which unlike non-logind vt switching has the > > nice property that if a compositor dies/goes rogue, logind can still force > > the switch. With vt-only switching you need the sysrq to reset the console > > to text and kill the foreground process for the same effect. > > -Daniel > > > > > > > > Christian. > > > > > > > -Daniel > > > > > > > > > Christian. > > > > > > > > > > > -Daniel > > > > > > > > > > > > > > > > > > > > > > > > > Regards, > > > > > > > Christian. > > > > > > > > > > > > > > > Regards, > > > > > > > > > > > > > > > > Tvrtko > > > > > > > > > > > > > > > > > -Daniel > > > > > > > > > > > > > > > > > > > > > > > > > > > > return 0; > > > > > > > > > > if (!capable(CAP_SYS_ADMIN)) > > > > > > > > > > diff --git a/drivers/gpu/drm/drm_debugfs.c > > > > > > > > > > b/drivers/gpu/drm/drm_debugfs.c > > > > > > > > > > index 42f657772025..9d4e3146a2b8 100644 > > > > > > > > > > --- a/drivers/gpu/drm/drm_debugfs.c > > > > > > > > > > +++ b/drivers/gpu/drm/drm_debugfs.c > > > > > > > > > > @@ -90,15 +90,17 @@ static int drm_clients_info(struct > > > > > > > > > > seq_file > > > > > > > >
Re: [RFC 3/3] drm: Update file owner during use
On 06/01/2023 18:00, Daniel Vetter wrote: On Fri, Jan 06, 2023 at 03:53:13PM +0100, Christian König wrote: Am 06.01.23 um 11:53 schrieb Daniel Vetter: On Fri, Jan 06, 2023 at 11:32:17AM +0100, Christian König wrote: Am 05.01.23 um 13:32 schrieb Daniel Vetter: [SNIP] For the case of an master fd I actually don't see the reason why we should limit that? And fd can become master if it either was master before or has CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid here? This is just info/debug printing, I don't see the connection to drm_auth/master stuff? Aside from the patch mixes up the master opener and the current user due to fd passing or something like that. That's exactly why my comment meant as well. The connect is that the drm_auth/master code currently the pid/tgid as indicator if the "owner" of the fd has changed and so if an access should be allowed or not. I find that approach a bit questionable. Note that we cannot do that (I think at least, after pondering this some more) because it would break the logind master fd passing scheme - there the receiving compositor is explicitly _not_ allowed to acquire master rights on its own. So the master priviledges must not move with the fd or things can go wrong. That could be the rational behind that, but why doesn't logind then just pass on a normal render node to the compositor? Because the compositor wants the kms node. We have 3 access levels in drm - render stuff - modeset stuff (needs a card* node and master rights for changing things) - set/drop master (needs root) logind wants to give the compositor modeset access, but not master drop/set access (because vt switching is controlled by logind). The pid check in drm_auth is for the use-case where you start your compositor on a root vt (or setuid-root), and then want to make sure that after cred dropping, set/drop master keeps working. Because in that case the vt switch dance is done by the compositor. Maybe we should document this stuff a bit better :-) Maybe add a friendly warning? E.g. like "Don't touch it, it works!" :) I think Tvrtko just volunteered for that :-) Maybe addition in the drm-uapi.rst section would be good that fills out the gaps we have. I can attempt to copy, paste and tidy what you wrote here, albeit with less than full degree of authority. Assuming into the existing comment above drm_master_check_perm? But in terms of where my series is going next I would need some clarification in the other sub-thread. Regards, Tvrtko So basically this is a valid use case where logind set/get the master status of a fd while the compositor uses the master functionality? Yup, and the compositor is _not_ allowed to call these. Despite that it's the exact sime struct file - it has to be the same struct file in both loging and compositor, otherwise logind cannot orchestratet the vt switch dance for the compositors. Which unlike non-logind vt switching has the nice property that if a compositor dies/goes rogue, logind can still force the switch. With vt-only switching you need the sysrq to reset the console to text and kill the foreground process for the same effect. -Daniel Christian. -Daniel Christian. -Daniel Regards, Christian. Regards, Tvrtko -Daniel return 0; if (!capable(CAP_SYS_ADMIN)) diff --git a/drivers/gpu/drm/drm_debugfs.c b/drivers/gpu/drm/drm_debugfs.c index 42f657772025..9d4e3146a2b8 100644 --- a/drivers/gpu/drm/drm_debugfs.c +++ b/drivers/gpu/drm/drm_debugfs.c @@ -90,15 +90,17 @@ static int drm_clients_info(struct seq_file *m, void *data) */ mutex_lock(>filelist_mutex); list_for_each_entry_reverse(priv, >filelist, lhead) { - struct task_struct *task; bool is_current_master = drm_is_current_master(priv); + struct task_struct *task; + struct pid *pid; - rcu_read_lock(); /* locks pid_task()->comm */ - task = pid_task(priv->pid, PIDTYPE_TGID); + rcu_read_lock(); /* Locks priv->pid and pid_task()->comm! */ + pid = rcu_dereference(priv->pid); + task = pid_task(pid, PIDTYPE_TGID); uid = task ? __task_cred(task)->euid : GLOBAL_ROOT_UID; seq_printf(m, "%20s %5d %3d %c %c %5d %10u\n", task ? task->comm : "", - pid_vnr(priv->pid), + pid_vnr(pid), priv->minor->index, is_current_master ? 'y' : 'n', priv->authenticated ? 'y' : 'n', diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c index 20a9aef2b398..3433f9610dba 100644 --- a/drivers/gpu/drm/drm_file.c +++ b/drivers/gpu/drm/drm_file.c @@ -156,7 +156,7 @@ struct drm_file *drm_file_alloc(struct drm_minor *minor) if (!file) return ERR_PTR(-ENOMEM); - file->pid = get_pid(task_tgid(current)); + rcu_assign_pointer(file->pid, get_pid(task_tgid(current)));
Re: [RFC 3/3] drm: Update file owner during use
On Fri, Jan 06, 2023 at 03:53:13PM +0100, Christian König wrote: > Am 06.01.23 um 11:53 schrieb Daniel Vetter: > > On Fri, Jan 06, 2023 at 11:32:17AM +0100, Christian König wrote: > > > Am 05.01.23 um 13:32 schrieb Daniel Vetter: > > > > [SNIP] > > > > > For the case of an master fd I actually don't see the reason why we > > > > > should > > > > > limit that? And fd can become master if it either was master before > > > > > or has > > > > > CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid here? > > > > This is just info/debug printing, I don't see the connection to > > > > drm_auth/master stuff? Aside from the patch mixes up the master opener > > > > and > > > > the current user due to fd passing or something like that. > > > That's exactly why my comment meant as well. > > > > > > The connect is that the drm_auth/master code currently the pid/tgid as > > > indicator if the "owner" of the fd has changed and so if an access should > > > be > > > allowed or not. I find that approach a bit questionable. > > > > > > > Note that we cannot do that (I think at least, after pondering this some > > > > more) because it would break the logind master fd passing scheme - there > > > > the receiving compositor is explicitly _not_ allowed to acquire master > > > > rights on its own. So the master priviledges must not move with the fd > > > > or > > > > things can go wrong. > > > That could be the rational behind that, but why doesn't logind then just > > > pass on a normal render node to the compositor? > > Because the compositor wants the kms node. We have 3 access levels in drm > > > > - render stuff > > - modeset stuff (needs a card* node and master rights for changing things) > > - set/drop master (needs root) > > > > logind wants to give the compositor modeset access, but not master > > drop/set access (because vt switching is controlled by logind). > > > > The pid check in drm_auth is for the use-case where you start your > > compositor on a root vt (or setuid-root), and then want to make sure > > that after cred dropping, set/drop master keeps working. Because in that > > case the vt switch dance is done by the compositor. > > > > Maybe we should document this stuff a bit better :-) > > Maybe add a friendly warning? E.g. like "Don't touch it, it works!" :) I think Tvrtko just volunteered for that :-) Maybe addition in the drm-uapi.rst section would be good that fills out the gaps we have. > So basically this is a valid use case where logind set/get the master status > of a fd while the compositor uses the master functionality? Yup, and the compositor is _not_ allowed to call these. Despite that it's the exact sime struct file - it has to be the same struct file in both loging and compositor, otherwise logind cannot orchestratet the vt switch dance for the compositors. Which unlike non-logind vt switching has the nice property that if a compositor dies/goes rogue, logind can still force the switch. With vt-only switching you need the sysrq to reset the console to text and kill the foreground process for the same effect. -Daniel > > Christian. > > > -Daniel > > > > > Christian. > > > > > > > -Daniel > > > > > > > > > > > > > > > > > Regards, > > > > > Christian. > > > > > > > > > > > Regards, > > > > > > > > > > > > Tvrtko > > > > > > > > > > > > > -Daniel > > > > > > > > > > > > > > > > > > > > > > return 0; > > > > > > > > if (!capable(CAP_SYS_ADMIN)) > > > > > > > > diff --git a/drivers/gpu/drm/drm_debugfs.c > > > > > > > > b/drivers/gpu/drm/drm_debugfs.c > > > > > > > > index 42f657772025..9d4e3146a2b8 100644 > > > > > > > > --- a/drivers/gpu/drm/drm_debugfs.c > > > > > > > > +++ b/drivers/gpu/drm/drm_debugfs.c > > > > > > > > @@ -90,15 +90,17 @@ static int drm_clients_info(struct seq_file > > > > > > > > *m, void *data) > > > > > > > > */ > > > > > > > > mutex_lock(>filelist_mutex); > > > > > > > > list_for_each_entry_reverse(priv, >filelist, > > > > > > > > lhead) { > > > > > > > > - struct task_struct *task; > > > > > > > > bool is_current_master = > > > > > > > > drm_is_current_master(priv); > > > > > > > > + struct task_struct *task; > > > > > > > > + struct pid *pid; > > > > > > > > - rcu_read_lock(); /* locks pid_task()->comm */ > > > > > > > > - task = pid_task(priv->pid, PIDTYPE_TGID); > > > > > > > > + rcu_read_lock(); /* Locks priv->pid and > > > > > > > > pid_task()->comm! */ > > > > > > > > + pid = rcu_dereference(priv->pid); > > > > > > > > + task = pid_task(pid, PIDTYPE_TGID); > > > > > > > > uid = task ? __task_cred(task)->euid : > > > > > > > > GLOBAL_ROOT_UID; > > > > > > > > seq_printf(m, "%20s %5d %3d %c %c %5d %10u\n", > > > > > > > > task ? task->comm : "", > > > > > > > > - pid_vnr(priv->pid), > > > > > > > > + pid_vnr(pid), > > > > > > > >
Re: [RFC 3/3] drm: Update file owner during use
Am 06.01.23 um 11:53 schrieb Daniel Vetter: On Fri, Jan 06, 2023 at 11:32:17AM +0100, Christian König wrote: Am 05.01.23 um 13:32 schrieb Daniel Vetter: [SNIP] For the case of an master fd I actually don't see the reason why we should limit that? And fd can become master if it either was master before or has CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid here? This is just info/debug printing, I don't see the connection to drm_auth/master stuff? Aside from the patch mixes up the master opener and the current user due to fd passing or something like that. That's exactly why my comment meant as well. The connect is that the drm_auth/master code currently the pid/tgid as indicator if the "owner" of the fd has changed and so if an access should be allowed or not. I find that approach a bit questionable. Note that we cannot do that (I think at least, after pondering this some more) because it would break the logind master fd passing scheme - there the receiving compositor is explicitly _not_ allowed to acquire master rights on its own. So the master priviledges must not move with the fd or things can go wrong. That could be the rational behind that, but why doesn't logind then just pass on a normal render node to the compositor? Because the compositor wants the kms node. We have 3 access levels in drm - render stuff - modeset stuff (needs a card* node and master rights for changing things) - set/drop master (needs root) logind wants to give the compositor modeset access, but not master drop/set access (because vt switching is controlled by logind). The pid check in drm_auth is for the use-case where you start your compositor on a root vt (or setuid-root), and then want to make sure that after cred dropping, set/drop master keeps working. Because in that case the vt switch dance is done by the compositor. Maybe we should document this stuff a bit better :-) Maybe add a friendly warning? E.g. like "Don't touch it, it works!" :) So basically this is a valid use case where logind set/get the master status of a fd while the compositor uses the master functionality? Christian. -Daniel Christian. -Daniel Regards, Christian. Regards, Tvrtko -Daniel return 0; if (!capable(CAP_SYS_ADMIN)) diff --git a/drivers/gpu/drm/drm_debugfs.c b/drivers/gpu/drm/drm_debugfs.c index 42f657772025..9d4e3146a2b8 100644 --- a/drivers/gpu/drm/drm_debugfs.c +++ b/drivers/gpu/drm/drm_debugfs.c @@ -90,15 +90,17 @@ static int drm_clients_info(struct seq_file *m, void *data) */ mutex_lock(>filelist_mutex); list_for_each_entry_reverse(priv, >filelist, lhead) { - struct task_struct *task; bool is_current_master = drm_is_current_master(priv); + struct task_struct *task; + struct pid *pid; - rcu_read_lock(); /* locks pid_task()->comm */ - task = pid_task(priv->pid, PIDTYPE_TGID); + rcu_read_lock(); /* Locks priv->pid and pid_task()->comm! */ + pid = rcu_dereference(priv->pid); + task = pid_task(pid, PIDTYPE_TGID); uid = task ? __task_cred(task)->euid : GLOBAL_ROOT_UID; seq_printf(m, "%20s %5d %3d %c %c %5d %10u\n", task ? task->comm : "", - pid_vnr(priv->pid), + pid_vnr(pid), priv->minor->index, is_current_master ? 'y' : 'n', priv->authenticated ? 'y' : 'n', diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c index 20a9aef2b398..3433f9610dba 100644 --- a/drivers/gpu/drm/drm_file.c +++ b/drivers/gpu/drm/drm_file.c @@ -156,7 +156,7 @@ struct drm_file *drm_file_alloc(struct drm_minor *minor) if (!file) return ERR_PTR(-ENOMEM); - file->pid = get_pid(task_tgid(current)); + rcu_assign_pointer(file->pid, get_pid(task_tgid(current))); file->minor = minor; /* for compatibility root is always authenticated */ @@ -502,6 +502,36 @@ int drm_release(struct inode *inode, struct file *filp) } EXPORT_SYMBOL(drm_release); +void drm_file_update_pid(struct drm_file *filp) +{ + struct drm_device *dev; + struct pid *pid, *old; + + /* Master nodes are not expected to be passed between processes. */ + if (filp->was_master) + return; + + pid = task_tgid(current); + + /* + * Quick unlocked check since the model is a single handover followed by + * exclusive repeated use. + */ + if (pid == rcu_access_pointer(filp->pid)) + return; + + dev = filp->minor->dev; + mutex_lock(>filelist_mutex); + old = rcu_replace_pointer(filp->pid, pid, 1); + mutex_unlock(>filelist_mutex); + + if (pid != old) { + get_pid(pid); + synchronize_rcu(); + put_pid(old); + } +} + /** * drm_release_noglobal - release method for DRM file * @inode: device inode diff --git a/drivers/gpu/drm/drm_ioctl.c b/drivers/gpu/drm/drm_ioctl.c
Re: [RFC 3/3] drm: Update file owner during use
On 05/01/2023 12:32, Daniel Vetter wrote: On Fri, Dec 02, 2022 at 10:01:01AM +0100, Christian König wrote: Am 01.12.22 um 12:09 schrieb Tvrtko Ursulin: On 30/11/2022 14:18, Daniel Vetter wrote: On Wed, Nov 30, 2022 at 01:34:07PM +, Tvrtko Ursulin wrote: From: Tvrtko Ursulin With the typical model where the display server opends the file descriptor and then hands it over to the client we were showing stale data in debugfs. Fix it by updating the drm_file->pid on ioctl access from a different process. The field is also made RCU protected to allow for lockless readers. Update side is protected with dev->filelist_mutex. Signed-off-by: Tvrtko Ursulin Cc: "Christian König" --- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 6 +++-- drivers/gpu/drm/drm_auth.c | 3 ++- drivers/gpu/drm/drm_debugfs.c | 10 drivers/gpu/drm/drm_file.c | 32 - drivers/gpu/drm/drm_ioctl.c | 3 +++ drivers/gpu/drm/nouveau/nouveau_drm.c | 5 +++- drivers/gpu/drm/vmwgfx/vmwgfx_gem.c | 6 +++-- include/drm/drm_file.h | 13 -- 8 files changed, 65 insertions(+), 13 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 30e24da1f398..385deb044058 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -959,6 +959,7 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) list_for_each_entry(file, >filelist, lhead) { struct task_struct *task; struct drm_gem_object *gobj; + struct pid *pid; int id; /* @@ -968,8 +969,9 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) * Therefore, we need to protect this ->comm access using RCU. */ rcu_read_lock(); - task = pid_task(file->pid, PIDTYPE_TGID); - seq_printf(m, "pid %8d command %s:\n", pid_nr(file->pid), + pid = rcu_dereference(file->pid); + task = pid_task(pid, PIDTYPE_TGID); + seq_printf(m, "pid %8d command %s:\n", pid_nr(pid), task ? task->comm : ""); rcu_read_unlock(); diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c index cf92a9ae8034..2ed2585ded37 100644 --- a/drivers/gpu/drm/drm_auth.c +++ b/drivers/gpu/drm/drm_auth.c @@ -235,7 +235,8 @@ static int drm_new_set_master(struct drm_device *dev, struct drm_file *fpriv) static int drm_master_check_perm(struct drm_device *dev, struct drm_file *file_priv) { - if (file_priv->pid == task_pid(current) && file_priv->was_master) + if (file_priv->was_master && + rcu_access_pointer(file_priv->pid) == task_pid(current)) This scares me, and also makes me wonder whether we really want to conflate the original owner with the rendering owner. And also, whether we really want to keep updating that, because for some of the "bind an fd to a pid" use-cases like svm we really do not want to ever again allow a change. So sligthly different idea: - we have a separate render pid/drm_file owner frome the open() owner that we track in drm_auth.c - that one is set the first time a driver specific ioctl is called (which for the "pass me the fd" dri3 mode should never be the compositor) - we start out with nothing and only set it once, which further simplifies the model (still need the mutex for concurrent first ioctl ofc) Simpler solution sounds plausible and mostly works for me. Certainly is attractive to simplify things. And as the disclaimer I put in the cover letter - I wasn't really sure at all if passing a master fd is a thing or not. Happy to implement your version if that will be the decision. The only downside I can think of right now with having two owners is if someone is "naughty" and actually uses the fd for rendering from two sides. That wouldn't conceptually work for what I am doing in the cgroup controller, where I need to attribute GPU usage to a process, which is a lookup from struct pid -> list of drm_files -> etc. So in the two owners scheme I would just need to ignore the "open owner" and rely that "render ownder" truly is the only one doing the rendering. Or maybe I'd need to add support for multiple owners as well.. would be a bit annoying probably. Hm now that I think about more.. the one shot nature of this scheme would have another downside. One could just send the fd back to itself via a throway forked helper, which only does one ioctl before sending it back, and then the "render owner" is forever lost. The proposal as I had it would be immune to this problem at least. Eventually we could then use that to enforce static binding to a pid, which is what we want for svm style models, i.e. if the pid changes, the fd does an -EACCESS or similar. Thoughts? This use case I am not familiar with at all so can't comment. Only
Re: [RFC 3/3] drm: Update file owner during use
On Fri, Jan 06, 2023 at 11:32:17AM +0100, Christian König wrote: > Am 05.01.23 um 13:32 schrieb Daniel Vetter: > > [SNIP] > > > For the case of an master fd I actually don't see the reason why we should > > > limit that? And fd can become master if it either was master before or has > > > CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid here? > > This is just info/debug printing, I don't see the connection to > > drm_auth/master stuff? Aside from the patch mixes up the master opener and > > the current user due to fd passing or something like that. > > That's exactly why my comment meant as well. > > The connect is that the drm_auth/master code currently the pid/tgid as > indicator if the "owner" of the fd has changed and so if an access should be > allowed or not. I find that approach a bit questionable. > > > Note that we cannot do that (I think at least, after pondering this some > > more) because it would break the logind master fd passing scheme - there > > the receiving compositor is explicitly _not_ allowed to acquire master > > rights on its own. So the master priviledges must not move with the fd or > > things can go wrong. > > That could be the rational behind that, but why doesn't logind then just > pass on a normal render node to the compositor? Because the compositor wants the kms node. We have 3 access levels in drm - render stuff - modeset stuff (needs a card* node and master rights for changing things) - set/drop master (needs root) logind wants to give the compositor modeset access, but not master drop/set access (because vt switching is controlled by logind). The pid check in drm_auth is for the use-case where you start your compositor on a root vt (or setuid-root), and then want to make sure that after cred dropping, set/drop master keeps working. Because in that case the vt switch dance is done by the compositor. Maybe we should document this stuff a bit better :-) -Daniel > > Christian. > > > -Daniel > > > > > > > > > Regards, > > > Christian. > > > > > > > Regards, > > > > > > > > Tvrtko > > > > > > > > > -Daniel > > > > > > > > > > > > > > > > return 0; > > > > > > if (!capable(CAP_SYS_ADMIN)) > > > > > > diff --git a/drivers/gpu/drm/drm_debugfs.c > > > > > > b/drivers/gpu/drm/drm_debugfs.c > > > > > > index 42f657772025..9d4e3146a2b8 100644 > > > > > > --- a/drivers/gpu/drm/drm_debugfs.c > > > > > > +++ b/drivers/gpu/drm/drm_debugfs.c > > > > > > @@ -90,15 +90,17 @@ static int drm_clients_info(struct seq_file > > > > > > *m, void *data) > > > > > > */ > > > > > > mutex_lock(>filelist_mutex); > > > > > > list_for_each_entry_reverse(priv, >filelist, lhead) { > > > > > > - struct task_struct *task; > > > > > > bool is_current_master = drm_is_current_master(priv); > > > > > > + struct task_struct *task; > > > > > > + struct pid *pid; > > > > > > - rcu_read_lock(); /* locks pid_task()->comm */ > > > > > > - task = pid_task(priv->pid, PIDTYPE_TGID); > > > > > > + rcu_read_lock(); /* Locks priv->pid and pid_task()->comm! > > > > > > */ > > > > > > + pid = rcu_dereference(priv->pid); > > > > > > + task = pid_task(pid, PIDTYPE_TGID); > > > > > > uid = task ? __task_cred(task)->euid : GLOBAL_ROOT_UID; > > > > > > seq_printf(m, "%20s %5d %3d %c %c %5d %10u\n", > > > > > > task ? task->comm : "", > > > > > > - pid_vnr(priv->pid), > > > > > > + pid_vnr(pid), > > > > > > priv->minor->index, > > > > > > is_current_master ? 'y' : 'n', > > > > > > priv->authenticated ? 'y' : 'n', > > > > > > diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c > > > > > > index 20a9aef2b398..3433f9610dba 100644 > > > > > > --- a/drivers/gpu/drm/drm_file.c > > > > > > +++ b/drivers/gpu/drm/drm_file.c > > > > > > @@ -156,7 +156,7 @@ struct drm_file *drm_file_alloc(struct > > > > > > drm_minor *minor) > > > > > > if (!file) > > > > > > return ERR_PTR(-ENOMEM); > > > > > > - file->pid = get_pid(task_tgid(current)); > > > > > > + rcu_assign_pointer(file->pid, get_pid(task_tgid(current))); > > > > > > file->minor = minor; > > > > > > /* for compatibility root is always authenticated */ > > > > > > @@ -502,6 +502,36 @@ int drm_release(struct inode *inode, struct > > > > > > file *filp) > > > > > > } > > > > > > EXPORT_SYMBOL(drm_release); > > > > > > +void drm_file_update_pid(struct drm_file *filp) > > > > > > +{ > > > > > > + struct drm_device *dev; > > > > > > + struct pid *pid, *old; > > > > > > + > > > > > > + /* Master nodes are not expected to be passed between > > > > > > processes. */ > > > > > > + if (filp->was_master) > > > > > > + return; > > > > > > + > > > > > > + pid = task_tgid(current); > > > > > > + > > > > > > + /* > > > > > > + * Quick
Re: [RFC 3/3] drm: Update file owner during use
Am 05.01.23 um 13:32 schrieb Daniel Vetter: [SNIP] For the case of an master fd I actually don't see the reason why we should limit that? And fd can become master if it either was master before or has CAP_SYS_ADMIN. Why would we want an extra check for the pid/tgid here? This is just info/debug printing, I don't see the connection to drm_auth/master stuff? Aside from the patch mixes up the master opener and the current user due to fd passing or something like that. That's exactly why my comment meant as well. The connect is that the drm_auth/master code currently the pid/tgid as indicator if the "owner" of the fd has changed and so if an access should be allowed or not. I find that approach a bit questionable. Note that we cannot do that (I think at least, after pondering this some more) because it would break the logind master fd passing scheme - there the receiving compositor is explicitly _not_ allowed to acquire master rights on its own. So the master priviledges must not move with the fd or things can go wrong. That could be the rational behind that, but why doesn't logind then just pass on a normal render node to the compositor? Christian. -Daniel Regards, Christian. Regards, Tvrtko -Daniel return 0; if (!capable(CAP_SYS_ADMIN)) diff --git a/drivers/gpu/drm/drm_debugfs.c b/drivers/gpu/drm/drm_debugfs.c index 42f657772025..9d4e3146a2b8 100644 --- a/drivers/gpu/drm/drm_debugfs.c +++ b/drivers/gpu/drm/drm_debugfs.c @@ -90,15 +90,17 @@ static int drm_clients_info(struct seq_file *m, void *data) */ mutex_lock(>filelist_mutex); list_for_each_entry_reverse(priv, >filelist, lhead) { - struct task_struct *task; bool is_current_master = drm_is_current_master(priv); + struct task_struct *task; + struct pid *pid; - rcu_read_lock(); /* locks pid_task()->comm */ - task = pid_task(priv->pid, PIDTYPE_TGID); + rcu_read_lock(); /* Locks priv->pid and pid_task()->comm! */ + pid = rcu_dereference(priv->pid); + task = pid_task(pid, PIDTYPE_TGID); uid = task ? __task_cred(task)->euid : GLOBAL_ROOT_UID; seq_printf(m, "%20s %5d %3d %c %c %5d %10u\n", task ? task->comm : "", - pid_vnr(priv->pid), + pid_vnr(pid), priv->minor->index, is_current_master ? 'y' : 'n', priv->authenticated ? 'y' : 'n', diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c index 20a9aef2b398..3433f9610dba 100644 --- a/drivers/gpu/drm/drm_file.c +++ b/drivers/gpu/drm/drm_file.c @@ -156,7 +156,7 @@ struct drm_file *drm_file_alloc(struct drm_minor *minor) if (!file) return ERR_PTR(-ENOMEM); - file->pid = get_pid(task_tgid(current)); + rcu_assign_pointer(file->pid, get_pid(task_tgid(current))); file->minor = minor; /* for compatibility root is always authenticated */ @@ -502,6 +502,36 @@ int drm_release(struct inode *inode, struct file *filp) } EXPORT_SYMBOL(drm_release); +void drm_file_update_pid(struct drm_file *filp) +{ + struct drm_device *dev; + struct pid *pid, *old; + + /* Master nodes are not expected to be passed between processes. */ + if (filp->was_master) + return; + + pid = task_tgid(current); + + /* + * Quick unlocked check since the model is a single handover followed by + * exclusive repeated use. + */ + if (pid == rcu_access_pointer(filp->pid)) + return; + + dev = filp->minor->dev; + mutex_lock(>filelist_mutex); + old = rcu_replace_pointer(filp->pid, pid, 1); + mutex_unlock(>filelist_mutex); + + if (pid != old) { + get_pid(pid); + synchronize_rcu(); + put_pid(old); + } +} + /** * drm_release_noglobal - release method for DRM file * @inode: device inode diff --git a/drivers/gpu/drm/drm_ioctl.c b/drivers/gpu/drm/drm_ioctl.c index 7c9d66ee917d..305b18d9d7b6 100644 --- a/drivers/gpu/drm/drm_ioctl.c +++ b/drivers/gpu/drm/drm_ioctl.c @@ -775,6 +775,9 @@ long drm_ioctl_kernel(struct file *file, drm_ioctl_t *func, void *kdata, struct drm_device *dev = file_priv->minor->dev; int retcode; + /* Update drm_file owner if fd was passed along. */ + drm_file_update_pid(file_priv); + if (drm_dev_is_unplugged(dev)) return -ENODEV; diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c index 80f154b6adab..a763d3ee61fb 100644 --- a/drivers/gpu/drm/nouveau/nouveau_drm.c +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c @@ -1097,7 +1097,10 @@ nouveau_drm_open(struct drm_device *dev, struct drm_file *fpriv) } get_task_comm(tmpname, current); - snprintf(name, sizeof(name), "%s[%d]", tmpname, pid_nr(fpriv->pid)); + rcu_read_lock(); + snprintf(name, sizeof(name), "%s[%d]", + tmpname, pid_nr(rcu_dereference(fpriv->pid))); +
Re: [RFC 3/3] drm: Update file owner during use
On Fri, Dec 02, 2022 at 10:01:01AM +0100, Christian König wrote: > Am 01.12.22 um 12:09 schrieb Tvrtko Ursulin: > > > > On 30/11/2022 14:18, Daniel Vetter wrote: > > > On Wed, Nov 30, 2022 at 01:34:07PM +, Tvrtko Ursulin wrote: > > > > From: Tvrtko Ursulin > > > > > > > > With the typical model where the display server opends the file > > > > descriptor > > > > and then hands it over to the client we were showing stale data in > > > > debugfs. > > > > > > > > Fix it by updating the drm_file->pid on ioctl access from a different > > > > process. > > > > > > > > The field is also made RCU protected to allow for lockless > > > > readers. Update > > > > side is protected with dev->filelist_mutex. > > > > > > > > Signed-off-by: Tvrtko Ursulin > > > > Cc: "Christian König" > > > > --- > > > > drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 6 +++-- > > > > drivers/gpu/drm/drm_auth.c | 3 ++- > > > > drivers/gpu/drm/drm_debugfs.c | 10 > > > > drivers/gpu/drm/drm_file.c | 32 > > > > - > > > > drivers/gpu/drm/drm_ioctl.c | 3 +++ > > > > drivers/gpu/drm/nouveau/nouveau_drm.c | 5 +++- > > > > drivers/gpu/drm/vmwgfx/vmwgfx_gem.c | 6 +++-- > > > > include/drm/drm_file.h | 13 -- > > > > 8 files changed, 65 insertions(+), 13 deletions(-) > > > > > > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > > > > index 30e24da1f398..385deb044058 100644 > > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > > > > @@ -959,6 +959,7 @@ static int > > > > amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) > > > > list_for_each_entry(file, >filelist, lhead) { > > > > struct task_struct *task; > > > > struct drm_gem_object *gobj; > > > > + struct pid *pid; > > > > int id; > > > > /* > > > > @@ -968,8 +969,9 @@ static int > > > > amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) > > > > * Therefore, we need to protect this ->comm access > > > > using RCU. > > > > */ > > > > rcu_read_lock(); > > > > - task = pid_task(file->pid, PIDTYPE_TGID); > > > > - seq_printf(m, "pid %8d command %s:\n", pid_nr(file->pid), > > > > + pid = rcu_dereference(file->pid); > > > > + task = pid_task(pid, PIDTYPE_TGID); > > > > + seq_printf(m, "pid %8d command %s:\n", pid_nr(pid), > > > > task ? task->comm : ""); > > > > rcu_read_unlock(); > > > > diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c > > > > index cf92a9ae8034..2ed2585ded37 100644 > > > > --- a/drivers/gpu/drm/drm_auth.c > > > > +++ b/drivers/gpu/drm/drm_auth.c > > > > @@ -235,7 +235,8 @@ static int drm_new_set_master(struct > > > > drm_device *dev, struct drm_file *fpriv) > > > > static int > > > > drm_master_check_perm(struct drm_device *dev, struct drm_file > > > > *file_priv) > > > > { > > > > - if (file_priv->pid == task_pid(current) && file_priv->was_master) > > > > + if (file_priv->was_master && > > > > + rcu_access_pointer(file_priv->pid) == task_pid(current)) > > > > > > This scares me, and also makes me wonder whether we really want to > > > conflate the original owner with the rendering owner. And also, > > > whether we > > > really want to keep updating that, because for some of the "bind an > > > fd to > > > a pid" use-cases like svm we really do not want to ever again allow a > > > change. > > > > > > So sligthly different idea: > > > - we have a separate render pid/drm_file owner frome the open() > > > owner that > > > we track in drm_auth.c > > > - that one is set the first time a driver specific ioctl is called > > > (which > > > for the "pass me the fd" dri3 mode should never be the compositor) > > > - we start out with nothing and only set it once, which further > > > simplifies > > > the model (still need the mutex for concurrent first ioctl ofc) > > > > Simpler solution sounds plausible and mostly works for me. Certainly is > > attractive to simplify things. And as the disclaimer I put in the cover > > letter - I wasn't really sure at all if passing a master fd is a thing > > or not. Happy to implement your version if that will be the decision. > > > > The only downside I can think of right now with having two owners is if > > someone is "naughty" and actually uses the fd for rendering from two > > sides. That wouldn't conceptually work for what I am doing in the cgroup > > controller, where I need to attribute GPU usage to a process, which is a > > lookup from struct pid -> list of drm_files -> etc. So in the two owners > > scheme I would just need to ignore the "open owner" and rely that > > "render ownder" truly is the only one doing the rendering. Or maybe I'd > > need to add support for multiple
Re: [RFC 3/3] drm: Update file owner during use
Am 01.12.22 um 12:09 schrieb Tvrtko Ursulin: On 30/11/2022 14:18, Daniel Vetter wrote: On Wed, Nov 30, 2022 at 01:34:07PM +, Tvrtko Ursulin wrote: From: Tvrtko Ursulin With the typical model where the display server opends the file descriptor and then hands it over to the client we were showing stale data in debugfs. Fix it by updating the drm_file->pid on ioctl access from a different process. The field is also made RCU protected to allow for lockless readers. Update side is protected with dev->filelist_mutex. Signed-off-by: Tvrtko Ursulin Cc: "Christian König" --- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 6 +++-- drivers/gpu/drm/drm_auth.c | 3 ++- drivers/gpu/drm/drm_debugfs.c | 10 drivers/gpu/drm/drm_file.c | 32 - drivers/gpu/drm/drm_ioctl.c | 3 +++ drivers/gpu/drm/nouveau/nouveau_drm.c | 5 +++- drivers/gpu/drm/vmwgfx/vmwgfx_gem.c | 6 +++-- include/drm/drm_file.h | 13 -- 8 files changed, 65 insertions(+), 13 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 30e24da1f398..385deb044058 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -959,6 +959,7 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) list_for_each_entry(file, >filelist, lhead) { struct task_struct *task; struct drm_gem_object *gobj; + struct pid *pid; int id; /* @@ -968,8 +969,9 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) * Therefore, we need to protect this ->comm access using RCU. */ rcu_read_lock(); - task = pid_task(file->pid, PIDTYPE_TGID); - seq_printf(m, "pid %8d command %s:\n", pid_nr(file->pid), + pid = rcu_dereference(file->pid); + task = pid_task(pid, PIDTYPE_TGID); + seq_printf(m, "pid %8d command %s:\n", pid_nr(pid), task ? task->comm : ""); rcu_read_unlock(); diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c index cf92a9ae8034..2ed2585ded37 100644 --- a/drivers/gpu/drm/drm_auth.c +++ b/drivers/gpu/drm/drm_auth.c @@ -235,7 +235,8 @@ static int drm_new_set_master(struct drm_device *dev, struct drm_file *fpriv) static int drm_master_check_perm(struct drm_device *dev, struct drm_file *file_priv) { - if (file_priv->pid == task_pid(current) && file_priv->was_master) + if (file_priv->was_master && + rcu_access_pointer(file_priv->pid) == task_pid(current)) This scares me, and also makes me wonder whether we really want to conflate the original owner with the rendering owner. And also, whether we really want to keep updating that, because for some of the "bind an fd to a pid" use-cases like svm we really do not want to ever again allow a change. So sligthly different idea: - we have a separate render pid/drm_file owner frome the open() owner that we track in drm_auth.c - that one is set the first time a driver specific ioctl is called (which for the "pass me the fd" dri3 mode should never be the compositor) - we start out with nothing and only set it once, which further simplifies the model (still need the mutex for concurrent first ioctl ofc) Simpler solution sounds plausible and mostly works for me. Certainly is attractive to simplify things. And as the disclaimer I put in the cover letter - I wasn't really sure at all if passing a master fd is a thing or not. Happy to implement your version if that will be the decision. The only downside I can think of right now with having two owners is if someone is "naughty" and actually uses the fd for rendering from two sides. That wouldn't conceptually work for what I am doing in the cgroup controller, where I need to attribute GPU usage to a process, which is a lookup from struct pid -> list of drm_files -> etc. So in the two owners scheme I would just need to ignore the "open owner" and rely that "render ownder" truly is the only one doing the rendering. Or maybe I'd need to add support for multiple owners as well.. would be a bit annoying probably. Hm now that I think about more.. the one shot nature of this scheme would have another downside. One could just send the fd back to itself via a throway forked helper, which only does one ioctl before sending it back, and then the "render owner" is forever lost. The proposal as I had it would be immune to this problem at least. Eventually we could then use that to enforce static binding to a pid, which is what we want for svm style models, i.e. if the pid changes, the fd does an -EACCESS or similar. Thoughts? This use case I am not familiar with at all so can't comment. Only intuitively I would ask - why is it something that needs to be solved at the DRM level? Because
Re: [RFC 3/3] drm: Update file owner during use
On 30/11/2022 14:18, Daniel Vetter wrote: On Wed, Nov 30, 2022 at 01:34:07PM +, Tvrtko Ursulin wrote: From: Tvrtko Ursulin With the typical model where the display server opends the file descriptor and then hands it over to the client we were showing stale data in debugfs. Fix it by updating the drm_file->pid on ioctl access from a different process. The field is also made RCU protected to allow for lockless readers. Update side is protected with dev->filelist_mutex. Signed-off-by: Tvrtko Ursulin Cc: "Christian König" --- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 6 +++-- drivers/gpu/drm/drm_auth.c | 3 ++- drivers/gpu/drm/drm_debugfs.c | 10 drivers/gpu/drm/drm_file.c | 32 - drivers/gpu/drm/drm_ioctl.c | 3 +++ drivers/gpu/drm/nouveau/nouveau_drm.c | 5 +++- drivers/gpu/drm/vmwgfx/vmwgfx_gem.c | 6 +++-- include/drm/drm_file.h | 13 -- 8 files changed, 65 insertions(+), 13 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 30e24da1f398..385deb044058 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -959,6 +959,7 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) list_for_each_entry(file, >filelist, lhead) { struct task_struct *task; struct drm_gem_object *gobj; + struct pid *pid; int id; /* @@ -968,8 +969,9 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) * Therefore, we need to protect this ->comm access using RCU. */ rcu_read_lock(); - task = pid_task(file->pid, PIDTYPE_TGID); - seq_printf(m, "pid %8d command %s:\n", pid_nr(file->pid), + pid = rcu_dereference(file->pid); + task = pid_task(pid, PIDTYPE_TGID); + seq_printf(m, "pid %8d command %s:\n", pid_nr(pid), task ? task->comm : ""); rcu_read_unlock(); diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c index cf92a9ae8034..2ed2585ded37 100644 --- a/drivers/gpu/drm/drm_auth.c +++ b/drivers/gpu/drm/drm_auth.c @@ -235,7 +235,8 @@ static int drm_new_set_master(struct drm_device *dev, struct drm_file *fpriv) static int drm_master_check_perm(struct drm_device *dev, struct drm_file *file_priv) { - if (file_priv->pid == task_pid(current) && file_priv->was_master) + if (file_priv->was_master && + rcu_access_pointer(file_priv->pid) == task_pid(current)) This scares me, and also makes me wonder whether we really want to conflate the original owner with the rendering owner. And also, whether we really want to keep updating that, because for some of the "bind an fd to a pid" use-cases like svm we really do not want to ever again allow a change. So sligthly different idea: - we have a separate render pid/drm_file owner frome the open() owner that we track in drm_auth.c - that one is set the first time a driver specific ioctl is called (which for the "pass me the fd" dri3 mode should never be the compositor) - we start out with nothing and only set it once, which further simplifies the model (still need the mutex for concurrent first ioctl ofc) Simpler solution sounds plausible and mostly works for me. Certainly is attractive to simplify things. And as the disclaimer I put in the cover letter - I wasn't really sure at all if passing a master fd is a thing or not. Happy to implement your version if that will be the decision. The only downside I can think of right now with having two owners is if someone is "naughty" and actually uses the fd for rendering from two sides. That wouldn't conceptually work for what I am doing in the cgroup controller, where I need to attribute GPU usage to a process, which is a lookup from struct pid -> list of drm_files -> etc. So in the two owners scheme I would just need to ignore the "open owner" and rely that "render ownder" truly is the only one doing the rendering. Or maybe I'd need to add support for multiple owners as well.. would be a bit annoying probably. Hm now that I think about more.. the one shot nature of this scheme would have another downside. One could just send the fd back to itself via a throway forked helper, which only does one ioctl before sending it back, and then the "render owner" is forever lost. The proposal as I had it would be immune to this problem at least. Eventually we could then use that to enforce static binding to a pid, which is what we want for svm style models, i.e. if the pid changes, the fd does an -EACCESS or similar. Thoughts? This use case I am not familiar with at all so can't comment. Only intuitively I would ask - why is it something that needs to be
Re: [RFC 3/3] drm: Update file owner during use
On Wed, Nov 30, 2022 at 01:34:07PM +, Tvrtko Ursulin wrote: > From: Tvrtko Ursulin > > With the typical model where the display server opends the file descriptor > and then hands it over to the client we were showing stale data in > debugfs. > > Fix it by updating the drm_file->pid on ioctl access from a different > process. > > The field is also made RCU protected to allow for lockless readers. Update > side is protected with dev->filelist_mutex. > > Signed-off-by: Tvrtko Ursulin > Cc: "Christian König" > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 6 +++-- > drivers/gpu/drm/drm_auth.c | 3 ++- > drivers/gpu/drm/drm_debugfs.c | 10 > drivers/gpu/drm/drm_file.c | 32 - > drivers/gpu/drm/drm_ioctl.c | 3 +++ > drivers/gpu/drm/nouveau/nouveau_drm.c | 5 +++- > drivers/gpu/drm/vmwgfx/vmwgfx_gem.c | 6 +++-- > include/drm/drm_file.h | 13 -- > 8 files changed, 65 insertions(+), 13 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > index 30e24da1f398..385deb044058 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > @@ -959,6 +959,7 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file > *m, void *unused) > list_for_each_entry(file, >filelist, lhead) { > struct task_struct *task; > struct drm_gem_object *gobj; > + struct pid *pid; > int id; > > /* > @@ -968,8 +969,9 @@ static int amdgpu_debugfs_gem_info_show(struct seq_file > *m, void *unused) >* Therefore, we need to protect this ->comm access using RCU. >*/ > rcu_read_lock(); > - task = pid_task(file->pid, PIDTYPE_TGID); > - seq_printf(m, "pid %8d command %s:\n", pid_nr(file->pid), > + pid = rcu_dereference(file->pid); > + task = pid_task(pid, PIDTYPE_TGID); > + seq_printf(m, "pid %8d command %s:\n", pid_nr(pid), > task ? task->comm : ""); > rcu_read_unlock(); > > diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c > index cf92a9ae8034..2ed2585ded37 100644 > --- a/drivers/gpu/drm/drm_auth.c > +++ b/drivers/gpu/drm/drm_auth.c > @@ -235,7 +235,8 @@ static int drm_new_set_master(struct drm_device *dev, > struct drm_file *fpriv) > static int > drm_master_check_perm(struct drm_device *dev, struct drm_file *file_priv) > { > - if (file_priv->pid == task_pid(current) && file_priv->was_master) > + if (file_priv->was_master && > + rcu_access_pointer(file_priv->pid) == task_pid(current)) This scares me, and also makes me wonder whether we really want to conflate the original owner with the rendering owner. And also, whether we really want to keep updating that, because for some of the "bind an fd to a pid" use-cases like svm we really do not want to ever again allow a change. So sligthly different idea: - we have a separate render pid/drm_file owner frome the open() owner that we track in drm_auth.c - that one is set the first time a driver specific ioctl is called (which for the "pass me the fd" dri3 mode should never be the compositor) - we start out with nothing and only set it once, which further simplifies the model (still need the mutex for concurrent first ioctl ofc) Eventually we could then use that to enforce static binding to a pid, which is what we want for svm style models, i.e. if the pid changes, the fd does an -EACCESS or similar. Thoughts? -Daniel > return 0; > > if (!capable(CAP_SYS_ADMIN)) > diff --git a/drivers/gpu/drm/drm_debugfs.c b/drivers/gpu/drm/drm_debugfs.c > index 42f657772025..9d4e3146a2b8 100644 > --- a/drivers/gpu/drm/drm_debugfs.c > +++ b/drivers/gpu/drm/drm_debugfs.c > @@ -90,15 +90,17 @@ static int drm_clients_info(struct seq_file *m, void > *data) >*/ > mutex_lock(>filelist_mutex); > list_for_each_entry_reverse(priv, >filelist, lhead) { > - struct task_struct *task; > bool is_current_master = drm_is_current_master(priv); > + struct task_struct *task; > + struct pid *pid; > > - rcu_read_lock(); /* locks pid_task()->comm */ > - task = pid_task(priv->pid, PIDTYPE_TGID); > + rcu_read_lock(); /* Locks priv->pid and pid_task()->comm! */ > + pid = rcu_dereference(priv->pid); > + task = pid_task(pid, PIDTYPE_TGID); > uid = task ? __task_cred(task)->euid : GLOBAL_ROOT_UID; > seq_printf(m, "%20s %5d %3d %c%c %5d %10u\n", > task ? task->comm : "", > -pid_vnr(priv->pid), > +pid_vnr(pid), > priv->minor->index, >