On 2017-05-26 05:02 AM, Mao, David wrote: > Hi Andres, > Why the fd is needed for this interface? The fd is used to identify the process for which we wish to raise the priority. It can be any fd from the target process, it doesn't have to be a drm file descriptor at all. The fd is used to retrieve the (struct pid*) of the target process on the kernel side. In effect, it is a replacement for passing a pid number across process boundaries. For reference, amdgpu_sched_process_priority_set() in patch 3 > Why not just using the dev->fd instead of > IIRC, if there are more than one fds opened in the process upon the same > device, they will share the same amdgpu_device_handle which is > guaranteed by amdgpu_device_initialize. Thanks for pointing that out. I wasn't aware that the amdgpu drm layer would always perform all command submission through the same fd (dev->fd) for the same amdgpu_device. Your suggestion actually makes it a lot simpler to deal with this issue at a file level instead of at a process level. Since only one fd per device is used for command submission. For a multi-gpu setup we would still need to share multiple fds, across the process boundaries. This also helped me realize that my current implementation doesn't deal with multi-gpu cases correctly. As I iterate over the fds belonging to a single drm device. > In other word, we should not run into the case that user creates more > contexts with newly opened fd after tuning the priority of existing > context in the same process unless the previous fd is closed. > > Thanks. > Best Regards, > David > >> On 25 May 2017, at 8:00 AM, Andres Rodriguez <andresx7 at gmail.com >> <mailto:andresx7 at gmail.com>> wrote: >> >> When multiple environments are running simultaneously on a system, e.g. >> an X desktop + a SteamVR game session, it may be useful to sacrifice >> performance in one environment in order to boost it on the other. >> >> This series provides a mechanism for a DRM_MASTER to provide exclusive >> gpu access to a group of processes. >> >> Note: This series is built on the assumption that the drm lease patch >> series >> will extend DRM_MASTER status to lesees. >> >> The libdrm we intend to provide is as follows: >> >> /** >> * Set the priority of all contexts in a process >> * >> * This function will change the priority of all contexts owned by >> * the process identified by fd. >> * >> * \param dev - \c [in] device handle >> * \param fd - \c [in] fd from target process >> * \param priority - \c [in] target priority AMDGPU_CTX_PRIORITY_* >> * >> * \return 0 on success\n >> * <0 - Negative POSIX error code >> * >> * \notes @fd can be *any* file descriptor from the target process. >> * \notes this function requires DRM_MASTER >> */ >> int amdgpu_sched_process_priority_set(amdgpu_device_handle dev, >> int fd, int32_t priority); >> >> /** >> * Request to raise the minimum required priority to schedule a gpu job >> * >> * Submit a request to increase the minimum required priority to schedule >> * a gpu job. Once this function returns, the gpu scheduler will no longer >> * consider jobs from contexts with priority lower than @priority. >> * >> * The minimum priority considered by the scheduler will be the highest >> from >> * all currently active requests. >> * >> * Requests are refcounted, and must be balanced using >> * amdgpu_sched_min_priority_put() >> * >> * \param dev - \c [in] device handle >> * \param priority - \c [in] target priority AMDGPU_CTX_PRIORITY_* >> * >> * \return 0 on success\n >> * <0 - Negative POSIX error code >> * >> * \notes this function requires DRM_MASTER >> */ >> int amdgpu_sched_min_priority_get(amdgpu_device_handle dev, >> int32_t priority); >> >> /** >> * Drop a request to raise the minimum required scheduler priority >> * >> * This call balances amdgpu_sched_min_priority_get() >> * >> * If no other active requests exists for @priority, the minimum required >> * priority will decay to a lower level until one is reached with an active >> * request or the lowest priority is reached. >> * >> * \param dev - \c [in] device handle >> * \param priority - \c [in] target priority AMDGPU_CTX_PRIORITY_* >> * >> * \return 0 on success\n >> * <0 - Negative POSIX error code >> * >> * \notes this function requires DRM_MASTER >> */ >> int amdgpu_sched_min_priority_put(amdgpu_device_handle dev, >> int32_t priority); >> >> Using this app, VRComposer can raise the priority of the VRapp and >> itself. Then >> it can restrict the minimum scheduler priority in order to become >> exclusive gpu >> clients. >> >> One of the areas I'd like feedback is the following scenario. If a >> VRapp opens >> a new fd and creates a new context after a call to set_priority, this >> specific >> context will be lower priority than the rest. If the minimum required >> priority >> is then raised, it is possible that this new context will be starved and >> deadlock the VRapp. >> >> One solution I had in mind to address this situation, is to make >> set_priority >> also raise the priority of future contexts created by the VRapp. >> However, that >> would require keeping track of the requested priority on a per-process >> data >> structure. The current design appears to steer clean of keeping any >> process >> specific data, and everything instead of stored on a per-file basis. >> Which is >> why I did not pursue this approach. But if this is something you'd >> like me to >> implement let me know. >> >> One could also argue that preventing an application deadlock should be >> handled >> between the VRComposer and the VRApp. It is not the kernel's >> responsibility to >> babysit userspace applications and prevent themselves from shooting >> themselves >> in the foot. The same could be achieved by improper usage of shared fences >> between processes. >> >> Thoughts/feedback/comments on this issue, or others, are appreciated. >> >> Regards, >> Andres >> >> >> _______________________________________________ >> amd-gfx mailing list >> amd-gfx at lists.freedesktop.org <mailto:amd-gfx at lists.freedesktop.org> >> https://lists.freedesktop.org/mailman/listinfo/amd-gfx >
