[RFC] Exclusive gpu access for SteamVR usecases

[Andres Rodriguez]

On 2017-05-30 11:19 AM, Christian König wrote:
> Looks like a good start, but a few notes in general:
> 
> 1. Split the patches into two sets.
> 
> One for implementing changing the priorities and one for limiting the 
> priorities.
> 

No problem.

> 2. How are the priorities from processes supposed to interact with the 
> per context priority?

Do you mean process niceness?

There isn't any relationship between niceness and gpu priority.

Let me know if you meant something different here.

> 
> 3. Thinking more about it we can't limit the minimum priority in the 
> scheduler.
> 
> For example a low priority job might block resources the high priority 
> job needs to run. E.g. VRAM memory.
> 

We avoid deadlocks by making sure that all dependencies of an exclusive 
task are also elevated to the same priority as said task. Usermode (the 
DRM_MASTER) is responsible to maintain this guarantee. The kernel does 
provide an ioctl that makes this task simple, 
amdgpu_sched_process_priority_set().

Lets take a look at this issue through three different scenarios.

(i) Job dependencies are all process internal, i.e. multiple contexts in 
one process.

This is the trivial case. A call to amdgpu_sched_process_priority_set() 
will change the priority of all contexts belonging to a process in lockstep.

Once amdgpu_sched_process_priority_set() returns, it is safe to raise 
the minimum priority using amdgpu_sched_min_priority_get(). At this 
point we have a guarantee that all contexts belonging to the process 
will be in a runnable state, or all the contexts will be in a 
not-runnable state. There won't be a mix of runnable and non-runnable 
processes.

Getting into that mixed state is what could cause a deadlock, a runnable 
context depends on a non-runnable one.

Note: the current patchset needs a fix to provide this guarantee in 
multi-gpu systems.

(ii) Job dependencies between two processes.

This case is mentioned separately as it is probably the most common use 
case we will encounter for this feature. Most graphics applications 
enter producer/consumer relationship with the compositor process (window 
swapchain).

In this case the compositor should already have all the information 
required to avoid a deadlock. It knows:
   - Itself (as a process)
   - The application process
   - The dependencies between both processes

At this stage it is simple for the compositor to understand that if it 
wishes to perform an exclusive mode transition, all dependencies (which 
are known) should also be part of the exclusive group.

We should be able to implement this feature without modifying a 
game/application.

(iii) Job dependencies between multiple (3+) processes.

This scenario is very uncommon for games. For example, if a game or 
application is split into multiple processes. Process A interacts with 
the compositor. Process B does some physics/compute calculations and 
send the results to Process A.

To support this use case, we would require an interface for the 
application to communicate to the compositor its dependencies. I.e. 
Process A would say, "Also keep Process B's priority in sync with mine". 
This should be a simple bit of plumbing to allow Process A to share an 
fd from Process B with the compositor.

B --[pipe_send(fdB)]--> A --[compositor_ext_priority_group_add(fdB)]--> 
Compositor

Once the compositor is aware of all of A's dependencies, this can be 
handled in the same fashion as (ii).

A special extension would be required for compositor protocols to 
communicate the dependencies fd. Applications would also need to be 
updated to use this extension.

I think this case would be very uncommon. But it is something that we 
would be able to handle if the need would arise.

 > We need something like blocking the submitter instead (bad) or detection
 > of dependencies in the scheduler (good, but tricky to implement).
 >

I definitely agree that detecting dependencies is tricky. Which is why I 
prefer an approach where usermode defines the dependencies. It is simple 
for both the kernel and usermode to implement.

 > Otherwise we can easily run into a deadlock situation with that approach.
 >

The current API does allow you to deadlock yourself pretty easily if 
misused. But so do many other APIs, like having a thread trying to grab 
the same lock twice :)

Thanks for the comments,
Andres

> Regards,
> Christian.
> 
> Am 25.05.2017 um 02:00 schrieb Andres Rodriguez:
>> 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
>>
>