We currently free sync when we find job list empty
If aggressive_sync is set to true, we try to free
sync during channel unbind() call
But we rarely free sync from channel_unbind() call
since freeing it when job list is empty is
aggressive enough
Hence remove sync free code from channel_unbind()
Implement refcounting for sync:
- get a refcount while submitting a job (and
allocate sync if it is not allocated already)
- put a refcount while freeing the job
- if refcount==0 and if aggressive_sync_destroy is
set, free the sync
- if aggressive_sync_destroy is not set, we will
free the sync during channel close time
Bug 200187553
Change-Id: I74e24adb15dc26a375ebca1fdd017b3ad6d57b61
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/1120410
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
Currently, we create sync_fence (from nvhost_sync_create_fence())
for every submit
But not all submits request for a sync_fence.
Also, nvhost_sync_create_fence() API takes about 1/3rd of the total
submit path.
Hence to optimize, we can allocate sync_fence
only when user explicitly asks for it using
(NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_GET &&
NVGPU_SUBMIT_GPFIFO_FLAGS_SYNC_FENCE)
Also, in CDE path from gk20a_prepare_compressible_read(),
we reuse existing fence stored in "state" and that can
result into not returning sync_fence_fd when user asked
for it
Hence, force allocation of sync_fence when job submission
comes from CDE path
Bug 200141116
Change-Id: Ia921701bf0e2432d6b8a5e8b7d91160e7f52db1e
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/812845
(cherry picked from commit 5fd47015eeed00352cc8473eff969a66c94fee98)
Reviewed-on: http://git-master/r/837662
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Sachin Nikam <snikam@nvidia.com>
gm20b has more channels than sync points. We use aggressive reclaim
of sync points to offset that. Disable aggressive reclaim for gk20a
because it is not needed there.
Bug 1583849
Change-Id: I2a74b0504150a54cb8a97016effe20c5d905ac95
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/657095
Reviewed-by: Deepak Nibade <dnibade@nvidia.com>
When moving compression state tracking and compbit management ops to
kernel, we need to attach a fence to dma-buf metadata, along with the
compbit state.
To make in-kernel fence management easier, introduce a new gk20a_fence
abstraction. A gk20a_fence may be backed by a semaphore or a syncpoint
(id, value) pair. If the kernel is configured with CONFIG_SYNC, it will
also contain a sync_fence. The gk20a_fence can easily be converted back
to a syncpoint (id, value) parir or sync FD when we need to return it to
user space.
Change gk20a_submit_channel_gpfifo to return a gk20a_fence instead of
nvhost_fence. This is to facilitate work submission initiated from
kernel.
Bug 1509620
Change-Id: I6154764a279dba83f5e91ba9e0cb5e227ca08e1b
Signed-off-by: Lauri Peltonen <lpeltonen@nvidia.com>
Reviewed-on: http://git-master/r/439846
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
Add semaphore implementation of the gk20a_channel_sync interface.
Each channel has one semaphore pool, which is mapped as read-write to
the channel vm. We allocate one or two semaphores from the pool for each
submit.
The first semaphore is only needed if we need to wait for an opaque sync
fd. In that case, we allocate the semaphore, and ask GPU to wait for
it's value to become 1 (semaphore acquire method). We also queue a
kernel work that waits on the fence fd, and subsequently releases the
semaphore (sets its value to 1) so that the command buffer can proceed.
The second semaphore is used on every submit, and is used for work
completion tracking. The GPU sets its value to 1 when the command buffer
has been processed.
The channel jobs need to hold references to both semaphores so that
their backing semaphore pool slots are not reused while the job is in
flight. Therefore gk20a_channel_fence will keep a reference to the
semaphore that it represents (channel fences are stored in the job
structure). This means that we must diligently close and dup the
gk20a_channel_fence objects to avoid leaking semaphores.
Bug 1450122
Bug 1445450
Change-Id: Ib61091a1b7632fa36efe0289011040ef7c4ae8f8
Signed-off-by: Lauri Peltonen <lpeltonen@nvidia.com>
Reviewed-on: http://git-master/r/374844
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Add infrastructure to allow freeing gpu channels' syncpts at
channel_unbind()
Currently, we free the syncpt at channel_free() only.
But we can free the syncpt when channel becomes idle.
When we do channel_unbind(), channel cannot be scheduled and
hence we can destroy its syncpt.
Channel will request again for new syncpt when it has new work
to do.
This feature will be enabled with boolean flag
syncpt_aggressive_destroy
Bug 1305024
Change-Id: Ib498e2c371e36ffc1430d4f95f2780b4c587e43e
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/395153
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
