for sub-scheduler authority checks. Add compat wrappers which fall back to
direct p->scx field writes on older kernels.
Suggested-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
In preparation for hierarchical schedulers, change scx_bpf_dsq_insert() and
scx_bpf_dsq_insert_vtime() to return bool instead of void. With
sub-schedulers, there will be no reliable way to guarantee a task is still
owned by the sub-scheduler at insertion time (e.g., the task may have been
migrated to another scheduler). The bool return value will enable
sub-schedulers to detect and gracefully handle insertion failures.
For the root scheduler, insertion failures will continue to trigger scheduler
abort via scx_error(), so existing code doesn't need to check the return
value. Backward compatibility is maintained through compat wrappers.
Also update scx_bpf_dsq_move() documentation to clarify that it can return
false for sub-schedulers when @dsq_id points to a disallowed local DSQ.
Reviewed-by: Changwoo Min <changwoo@igalia.com>
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_bpf_dsq_insert_vtime() and scx_bpf_select_cpu_and() currently have 5
parameters. An upcoming change will add aux__prog parameter which will exceed
BPF's 5 argument limit.
Prepare by adding new kfuncs __scx_bpf_dsq_insert_vtime() and
__scx_bpf_select_cpu_and() that take args structs. The existing kfuncs are
kept as compatibility wrappers. BPF programs use inline wrappers that detect
kernel API version via bpf_core_type_exists() and use the new struct-based
kfuncs when available, falling back to compat kfuncs otherwise. This allows
BPF programs to work with both old and new kernels.
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
With the planned hierarchical scheduler support, sub-schedulers will need to
be verified for authority before being allowed to modify task->scx.slice and
task->scx.dsq_vtime. Add scx_bpf_task_set_slice() and
scx_bpf_task_set_dsq_vtime() which will perform the necessary permission
checks.
Root schedulers can still directly write to these fields, so this doesn't
affect existing schedulers.
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Enough time has passed since the introduction of scx_bpf_task_cgroup() and
the scx_bpf_dispatch* -> scx_bpf_dsq* kfunc renaming. Strip the compatibility
macros.
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
There is no need to complete the entire scx initialization if a
scheduler is failing to be attached due to a hotplug event.
Exit early to avoid unnecessary work and simplify the attach flow.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
On systems with >4096 CPUs, scx_kick_cpus_pnt_seqs allocation fails during
boot because it exceeds the 32,768 byte percpu allocator limit.
Restructure to use DEFINE_PER_CPU() for the per-CPU pointers, with each CPU
pointing to its own kvzalloc'd array. Move allocation from boot time to
scx_enable() and free in scx_disable(), so the O(nr_cpu_ids^2) memory is only
consumed when sched_ext is active.
Use RCU to guard against racing with free. Arrays are freed via call_rcu()
and kick_cpus_irq_workfn() uses rcu_dereference_bh() with a NULL check.
While at it, rename to scx_kick_pseqs for brevity and update comments to
clarify these are pick_task sequence numbers.
v2: RCU protect scx_kick_seqs to manage kick_cpus_irq_workfn() racing
against disable as per Andrea.
v3: Fix bugs notcied by Andrea.
Reported-by: Phil Auld <pauld@redhat.com>
Link: http://lkml.kernel.org/r/20251007133523.GA93086@pauld.westford.csb
Cc: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Reviewed-by: Phil Auld <pauld@redhat.com>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The sched_ext code calls queue_balance_callback() during enqueue_task()
to defer operations that drop multiple locks until we can unpin them.
The call assumes that the rq lock is held until the callbacks are
invoked, and the pending callbacks will not be visible to any other
threads. This is enforced by a WARN_ON_ONCE() in rq_pin_lock().
However, balance_one() may actually drop the lock during a BPF dispatch
call. Another thread may win the race to get the rq lock and see the
pending callback. To avoid this, sched_ext must only queue the callback
after the dispatch calls have completed.
CPU 0 CPU 1 CPU 2
scx_balance()
rq_unpin_lock()
scx_balance_one()
|= IN_BALANCE scx_enqueue()
ops.dispatch()
rq_unlock()
rq_lock()
queue_balance_callback()
rq_unlock()
[WARN] rq_pin_lock()
rq_lock()
&= ~IN_BALANCE
rq_repin_lock()
Changelog
v2-> v1 (https://lore.kernel.org/sched-ext/aOgOxtHCeyRT_7jn@gpd4)
- Fixed explanation in patch description (Andrea)
- Fixed scx_rq mask state updates (Andrea)
- Added Reviewed-by tag from Andrea
Reported-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Emil Tsalapatis (Meta) <emil@etsalapatis.com>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Since the last renewal time was initialized to 0 and jiffies start
counting at -5 minutes, any clients connected in the first 5 minutes
after a reboot would have their renewal timer set to a very long
interval. If the connection was idle, this would result in the client
state timing out on the server and the next call to the server would
return NFS4ERR_BADSESSION.
Fix this by initializing the last renewal time to the current jiffies
instead of 0.
Signed-off-by: Joshua Watt <jpewhacker@gmail.com>
Signed-off-by: Anna Schumaker <anna.schumaker@oracle.com>
I noticed xfstests generic/193 and generic/355 started failing against
knfsd after commit e7a8ebc305 ("NFSD: Offer write delegation for OPEN
with OPEN4_SHARE_ACCESS_WRITE").
I ran those same tests against ONTAP (which has had write delegation
support for a lot longer than knfsd) and they fail there too... so
while it's a new failure against knfsd, it isn't an entirely new
failure.
Add the NFS_INO_REVAL_FORCED flag so that the presence of a delegation
doesn't keep the inode from being revalidated to fetch the updated mode.
Signed-off-by: Scott Mayhew <smayhew@redhat.com>
Signed-off-by: Anna Schumaker <anna.schumaker@oracle.com>
The setting of delay_retrans is applied to synchronous RPC operations
because the retransmit count is stored in same struct nfs4_exception
that is passed each time an error is checked. However, for asynchronous
operations (READ, WRITE, LOCKU, CLOSE, DELEGRETURN), a new struct
nfs4_exception is made on the stack each time the task callback is
invoked. This means that the retransmit count is always zero and thus
delay_retrans never takes effect.
Apply delay_retrans to these operations by tracking and updating their
retransmit count.
Change-Id: Ieb33e046c2b277cb979caa3faca7f52faf0568c9
Signed-off-by: Joshua Watt <jpewhacker@gmail.com>
Reviewed-by: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: Anna Schumaker <anna.schumaker@oracle.com>
Move mirror_array's dss_count initialization and dss allocation to
ff_layout_alloc_mirror(), just before the loop that initializes each
nfs4_ff_layout_ds_stripe's nfs_file_localio.
Also handle NULL return from kcalloc() and remove one level of indent
in ff_layout_alloc_mirror().
This commit fixes dangling nfsd_serv refcount issues seen when using
NFS LOCALIO and then attempting to stop the NFSD service.
Fixes: 20b1d75fb8 ("NFSv4/flexfiles: Add support for striped layouts")
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
Signed-off-by: Anna Schumaker <anna.schumaker@oracle.com>
By the time scx_sched_free_rcu_work() runs, the scx_sched is no longer
reachable. However, a previously queued error_irq_work may still be pending or
running. Ensure it completes before proceeding with teardown.
Fixes: bff3b5aec1 ("sched_ext: Move disable machinery into scx_sched")
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This reverts commit 90bfb28d5f.
Kevin reports that this commit causes an issue for him with LVM
snapshots, most likely because of turning off NOWAIT support while a
snapshot is being created. This makes -EOPNOTSUPP bubble back through
the completion handler, where io_uring read/write handling should just
retry it.
Reinstate the previous check removed by the referenced commit.
Cc: stable@vger.kernel.org
Fixes: 90bfb28d5f ("io_uring/rw: drop -EOPNOTSUPP check in __io_complete_rw_common()")
Reported-by: Salvatore Bonaccorso <carnil@debian.org>
Reported-by: Kevin Lumik <kevin@xf.ee>
Link: https://lore.kernel.org/io-uring/cceb723c-051b-4de2-9a4c-4aa82e1619ee@kernel.dk/
Signed-off-by: Jens Axboe <axboe@kernel.dk>
User requested power limits and clock settings are already restored as
part of smu_restore_dpm_user_profile(). It's unnecessary to call the
same restore as part of smu_resume().
Revert the following commits to drop that extra restore:
commit ed4efe426a ("drm/amd: Restore cached power limit during resume")
commit 796ff8a7e0 ("drm/amd: Restore cached manual clock settings during resume")
commit f9b80514a7 ("drm/amd: Only restore cached manual clock settings in restore if OD enabled")
Suggested-by: Lijo Lazar <Lijo.Lazar@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Remove queue after reset flag is required to remove a queue that has
been successfully reset to clean up the MES' internal state.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
userptrs could be changed by the user at any time and
hence while locking all the bos before GPU start processing
validate all the userptr bos.
Signed-off-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
At times we need a bo reference for hmm and for that add
a new struct amdgpu_hmm_range which will hold an optional
bo member and hmm_range.
Use amdgpu_hmm_range instead of hmm_range and let the bo
as an optional argument for the caller if they want to
the bo reference to be taken or they want to handle that
explicitly.
Signed-off-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For legacy SOCs, discovery binary is sideloaded. Instead of checking for
binary blob, use a flag to determine if discovery region needs to be
reserved.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
provide a interface that allows ras client send msg to smu/pmfw directly.
Signed-off-by: Yang Wang <kevinyang.wang@amd.com>
Signed-off-by: YiPeng Chai <YiPeng.Chai@amd.com>
Reviewed-by: Tao Zhou <tao.zhou1@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Allow access to user interfaces like sysfs/hwmon only after full
initialization of the device. When device is part of XGMI hive and a
reset is required during initialization, the inteface files will be
created as part of minimal device initialization. Full initialization of
the device will be done only after all devices in XGMI hive are probed
and a reset is done together on all.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Complete reset-on-init sequence before sysfs interfaces are created.
Devices get properly initiaized only after reset, and then only sysfs
interfaces should be made available.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add amdgpu_discovery_info structure to keep all discovery related
information.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Aggregate sysfs ini/fini calls into separate functions. No functional
change.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Decouple the amdgpu fence from the amdgpu_job structure.
This lets us clean up the separate fence ops for the embedded
fence and other fences. This also allows us to allocate the
vm fence up front when we allocate the job.
v2: Additional cleanup suggested by Christian
v3: Additional cleanups suggested by Christian
v4: Additional cleanups suggested by David and
vm fence fix
v5: cast seqno (David)
Cc: David.Wu3@amd.com
Cc: christian.koenig@amd.com
Tested-by: David (Ming Qiang) Wu <David.Wu3@amd.com>
Reviewed-by: David (Ming Qiang) Wu <David.Wu3@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Vangogh has separate limits for default PPT and fast PPT. Add
infrastructure to save both of these limits and restore both of them.
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The min/max limits only make sense for default PPT. Restructure
smu_set_power_limit() to only use them in that case.
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When passed around internally the upper 8 bits of power limit include
the limit type. This is non-obvious without digging into the nuances
of each function. Instead pass the limit type as an argument to all
applicable layers.
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
VCN and VPE userqs are not yet supported and this code is
not correct. Userspace should provide the correct
doorbell offset with in their doorbell page for the IP.
Adjusting it here will not work as expected as userspace
and the queue itself will have different offsets.
We need to add a INFO IOCTL query to get the offset and
range for each IP within the doorbell page to handle this
properly.
Cc: Saleemkhan Jamadar <saleemkhan.jamadar@amd.com>
Reviewed-by: Saleemkhan Jamadar <saleemkhan.jamadar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If a userq failed to suspend the rest of the suspend sequence may
have problems. Pass the error code up to the caller for a decision
on what to do.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If multiple userq IDR are in use and there is an error handling one
at suspend or resume it will be silently discarded.
Switch the suspend/resume() code to use guards and return immediately.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If IP suspend fails the callers should be notified so that they can
potentially react.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_device_ip_suspend_phase2() calls amdgpu_ip_block_suspend()
which already sets HW block status to false when succeeding with
IP suspend. Remove the explicit call in
amdgpu_device_ip_suspend_phase2() so that the status is accurate.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Error handling was introduced in commit e095026f00 ("drm/amdgpu:
validate suspend before function call") so the comment about TODO is no
longer needed.
Fixes: e095026f00 ("drm/amdgpu: validate suspend before function call")
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_device_ip_suspend() doesn't have a caller outside of
amdgpu_device.c. Make it static.
No intended functional changes.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
The shutdown() callback uses amdgpu_ip_suspend() which doesn't notify
drm clients during shutdown. This could lead to hangs.
[How]
Change amdgpu_pci_shutdown() to call the same sequence as suspend/resume.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When a user unmaps a userq VA, the driver must ensure
the queue has no in-flight jobs. If there is pending work,
the kernel should wait for the attached eviction (bookkeeping)
fence to signal before deleting the mapping.
Suggested-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
It requires validating the userq VA whether is mapped before
trying to resume the queue.
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Keeping waiting the userq fence infinitely until
hang detection, and then suspend the hang queue and
set the fence error.
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Track the userq obj for its life time, and reference and
dereference the buffer flag at its creating and destroying
period.
Suggested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add the userq object virtual address list_add() helpers
for tracking the userq obj va address usage.
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
There has been multiple complains that 10 seconds are usually to long.
The original requirement for longer timeout came from compute tests on
AMDVLK, since that is no longer a topic reduce the timeout back to 2
seconds for all queues.
While at it also remove any special handling for compute queues under
SRIOV or pass through.
v2: fix checkpatch warning.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Unnecessary headers can slow down the build, drop em.
No intended functional changes.
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Tested-by: Robert Beckett <bob.beckett@collabora.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
A few changes have more whitespace than needed. Clean them up.
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Tested-by: Robert Beckett <bob.beckett@collabora.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The firmware limits the max vmid, but align the
settings with the hw limits as well just to be safe.
Reviewed-by: Shaoyun liu <Shaoyun.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For SDMA IP versions >= v4.4.2, firmware will take care of quiescing SDMA
before mode-2 reset.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This lock was part of the SDMA workaround originally implemented in
gmc_v10_0_flush_gpu_tlb (a70cb2176f moved it to
amdgpu_gmc_flush_gpu_tlb).
This means this lock is useless and be safely dropped.
Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_ttm_copy_mem_to_mem has a single caller, make sure the out
fence is non-NULL to simplify the code.
Since none of the pointers should be NULL, we can enable
__attribute__((nonnull))__.
While at it make the function static since it's only used from
amdgpuu_ttm.c.
Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
gfx_v9_4_3_xcc_kcq_init_queue doesn't have a fail condition.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>