On 2021-10-26 6:54 a.m., Christian König wrote:
Am 26.10.21 um 04:33 schrieb Andrey Grodzovsky:
On 2021-10-25 3:56 p.m., Christian König wrote:
In general I'm all there to get this fixed, but there is one major
problem: Drivers don't expect the lock to be dropped.
I am probably missing something but in my approach we only modify the
code for those clients that call dma_fence_signal,
not dma_fence_signal_locked. In those cases the drivers are agnostic
to lock behavior (or should be at least) since the lock
is acquired within the dma fence code. Note that if you are worried
about calling the callback without lock then same exact
concern is relevant to using the irq_work directly in the fence code
since the irq_work will execute at a later time without locked
fence->lock (which is the point of using irq_work).
Yeah, I've seen that it just doesn't make much sense to me.
Not clear what doesn't make sense ?
What we could do is to change all drivers so they call always call
the dma_fence_signal functions and drop the _locked variants. This
way we could move calling the callback out of the spinlock.
But that requires audit of all drivers, so quite a lot of work to do.
As i said earlier - if we only modify dma_fence_signal and don't
touch dma_fence_signal_locked then our only concern should the users
of dma_fence_signal.
Yes, but what do you do with the drivers who call the _locked variant?
IMHO we don't touch them at all, they stay as is, we only re-implement
dma_fence_signal because drivers that use the locked variant take the
lock explicitly and so they intend for their callbacks to run
under the lock and if they don't , it's their own problem within their
code and they should fix it locally there. Driver that call
dma_fence_signal are our only problem because they didn't take the lock
explicitly but were forced to run the callback under lock by the
dma-fence framework and that something we can change.
Let me please know if I am still missing some point of yours.
Well, I mean we need to be able to handle this for all drivers.
For sure, but as i said above in my opinion we need to change only for
those drivers that don't use the _locked version.
Andrey
Regards,
Christian.
Andrey
Regards,
Christian.
Am 25.10.21 um 21:10 schrieb Andrey Grodzovsky:
Adding back Daniel (somehow he got off the addresses list) and
Chris who worked a lot in this area.
On 2021-10-21 2:34 a.m., Christian König wrote:
Am 20.10.21 um 21:32 schrieb Andrey Grodzovsky:
On 2021-10-04 4:14 a.m., Christian König wrote:
The problem is a bit different.
The callback is on the dependent fence, while we need to signal
the scheduler fence.
Daniel is right that this needs an irq_work struct to handle
this properly.
Christian.
So we had some discussions with Christian regarding irq_work and
agreed I should look into doing it but stepping back for a sec -
Why we insist on calling the dma_fence_cb with fence->lock
locked ? Is it because of dma_fence_add_callback ?
Because we first test for DMA_FENCE_FLAG_SIGNALED_BIT and only
after that lock the fence->lock ? If so, can't we
move DMA_FENCE_FLAG_SIGNALED_BIT check inside the locked section
? Because if in theory
we could call the cb with unlocked fence->lock (i.e. this kind of
iteration
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.bootlin.com%2Flinux%2Fv5.15-rc6%2Fsource%2Fdrivers%2Fgpu%2Fdrm%2Fttm%2Fttm_resource.c%23L117&data=04%7C01%7Candrey.grodzovsky%40amd.com%7Cc8a4525f94c244bebbd208d997f19242%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637707886075917091%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=YBq%2BNkDuYKERc8XJDWTfeM%2FSknpuCBHQYgN8Uo5PFv0%3D&reserved=0)
we wouldn't have the lockdep splat. And in general, is it really
the correct approach to call a third party code from a call back
with locked spinlock ? We don't know what the cb does inside
and I don't see any explicit restrictions in documentation of
dma_fence_func_t what can and cannot be done there.
Yeah, that's exactly what I meant with using the irq_work directly
in the fence code.
My idea is not to use irq work at all but instead to implement
unlocked dma_fence cb execution using iteration
which drops the spinlock each time next cb is executed and
acquiring it again after (until cb_list is empy).
The problem is dma_fence_signal_locked() which is used by quite a
number of drivers to signal the fence while holding the lock.
For this I think we should not reuse dma_fence_signal_locked inside
dma_fence_signal and instead implement it using the
unlocked iteration I mentioned above. I looked a bit in the code
and the history and I see that until some time ago
(this commit by Chris 0fc89b6802ba1fcc561b0c906e0cefd384e3b2e5),
indeed dma_fence_signal was doing it's own, locked iteration
and wasn't reusing dma_fence_signal_locked. This way whoever relies
on the dma_fence_signal_locked won't be impacted
an who is not (like us in
drm_sched_fence_scheduled/drm_sched_fence_finished) should also not
be impacted by more narrow
scope of the lock. I also looked at dma_fence_default_wait and how
it locks the fence->lock and check if fence is signaled
before wait start and I don't see a problem there either.
I attached quick draft of this proposal to clarify.
Andrey
Otherwise we could indeed simplify the fence handling a lot.
Christian.
Andrey
Am 01.10.21 um 17:10 schrieb Andrey Grodzovsky:
From what I see here you supposed to have actual deadlock and
not only warning, sched_fence->finished is first signaled from
within
hw fence done callback (drm_sched_job_done_cb) but then again
from within it's own callback (drm_sched_entity_kill_jobs_cb)
and so
looks like same fence object is recursively signaled twice.
This leads to attempt to lock fence->lock second time while
it's already
locked. I don't see a need to call drm_sched_fence_finished
from within drm_sched_entity_kill_jobs_cb as this callback
already registered
on sched_fence->finished fence (entity->last_scheduled ==
s_fence->finished) and hence the signaling already took place.
Andrey
On 2021-10-01 6:50 a.m., Christian König wrote:
Hey, Andrey.
while investigating some memory management problems I've got
the logdep splat below.
Looks like something is wrong with
drm_sched_entity_kill_jobs_cb(), can you investigate?
Thanks,
Christian.
[11176.741052] ============================================
[11176.741056] WARNING: possible recursive locking detected
[11176.741060] 5.15.0-rc1-00031-g9d546d600800 #171 Not tainted
[11176.741066] --------------------------------------------
[11176.741070] swapper/12/0 is trying to acquire lock:
[11176.741074] ffff9c337ed175a8 (&fence->lock){-.-.}-{3:3},
at: dma_fence_signal+0x28/0x80
[11176.741088]
but task is already holding lock:
[11176.741092] ffff9c337ed172a8 (&fence->lock){-.-.}-{3:3},
at: dma_fence_signal+0x28/0x80
[11176.741100]
other info that might help us debug this:
[11176.741104] Possible unsafe locking scenario:
[11176.741108] CPU0
[11176.741110] ----
[11176.741113] lock(&fence->lock);
[11176.741118] lock(&fence->lock);
[11176.741122]
*** DEADLOCK ***
[11176.741125] May be due to missing lock nesting notation
[11176.741128] 2 locks held by swapper/12/0:
[11176.741133] #0: ffff9c339c30f768
(&ring->fence_drv.lock){-.-.}-{3:3}, at:
dma_fence_signal+0x28/0x80
[11176.741142] #1: ffff9c337ed172a8
(&fence->lock){-.-.}-{3:3}, at: dma_fence_signal+0x28/0x80
[11176.741151]
stack backtrace:
[11176.741155] CPU: 12 PID: 0 Comm: swapper/12 Not tainted
5.15.0-rc1-00031-g9d546d600800 #171
[11176.741160] Hardware name: System manufacturer System
Product Name/PRIME X399-A, BIOS 0808 10/12/2018
[11176.741165] Call Trace:
[11176.741169] <IRQ>
[11176.741173] dump_stack_lvl+0x5b/0x74
[11176.741181] dump_stack+0x10/0x12
[11176.741186] __lock_acquire.cold+0x208/0x2df
[11176.741197] lock_acquire+0xc6/0x2d0
[11176.741204] ? dma_fence_signal+0x28/0x80
[11176.741212] _raw_spin_lock_irqsave+0x4d/0x70
[11176.741219] ? dma_fence_signal+0x28/0x80
[11176.741225] dma_fence_signal+0x28/0x80
[11176.741230] drm_sched_fence_finished+0x12/0x20 [gpu_sched]
[11176.741240] drm_sched_entity_kill_jobs_cb+0x1c/0x50
[gpu_sched]
[11176.741248] dma_fence_signal_timestamp_locked+0xac/0x1a0
[11176.741254] dma_fence_signal+0x3b/0x80
[11176.741260] drm_sched_fence_finished+0x12/0x20 [gpu_sched]
[11176.741268] drm_sched_job_done.isra.0+0x7f/0x1a0 [gpu_sched]
[11176.741277] drm_sched_job_done_cb+0x12/0x20 [gpu_sched]
[11176.741284] dma_fence_signal_timestamp_locked+0xac/0x1a0
[11176.741290] dma_fence_signal+0x3b/0x80
[11176.741296] amdgpu_fence_process+0xd1/0x140 [amdgpu]
[11176.741504] sdma_v4_0_process_trap_irq+0x8c/0xb0 [amdgpu]
[11176.741731] amdgpu_irq_dispatch+0xce/0x250 [amdgpu]
[11176.741954] amdgpu_ih_process+0x81/0x100 [amdgpu]
[11176.742174] amdgpu_irq_handler+0x26/0xa0 [amdgpu]
[11176.742393] __handle_irq_event_percpu+0x4f/0x2c0
[11176.742402] handle_irq_event_percpu+0x33/0x80
[11176.742408] handle_irq_event+0x39/0x60
[11176.742414] handle_edge_irq+0x93/0x1d0
[11176.742419] __common_interrupt+0x50/0xe0
[11176.742426] common_interrupt+0x80/0x90
[11176.742431] </IRQ>
[11176.742436] asm_common_interrupt+0x1e/0x40
[11176.742442] RIP: 0010:cpuidle_enter_state+0xff/0x470
[11176.742449] Code: 0f a3 05 04 54 24 01 0f 82 70 02 00 00 31
ff e8 37 5d 6f ff 80 7d d7 00 0f 85 e9 01 00 00 e8 58 a2 7f ff
fb 66 0f 1f 44 00 00 <45> 85 ff 0f 88 01 01 00 00 49 63 c7 4c
2b 75 c8 48 8d 14 40 48 8d
[11176.742455] RSP: 0018:ffffb6970021fe48 EFLAGS: 00000202
[11176.742461] RAX: 000000000059be25 RBX: 0000000000000002
RCX: 0000000000000000
[11176.742465] RDX: 0000000000000000 RSI: 0000000000000000
RDI: ffffffff9efeed78
[11176.742470] RBP: ffffb6970021fe80 R08: 0000000000000001
R09: 0000000000000001
[11176.742473] R10: 0000000000000001 R11: 0000000000000001
R12: ffff9c3350b0e800
[11176.742477] R13: ffffffffa00e9680 R14: 00000a2a49ada060
R15: 0000000000000002
[11176.742483] ? cpuidle_enter_state+0xf8/0x470
[11176.742489] ? cpuidle_enter_state+0xf8/0x470
[11176.742495] cpuidle_enter+0x2e/0x40
[11176.742500] call_cpuidle+0x23/0x40
[11176.742506] do_idle+0x201/0x280
[11176.742512] cpu_startup_entry+0x20/0x30
[11176.742517] start_secondary+0x11f/0x160
[11176.742523] secondary_startup_64_no_verify+0xb0/0xbb
