John mentioned 2 options:
1. “add the sync/async flag to _submission_disable()”
Consider to be a small change
1. add an 'are busyness stats enabled' boolean to the guc structure
Seems effected area among the flow and is much more than option 1.
I would like to discuss a bit more before moving forward.
Regards,
Zhanjun Dong
From: Harrison, John C <john.c.harri...@intel.com>
Sent: June 7, 2023 4:17 PM
To: Dong, Zhanjun <zhanjun.d...@intel.com>; intel-gfx@lists.freedesktop.org;
dri-de...@lists.freedesktop.org; Nerlige Ramappa, Umesh
<umesh.nerlige.rama...@intel.com>; Ceraolo Spurio, Daniele
<daniele.ceraolospu...@intel.com>
Subject: Re: [Intel-gfx] [PATCH] drm/i915: Avoid circular locking dependency
when flush delayed work on gt reset
On 6/7/2023 12:03, Zhanjun Dong wrote:
This attempts to avoid circular locking dependency between flush delayed work
and intel_gt_reset.
Switched from cancel_delayed_work_sync to cancel_delayed_work, the non-sync
version for reset path, it is safe as the worker has the trylock code to handle
the lock; Meanwhile keep the sync version for park/fini to ensure the worker is
not still running during suspend or shutdown.
WARNING: possible circular locking dependency detected
6.4.0-rc1-drmtip_1340-g31e3463b0edb+ #1 Not tainted
------------------------------------------------------
kms_pipe_crc_ba/6415 is trying to acquire lock:
ffff88813e6cc640
((work_completion)(&(&guc->timestamp.work)->work)){+.+.}-{0:0}, at:
__flush_work+0x42/0x530
but task is already holding lock:
ffff88813e6cce90 (>->reset.mutex){+.+.}-{3:3}, at: intel_gt_reset+0x19e/0x470
[i915]
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #3 (>->reset.mutex){+.+.}-{3:3}:
lock_acquire+0xd8/0x2d0
i915_gem_shrinker_taints_mutex+0x31/0x50 [i915]
intel_gt_init_reset+0x65/0x80 [i915]
intel_gt_common_init_early+0xe1/0x170 [i915]
intel_root_gt_init_early+0x48/0x60 [i915]
i915_driver_probe+0x671/0xcb0 [i915]
i915_pci_probe+0xdc/0x210 [i915]
pci_device_probe+0x95/0x120
really_probe+0x164/0x3c0
__driver_probe_device+0x73/0x160
driver_probe_device+0x19/0xa0
__driver_attach+0xb6/0x180
bus_for_each_dev+0x77/0xd0
bus_add_driver+0x114/0x210
driver_register+0x5b/0x110
__pfx_vgem_open+0x3/0x10 [vgem]
do_one_initcall+0x57/0x270
do_init_module+0x5f/0x220
load_module+0x1ca4/0x1f00
__do_sys_finit_module+0xb4/0x130
do_syscall_64+0x3c/0x90
entry_SYSCALL_64_after_hwframe+0x72/0xdc
-> #2 (fs_reclaim){+.+.}-{0:0}:
lock_acquire+0xd8/0x2d0
fs_reclaim_acquire+0xac/0xe0
kmem_cache_alloc+0x32/0x260
i915_vma_instance+0xb2/0xc60 [i915]
i915_gem_object_ggtt_pin_ww+0x175/0x370 [i915]
vm_fault_gtt+0x22d/0xf60 [i915]
__do_fault+0x2f/0x1d0
do_pte_missing+0x4a/0xd20
__handle_mm_fault+0x5b0/0x790
handle_mm_fault+0xa2/0x230
do_user_addr_fault+0x3ea/0xa10
exc_page_fault+0x68/0x1a0
asm_exc_page_fault+0x26/0x30
-> #1 (>->reset.backoff_srcu){++++}-{0:0}:
lock_acquire+0xd8/0x2d0
_intel_gt_reset_lock+0x57/0x330 [i915]
guc_timestamp_ping+0x35/0x130 [i915]
process_one_work+0x250/0x510
worker_thread+0x4f/0x3a0
kthread+0xff/0x130
ret_from_fork+0x29/0x50
-> #0 ((work_completion)(&(&guc->timestamp.work)->work)){+.+.}-{0:0}:
check_prev_add+0x90/0xc60
__lock_acquire+0x1998/0x2590
lock_acquire+0xd8/0x2d0
__flush_work+0x74/0x530
__cancel_work_timer+0x14f/0x1f0
intel_guc_submission_reset_prepare+0x81/0x4b0 [i915]
intel_uc_reset_prepare+0x9c/0x120 [i915]
reset_prepare+0x21/0x60 [i915]
intel_gt_reset+0x1dd/0x470 [i915]
intel_gt_reset_global+0xfb/0x170 [i915]
intel_gt_handle_error+0x368/0x420 [i915]
intel_gt_debugfs_reset_store+0x5c/0xc0 [i915]
i915_wedged_set+0x29/0x40 [i915]
simple_attr_write_xsigned.constprop.0+0xb4/0x110
full_proxy_write+0x52/0x80
vfs_write+0xc5/0x4f0
ksys_write+0x64/0xe0
do_syscall_64+0x3c/0x90
entry_SYSCALL_64_after_hwframe+0x72/0xdc
other info that might help us debug this:
Chain exists of:
(work_completion)(&(&guc->timestamp.work)->work) --> fs_reclaim -->
>->reset.mutex
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(>->reset.mutex);
lock(fs_reclaim);
lock(>->reset.mutex);
lock((work_completion)(&(&guc->timestamp.work)->work));
*** DEADLOCK ***
3 locks held by kms_pipe_crc_ba/6415:
#0: ffff888101541430 (sb_writers#15){.+.+}-{0:0}, at: ksys_write+0x64/0xe0
#1: ffff888136c7eab8 (&attr->mutex){+.+.}-{3:3}, at:
simple_attr_write_xsigned.constprop.0+0x47/0x110
#2: ffff88813e6cce90 (>->reset.mutex){+.+.}-{3:3}, at:
intel_gt_reset+0x19e/0x470 [i915]
Signed-off-by: Zhanjun Dong
<zhanjun.d...@intel.com><mailto:zhanjun.d...@intel.com>
---
drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c | 15 +++++++++------
1 file changed, 9 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
index a0e3ef1c65d2..cca6960d3490 100644
--- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
@@ -1357,9 +1357,12 @@ static void guc_enable_busyness_worker(struct intel_guc
*guc)
mod_delayed_work(system_highpri_wq, &guc->timestamp.work,
guc->timestamp.ping_delay);
}
-static void guc_cancel_busyness_worker(struct intel_guc *guc)
+static void guc_cancel_busyness_worker(struct intel_guc *guc, bool sync)
{
- cancel_delayed_work_sync(&guc->timestamp.work);
+ if (sync)
+ cancel_delayed_work_sync(&guc->timestamp.work);
+ else
+ cancel_delayed_work(&guc->timestamp.work);
}
static void __reset_guc_busyness_stats(struct intel_guc *guc)
@@ -1370,7 +1373,7 @@ static void __reset_guc_busyness_stats(struct intel_guc
*guc)
unsigned long flags;
ktime_t unused;
- guc_cancel_busyness_worker(guc);
+ guc_cancel_busyness_worker(guc, 0);
Should use true/false rather than 1/0 for bool values.
Also, this needs a comment actually in the code, not just in the patch
description. E.g.:
Attempting a synchronous cancel within the reset path leads to a circular mutex
locking complaint by lockdep. However, it is safe to use an asynchronous cancel
here. If the worker does actually run concurrently with a reset then it will
early exit due to the mutex_trylock call rather than block.
spin_lock_irqsave(&guc->timestamp.lock, flags);
@@ -1485,7 +1488,7 @@ static int guc_init_engine_stats(struct intel_guc *guc)
static void guc_fini_engine_stats(struct intel_guc *guc)
{
- guc_cancel_busyness_worker(guc);
+ guc_cancel_busyness_worker(guc, 1);
}
void intel_guc_busyness_park(struct intel_gt *gt)
@@ -1500,7 +1503,7 @@ void intel_guc_busyness_park(struct intel_gt *gt)
* and causes an unclaimed register access warning. Cancel the worker
* synchronously here.
*/
- guc_cancel_busyness_worker(guc);
+ guc_cancel_busyness_worker(guc, 1);
/*
* Before parking, we should sample engine busyness stats if we need to.
@@ -4503,7 +4506,7 @@ int intel_guc_submission_enable(struct intel_guc *guc)
/* Note: By the time we're here, GuC may have already been reset */
void intel_guc_submission_disable(struct intel_guc *guc)
{
- guc_cancel_busyness_worker(guc);
+ guc_cancel_busyness_worker(guc, 0);
Hmm. I think this is going to need breaking up further back in the stack.
We definitely want to be doing a synchronous cancel in the general case of
disabling submission (e.g. due to suspend or driver unload). But if this is
happening as part of a reset call stack, then it is a problem.
AFAICT, the only way _submission_disable would be called within a reset is if
__uc_init_hw() failed for some reason. So one option would be to add the
sync/async flag to _submission_disable() as well and just make the init failure
case async with other callers being sync.
A better option might be to add an 'are busyness stats enabled' boolean to the
guc structure. And inside the cancel function, early exit if the worker is not
actually enabled (and set the flag anywhere and everywhere that does a mod_work
to enable it, which is just guc_enable_busyness_worker() I think?). That would
mean that multiple disables do nothing. So e.g. a failed reset will cancel the
worker asynchronously in reset prepare but then not try to cancel it again
synchronously in submission disable. Hmm, except that init_hw has already
enabled the worker by that point :(.
FYC: Umesh and Daniele... any thoughts?
I would be tempted to say is there any way we can just add a lockdep annotation
to ignore this issue? The lockdep splat described in the patch description
above seems like a false positive to me. Sure the reset lock is held by the
reset code which is now trying to synchronously flush the busyness worker
thread which also takes the reset lock. But the busyness worker thread does a
trylock and will abort if the lock is already held. So no issue...
However. I think we do have a genuine issue with the internal delayed worker
lock itself, which has maybe shown up in other lockdep splat reports. For
example, if a worker thread triggers a reset (e.g. anything reset related
coming in via a G2H) then the reset code is running inside a worker thread.
Which is maybe holding an internal kernel worker thread lock? So if the reset
path does a synchronous cancel of another worker thread, that also requires
taking the worker thread lock and thus a deadlock occurs.
John.
/* Semaphore interrupt disable and route to host */
guc_route_semaphores(guc, false);
