<html><head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  </head>
  <body>
    On 6/7/2023 12:03, Zhanjun Dong wrote:<br>
    <blockquote type="cite" cite="mid:20230607190350.287644-1-zhanjun.dong@intel.com">
      <pre class="moz-quote-pre" wrap="">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 (&gt->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 (&gt->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 (&gt->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 --> &gt->reset.mutex
  Possible unsafe locking scenario:
        CPU0                    CPU1
        ----                    ----
   lock(&gt->reset.mutex);
                                lock(fs_reclaim);
                                lock(&gt->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 (&gt->reset.mutex){+.+.}-{3:3}, at: intel_gt_reset+0x19e/0x470 [i915]

Signed-off-by: Zhanjun Dong <a class="moz-txt-link-rfc2396E" href="mailto:zhanjun.dong@intel.com"><zhanjun.dong@intel.com></a>
---
 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);</pre>
    </blockquote>
    Should use true/false rather than 1/0 for bool values.<br>
    <br>
    Also, this needs a comment actually in the code, not just in the
    patch description. E.g.:<br>
    <blockquote>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.<br>
    </blockquote>
    <br>
    <blockquote type="cite" cite="mid:20230607190350.287644-1-zhanjun.dong@intel.com">
      <pre class="moz-quote-pre" wrap="">
 
        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);</pre>
    </blockquote>
    Hmm. I think this is going to need breaking up further back in the
    stack.<br>
    <br>
    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.<br>
    <br>
    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.<br>
    <br>
    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 :(.<br>
    <br>
    FYC: Umesh and Daniele... any thoughts?<br>
    <br>
    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...<br>
    <br>
    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.<br>
    <br>
    John.<br>
    <br>
    <br>
    <blockquote type="cite" cite="mid:20230607190350.287644-1-zhanjun.dong@intel.com">
      <pre class="moz-quote-pre" wrap="">
 
        /* Semaphore interrupt disable and route to host */
        guc_route_semaphores(guc, false);
</pre>
    </blockquote>
    <br>
  </body>
</html>