interconnect: Fix locking for runpm vs reclaim
authorRob Clark <robdclark@chromium.org>
Mon, 7 Aug 2023 17:11:40 +0000 (10:11 -0700)
committerGeorgi Djakov <djakov@kernel.org>
Mon, 21 Aug 2023 22:24:59 +0000 (01:24 +0300)
For cases where icc_bw_set() can be called in callbaths that could
deadlock against shrinker/reclaim, such as runpm resume, we need to
decouple the icc locking.  Introduce a new icc_bw_lock for cases where
we need to serialize bw aggregation and update to decouple that from
paths that require memory allocation such as node/link creation/
destruction.

Fixes this lockdep splat:

   ======================================================
   WARNING: possible circular locking dependency detected
   6.2.0-rc8-debug+ #554 Not tainted
   ------------------------------------------------------
   ring0/132 is trying to acquire lock:
   ffffff80871916d0 (&gmu->lock){+.+.}-{3:3}, at: a6xx_pm_resume+0xf0/0x234

   but task is already holding lock:
   ffffffdb5aee57e8 (dma_fence_map){++++}-{0:0}, at: msm_job_run+0x68/0x150

   which lock already depends on the new lock.

   the existing dependency chain (in reverse order) is:

   -> #4 (dma_fence_map){++++}-{0:0}:
          __dma_fence_might_wait+0x74/0xc0
          dma_resv_lockdep+0x1f4/0x2f4
          do_one_initcall+0x104/0x2bc
          kernel_init_freeable+0x344/0x34c
          kernel_init+0x30/0x134
          ret_from_fork+0x10/0x20

   -> #3 (mmu_notifier_invalidate_range_start){+.+.}-{0:0}:
          fs_reclaim_acquire+0x80/0xa8
          slab_pre_alloc_hook.constprop.0+0x40/0x25c
          __kmem_cache_alloc_node+0x60/0x1cc
          __kmalloc+0xd8/0x100
          topology_parse_cpu_capacity+0x8c/0x178
          get_cpu_for_node+0x88/0xc4
          parse_cluster+0x1b0/0x28c
          parse_cluster+0x8c/0x28c
          init_cpu_topology+0x168/0x188
          smp_prepare_cpus+0x24/0xf8
          kernel_init_freeable+0x18c/0x34c
          kernel_init+0x30/0x134
          ret_from_fork+0x10/0x20

   -> #2 (fs_reclaim){+.+.}-{0:0}:
          __fs_reclaim_acquire+0x3c/0x48
          fs_reclaim_acquire+0x54/0xa8
          slab_pre_alloc_hook.constprop.0+0x40/0x25c
          __kmem_cache_alloc_node+0x60/0x1cc
          __kmalloc+0xd8/0x100
          kzalloc.constprop.0+0x14/0x20
          icc_node_create_nolock+0x4c/0xc4
          icc_node_create+0x38/0x58
          qcom_icc_rpmh_probe+0x1b8/0x248
          platform_probe+0x70/0xc4
          really_probe+0x158/0x290
          __driver_probe_device+0xc8/0xe0
          driver_probe_device+0x44/0x100
          __driver_attach+0xf8/0x108
          bus_for_each_dev+0x78/0xc4
          driver_attach+0x2c/0x38
          bus_add_driver+0xd0/0x1d8
          driver_register+0xbc/0xf8
          __platform_driver_register+0x30/0x3c
          qnoc_driver_init+0x24/0x30
          do_one_initcall+0x104/0x2bc
          kernel_init_freeable+0x344/0x34c
          kernel_init+0x30/0x134
          ret_from_fork+0x10/0x20

   -> #1 (icc_lock){+.+.}-{3:3}:
          __mutex_lock+0xcc/0x3c8
          mutex_lock_nested+0x30/0x44
          icc_set_bw+0x88/0x2b4
          _set_opp_bw+0x8c/0xd8
          _set_opp+0x19c/0x300
          dev_pm_opp_set_opp+0x84/0x94
          a6xx_gmu_resume+0x18c/0x804
          a6xx_pm_resume+0xf8/0x234
          adreno_runtime_resume+0x2c/0x38
          pm_generic_runtime_resume+0x30/0x44
          __rpm_callback+0x15c/0x174
          rpm_callback+0x78/0x7c
          rpm_resume+0x318/0x524
          __pm_runtime_resume+0x78/0xbc
          adreno_load_gpu+0xc4/0x17c
          msm_open+0x50/0x120
          drm_file_alloc+0x17c/0x228
          drm_open_helper+0x74/0x118
          drm_open+0xa0/0x144
          drm_stub_open+0xd4/0xe4
          chrdev_open+0x1b8/0x1e4
          do_dentry_open+0x2f8/0x38c
          vfs_open+0x34/0x40
          path_openat+0x64c/0x7b4
          do_filp_open+0x54/0xc4
          do_sys_openat2+0x9c/0x100
          do_sys_open+0x50/0x7c
          __arm64_sys_openat+0x28/0x34
          invoke_syscall+0x8c/0x128
          el0_svc_common.constprop.0+0xa0/0x11c
          do_el0_svc+0xac/0xbc
          el0_svc+0x48/0xa0
          el0t_64_sync_handler+0xac/0x13c
          el0t_64_sync+0x190/0x194

   -> #0 (&gmu->lock){+.+.}-{3:3}:
          __lock_acquire+0xe00/0x1060
          lock_acquire+0x1e0/0x2f8
          __mutex_lock+0xcc/0x3c8
          mutex_lock_nested+0x30/0x44
          a6xx_pm_resume+0xf0/0x234
          adreno_runtime_resume+0x2c/0x38
          pm_generic_runtime_resume+0x30/0x44
          __rpm_callback+0x15c/0x174
          rpm_callback+0x78/0x7c
          rpm_resume+0x318/0x524
          __pm_runtime_resume+0x78/0xbc
          pm_runtime_get_sync.isra.0+0x14/0x20
          msm_gpu_submit+0x58/0x178
          msm_job_run+0x78/0x150
          drm_sched_main+0x290/0x370
          kthread+0xf0/0x100
          ret_from_fork+0x10/0x20

   other info that might help us debug this:

   Chain exists of:
     &gmu->lock --> mmu_notifier_invalidate_range_start --> dma_fence_map

    Possible unsafe locking scenario:

          CPU0                    CPU1
          ----                    ----
     lock(dma_fence_map);
                                  lock(mmu_notifier_invalidate_range_start);
                                  lock(dma_fence_map);
     lock(&gmu->lock);

    *** DEADLOCK ***

   2 locks held by ring0/132:
    #0: ffffff8087191170 (&gpu->lock){+.+.}-{3:3}, at: msm_job_run+0x64/0x150
    #1: ffffffdb5aee57e8 (dma_fence_map){++++}-{0:0}, at: msm_job_run+0x68/0x150

   stack backtrace:
   CPU: 7 PID: 132 Comm: ring0 Not tainted 6.2.0-rc8-debug+ #554
   Hardware name: Google Lazor (rev1 - 2) with LTE (DT)
   Call trace:
    dump_backtrace.part.0+0xb4/0xf8
    show_stack+0x20/0x38
    dump_stack_lvl+0x9c/0xd0
    dump_stack+0x18/0x34
    print_circular_bug+0x1b4/0x1f0
    check_noncircular+0x78/0xac
    __lock_acquire+0xe00/0x1060
    lock_acquire+0x1e0/0x2f8
    __mutex_lock+0xcc/0x3c8
    mutex_lock_nested+0x30/0x44
    a6xx_pm_resume+0xf0/0x234
    adreno_runtime_resume+0x2c/0x38
    pm_generic_runtime_resume+0x30/0x44
    __rpm_callback+0x15c/0x174
    rpm_callback+0x78/0x7c
    rpm_resume+0x318/0x524
    __pm_runtime_resume+0x78/0xbc
    pm_runtime_get_sync.isra.0+0x14/0x20
    msm_gpu_submit+0x58/0x178
    msm_job_run+0x78/0x150
    drm_sched_main+0x290/0x370
    kthread+0xf0/0x100
    ret_from_fork+0x10/0x20

Signed-off-by: Rob Clark <robdclark@chromium.org>
Link: https://lore.kernel.org/r/20230807171148.210181-7-robdclark@gmail.com
Signed-off-by: Georgi Djakov <djakov@kernel.org>
drivers/interconnect/core.c

index 5fac448..e15a92a 100644 (file)
@@ -28,6 +28,7 @@ static LIST_HEAD(icc_providers);
 static int providers_count;
 static bool synced_state;
 static DEFINE_MUTEX(icc_lock);
+static DEFINE_MUTEX(icc_bw_lock);
 static struct dentry *icc_debugfs_dir;
 
 static void icc_summary_show_one(struct seq_file *s, struct icc_node *n)
@@ -631,7 +632,7 @@ int icc_set_bw(struct icc_path *path, u32 avg_bw, u32 peak_bw)
        if (WARN_ON(IS_ERR(path) || !path->num_nodes))
                return -EINVAL;
 
-       mutex_lock(&icc_lock);
+       mutex_lock(&icc_bw_lock);
 
        old_avg = path->reqs[0].avg_bw;
        old_peak = path->reqs[0].peak_bw;
@@ -663,7 +664,7 @@ int icc_set_bw(struct icc_path *path, u32 avg_bw, u32 peak_bw)
                apply_constraints(path);
        }
 
-       mutex_unlock(&icc_lock);
+       mutex_unlock(&icc_bw_lock);
 
        trace_icc_set_bw_end(path, ret);
 
@@ -872,6 +873,7 @@ void icc_node_add(struct icc_node *node, struct icc_provider *provider)
                return;
 
        mutex_lock(&icc_lock);
+       mutex_lock(&icc_bw_lock);
 
        node->provider = provider;
        list_add_tail(&node->node_list, &provider->nodes);
@@ -900,6 +902,7 @@ void icc_node_add(struct icc_node *node, struct icc_provider *provider)
        node->avg_bw = 0;
        node->peak_bw = 0;
 
+       mutex_unlock(&icc_bw_lock);
        mutex_unlock(&icc_lock);
 }
 EXPORT_SYMBOL_GPL(icc_node_add);
@@ -1025,6 +1028,7 @@ void icc_sync_state(struct device *dev)
                return;
 
        mutex_lock(&icc_lock);
+       mutex_lock(&icc_bw_lock);
        synced_state = true;
        list_for_each_entry(p, &icc_providers, provider_list) {
                dev_dbg(p->dev, "interconnect provider is in synced state\n");