On Tue, Aug 22, 2023 at 8:02 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
>
> From: Rob Clark <robdclark@xxxxxxxxxxxx>
>
> In the process of adding lockdep annotation for drm GPU scheduler's
> job_run() to detect potential deadlock against shrinker/reclaim, I hit
> this lockdep splat:
>
> ======================================================
> WARNING: possible circular locking dependency detected
> 6.2.0-rc8-debug+ #558 Tainted: G W
> ------------------------------------------------------
> ring0/125 is trying to acquire lock:
> ffffffd6d6ce0f28 (dev_pm_qos_mtx){+.+.}-{3:3}, at: dev_pm_qos_update_request+0x38/0x68
>
> but task is already holding lock:
> ffffff8087239208 (&gpu->active_lock){+.+.}-{3:3}, at: msm_gpu_submit+0xec/0x178
>
> which lock already depends on the new lock.
>
> the existing dependency chain (in reverse order) is:
>
> -> #4 (&gpu->active_lock){+.+.}-{3:3}:
> __mutex_lock+0xcc/0x3c8
> mutex_lock_nested+0x30/0x44
> msm_gpu_submit+0xec/0x178
> msm_job_run+0x78/0x150
> drm_sched_main+0x290/0x370
> kthread+0xf0/0x100
> ret_from_fork+0x10/0x20
>
> -> #3 (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
>
> -> #2 (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
>
> -> #1 (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_trace+0x50/0xa8
> dev_pm_qos_constraints_allocate+0x38/0x100
> __dev_pm_qos_add_request+0xb0/0x1e8
> dev_pm_qos_add_request+0x58/0x80
> dev_pm_qos_expose_latency_limit+0x60/0x13c
> register_cpu+0x12c/0x130
> topology_init+0xac/0xbc
> do_one_initcall+0x104/0x2bc
> kernel_init_freeable+0x344/0x34c
> kernel_init+0x30/0x134
> ret_from_fork+0x10/0x20
>
> -> #0 (dev_pm_qos_mtx){+.+.}-{3:3}:
> __lock_acquire+0xe00/0x1060
> lock_acquire+0x1e0/0x2f8
> __mutex_lock+0xcc/0x3c8
> mutex_lock_nested+0x30/0x44
> dev_pm_qos_update_request+0x38/0x68
> msm_devfreq_boost+0x40/0x70
> msm_devfreq_active+0xc0/0xf0
> msm_gpu_submit+0x10c/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:
> dev_pm_qos_mtx --> dma_fence_map --> &gpu->active_lock
>
> Possible unsafe locking scenario:
>
> CPU0 CPU1
> ---- ----
> lock(&gpu->active_lock);
> lock(dma_fence_map);
> lock(&gpu->active_lock);
> lock(dev_pm_qos_mtx);
>
> *** DEADLOCK ***
>
> 3 locks held by ring0/123:
> #0: ffffff8087251170 (&gpu->lock){+.+.}-{3:3}, at: msm_job_run+0x64/0x150
> #1: ffffffd00b0e57e8 (dma_fence_map){++++}-{0:0}, at: msm_job_run+0x68/0x150
> #2: ffffff8087251208 (&gpu->active_lock){+.+.}-{3:3}, at: msm_gpu_submit+0xec/0x178
>
> stack backtrace:
> CPU: 6 PID: 123 Comm: ring0 Not tainted 6.2.0-rc8-debug+ #559
> 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
> dev_pm_qos_update_request+0x38/0x68
> msm_devfreq_boost+0x40/0x70
> msm_devfreq_active+0xc0/0xf0
> msm_gpu_submit+0x10c/0x178
> msm_job_run+0x78/0x150
> drm_sched_main+0x290/0x370
> kthread+0xf0/0x100
> ret_from_fork+0x10/0x20
>
> The issue is that dev_pm_qos_mtx is held in the runpm suspend/resume (or
> freq change) path, but it is also held across allocations that could
> recurse into shrinker.
>
> Solve this by changing dev_pm_qos_constraints_allocate() into a function
> that can be called unconditionally before the device qos object is
> needed and before aquiring dev_pm_qos_mtx. This way the allocations can
acquiring
> be done without holding the mutex. In the case that we raced with
> another thread to allocate the qos object, detect this *after* acquiring
> the dev_pm_qos_mtx and simply free the redundant allocations.
>
> Suggested-by: Rafael J. Wysocki <rafael@xxxxxxxxxx>
> Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
Please feel free to add
Acked-by: Rafael J. Wysocki <rafael@xxxxxxxxxx>
to this patch and the next 2 PM QoS ones in this series.
Thanks!
> ---
> drivers/base/power/qos.c | 76 +++++++++++++++++++++++++++++-----------
> 1 file changed, 56 insertions(+), 20 deletions(-)
>
> diff --git a/drivers/base/power/qos.c b/drivers/base/power/qos.c
> index 8e93167f1783..7e95760d16dc 100644
> --- a/drivers/base/power/qos.c
> +++ b/drivers/base/power/qos.c
> @@ -185,27 +185,33 @@ static int apply_constraint(struct dev_pm_qos_request *req,
> }
>
> /*
> - * dev_pm_qos_constraints_allocate
> + * dev_pm_qos_constraints_allocate: Allocate and initializes qos constraints
> * @dev: device to allocate data for
> *
> - * Called at the first call to add_request, for constraint data allocation
> - * Must be called with the dev_pm_qos_mtx mutex held
> + * Called to allocate constraints before dev_pm_qos_mtx mutex is held. Should
> + * be matched with a call to dev_pm_qos_constraints_set() once dev_pm_qos_mtx
> + * is held.
> */
> -static int dev_pm_qos_constraints_allocate(struct device *dev)
> +static struct dev_pm_qos *dev_pm_qos_constraints_allocate(struct device *dev)
> {
> struct dev_pm_qos *qos;
> struct pm_qos_constraints *c;
> struct blocking_notifier_head *n;
>
> - qos = kzalloc(sizeof(*qos), GFP_KERNEL);
> + /*
> + * If constraints are already allocated, we can skip speculatively
> + * allocating a new one, as we don't have to work about qos transitioning
> + * from non-null to null. The constraints are only freed on device
> + * removal.
> + */
> + if (dev->power.qos)
> + return NULL;
> +
> + qos = kzalloc(sizeof(*qos) + 3 * sizeof(*n), GFP_KERNEL);
> if (!qos)
> - return -ENOMEM;
> + return NULL;
>
> - n = kzalloc(3 * sizeof(*n), GFP_KERNEL);
> - if (!n) {
> - kfree(qos);
> - return -ENOMEM;
> - }
> + n = (struct blocking_notifier_head *)(qos + 1);
>
> c = &qos->resume_latency;
> plist_head_init(&c->list);
> @@ -227,11 +233,29 @@ static int dev_pm_qos_constraints_allocate(struct device *dev)
>
> INIT_LIST_HEAD(&qos->flags.list);
>
> + return qos;
> +}
> +
> +/*
> + * dev_pm_qos_constraints_set: Ensure dev->power.qos is set
> + *
> + * If dev->power.qos is already set, free the newly allocated qos constraints.
> + * Otherwise set dev->power.qos. Must be called with dev_pm_qos_mtx held.
> + *
> + * This split unsynchronized allocation and synchronized set moves allocation
> + * out from under dev_pm_qos_mtx, so that lockdep does does not get angry about
> + * drivers which use dev_pm_qos in paths related to shrinker/reclaim.
> + */
> +static void dev_pm_qos_constraints_set(struct device *dev, struct dev_pm_qos *qos)
> +{
> + if (dev->power.qos) {
> + kfree(qos);
> + return;
> + }
> +
> spin_lock_irq(&dev->power.lock);
> dev->power.qos = qos;
> spin_unlock_irq(&dev->power.lock);
> -
> - return 0;
> }
>
> static void __dev_pm_qos_hide_latency_limit(struct device *dev);
> @@ -309,7 +333,6 @@ void dev_pm_qos_constraints_destroy(struct device *dev)
> dev->power.qos = ERR_PTR(-ENODEV);
> spin_unlock_irq(&dev->power.lock);
>
> - kfree(qos->resume_latency.notifiers);
> kfree(qos);
>
> out:
> @@ -341,7 +364,7 @@ static int __dev_pm_qos_add_request(struct device *dev,
> if (IS_ERR(dev->power.qos))
> ret = -ENODEV;
> else if (!dev->power.qos)
> - ret = dev_pm_qos_constraints_allocate(dev);
> + ret = -ENOMEM;
>
> trace_dev_pm_qos_add_request(dev_name(dev), type, value);
> if (ret)
> @@ -388,9 +411,11 @@ static int __dev_pm_qos_add_request(struct device *dev,
> int dev_pm_qos_add_request(struct device *dev, struct dev_pm_qos_request *req,
> enum dev_pm_qos_req_type type, s32 value)
> {
> + struct dev_pm_qos *qos = dev_pm_qos_constraints_allocate(dev);
> int ret;
>
> mutex_lock(&dev_pm_qos_mtx);
> + dev_pm_qos_constraints_set(dev, qos);
> ret = __dev_pm_qos_add_request(dev, req, type, value);
> mutex_unlock(&dev_pm_qos_mtx);
> return ret;
> @@ -535,14 +560,15 @@ EXPORT_SYMBOL_GPL(dev_pm_qos_remove_request);
> int dev_pm_qos_add_notifier(struct device *dev, struct notifier_block *notifier,
> enum dev_pm_qos_req_type type)
> {
> + struct dev_pm_qos *qos = dev_pm_qos_constraints_allocate(dev);
> int ret = 0;
>
> mutex_lock(&dev_pm_qos_mtx);
>
> + dev_pm_qos_constraints_set(dev, qos);
> +
> if (IS_ERR(dev->power.qos))
> ret = -ENODEV;
> - else if (!dev->power.qos)
> - ret = dev_pm_qos_constraints_allocate(dev);
>
> if (ret)
> goto unlock;
> @@ -903,12 +929,22 @@ s32 dev_pm_qos_get_user_latency_tolerance(struct device *dev)
> */
> int dev_pm_qos_update_user_latency_tolerance(struct device *dev, s32 val)
> {
> - int ret;
> + struct dev_pm_qos *qos = dev_pm_qos_constraints_allocate(dev);
> + int ret = 0;
>
> mutex_lock(&dev_pm_qos_mtx);
>
> - if (IS_ERR_OR_NULL(dev->power.qos)
> - || !dev->power.qos->latency_tolerance_req) {
> + dev_pm_qos_constraints_set(dev, qos);
> +
> + if (IS_ERR(dev->power.qos))
> + ret = -ENODEV;
> + else if (!dev->power.qos)
> + ret = -ENOMEM;
> +
> + if (ret)
> + goto out;
> +
> + if (!dev->power.qos->latency_tolerance_req) {
> struct dev_pm_qos_request *req;
>
> if (val < 0) {
> --
> 2.41.0
>
