From efeaed4d98eb4dc9ce01e1dca6d3778d180b272c Mon Sep 17 00:00:00 2001 From: Felix Kuehling Date: Wed, 11 Jul 2018 22:32:44 -0400 Subject: [PATCH] drm/amdkfd: Reliably prevent reclaim-FS while holding DQM lock MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit This is needed to prevent deadlocks when MMU notifiers run in reclaim-FS context and take the DQM lock for userptr evictions. Previously this was done by making all memory allocations under DQM locks GFP_NOIO. This is error prone. Using memalloc_nofs_save/restore will reliably affect all memory allocations anywhere in the kernel while the DQM lock is held. Signed-off-by: Felix Kuehling Acked-by: Christian König Signed-off-by: Oded Gabbay --- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 85 +++++++++++----------- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 20 ++++- 2 files changed, 62 insertions(+), 43 deletions(-) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 668ad07..f2f81d2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -240,7 +240,7 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm, print_queue(q); - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (dqm->total_queue_count >= max_num_of_queues_per_device) { pr_warn("Can't create new usermode queue because %d queues were already created\n", @@ -297,7 +297,7 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm, dqm->total_queue_count); out_unlock: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -457,9 +457,9 @@ static int destroy_queue_nocpsch(struct device_queue_manager *dqm, { int retval; - mutex_lock(&dqm->lock); + dqm_lock(dqm); retval = destroy_queue_nocpsch_locked(dqm, qpd, q); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -471,7 +471,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q) struct kfd_process_device *pdd; bool prev_active = false; - mutex_lock(&dqm->lock); + dqm_lock(dqm); pdd = kfd_get_process_device_data(q->device, q->process); if (!pdd) { retval = -ENODEV; @@ -537,7 +537,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q) &q->properties, q->process->mm); out_unlock: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -570,7 +570,7 @@ static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, struct kfd_process_device *pdd; int retval = 0; - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (qpd->evicted++ > 0) /* already evicted, do nothing */ goto out; @@ -600,7 +600,7 @@ static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, } out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -611,7 +611,7 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, struct kfd_process_device *pdd; int retval = 0; - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (qpd->evicted++ > 0) /* already evicted, do nothing */ goto out; @@ -633,7 +633,7 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -650,7 +650,7 @@ static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, /* Retrieve PD base */ pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm); - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ goto out; if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ @@ -695,7 +695,7 @@ static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, } qpd->evicted = 0; out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -711,7 +711,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, /* Retrieve PD base */ pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm); - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ goto out; if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ @@ -739,7 +739,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, if (!retval) qpd->evicted = 0; out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -761,7 +761,7 @@ static int register_process(struct device_queue_manager *dqm, /* Retrieve PD base */ pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm); - mutex_lock(&dqm->lock); + dqm_lock(dqm); list_add(&n->list, &dqm->queues); /* Update PD Base in QPD */ @@ -771,7 +771,7 @@ static int register_process(struct device_queue_manager *dqm, dqm->processes_count++; - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -786,7 +786,7 @@ static int unregister_process(struct device_queue_manager *dqm, list_empty(&qpd->queues_list) ? "empty" : "not empty"); retval = 0; - mutex_lock(&dqm->lock); + dqm_lock(dqm); list_for_each_entry_safe(cur, next, &dqm->queues, list) { if (qpd == cur->qpd) { @@ -799,7 +799,7 @@ static int unregister_process(struct device_queue_manager *dqm, /* qpd not found in dqm list */ retval = 1; out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -838,7 +838,7 @@ static int initialize_nocpsch(struct device_queue_manager *dqm) if (!dqm->allocated_queues) return -ENOMEM; - mutex_init(&dqm->lock); + mutex_init(&dqm->lock_hidden); INIT_LIST_HEAD(&dqm->queues); dqm->queue_count = dqm->next_pipe_to_allocate = 0; dqm->sdma_queue_count = 0; @@ -867,7 +867,7 @@ static void uninitialize(struct device_queue_manager *dqm) kfree(dqm->allocated_queues); for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++) kfree(dqm->mqds[i]); - mutex_destroy(&dqm->lock); + mutex_destroy(&dqm->lock_hidden); kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem); } @@ -1003,7 +1003,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm) { pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); - mutex_init(&dqm->lock); + mutex_init(&dqm->lock_hidden); INIT_LIST_HEAD(&dqm->queues); dqm->queue_count = dqm->processes_count = 0; dqm->sdma_queue_count = 0; @@ -1041,9 +1041,9 @@ static int start_cpsch(struct device_queue_manager *dqm) init_interrupts(dqm); - mutex_lock(&dqm->lock); + dqm_lock(dqm); execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return 0; fail_allocate_vidmem: @@ -1055,9 +1055,9 @@ fail_packet_manager_init: static int stop_cpsch(struct device_queue_manager *dqm) { - mutex_lock(&dqm->lock); + dqm_lock(dqm); unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); pm_uninit(&dqm->packets); @@ -1069,11 +1069,11 @@ static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, struct kernel_queue *kq, struct qcm_process_device *qpd) { - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (dqm->total_queue_count >= max_num_of_queues_per_device) { pr_warn("Can't create new kernel queue because %d queues were already created\n", dqm->total_queue_count); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return -EPERM; } @@ -1089,7 +1089,7 @@ static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, dqm->queue_count++; qpd->is_debug = true; execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return 0; } @@ -1098,7 +1098,7 @@ static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, struct kernel_queue *kq, struct qcm_process_device *qpd) { - mutex_lock(&dqm->lock); + dqm_lock(dqm); list_del(&kq->list); dqm->queue_count--; qpd->is_debug = false; @@ -1110,7 +1110,7 @@ static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, dqm->total_queue_count--; pr_debug("Total of %d queues are accountable so far\n", dqm->total_queue_count); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); } static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, @@ -1121,7 +1121,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, retval = 0; - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (dqm->total_queue_count >= max_num_of_queues_per_device) { pr_warn("Can't create new usermode queue because %d queues were already created\n", @@ -1188,7 +1188,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, pr_debug("Total of %d queues are accountable so far\n", dqm->total_queue_count); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; out_deallocate_doorbell: @@ -1197,7 +1197,8 @@ out_deallocate_sdma_queue: if (q->properties.type == KFD_QUEUE_TYPE_SDMA) deallocate_sdma_queue(dqm, q->sdma_id); out_unlock: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); + return retval; } @@ -1314,7 +1315,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, retval = 0; /* remove queue from list to prevent rescheduling after preemption */ - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (qpd->is_debug) { /* @@ -1360,14 +1361,14 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, pr_debug("Total of %d queues are accountable so far\n", dqm->total_queue_count); - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; failed: failed_try_destroy_debugged_queue: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -1391,7 +1392,7 @@ static bool set_cache_memory_policy(struct device_queue_manager *dqm, if (!dqm->asic_ops.set_cache_memory_policy) return retval; - mutex_lock(&dqm->lock); + dqm_lock(dqm); if (alternate_aperture_size == 0) { /* base > limit disables APE1 */ @@ -1437,7 +1438,7 @@ static bool set_cache_memory_policy(struct device_queue_manager *dqm, qpd->sh_mem_ape1_limit); out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -1468,7 +1469,7 @@ static int process_termination_nocpsch(struct device_queue_manager *dqm, struct device_process_node *cur, *next_dpn; int retval = 0; - mutex_lock(&dqm->lock); + dqm_lock(dqm); /* Clear all user mode queues */ list_for_each_entry_safe(q, next, &qpd->queues_list, list) { @@ -1489,7 +1490,7 @@ static int process_termination_nocpsch(struct device_queue_manager *dqm, } } - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } @@ -1507,7 +1508,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, retval = 0; - mutex_lock(&dqm->lock); + dqm_lock(dqm); /* Clean all kernel queues */ list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) { @@ -1562,7 +1563,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, } out: - mutex_unlock(&dqm->lock); + dqm_unlock(dqm); return retval; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index 59a6b19..0a23dda 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -26,6 +26,8 @@ #include #include +#include +#include #include "kfd_priv.h" #include "kfd_mqd_manager.h" @@ -173,8 +175,9 @@ struct device_queue_manager { struct mqd_manager *mqds[KFD_MQD_TYPE_MAX]; struct packet_manager packets; struct kfd_dev *dev; - struct mutex lock; + struct mutex lock_hidden; /* use dqm_lock/unlock(dqm) */ struct list_head queues; + unsigned int saved_flags; unsigned int processes_count; unsigned int queue_count; unsigned int sdma_queue_count; @@ -219,4 +222,19 @@ get_sh_mem_bases_nybble_64(struct kfd_process_device *pdd) return (pdd->lds_base >> 60) & 0x0E; } +/* The DQM lock can be taken in MMU notifiers. Make sure no reclaim-FS + * happens while holding this lock anywhere to prevent deadlocks when + * an MMU notifier runs in reclaim-FS context. + */ +static inline void dqm_lock(struct device_queue_manager *dqm) +{ + mutex_lock(&dqm->lock_hidden); + dqm->saved_flags = memalloc_nofs_save(); +} +static inline void dqm_unlock(struct device_queue_manager *dqm) +{ + memalloc_nofs_restore(dqm->saved_flags); + mutex_unlock(&dqm->lock_hidden); +} + #endif /* KFD_DEVICE_QUEUE_MANAGER_H_ */ -- 2.7.4