Use xa_insert_irq() to do the allocation before grabbing the other locks. This user appears to be able to race, so use xa_cmpxchg() to handle the race effectively. Signed-off-by: Matthew Wilcox <willy@xxxxxxxxxxxxx> --- block/blk-ioc.c | 23 +++++++++++++---------- include/linux/iocontext.h | 6 +++--- 2 files changed, 16 insertions(+), 13 deletions(-) diff --git a/block/blk-ioc.c b/block/blk-ioc.c index 5ed59ac6ae58..1db53c371b14 100644 --- a/block/blk-ioc.c +++ b/block/blk-ioc.c @@ -65,7 +65,7 @@ static void ioc_destroy_icq(struct io_cq *icq) lockdep_assert_held(&ioc->lock); - radix_tree_delete(&ioc->icq_tree, icq->q->id); + xa_erase(&ioc->icq_array, icq->q->id); hlist_del_init(&icq->ioc_node); list_del_init(&icq->q_node); @@ -255,7 +255,7 @@ int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node) atomic_set(&ioc->nr_tasks, 1); atomic_set(&ioc->active_ref, 1); spin_lock_init(&ioc->lock); - INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC); + xa_init_flags(&ioc->icq_array, XA_FLAGS_LOCK_IRQ); INIT_HLIST_HEAD(&ioc->icq_list); INIT_WORK(&ioc->release_work, ioc_release_fn); @@ -339,7 +339,7 @@ struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q) if (icq && icq->q == q) goto out; - icq = radix_tree_lookup(&ioc->icq_tree, q->id); + icq = xa_load(&ioc->icq_array, q->id); if (icq && icq->q == q) rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */ else @@ -366,7 +366,7 @@ struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q, gfp_t gfp_mask) { struct elevator_type *et = q->elevator->type; - struct io_cq *icq; + struct io_cq *icq, *curr; /* allocate stuff */ icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO, @@ -374,10 +374,14 @@ struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q, if (!icq) return NULL; - if (radix_tree_maybe_preload(gfp_mask) < 0) { + if (xa_insert_irq(&ioc->icq_array, q->id, NULL, gfp_mask) == -ENOMEM) { kmem_cache_free(et->icq_cache, icq); return NULL; } + /* + * If we get -EBUSY, we're racing with another caller; we'll see + * who wins the race below. + */ icq->ioc = ioc; icq->q = q; @@ -388,21 +392,20 @@ struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q, spin_lock_irq(&q->queue_lock); spin_lock(&ioc->lock); - if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) { + curr = xa_cmpxchg(&ioc->icq_array, q->id, XA_ZERO_ENTRY, icq, + GFP_ATOMIC); + if (likely(!curr)) { hlist_add_head(&icq->ioc_node, &ioc->icq_list); list_add(&icq->q_node, &q->icq_list); if (et->ops.init_icq) et->ops.init_icq(icq); } else { kmem_cache_free(et->icq_cache, icq); - icq = ioc_lookup_icq(ioc, q); - if (!icq) - printk(KERN_ERR "cfq: icq link failed!\n"); + icq = curr; } spin_unlock(&ioc->lock); spin_unlock_irq(&q->queue_lock); - radix_tree_preload_end(); return icq; } diff --git a/include/linux/iocontext.h b/include/linux/iocontext.h index dba15ca8e60b..e16224f70084 100644 --- a/include/linux/iocontext.h +++ b/include/linux/iocontext.h @@ -2,9 +2,9 @@ #ifndef IOCONTEXT_H #define IOCONTEXT_H -#include <linux/radix-tree.h> #include <linux/rcupdate.h> #include <linux/workqueue.h> +#include <linux/xarray.h> enum { ICQ_EXITED = 1 << 2, @@ -56,7 +56,7 @@ enum { * - ioc->icq_list and icq->ioc_node are protected by ioc lock. * q->icq_list and icq->q_node by q lock. * - * - ioc->icq_tree and ioc->icq_hint are protected by ioc lock, while icq + * - ioc->icq_array and ioc->icq_hint are protected by ioc lock, while icq * itself is protected by q lock. However, both the indexes and icq * itself are also RCU managed and lookup can be performed holding only * the q lock. @@ -111,7 +111,7 @@ struct io_context { int nr_batch_requests; /* Number of requests left in the batch */ unsigned long last_waited; /* Time last woken after wait for request */ - struct radix_tree_root icq_tree; + struct xarray icq_array; struct io_cq __rcu *icq_hint; struct hlist_head icq_list; -- 2.20.1