Merge tag 'io_uring-6.9-20240405' of git://git.kernel.dk/linux

Pull io_uring fixes from Jens Axboe:

 - Backport of some fixes that came up during development of the 6.10
   io_uring patches. This includes some kbuf cleanups and reference
   fixes.

 - Disable multishot read if we don't have NOWAIT support on the target

 - Fix for a dependency issue with workqueue flushing

* tag 'io_uring-6.9-20240405' of git://git.kernel.dk/linux:
  io_uring/kbuf: hold io_buffer_list reference over mmap
  io_uring/kbuf: protect io_buffer_list teardown with a reference
  io_uring/kbuf: get rid of bl->is_ready
  io_uring/kbuf: get rid of lower BGID lists
  io_uring: use private workqueue for exit work
  io_uring: disable io-wq execution of multishot NOWAIT requests
  io_uring/rw: don't allow multishot reads without NOWAIT support

include/linux/io_uring_types.h

@@ -294,7 +294,6 @@ struct io_ring_ctx {
 
 		struct io_submit_state	submit_state;
 
-		struct io_buffer_list	*io_bl;
 		struct xarray		io_bl_xa;
 
 		struct io_hash_table	cancel_table_locked;

io_uring/io_uring.c

@@ -147,6 +147,7 @@ static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 static void io_queue_sqe(struct io_kiocb *req);
 
 struct kmem_cache *req_cachep;
+static struct workqueue_struct *iou_wq __ro_after_init;
 
 static int __read_mostly sysctl_io_uring_disabled;
 static int __read_mostly sysctl_io_uring_group = -1;
@@ -350,7 +351,6 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 err:
 	kfree(ctx->cancel_table.hbs);
 	kfree(ctx->cancel_table_locked.hbs);
-	kfree(ctx->io_bl);
 	xa_destroy(&ctx->io_bl_xa);
 	kfree(ctx);
 	return NULL;
@@ -1982,10 +1982,15 @@ void io_wq_submit_work(struct io_wq_work *work)
 		err = -EBADFD;
 		if (!io_file_can_poll(req))
 			goto fail;
-		err = -ECANCELED;
-		if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK)
-			goto fail;
-		return;
+		if (req->file->f_flags & O_NONBLOCK ||
+		    req->file->f_mode & FMODE_NOWAIT) {
+			err = -ECANCELED;
+			if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK)
+				goto fail;
+			return;
+		} else {
+			req->flags &= ~REQ_F_APOLL_MULTISHOT;
+		}
 	}
 
 	if (req->flags & REQ_F_FORCE_ASYNC) {
@@ -2926,7 +2931,6 @@ static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
 	io_napi_free(ctx);
 	kfree(ctx->cancel_table.hbs);
 	kfree(ctx->cancel_table_locked.hbs);
-	kfree(ctx->io_bl);
 	xa_destroy(&ctx->io_bl_xa);
 	kfree(ctx);
 }
@@ -3161,7 +3165,7 @@ static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
 	 * noise and overhead, there's no discernable change in runtime
 	 * over using system_wq.
 	 */
-	queue_work(system_unbound_wq, &ctx->exit_work);
+	queue_work(iou_wq, &ctx->exit_work);
 }
 
 static int io_uring_release(struct inode *inode, struct file *file)
@@ -3443,14 +3447,15 @@ static void *io_uring_validate_mmap_request(struct file *file,
 		ptr = ctx->sq_sqes;
 		break;
 	case IORING_OFF_PBUF_RING: {
+		struct io_buffer_list *bl;
 		unsigned int bgid;
 
 		bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
-		rcu_read_lock();
-		ptr = io_pbuf_get_address(ctx, bgid);
-		rcu_read_unlock();
-		if (!ptr)
-			return ERR_PTR(-EINVAL);
+		bl = io_pbuf_get_bl(ctx, bgid);
+		if (IS_ERR(bl))
+			return bl;
+		ptr = bl->buf_ring;
+		io_put_bl(ctx, bl);
 		break;
 		}
 	default:
@@ -4185,6 +4190,8 @@ static int __init io_uring_init(void)
 	io_buf_cachep = KMEM_CACHE(io_buffer,
 					  SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
 
+	iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64);
+
 #ifdef CONFIG_SYSCTL
 	register_sysctl_init("kernel", kernel_io_uring_disabled_table);
 #endif

io_uring/kbuf.c

@@ -17,8 +17,6 @@
 
 #define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
 
-#define BGID_ARRAY	64
-
 /* BIDs are addressed by a 16-bit field in a CQE */
 #define MAX_BIDS_PER_BGID (1 << 16)
 
@@ -40,13 +38,9 @@ struct io_buf_free {
 	int				inuse;
 };
 
-static struct io_buffer_list *__io_buffer_get_list(struct io_ring_ctx *ctx,
-						   struct io_buffer_list *bl,
-						   unsigned int bgid)
+static inline struct io_buffer_list *__io_buffer_get_list(struct io_ring_ctx *ctx,
+							  unsigned int bgid)
 {
-	if (bl && bgid < BGID_ARRAY)
-		return &bl[bgid];
-
 	return xa_load(&ctx->io_bl_xa, bgid);
 }
 
@@ -55,7 +49,7 @@ static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
 {
 	lockdep_assert_held(&ctx->uring_lock);
 
-	return __io_buffer_get_list(ctx, ctx->io_bl, bgid);
+	return __io_buffer_get_list(ctx, bgid);
 }
 
 static int io_buffer_add_list(struct io_ring_ctx *ctx,
@@ -67,11 +61,7 @@ static int io_buffer_add_list(struct io_ring_ctx *ctx,
 	 * always under the ->uring_lock, but the RCU lookup from mmap does.
 	 */
 	bl->bgid = bgid;
-	smp_store_release(&bl->is_ready, 1);
-
-	if (bgid < BGID_ARRAY)
-		return 0;
-
+	atomic_set(&bl->refs, 1);
 	return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
 }
 
@@ -208,24 +198,6 @@ void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 	return ret;
 }
 
-static __cold int io_init_bl_list(struct io_ring_ctx *ctx)
-{
-	struct io_buffer_list *bl;
-	int i;
-
-	bl = kcalloc(BGID_ARRAY, sizeof(struct io_buffer_list), GFP_KERNEL);
-	if (!bl)
-		return -ENOMEM;
-
-	for (i = 0; i < BGID_ARRAY; i++) {
-		INIT_LIST_HEAD(&bl[i].buf_list);
-		bl[i].bgid = i;
-	}
-
-	smp_store_release(&ctx->io_bl, bl);
-	return 0;
-}
-
 /*
  * Mark the given mapped range as free for reuse
  */
@@ -294,24 +266,24 @@ static int __io_remove_buffers(struct io_ring_ctx *ctx,
 	return i;
 }
 
+void io_put_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
+{
+	if (atomic_dec_and_test(&bl->refs)) {
+		__io_remove_buffers(ctx, bl, -1U);
+		kfree_rcu(bl, rcu);
+	}
+}
+
 void io_destroy_buffers(struct io_ring_ctx *ctx)
 {
 	struct io_buffer_list *bl;
 	struct list_head *item, *tmp;
 	struct io_buffer *buf;
 	unsigned long index;
-	int i;
-
-	for (i = 0; i < BGID_ARRAY; i++) {
-		if (!ctx->io_bl)
-			break;
-		__io_remove_buffers(ctx, &ctx->io_bl[i], -1U);
-	}
 
 	xa_for_each(&ctx->io_bl_xa, index, bl) {
 		xa_erase(&ctx->io_bl_xa, bl->bgid);
-		__io_remove_buffers(ctx, bl, -1U);
-		kfree_rcu(bl, rcu);
+		io_put_bl(ctx, bl);
 	}
 
 	/*
@@ -489,12 +461,6 @@ int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
 
 	io_ring_submit_lock(ctx, issue_flags);
 
-	if (unlikely(p->bgid < BGID_ARRAY && !ctx->io_bl)) {
-		ret = io_init_bl_list(ctx);
-		if (ret)
-			goto err;
-	}
-
 	bl = io_buffer_get_list(ctx, p->bgid);
 	if (unlikely(!bl)) {
 		bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
@@ -507,14 +473,9 @@ int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
 		if (ret) {
 			/*
 			 * Doesn't need rcu free as it was never visible, but
-			 * let's keep it consistent throughout. Also can't
-			 * be a lower indexed array group, as adding one
-			 * where lookup failed cannot happen.
+			 * let's keep it consistent throughout.
 			 */
-			if (p->bgid >= BGID_ARRAY)
-				kfree_rcu(bl, rcu);
-			else
-				WARN_ON_ONCE(1);
+			kfree_rcu(bl, rcu);
 			goto err;
 		}
 	}
@@ -679,12 +640,6 @@ int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
 	if (reg.ring_entries >= 65536)
 		return -EINVAL;
 
-	if (unlikely(reg.bgid < BGID_ARRAY && !ctx->io_bl)) {
-		int ret = io_init_bl_list(ctx);
-		if (ret)
-			return ret;
-	}
-
 	bl = io_buffer_get_list(ctx, reg.bgid);
 	if (bl) {
 		/* if mapped buffer ring OR classic exists, don't allow */
@@ -733,11 +688,8 @@ int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
 	if (!bl->is_buf_ring)
 		return -EINVAL;
 
-	__io_remove_buffers(ctx, bl, -1U);
-	if (bl->bgid >= BGID_ARRAY) {
-		xa_erase(&ctx->io_bl_xa, bl->bgid);
-		kfree_rcu(bl, rcu);
-	}
+	xa_erase(&ctx->io_bl_xa, bl->bgid);
+	io_put_bl(ctx, bl);
 	return 0;
 }
 
@@ -767,23 +719,35 @@ int io_register_pbuf_status(struct io_ring_ctx *ctx, void __user *arg)
 	return 0;
 }
 
-void *io_pbuf_get_address(struct io_ring_ctx *ctx, unsigned long bgid)
+struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
+				      unsigned long bgid)
 {
 	struct io_buffer_list *bl;
+	bool ret;
 
-	bl = __io_buffer_get_list(ctx, smp_load_acquire(&ctx->io_bl), bgid);
-
-	if (!bl || !bl->is_mmap)
-		return NULL;
 	/*
-	 * Ensure the list is fully setup. Only strictly needed for RCU lookup
-	 * via mmap, and in that case only for the array indexed groups. For
-	 * the xarray lookups, it's either visible and ready, or not at all.
+	 * We have to be a bit careful here - we're inside mmap and cannot grab
+	 * the uring_lock. This means the buffer_list could be simultaneously
+	 * going away, if someone is trying to be sneaky. Look it up under rcu
+	 * so we know it's not going away, and attempt to grab a reference to
+	 * it. If the ref is already zero, then fail the mapping. If successful,
+	 * the caller will call io_put_bl() to drop the the reference at at the
+	 * end. This may then safely free the buffer_list (and drop the pages)
+	 * at that point, vm_insert_pages() would've already grabbed the
+	 * necessary vma references.
 	 */
-	if (!smp_load_acquire(&bl->is_ready))
-		return NULL;
-
-	return bl->buf_ring;
+	rcu_read_lock();
+	bl = xa_load(&ctx->io_bl_xa, bgid);
+	/* must be a mmap'able buffer ring and have pages */
+	ret = false;
+	if (bl && bl->is_mmap)
+		ret = atomic_inc_not_zero(&bl->refs);
+	rcu_read_unlock();
+
+	if (ret)
+		return bl;
+
+	return ERR_PTR(-EINVAL);
 }
 
 /*

io_uring/kbuf.h

@@ -25,12 +25,12 @@ struct io_buffer_list {
 	__u16 head;
 	__u16 mask;
 
+	atomic_t refs;
+
 	/* ring mapped provided buffers */
 	__u8 is_buf_ring;
 	/* ring mapped provided buffers, but mmap'ed by application */
 	__u8 is_mmap;
-	/* bl is visible from an RCU point of view for lookup */
-	__u8 is_ready;
 };
 
 struct io_buffer {
@@ -61,7 +61,9 @@ void __io_put_kbuf(struct io_kiocb *req, unsigned issue_flags);
 
 bool io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags);
 
-void *io_pbuf_get_address(struct io_ring_ctx *ctx, unsigned long bgid);
+void io_put_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl);
+struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
+				      unsigned long bgid);
 
 static inline bool io_kbuf_recycle_ring(struct io_kiocb *req)
 {

io_uring/rw.c

@@ -936,6 +936,13 @@ int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
 
 	ret = __io_read(req, issue_flags);
 
+	/*
+	 * If the file doesn't support proper NOWAIT, then disable multishot
+	 * and stay in single shot mode.
+	 */
+	if (!io_file_supports_nowait(req))
+		req->flags &= ~REQ_F_APOLL_MULTISHOT;
+
 	/*
 	 * If we get -EAGAIN, recycle our buffer and just let normal poll
 	 * handling arm it.
@@ -955,7 +962,7 @@ int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
 	/*
 	 * Any successful return value will keep the multishot read armed.
 	 */
-	if (ret > 0) {
+	if (ret > 0 && req->flags & REQ_F_APOLL_MULTISHOT) {
 		/*
 		 * Put our buffer and post a CQE. If we fail to post a CQE, then
 		 * jump to the termination path. This request is then done.