[PATCHSET v3 next 0/4] Extend cancelation support

[PATCHSET v3 next 0/4] Extend cancelation support

[Jens Axboe]

Hi,

We currently only support looking up and canceling requests based on
the user_data of the original request. Sometimes it can be useful to
instead key off the fd used in the original request, eg if a socket
goes away.

Patch 1 is just a cleanup spotted while doing this, 2 is a prep patch,
patch 3 adds support for IORING_ASYNC_CANCEL_ALL, and finally patch 4
adds support for IORING_ASYNC_CANCEL_FD.

If IORING_ASYNC_CANCEL_ALL is set, all requests matching the given
criteria are canceled. Return value is number of requests canceled,
and 0 if none were found, or any error encountered canceling requests.

If IORING_ASYNC_CANCEL_FD is set, requests matching sqe->fd are
canceled rather than matching on sqe->addr for user_data.

v3:
- Fixup issues with CANCEL_ALL
- Combine two prep patches, and extent io_cancel_data propagation
  further.
- Get rid of buggy task_work running, we don't need it anymore.

-- 
Jens Axboe

[PATCH 1/4] io_uring: remove dead 'poll_only' argument to io_poll_cancel()

[Jens Axboe]

It's only called from one location, and it always passes in 'false'.
Kill the argument, and just pass in 'false' to io_poll_find().

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index d3fc0c5b4e82..878d30a31606 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -6316,11 +6316,10 @@ static bool io_poll_disarm(struct io_kiocb *req)
 	return true;
 }
 
-static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
-			  bool poll_only)
+static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
 	__must_hold(&ctx->completion_lock)
 {
-	struct io_kiocb *req = io_poll_find(ctx, sqe_addr, poll_only);
+	struct io_kiocb *req = io_poll_find(ctx, sqe_addr, false);
 
 	if (!req)
 		return -ENOENT;
@@ -6808,7 +6807,7 @@ static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
 		return 0;
 
 	spin_lock(&ctx->completion_lock);
-	ret = io_poll_cancel(ctx, sqe_addr, false);
+	ret = io_poll_cancel(ctx, sqe_addr);
 	if (ret != -ENOENT)
 		goto out;
 
-- 
2.35.1

[PATCH 2/4] io_uring: pass in struct io_cancel_data consistently

[Jens Axboe]

In preparation for being able to not only key cancel off the user_data,
pass in the io_cancel_data struct for the various functions that deal
with request cancelation.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 76 +++++++++++++++++++++++++++++----------------------
 1 file changed, 44 insertions(+), 32 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 878d30a31606..16f8e1f7dcae 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -971,6 +971,11 @@ struct io_defer_entry {
 	u32			seq;
 };
 
+struct io_cancel_data {
+	struct io_ring_ctx *ctx;
+	u64 data;
+};
+
 struct io_op_def {
 	/* needs req->file assigned */
 	unsigned		needs_file : 1;
@@ -6288,16 +6293,16 @@ static __cold bool io_poll_remove_all(struct io_ring_ctx *ctx,
 	return found;
 }
 
-static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
-				     bool poll_only)
+static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
+				     struct io_cancel_data *cd)
 	__must_hold(&ctx->completion_lock)
 {
 	struct hlist_head *list;
 	struct io_kiocb *req;
 
-	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
+	list = &ctx->cancel_hash[hash_long(cd->data, ctx->cancel_hash_bits)];
 	hlist_for_each_entry(req, list, hash_node) {
-		if (sqe_addr != req->cqe.user_data)
+		if (cd->data != req->cqe.user_data)
 			continue;
 		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
 			continue;
@@ -6316,10 +6321,10 @@ static bool io_poll_disarm(struct io_kiocb *req)
 	return true;
 }
 
-static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
+static int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
 	__must_hold(&ctx->completion_lock)
 {
-	struct io_kiocb *req = io_poll_find(ctx, sqe_addr, false);
+	struct io_kiocb *req = io_poll_find(ctx, false, cd);
 
 	if (!req)
 		return -ENOENT;
@@ -6411,13 +6416,14 @@ static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
 
 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
 {
+	struct io_cancel_data cd = { .data = req->poll_update.old_user_data, };
 	struct io_ring_ctx *ctx = req->ctx;
 	struct io_kiocb *preq;
 	int ret2, ret = 0;
 	bool locked;
 
 	spin_lock(&ctx->completion_lock);
-	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
+	preq = io_poll_find(ctx, true, &cd);
 	if (!preq || !io_poll_disarm(preq)) {
 		spin_unlock(&ctx->completion_lock);
 		ret = preq ? -EALREADY : -ENOENT;
@@ -6477,7 +6483,7 @@ static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
 }
 
 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
-					   __u64 user_data)
+					   struct io_cancel_data *cd)
 	__must_hold(&ctx->timeout_lock)
 {
 	struct io_timeout_data *io;
@@ -6485,7 +6491,7 @@ static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
 	bool found = false;
 
 	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
-		found = user_data == req->cqe.user_data;
+		found = cd->data == req->cqe.user_data;
 		if (found)
 			break;
 	}
@@ -6499,11 +6505,11 @@ static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
 	return req;
 }
 
-static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
+static int io_timeout_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
 	__must_hold(&ctx->completion_lock)
 	__must_hold(&ctx->timeout_lock)
 {
-	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
+	struct io_kiocb *req = io_timeout_extract(ctx, cd);
 
 	if (IS_ERR(req))
 		return PTR_ERR(req);
@@ -6556,7 +6562,8 @@ static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
 			     struct timespec64 *ts, enum hrtimer_mode mode)
 	__must_hold(&ctx->timeout_lock)
 {
-	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
+	struct io_cancel_data cd = { .data = user_data, };
+	struct io_kiocb *req = io_timeout_extract(ctx, &cd);
 	struct io_timeout_data *data;
 
 	if (IS_ERR(req))
@@ -6621,9 +6628,11 @@ static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
 	int ret;
 
 	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
+		struct io_cancel_data cd = { .data = tr->addr, };
+
 		spin_lock(&ctx->completion_lock);
 		spin_lock_irq(&ctx->timeout_lock);
-		ret = io_timeout_cancel(ctx, tr->addr);
+		ret = io_timeout_cancel(ctx, &cd);
 		spin_unlock_irq(&ctx->timeout_lock);
 		spin_unlock(&ctx->completion_lock);
 	} else {
@@ -6752,30 +6761,24 @@ static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
 	return 0;
 }
 
-struct io_cancel_data {
-	struct io_ring_ctx *ctx;
-	u64 user_data;
-};
-
 static bool io_cancel_cb(struct io_wq_work *work, void *data)
 {
 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
 	struct io_cancel_data *cd = data;
 
-	return req->ctx == cd->ctx && req->cqe.user_data == cd->user_data;
+	return req->ctx == cd->ctx && req->cqe.user_data == cd->data;
 }
 
-static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
-			       struct io_ring_ctx *ctx)
+static int io_async_cancel_one(struct io_uring_task *tctx,
+			       struct io_cancel_data *cd)
 {
-	struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
 	enum io_wq_cancel cancel_ret;
 	int ret = 0;
 
 	if (!tctx || !tctx->io_wq)
 		return -ENOENT;
 
-	cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
+	cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, cd, false);
 	switch (cancel_ret) {
 	case IO_WQ_CANCEL_OK:
 		ret = 0;
@@ -6791,14 +6794,14 @@ static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
 	return ret;
 }
 
-static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
+static int io_try_cancel(struct io_kiocb *req, struct io_cancel_data *cd)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 	int ret;
 
 	WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
 
-	ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
+	ret = io_async_cancel_one(req->task->io_uring, cd);
 	/*
 	 * Fall-through even for -EALREADY, as we may have poll armed
 	 * that need unarming.
@@ -6807,12 +6810,12 @@ static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
 		return 0;
 
 	spin_lock(&ctx->completion_lock);
-	ret = io_poll_cancel(ctx, sqe_addr);
+	ret = io_poll_cancel(ctx, cd);
 	if (ret != -ENOENT)
 		goto out;
 
 	spin_lock_irq(&ctx->timeout_lock);
-	ret = io_timeout_cancel(ctx, sqe_addr);
+	ret = io_timeout_cancel(ctx, cd);
 	spin_unlock_irq(&ctx->timeout_lock);
 out:
 	spin_unlock(&ctx->completion_lock);
@@ -6837,11 +6840,14 @@ static int io_async_cancel_prep(struct io_kiocb *req,
 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
 {
 	struct io_ring_ctx *ctx = req->ctx;
-	u64 sqe_addr = req->cancel.addr;
+	struct io_cancel_data cd = {
+		.ctx		= ctx,
+		.data		= req->cancel.addr,
+	};
 	struct io_tctx_node *node;
 	int ret;
 
-	ret = io_try_cancel_userdata(req, sqe_addr);
+	ret = io_try_cancel(req, &cd);
 	if (ret != -ENOENT)
 		goto done;
 
@@ -6851,7 +6857,7 @@ static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
 	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
 		struct io_uring_task *tctx = node->task->io_uring;
 
-		ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
+		ret = io_async_cancel_one(tctx, &cd);
 		if (ret != -ENOENT)
 			break;
 	}
@@ -7455,8 +7461,14 @@ static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
 	int ret = -ENOENT;
 
 	if (prev) {
-		if (!(req->task->flags & PF_EXITING))
-			ret = io_try_cancel_userdata(req, prev->cqe.user_data);
+		if (!(req->task->flags & PF_EXITING)) {
+			struct io_cancel_data cd = {
+				.ctx		= req->ctx,
+				.data		= prev->cqe.user_data,
+			};
+
+			ret = io_try_cancel(req, &cd);
+		}
 		io_req_complete_post(req, ret ?: -ETIME, 0);
 		io_put_req(prev);
 	} else {
-- 
2.35.1

[PATCH 3/4] io_uring: add support for IORING_ASYNC_CANCEL_ALL

[Jens Axboe]

The current cancelation will lookup and cancel the first request it
finds based on the key passed in. Add a flag that allows to cancel any
request that matches they key. It completes with the number of requests
found and canceled, or res < 0 if an error occured.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io-wq.h                    |  1 +
 fs/io_uring.c                 | 91 +++++++++++++++++++++++++++--------
 include/uapi/linux/io_uring.h |  7 +++
 3 files changed, 78 insertions(+), 21 deletions(-)

diff --git a/fs/io-wq.h b/fs/io-wq.h
index dbecd27656c7..ba6eee76d028 100644
--- a/fs/io-wq.h
+++ b/fs/io-wq.h
@@ -155,6 +155,7 @@ struct io_wq_work_node *wq_stack_extract(struct io_wq_work_node *stack)
 struct io_wq_work {
 	struct io_wq_work_node list;
 	unsigned flags;
+	int cancel_seq;
 };
 
 static inline struct io_wq_work *wq_next_work(struct io_wq_work *work)
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 16f8e1f7dcae..79601a333903 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -401,6 +401,7 @@ struct io_ring_ctx {
 		struct xarray		personalities;
 		u32			pers_next;
 		unsigned		sq_thread_idle;
+		atomic_t		cancel_seq;
 	} ____cacheline_aligned_in_smp;
 
 	/* IRQ completion list, under ->completion_lock */
@@ -568,6 +569,7 @@ struct io_sync {
 struct io_cancel {
 	struct file			*file;
 	u64				addr;
+	u32				flags;
 };
 
 struct io_timeout {
@@ -974,6 +976,8 @@ struct io_defer_entry {
 struct io_cancel_data {
 	struct io_ring_ctx *ctx;
 	u64 data;
+	u32 flags;
+	int seq;
 };
 
 struct io_op_def {
@@ -1694,6 +1698,7 @@ static void io_prep_async_work(struct io_kiocb *req)
 
 	req->work.list.next = NULL;
 	req->work.flags = 0;
+	req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
 	if (req->flags & REQ_F_FORCE_ASYNC)
 		req->work.flags |= IO_WQ_WORK_CONCURRENT;
 
@@ -6149,6 +6154,7 @@ static int __io_arm_poll_handler(struct io_kiocb *req,
 	int v;
 
 	INIT_HLIST_NODE(&req->hash_node);
+	req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
 	io_init_poll_iocb(poll, mask, io_poll_wake);
 	poll->file = req->file;
 
@@ -6306,6 +6312,11 @@ static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
 			continue;
 		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
 			continue;
+		if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
+			if (cd->seq == req->work.cancel_seq)
+				continue;
+			req->work.cancel_seq = cd->seq;
+		}
 		return req;
 	}
 	return NULL;
@@ -6491,9 +6502,15 @@ static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
 	bool found = false;
 
 	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
-		found = cd->data == req->cqe.user_data;
-		if (found)
-			break;
+		if (cd->data != req->cqe.user_data)
+			continue;
+		if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
+			if (cd->seq == req->work.cancel_seq)
+				continue;
+			req->work.cancel_seq = cd->seq;
+		}
+		found = true;
+		break;
 	}
 	if (!found)
 		return ERR_PTR(-ENOENT);
@@ -6766,7 +6783,16 @@ static bool io_cancel_cb(struct io_wq_work *work, void *data)
 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
 	struct io_cancel_data *cd = data;
 
-	return req->ctx == cd->ctx && req->cqe.user_data == cd->data;
+	if (req->ctx != cd->ctx)
+		return false;
+	if (req->cqe.user_data != cd->data)
+		return false;
+	if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
+		if (cd->seq == req->work.cancel_seq)
+			return false;
+		req->work.cancel_seq = cd->seq;
+	}
+	return true;
 }
 
 static int io_async_cancel_one(struct io_uring_task *tctx,
@@ -6778,7 +6804,8 @@ static int io_async_cancel_one(struct io_uring_task *tctx,
 	if (!tctx || !tctx->io_wq)
 		return -ENOENT;
 
-	cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, cd, false);
+	cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, cd,
+					cd->flags & IORING_ASYNC_CANCEL_ALL);
 	switch (cancel_ret) {
 	case IO_WQ_CANCEL_OK:
 		ret = 0;
@@ -6829,27 +6856,33 @@ static int io_async_cancel_prep(struct io_kiocb *req,
 		return -EINVAL;
 	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
 		return -EINVAL;
-	if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
-	    sqe->splice_fd_in)
+	if (sqe->ioprio || sqe->off || sqe->len || sqe->splice_fd_in)
 		return -EINVAL;
 
 	req->cancel.addr = READ_ONCE(sqe->addr);
+	req->cancel.flags = READ_ONCE(sqe->cancel_flags);
+	if (req->cancel.flags & ~IORING_ASYNC_CANCEL_ALL)
+		return -EINVAL;
+
 	return 0;
 }
 
-static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
+static int __io_async_cancel(struct io_cancel_data *cd, struct io_kiocb *req,
+			     unsigned int issue_flags)
 {
-	struct io_ring_ctx *ctx = req->ctx;
-	struct io_cancel_data cd = {
-		.ctx		= ctx,
-		.data		= req->cancel.addr,
-	};
+	bool cancel_all = cd->flags & IORING_ASYNC_CANCEL_ALL;
+	struct io_ring_ctx *ctx = cd->ctx;
 	struct io_tctx_node *node;
-	int ret;
+	int ret, nr = 0;
 
-	ret = io_try_cancel(req, &cd);
-	if (ret != -ENOENT)
-		goto done;
+	do {
+		ret = io_try_cancel(req, cd);
+		if (ret == -ENOENT)
+			break;
+		if (!cancel_all)
+			return ret;
+		nr++;
+	} while (1);
 
 	/* slow path, try all io-wq's */
 	io_ring_submit_lock(ctx, issue_flags);
@@ -6857,12 +6890,28 @@ static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
 	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
 		struct io_uring_task *tctx = node->task->io_uring;
 
-		ret = io_async_cancel_one(tctx, &cd);
-		if (ret != -ENOENT)
-			break;
+		ret = io_async_cancel_one(tctx, cd);
+		if (ret != -ENOENT) {
+			if (!cancel_all)
+				break;
+			nr++;
+		}
 	}
 	io_ring_submit_unlock(ctx, issue_flags);
-done:
+	return cancel_all ? nr : ret;
+}
+
+static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
+{
+	struct io_cancel_data cd = {
+		.ctx	= req->ctx,
+		.data	= req->cancel.addr,
+		.flags	= req->cancel.flags,
+		.seq	= atomic_inc_return(&req->ctx->cancel_seq),
+	};
+	int ret;
+
+	ret = __io_async_cancel(&cd, req, issue_flags);
 	if (ret < 0)
 		req_set_fail(req);
 	io_req_complete_post(req, ret, 0);
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 1845cf7c80ba..476e58a2837f 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -187,6 +187,13 @@ enum {
 #define IORING_POLL_UPDATE_EVENTS	(1U << 1)
 #define IORING_POLL_UPDATE_USER_DATA	(1U << 2)
 
+/*
+ * ASYNC_CANCEL flags.
+ *
+ * IORING_ASYNC_CANCEL_ALL	Cancel all requests that match the given key
+ */
+#define IORING_ASYNC_CANCEL_ALL	(1U << 0)
+
 /*
  * IO completion data structure (Completion Queue Entry)
  */
-- 
2.35.1

[PATCH 4/4] io_uring: allow IORING_OP_ASYNC_CANCEL with 'fd' key

[Jens Axboe]

Currently sqe->addr must contain the user_data of the request being
canceled. Introduce the IORING_ASYNC_CANCEL_FD flag, which tells the
kernel that we're keying off the file fd instead for cancelation. This
allows canceling any request that a) uses a file, and b) was assigned the
file based on the value being passed in.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 71 ++++++++++++++++++++++++++++++-----
 include/uapi/linux/io_uring.h |  3 ++
 2 files changed, 65 insertions(+), 9 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 79601a333903..07acf7199fa7 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -570,6 +570,7 @@ struct io_cancel {
 	struct file			*file;
 	u64				addr;
 	u32				flags;
+	s32				fd;
 };
 
 struct io_timeout {
@@ -975,7 +976,10 @@ struct io_defer_entry {
 
 struct io_cancel_data {
 	struct io_ring_ctx *ctx;
-	u64 data;
+	union {
+		u64 data;
+		struct file *file;
+	};
 	u32 flags;
 	int seq;
 };
@@ -6322,6 +6326,29 @@ static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
 	return NULL;
 }
 
+static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
+					  struct io_cancel_data *cd)
+	__must_hold(&ctx->completion_lock)
+{
+	struct io_kiocb *req;
+	int i;
+
+	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
+		struct hlist_head *list;
+
+		list = &ctx->cancel_hash[i];
+		hlist_for_each_entry(req, list, hash_node) {
+			if (req->file != cd->file)
+				continue;
+			if (cd->seq == req->work.cancel_seq)
+				continue;
+			req->work.cancel_seq = cd->seq;
+			return req;
+		}
+	}
+	return NULL;
+}
+
 static bool io_poll_disarm(struct io_kiocb *req)
 	__must_hold(&ctx->completion_lock)
 {
@@ -6335,8 +6362,12 @@ static bool io_poll_disarm(struct io_kiocb *req)
 static int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
 	__must_hold(&ctx->completion_lock)
 {
-	struct io_kiocb *req = io_poll_find(ctx, false, cd);
+	struct io_kiocb *req;
 
+	if (cd->flags & IORING_ASYNC_CANCEL_FD)
+		req = io_poll_file_find(ctx, cd);
+	else
+		req = io_poll_find(ctx, false, cd);
 	if (!req)
 		return -ENOENT;
 	io_poll_cancel_req(req);
@@ -6785,8 +6816,13 @@ static bool io_cancel_cb(struct io_wq_work *work, void *data)
 
 	if (req->ctx != cd->ctx)
 		return false;
-	if (req->cqe.user_data != cd->data)
-		return false;
+	if (cd->flags & IORING_ASYNC_CANCEL_FD) {
+		if (req->file != cd->file)
+			return false;
+	} else {
+		if (req->cqe.user_data != cd->data)
+			return false;
+	}
 	if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
 		if (cd->seq == req->work.cancel_seq)
 			return false;
@@ -6841,9 +6877,11 @@ static int io_try_cancel(struct io_kiocb *req, struct io_cancel_data *cd)
 	if (ret != -ENOENT)
 		goto out;
 
-	spin_lock_irq(&ctx->timeout_lock);
-	ret = io_timeout_cancel(ctx, cd);
-	spin_unlock_irq(&ctx->timeout_lock);
+	if (!(cd->flags & IORING_ASYNC_CANCEL_FD)) {
+		spin_lock_irq(&ctx->timeout_lock);
+		ret = io_timeout_cancel(ctx, cd);
+		spin_unlock_irq(&ctx->timeout_lock);
+	}
 out:
 	spin_unlock(&ctx->completion_lock);
 	return ret;
@@ -6854,15 +6892,17 @@ static int io_async_cancel_prep(struct io_kiocb *req,
 {
 	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
 		return -EINVAL;
-	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
+	if (unlikely(req->flags & REQ_F_BUFFER_SELECT))
 		return -EINVAL;
 	if (sqe->ioprio || sqe->off || sqe->len || sqe->splice_fd_in)
 		return -EINVAL;
 
 	req->cancel.addr = READ_ONCE(sqe->addr);
 	req->cancel.flags = READ_ONCE(sqe->cancel_flags);
-	if (req->cancel.flags & ~IORING_ASYNC_CANCEL_ALL)
+	if (req->cancel.flags & ~(IORING_ASYNC_CANCEL_ALL|IORING_ASYNC_CANCEL_FD))
 		return -EINVAL;
+	if (req->cancel.flags & IORING_ASYNC_CANCEL_FD)
+		req->cancel.fd = READ_ONCE(sqe->fd);
 
 	return 0;
 }
@@ -6911,7 +6951,20 @@ static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
 	};
 	int ret;
 
+	if (cd.flags & IORING_ASYNC_CANCEL_FD) {
+		if (req->flags & REQ_F_FIXED_FILE)
+			req->file = io_file_get_fixed(req, req->cancel.fd, issue_flags);
+		else
+			req->file = io_file_get_normal(req, req->cancel.fd);
+		if (!req->file) {
+			ret = -EBADF;
+			goto done;
+		}
+		cd.file = req->file;
+	}
+
 	ret = __io_async_cancel(&cd, req, issue_flags);
+done:
 	if (ret < 0)
 		req_set_fail(req);
 	io_req_complete_post(req, ret, 0);
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 476e58a2837f..cc7fe82a1798 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -191,8 +191,11 @@ enum {
  * ASYNC_CANCEL flags.
  *
  * IORING_ASYNC_CANCEL_ALL	Cancel all requests that match the given key
+ * IORING_ASYNC_CANCEL_FD	Key off 'fd' for cancelation rather than the
+ *				request 'user_data'
  */
 #define IORING_ASYNC_CANCEL_ALL	(1U << 0)
+#define IORING_ASYNC_CANCEL_FD	(1U << 1)
 
 /*
  * IO completion data structure (Completion Queue Entry)
-- 
2.35.1

[PATCHSET next 0/4] Allow cancelation based on fd

[Jens Axboe]

Hi,

We currently only support looking up and canceling requests based on
the user_data of the original request. Sometimes it can be useful to
instead key off the fd used in the original request, eg if a socket
goes away.

Patch 1 is just a cleanup spotted while doing this, 2+3 are prep patches,
and patch 4 then finally is the meat of it.

Adds IORING_ASYNC_CANCEL_FD as a IORING_OP_ASYNC_CANCEL request flag,
which tells the kernel that we're looking up by fd rather than user_data.

-- 
Jens Axboe

[PATCH 4/4] io_uring: allow IORING_OP_ASYNC_CANCEL with 'fd' key

[Jens Axboe]

Currently sqe->addr must contain the user_data of the request being
canceled. Introduce the IORING_ASYNC_CANCEL_FD flag, which tells the
kernel that we're keying off the file fd instead for cancelation. This
allows canceling any request that a) uses a file, and b) was assigned the
file based on the value being passed in.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 58 +++++++++++++++++++++++++++++------
 include/uapi/linux/io_uring.h |  8 +++++
 2 files changed, 56 insertions(+), 10 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 0ef8401b6552..c86a92a975b7 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -567,7 +567,8 @@ struct io_sync {
 
 struct io_cancel {
 	struct file			*file;
-	u64				addr;
+	u64				data;
+	u32				flags;
 };
 
 struct io_timeout {
@@ -6306,6 +6307,25 @@ static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
 	return NULL;
 }
 
+static struct io_kiocb *io_poll_fd_find(struct io_ring_ctx *ctx, int fd)
+	__must_hold(&ctx->completion_lock)
+{
+	struct io_kiocb *req;
+	int i;
+
+	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
+		struct hlist_head *list;
+
+		list = &ctx->cancel_hash[i];
+		hlist_for_each_entry(req, list, hash_node) {
+			if (!req->file || fd != req->cqe.fd)
+				continue;
+			return req;
+		}
+	}
+	return NULL;
+}
+
 static bool io_poll_disarm(struct io_kiocb *req)
 	__must_hold(&ctx->completion_lock)
 {
@@ -6319,13 +6339,18 @@ static bool io_poll_disarm(struct io_kiocb *req)
 struct io_cancel_data {
 	struct io_ring_ctx *ctx;
 	u64 data;
+	unsigned int flags;
 };
 
 static int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
 	__must_hold(&ctx->completion_lock)
 {
-	struct io_kiocb *req = io_poll_find(ctx, cd->data, false);
+	struct io_kiocb *req;
 
+	if (cd->flags & IORING_ASYNC_CANCEL_FD)
+		req = io_poll_fd_find(ctx, cd->data);
+	else
+		req = io_poll_find(ctx, cd->data, false);
 	if (!req)
 		return -ENOENT;
 	io_poll_cancel_req(req);
@@ -6762,7 +6787,11 @@ static bool io_cancel_cb(struct io_wq_work *work, void *data)
 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
 	struct io_cancel_data *cd = data;
 
-	return req->ctx == cd->ctx && req->cqe.user_data == cd->data;
+	if (req->ctx != cd->ctx)
+		return false;
+	if (cd->flags & IORING_ASYNC_CANCEL_FD)
+		return req->file && cd->data == req->cqe.fd;
+	return req->cqe.user_data == cd->data;
 }
 
 static int io_async_cancel_one(struct io_uring_task *tctx,
@@ -6811,9 +6840,11 @@ static int io_try_cancel_userdata(struct io_kiocb *req,
 	if (ret != -ENOENT)
 		goto out;
 
-	spin_lock_irq(&ctx->timeout_lock);
-	ret = io_timeout_cancel(ctx, cd->data);
-	spin_unlock_irq(&ctx->timeout_lock);
+	if (!(cd->flags & IORING_ASYNC_CANCEL_FD)) {
+		spin_lock_irq(&ctx->timeout_lock);
+		ret = io_timeout_cancel(ctx, cd->data);
+		spin_unlock_irq(&ctx->timeout_lock);
+	}
 out:
 	spin_unlock(&ctx->completion_lock);
 	return ret;
@@ -6826,11 +6857,17 @@ static int io_async_cancel_prep(struct io_kiocb *req,
 		return -EINVAL;
 	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
 		return -EINVAL;
-	if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
-	    sqe->splice_fd_in)
+	if (sqe->ioprio || sqe->off || sqe->len || sqe->splice_fd_in)
+		return -EINVAL;
+
+	req->cancel.data = READ_ONCE(sqe->addr);
+	req->cancel.flags = READ_ONCE(sqe->cancel_flags);
+	if (req->cancel.flags & ~IORING_ASYNC_CANCEL_FD)
+		return -EINVAL;
+	else if ((req->cancel.flags & IORING_ASYNC_CANCEL_FD) &&
+		 req->cancel.data > INT_MAX)
 		return -EINVAL;
 
-	req->cancel.addr = READ_ONCE(sqe->addr);
 	return 0;
 }
 
@@ -6839,7 +6876,8 @@ static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
 	struct io_ring_ctx *ctx = req->ctx;
 	struct io_cancel_data cd = {
 		.ctx	= ctx,
-		.data	= req->cancel.addr,
+		.data	= req->cancel.data,
+		.flags	= req->cancel.flags,
 	};
 	struct io_tctx_node *node;
 	int ret;
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 1845cf7c80ba..806c473dde9f 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -187,6 +187,14 @@ enum {
 #define IORING_POLL_UPDATE_EVENTS	(1U << 1)
 #define IORING_POLL_UPDATE_USER_DATA	(1U << 2)
 
+/*
+ * ASYNC_CANCEL flags.
+ *
+ * IORING_ASYNC_CANCEL_FD	Key off 'fd' for cancelation rather than the
+ *				request 'user_data'
+ */
+#define IORING_ASYNC_CANCEL_FD	(1U << 0)
+
 /*
  * IO completion data structure (Completion Queue Entry)
  */
-- 
2.35.1