Skip to contentSkip to conversation

io_uring: fix short read/write with linked ops

PATCH In review David Stevens 2 participants 3 messages Oct 3, 2022 – Oct 4, 2022 io_uring

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 io_uring: fix short read/write with linked ops

From: David Stevens <[email protected]>

When continuing a short read/write, account for any completed bytes when calculating the operation's target length. The operation's actual accumulated length is fixed up by kiocb_done, and the target and actual lengths are then compared by __io_complete_rw_common. That function already propagated the actual length to userspace, but the incorrect target length was causing it to always cancel linked operations, even with a successfully completed read/write.

Fixes 227c0c9673d8 ("io_uring: internally retry short reads")
---
 io_uring/rw.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
io_uring/rw.c +8 -4
index 76ebcfebc9a6..aa9967a52dfd 100644
@@ -706,13 +706,14 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
 	struct kiocb *kiocb = &rw->kiocb;
 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
 	struct io_async_rw *io;
-	ssize_t ret, ret2;
+	ssize_t ret, ret2, target_len;
 	loff_t *ppos;
 
 	if (!req_has_async_data(req)) {
 		ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
 		if (unlikely(ret < 0))
 			return ret;
+		target_len = iov_iter_count(&s->iter);
 	} else {
 		io = req->async_data;
 		s = &io->s;
@@ -733,6 +734,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
 		 * need to make this conditional.
 		 */
 		iov_iter_restore(&s->iter, &s->iter_state);
+		target_len = iov_iter_count(&s->iter) + io->bytes_done;
 		iovec = NULL;
 	}
 	ret = io_rw_init_file(req, FMODE_READ);
@@ -740,7 +742,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
 		kfree(iovec);
 		return ret;
 	}
-	req->cqe.res = iov_iter_count(&s->iter);
+	req->cqe.res = target_len;
 
 	if (force_nonblock) {
 		/* If the file doesn't support async, just async punt */
@@ -850,18 +852,20 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
 	struct iovec *iovec;
 	struct kiocb *kiocb = &rw->kiocb;
 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
-	ssize_t ret, ret2;
+	ssize_t ret, ret2, target_len;
 	loff_t *ppos;
 
 	if (!req_has_async_data(req)) {
 		ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
 		if (unlikely(ret < 0))
 			return ret;
+		target_len = iov_iter_count(&s->iter);
 	} else {
 		struct io_async_rw *io = req->async_data;
 
 		s = &io->s;
 		iov_iter_restore(&s->iter, &s->iter_state);
+		target_len = iov_iter_count(&s->iter) + io->bytes_done;
 		iovec = NULL;
 	}
 	ret = io_rw_init_file(req, FMODE_WRITE);
@@ -869,7 +873,7 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
 		kfree(iovec);
 		return ret;
 	}
-	req->cqe.res = iov_iter_count(&s->iter);
+	req->cqe.res = target_len;
 
 	if (force_nonblock) {
 		/* If the file doesn't support async, just async punt */
Signature · 1 line
2.38.0.rc1.362.ged0d419d3c-goog
#1 lore raw
Quoted 9 lines · David Stevens

From: David Stevens <[email protected]>

When continuing a short read/write, account for any completed bytes when calculating the operation's target length. The operation's actual accumulated length is fixed up by kiocb_done, and the target and actual lengths are then compared by __io_complete_rw_common. That function already propagated the actual length to userspace, but the incorrect target length was causing it to always cancel linked operations, even with a successfully completed read/write.

The issue looks same as fixed with

https://git.kernel.dk/cgit/linux-block/commit/?h=for-6.1/io_uring&id=bf68b5b34311ee57ed40749a1257a30b46127556

Can you check if for-6.1 works for you?

git://git.kernel.dk/linux.git for-6.1/io_uring
https://git.kernel.dk/cgit/linux-block/log/?h=for-6.1/io_uring

Quoted 74 lines
Fixes 227c0c9673d8 ("io_uring: internally retry short reads")
---
  io_uring/rw.c | 12 ++++++++----
  1 file changed, 8 insertions(+), 4 deletions(-)
io_uring/rw.c +8 -4
index 76ebcfebc9a6..aa9967a52dfd 100644
@@ -706,13 +706,14 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
  	struct kiocb *kiocb = &rw->kiocb;
  	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
  	struct io_async_rw *io;
-	ssize_t ret, ret2;
+	ssize_t ret, ret2, target_len;
  	loff_t *ppos;
 
  	if (!req_has_async_data(req)) {
  		ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
  		if (unlikely(ret < 0))
  			return ret;
+		target_len = iov_iter_count(&s->iter);
  	} else {
  		io = req->async_data;
  		s = &io->s;
@@ -733,6 +734,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
  		 * need to make this conditional.
  		 */
  		iov_iter_restore(&s->iter, &s->iter_state);
+		target_len = iov_iter_count(&s->iter) + io->bytes_done;
  		iovec = NULL;
  	}
  	ret = io_rw_init_file(req, FMODE_READ);
@@ -740,7 +742,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
  		kfree(iovec);
  		return ret;
  	}
-	req->cqe.res = iov_iter_count(&s->iter);
+	req->cqe.res = target_len;
 
  	if (force_nonblock) {
  		/* If the file doesn't support async, just async punt */
@@ -850,18 +852,20 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
  	struct iovec *iovec;
  	struct kiocb *kiocb = &rw->kiocb;
  	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
-	ssize_t ret, ret2;
+	ssize_t ret, ret2, target_len;
  	loff_t *ppos;
 
  	if (!req_has_async_data(req)) {
  		ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
  		if (unlikely(ret < 0))
  			return ret;
+		target_len = iov_iter_count(&s->iter);
  	} else {
  		struct io_async_rw *io = req->async_data;
 
  		s = &io->s;
  		iov_iter_restore(&s->iter, &s->iter_state);
+		target_len = iov_iter_count(&s->iter) + io->bytes_done;
  		iovec = NULL;
  	}
  	ret = io_rw_init_file(req, FMODE_WRITE);
@@ -869,7 +873,7 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
  		kfree(iovec);
  		return ret;
  	}
-	req->cqe.res = iov_iter_count(&s->iter);
+	req->cqe.res = target_len;
 
  	if (force_nonblock) {
  		/* If the file doesn't support async, just async punt */
Signature · 1 line
Pavel Begunkov
#2 lore raw
Quoted 12 lines · Pavel Begunkov
Quoted 9 lines · David Stevens

From: David Stevens <[email protected]>

When continuing a short read/write, account for any completed bytes when calculating the operation's target length. The operation's actual accumulated length is fixed up by kiocb_done, and the target and actual lengths are then compared by __io_complete_rw_common. That function already propagated the actual length to userspace, but the incorrect target length was causing it to always cancel linked operations, even with a successfully completed read/write.

Quoted 8 lines

The issue looks same as fixed with

https://git.kernel.dk/cgit/linux-block/commit/?h=for-6.1/io_uring&id=bf68b5b34311ee57ed40749a1257a30b46127556

Can you check if for-6.1 works for you?

git://git.kernel.dk/linux.git for-6.1/io_uring
https://git.kernel.dk/cgit/linux-block/log/?h=for-6.1/io_uring

Yes, it looks like that fixes the bug I was running into.

-David

Quoted 77 lines
Quoted 66 lines
Fixes 227c0c9673d8 ("io_uring: internally retry short reads")
---
  io_uring/rw.c | 12 ++++++++----
  1 file changed, 8 insertions(+), 4 deletions(-)
io_uring/rw.c +7 -3
index 76ebcfebc9a6..aa9967a52dfd 100644
@@ -706,13 +706,14 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
      struct kiocb *kiocb = &rw->kiocb;
      bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
      struct io_async_rw *io;
-     ssize_t ret, ret2;
+     ssize_t ret, ret2, target_len;
      loff_t *ppos;
 
      if (!req_has_async_data(req)) {
              ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
              if (unlikely(ret < 0))
                      return ret;
+             target_len = iov_iter_count(&s->iter);
      } else {
              io = req->async_data;
              s = &io->s;
@@ -733,6 +734,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
               * need to make this conditional.
               */
              iov_iter_restore(&s->iter, &s->iter_state);
+             target_len = iov_iter_count(&s->iter) + io->bytes_done;
              iovec = NULL;
      }
      ret = io_rw_init_file(req, FMODE_READ);
@@ -740,7 +742,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags)
              kfree(iovec);
              return ret;
      }
-     req->cqe.res = iov_iter_count(&s->iter);
+     req->cqe.res = target_len;
 
      if (force_nonblock) {
              /* If the file doesn't support async, just async punt */
@@ -850,18 +852,20 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
      struct iovec *iovec;
      struct kiocb *kiocb = &rw->kiocb;
      bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
-     ssize_t ret, ret2;
+     ssize_t ret, ret2, target_len;
      loff_t *ppos;
 
      if (!req_has_async_data(req)) {
              ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
              if (unlikely(ret < 0))
                      return ret;
+             target_len = iov_iter_count(&s->iter);
      } else {
              struct io_async_rw *io = req->async_data;
 
              s = &io->s;
              iov_iter_restore(&s->iter, &s->iter_state);
+             target_len = iov_iter_count(&s->iter) + io->bytes_done;
              iovec = NULL;
      }
      ret = io_rw_init_file(req, FMODE_WRITE);

io_uring/rw.c int io_write(struct io_kiocb *req, unsigned int issue_flags) · line 879

Quoted 8 lines
io_uring/rw.c +1 -1
@@ -869,7 +873,7 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
              kfree(iovec);
              return ret;
      }
-     req->cqe.res = iov_iter_count(&s->iter);
+     req->cqe.res = target_len;
 
      if (force_nonblock) {
              /* If the file doesn't support async, just async punt */

--
Pavel Begunkov