io_uring/net: punt IORING_OP_BIND async if it needs file create
Jens Axboe
maintainerPATCH 1/1 io_uring/net: punt IORING_OP_BIND async if it needs file create
For two reasons:
1) An opcode cannot block inside io_uring_enter() doing submissions, as it'll stall the submission side pipeline.
2) Ending up in sb_start_write() -> __sb_start_write() -> percpu_down_read_freezable() introduces a new lockdep edge, which it correctly complains about.
Check if the socket type is AF_UNIX and has a non-empty pathname. If it does, mark it REQ_F_FORCE_ASYNC to punt the submission to io-wq rather than attempt to do it inline.
---
io_uring/net.c +25 -1
@@ -4,6 +4,7 @@#include <linux/file.h> #include <linux/slab.h> #include <linux/net.h> +#include <linux/un.h> #include <linux/compat.h> #include <net/compat.h> #include <linux/io_uring.h>
@@ -1799,11 +1800,29 @@ int io_connect(struct io_kiocb *req, unsigned int issue_flags)return IOU_COMPLETE ; } +/* +* Check if bind request would potentiall end up with filename_create(), +* which in turn end up in mnt_want_write() which will grab the fs +* percpu start write sem. This can trigger a lockdep warning. +*/ +static int io_bind_file_create (const struct io_async_msghdr *io,int addr_len) +{ +const struct sockaddr_un *sun; + +if (io->addr.ss_family !=AF_UNIX ) +return 0 ; +if (addr_len <= offsetof(struct sockaddr_un , sun_path)) +return 0 ; + sun = (const struct sockaddr_un *) &io->addr; +return sun->sun_path[0 ] !='\0' ; +} +int io_bind_prep (struct io_kiocb *req,const struct io_uring_sqe *sqe) { struct io_bind *bind = io_kiocb_to_cmd(req, struct io_bind);struct sockaddr __user *uaddr;struct io_async_msghdr *io; +int ret;if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)return -EINVAL ;
@@ -1814,7 +1833,12 @@ int io_bind_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) io =io_msg_alloc_async (req);if (unlikely (!io))return -ENOMEM ; -return move_addr_to_kernel (uaddr, bind->addr_len, &io->addr); + ret =move_addr_to_kernel (uaddr, bind->addr_len, &io->addr); +if (unlikely (ret)) +return ret; +if (io_bind_file_create (io, bind->addr_len)) + req->flags |=REQ_F_FORCE_ASYNC ; +return 0 ; }int io_bind (struct io_kiocb *req,unsigned int issue_flags)
Signature · 1 line
Jens Axboe
Quoted 31 lines · Jens Axboe
For two reasons:
1) An opcode cannot block inside
io_uring_enter()doing submissions, as it'll stall the submission side pipeline.2) Ending up in
sb_start_write()->__sb_start_write()->percpu_down_read_freezable()introduces a new lockdep edge, which it correctly complains about.Check if the socket type is
AF_UNIXand has a non-empty pathname. If it does, mark itREQ_F_FORCE_ASYNCto punt the submission to io-wq rather than attempt to do it inline.Fixes 7481fd93fa0a ("io_uring: Introduce IORING_OP_BIND")Signed-off-by Jens Axboe <[email protected]>---
io_uring/net.c +1
-0@@ -4,6 +4,7 @@#include <linux/file.h> #include <linux/slab.h> #include <linux/net.h> +#include <linux/un.h> #include <linux/compat.h> #include <net/compat.h> #include <linux/io_uring.h>
io_uring/net.c int io_connect(struct io_kiocb *req, unsigned int issue_flags) · line 1804
Quoted 5 lines
io_uring/net.c +2
-0@@ -1799,11 +1800,29 @@ int io_connect(struct io_kiocb *req, unsigned int issue_flags)return IOU_COMPLETE ; } +/* +* Check if bind request would potentiall end up with filename_create(),
potentially
Quoted 26 lines
io_uring/net.c +16
-0+* which in turn end up in mnt_want_write() which will grab the fs +* percpu start write sem. This can trigger a lockdep warning. +*/ +static int io_bind_file_create (const struct io_async_msghdr *io,int addr_len) +{ +const struct sockaddr_un *sun; + +if (io->addr.ss_family !=AF_UNIX ) +return 0 ; +if (addr_len <= offsetof(struct sockaddr_un , sun_path)) +return 0 ; + sun = (const struct sockaddr_un *) &io->addr; +return sun->sun_path[0 ] !='\0' ; +} +int io_bind_prep (struct io_kiocb *req,const struct io_uring_sqe *sqe) { struct io_bind *bind = io_kiocb_to_cmd(req, struct io_bind);struct sockaddr __user *uaddr;struct io_async_msghdr *io; +int ret;if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)return -EINVAL ;@@ -1814,7 +1833,12 @@ int io_bind_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) io =io_msg_alloc_async (req);
io_uring/net.c int io_bind_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · line 1842
Quoted 10 lines
io_uring/net.c +6
-1if (unlikely (!io))return -ENOMEM ; -return move_addr_to_kernel (uaddr, bind->addr_len, &io->addr); + ret =move_addr_to_kernel (uaddr, bind->addr_len, &io->addr); +if (unlikely (ret)) +return ret; +if (io_bind_file_create (io, bind->addr_len)) + req->flags |=REQ_F_FORCE_ASYNC ; +return 0 ; }
Patch looks ok. I will add a TODO to myself to look into the bind call getting a flag and returning -EWOULDBLOCK, so we can eventually drop the workaround.
Quoted 2 lines
io_uring/net.c +0
-0int io_bind (struct io_kiocb *req,unsigned int issue_flags)
Signature · 1 line
Gabriel Krisman Bertazi
Jens Axboe
maintainerQuoted 16 lines · Gabriel Krisman Bertazi
Quoted 16 lines
io_uring/net.c +1
-0@@ -4,6 +4,7 @@#include <linux/file.h> #include <linux/slab.h> #include <linux/net.h> +#include <linux/un.h> #include <linux/compat.h> #include <net/compat.h> #include <linux/io_uring.h> @@ -1799,11 +1800,29 @@ int io_connect(struct io_kiocb *req, unsigned int issue_flags)return IOU_COMPLETE ; }
Quoted 4 lines
io_uring/net.c int io_connect(struct io_kiocb *req, unsigned int issue_flags) · line 1804
Quoted 2 lines
io_uring/net.c +2
-0+/* +* Check if bind request would potentiall end up with filename_create(), potentially
thanks, fixed.
Quoted 32 lines
Quoted 32 lines
io_uring/net.c +19
-1+* which in turn end up in mnt_want_write() which will grab the fs +* percpu start write sem. This can trigger a lockdep warning. +*/ +static int io_bind_file_create (const struct io_async_msghdr *io,int addr_len) +{ +const struct sockaddr_un *sun; + +if (io->addr.ss_family !=AF_UNIX ) +return 0 ; +if (addr_len <= offsetof(struct sockaddr_un , sun_path)) +return 0 ; + sun = (const struct sockaddr_un *) &io->addr; +return sun->sun_path[0 ] !='\0' ; +} +int io_bind_prep (struct io_kiocb *req,const struct io_uring_sqe *sqe) { struct io_bind *bind = io_kiocb_to_cmd(req, struct io_bind);struct sockaddr __user *uaddr;struct io_async_msghdr *io; +int ret;if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)return -EINVAL ;@@ -1814,7 +1833,12 @@ int io_bind_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) io =io_msg_alloc_async (req);if (unlikely (!io))return -ENOMEM ; -return move_addr_to_kernel (uaddr, bind->addr_len, &io->addr); + ret =move_addr_to_kernel (uaddr, bind->addr_len, &io->addr); +if (unlikely (ret)) +return ret;
Quoted 8 lines
io_uring/net.c int io_bind_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · line 1842
Quoted 4 lines
io_uring/net.c +3
-0+if (io_bind_file_create (io, bind->addr_len)) + req->flags |=REQ_F_FORCE_ASYNC ; +return 0 ; }Patch looks ok. I will add a TODO to myself to look into the bind call getting a flag and returning
-EWOULDBLOCK, so we can eventually drop the workaround.
That would indeed be better, but probably better darn far down, and it'd need refactoring to pass down 'flags' or similar as msghdr isn't a thing there... So not sure it's going to be super viable.
Quoted 1 line
Reviewed-by Gabriel Krisman Bertazi <[email protected]>
Thanks, added!
Signature · 1 line
Jens Axboe