Pull io_uring fixes from Jens Axboe: "Fix for an issue reported where reading fdinfo could find a NULL thread as we didn't properly synchronize, and then a disable for the IOCB_DIO_CALLER_COMP optimization as a recent reported highlighted how that could lead to deadlocks if the task issued async O_DIRECT writes and then proceeded to do sync fallocate() calls" * tag 'io_uring-6.6-2023-10-27' of git://git.kernel.dk/linux: io_uring/rw: disable IOCB_DIO_CALLER_COMP io_uring/fdinfo: lock SQ thread while retrieving thread cpu/pid
Unified · Split
| @@ -53,7 +53,6 @@ static __cold int io_uring_show_cred(struct seq_file *m, unsigned int id, · 52 unmodified lines | ||
| 53 | 53 | __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f) |
| 54 | 54 | { |
| 55 | 55 | struct io_ring_ctx *ctx = f->private_data; |
| 56 | − struct io_sq_data *sq = NULL; | |
| 57 | 56 | struct io_overflow_cqe *ocqe; |
| 58 | 57 | struct io_rings *r = ctx->rings; |
| 59 | 58 | unsigned int sq_mask = ctx->sq_entries - 1, cq_mask = ctx->cq_entries - 1; |
| @@ -64,6 +63,7 @@ __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f) · 4 unmodified lines | ||
| 64 | 63 | unsigned int cq_shift = 0; |
| 65 | 64 | unsigned int sq_shift = 0; |
| 66 | 65 | unsigned int sq_entries, cq_entries; |
| 66 | + int sq_pid = -1, sq_cpu = -1; | |
| 67 | 67 | bool has_lock; |
| 68 | 68 | unsigned int i; |
| 69 | 69 | |
| @@ -143,13 +143,19 @@ __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f) · 73 unmodified lines | ||
| 143 | 143 | has_lock = mutex_trylock(&ctx->uring_lock); |
| 144 | 144 | |
| 145 | 145 | if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) { |
| 146 | − sq = ctx->sq_data; | |
| 147 | − if (!sq->thread) | |
| 148 | − sq = NULL; | |
| 146 | + struct io_sq_data *sq = ctx->sq_data; | |
| 147 | + | |
| 148 | + if (mutex_trylock(&sq->lock)) { | |
| 149 | + if (sq->thread) { | |
| 150 | + sq_pid = task_pid_nr(sq->thread); | |
| 151 | + sq_cpu = task_cpu(sq->thread); | |
| 152 | + } | |
| 153 | + mutex_unlock(&sq->lock); | |
| 154 | + } | |
| 149 | 155 | } |
| 150 | 156 | |
| 151 | − seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1); | |
| 152 | − seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1); | |
| 157 | + seq_printf(m, "SqThread:\t%d\n", sq_pid); | |
| 158 | + seq_printf(m, "SqThreadCpu:\t%d\n", sq_cpu); | |
| 153 | 159 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
| 154 | 160 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
| 155 | 161 | struct file *f = io_file_from_index(&ctx->file_table, i); |
| @@ -913,15 +913,6 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags) · 912 unmodified lines | ||
| 913 | 913 | kiocb_start_write(kiocb); |
| 914 | 914 | kiocb->ki_flags |= IOCB_WRITE; |
| 915 | 915 | |
| 916 | − /* | |
| 917 | − * For non-polled IO, set IOCB_DIO_CALLER_COMP, stating that our handler | |
| 918 | − * groks deferring the completion to task context. This isn't | |
| 919 | − * necessary and useful for polled IO as that can always complete | |
| 920 | − * directly. | |
| 921 | − */ | |
| 922 | − if (!(kiocb->ki_flags & IOCB_HIPRI)) | |
| 923 | − kiocb->ki_flags |= IOCB_DIO_CALLER_COMP; | |
| 924 | − | |
| 925 | 916 | if (likely(req->file->f_op->write_iter)) |
| 926 | 917 | ret2 = call_write_iter(req->file, kiocb, &s->iter); |
| 927 | 918 | else if (req->file->f_op->write) |