| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | #include <linux/kernel.h> |
| 3 | #include <linux/errno.h> |
| 4 | #include <linux/fs.h> |
| 5 | #include <linux/file.h> |
| 6 | #include <linux/proc_fs.h> |
| 7 | #include <linux/seq_file.h> |
| 8 | #include <linux/io_uring.h> |
| 9 | |
| 10 | #include <uapi/linux/io_uring.h> |
| 11 | |
| 12 | #include "filetable.h" |
| 13 | #include "sqpoll.h" |
| 14 | #include "fdinfo.h" |
| 15 | #include "cancel.h" |
| 16 | #include "rsrc.h" |
| 17 | |
| 18 | #ifdef CONFIG_NET_RX_BUSY_POLL |
| 19 | static __cold void common_tracking_show_fdinfo(struct io_ring_ctx *ctx, |
| 20 | struct seq_file *m, |
| 21 | const char *tracking_strategy) |
| 22 | { |
| 23 | seq_puts(m, s: "NAPI:\tenabled\n" ); |
| 24 | seq_printf(m, fmt: "napi tracking:\t%s\n" , tracking_strategy); |
| 25 | seq_printf(m, fmt: "napi_busy_poll_dt:\t%llu\n" , ctx->napi_busy_poll_dt); |
| 26 | if (ctx->napi_prefer_busy_poll) |
| 27 | seq_puts(m, s: "napi_prefer_busy_poll:\ttrue\n" ); |
| 28 | else |
| 29 | seq_puts(m, s: "napi_prefer_busy_poll:\tfalse\n" ); |
| 30 | } |
| 31 | |
| 32 | static __cold void napi_show_fdinfo(struct io_ring_ctx *ctx, |
| 33 | struct seq_file *m) |
| 34 | { |
| 35 | unsigned int mode = READ_ONCE(ctx->napi_track_mode); |
| 36 | |
| 37 | switch (mode) { |
| 38 | case IO_URING_NAPI_TRACKING_INACTIVE: |
| 39 | seq_puts(m, s: "NAPI:\tdisabled\n" ); |
| 40 | break; |
| 41 | case IO_URING_NAPI_TRACKING_DYNAMIC: |
| 42 | common_tracking_show_fdinfo(ctx, m, tracking_strategy: "dynamic" ); |
| 43 | break; |
| 44 | case IO_URING_NAPI_TRACKING_STATIC: |
| 45 | common_tracking_show_fdinfo(ctx, m, tracking_strategy: "static" ); |
| 46 | break; |
| 47 | default: |
| 48 | seq_printf(m, fmt: "NAPI:\tunknown mode (%u)\n" , mode); |
| 49 | } |
| 50 | } |
| 51 | #else |
| 52 | static inline void napi_show_fdinfo(struct io_ring_ctx *ctx, |
| 53 | struct seq_file *m) |
| 54 | { |
| 55 | } |
| 56 | #endif |
| 57 | |
| 58 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) |
| 59 | { |
| 60 | struct io_overflow_cqe *ocqe; |
| 61 | struct io_rings *r = ctx->rings; |
| 62 | struct rusage sq_usage; |
| 63 | unsigned int sq_mask = ctx->sq_entries - 1, cq_mask = ctx->cq_entries - 1; |
| 64 | unsigned int sq_head = READ_ONCE(r->sq.head); |
| 65 | unsigned int sq_tail = READ_ONCE(r->sq.tail); |
| 66 | unsigned int cq_head = READ_ONCE(r->cq.head); |
| 67 | unsigned int cq_tail = READ_ONCE(r->cq.tail); |
| 68 | unsigned int sq_shift = 0; |
| 69 | unsigned int sq_entries; |
| 70 | int sq_pid = -1, sq_cpu = -1; |
| 71 | u64 sq_total_time = 0, sq_work_time = 0; |
| 72 | unsigned int i; |
| 73 | |
| 74 | if (ctx->flags & IORING_SETUP_SQE128) |
| 75 | sq_shift = 1; |
| 76 | |
| 77 | /* |
| 78 | * we may get imprecise sqe and cqe info if uring is actively running |
| 79 | * since we get cached_sq_head and cached_cq_tail without uring_lock |
| 80 | * and sq_tail and cq_head are changed by userspace. But it's ok since |
| 81 | * we usually use these info when it is stuck. |
| 82 | */ |
| 83 | seq_printf(m, fmt: "SqMask:\t0x%x\n" , sq_mask); |
| 84 | seq_printf(m, fmt: "SqHead:\t%u\n" , sq_head); |
| 85 | seq_printf(m, fmt: "SqTail:\t%u\n" , sq_tail); |
| 86 | seq_printf(m, fmt: "CachedSqHead:\t%u\n" , data_race(ctx->cached_sq_head)); |
| 87 | seq_printf(m, fmt: "CqMask:\t0x%x\n" , cq_mask); |
| 88 | seq_printf(m, fmt: "CqHead:\t%u\n" , cq_head); |
| 89 | seq_printf(m, fmt: "CqTail:\t%u\n" , cq_tail); |
| 90 | seq_printf(m, fmt: "CachedCqTail:\t%u\n" , data_race(ctx->cached_cq_tail)); |
| 91 | seq_printf(m, fmt: "SQEs:\t%u\n" , sq_tail - sq_head); |
| 92 | sq_entries = min(sq_tail - sq_head, ctx->sq_entries); |
| 93 | for (i = 0; i < sq_entries; i++) { |
| 94 | unsigned int entry = i + sq_head; |
| 95 | struct io_uring_sqe *sqe; |
| 96 | unsigned int sq_idx; |
| 97 | |
| 98 | if (ctx->flags & IORING_SETUP_NO_SQARRAY) |
| 99 | break; |
| 100 | sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]); |
| 101 | if (sq_idx > sq_mask) |
| 102 | continue; |
| 103 | sqe = &ctx->sq_sqes[sq_idx << sq_shift]; |
| 104 | seq_printf(m, fmt: "%5u: opcode:%s, fd:%d, flags:%x, off:%llu, " |
| 105 | "addr:0x%llx, rw_flags:0x%x, buf_index:%d " |
| 106 | "user_data:%llu" , |
| 107 | sq_idx, io_uring_get_opcode(opcode: sqe->opcode), sqe->fd, |
| 108 | sqe->flags, (unsigned long long) sqe->off, |
| 109 | (unsigned long long) sqe->addr, sqe->rw_flags, |
| 110 | sqe->buf_index, sqe->user_data); |
| 111 | if (sq_shift) { |
| 112 | u64 *sqeb = (void *) (sqe + 1); |
| 113 | int size = sizeof(struct io_uring_sqe) / sizeof(u64); |
| 114 | int j; |
| 115 | |
| 116 | for (j = 0; j < size; j++) { |
| 117 | seq_printf(m, fmt: ", e%d:0x%llx" , j, |
| 118 | (unsigned long long) *sqeb); |
| 119 | sqeb++; |
| 120 | } |
| 121 | } |
| 122 | seq_printf(m, fmt: "\n" ); |
| 123 | } |
| 124 | seq_printf(m, fmt: "CQEs:\t%u\n" , cq_tail - cq_head); |
| 125 | while (cq_head < cq_tail) { |
| 126 | struct io_uring_cqe *cqe; |
| 127 | bool cqe32 = false; |
| 128 | |
| 129 | cqe = &r->cqes[(cq_head & cq_mask)]; |
| 130 | if (cqe->flags & IORING_CQE_F_32 || ctx->flags & IORING_SETUP_CQE32) |
| 131 | cqe32 = true; |
| 132 | seq_printf(m, fmt: "%5u: user_data:%llu, res:%d, flag:%x" , |
| 133 | cq_head & cq_mask, cqe->user_data, cqe->res, |
| 134 | cqe->flags); |
| 135 | if (cqe32) |
| 136 | seq_printf(m, fmt: ", extra1:%llu, extra2:%llu\n" , |
| 137 | cqe->big_cqe[0], cqe->big_cqe[1]); |
| 138 | seq_printf(m, fmt: "\n" ); |
| 139 | cq_head++; |
| 140 | if (cqe32) |
| 141 | cq_head++; |
| 142 | } |
| 143 | |
| 144 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
| 145 | struct io_sq_data *sq = ctx->sq_data; |
| 146 | struct task_struct *tsk; |
| 147 | |
| 148 | rcu_read_lock(); |
| 149 | tsk = rcu_dereference(sq->thread); |
| 150 | /* |
| 151 | * sq->thread might be NULL if we raced with the sqpoll |
| 152 | * thread termination. |
| 153 | */ |
| 154 | if (tsk) { |
| 155 | get_task_struct(t: tsk); |
| 156 | rcu_read_unlock(); |
| 157 | getrusage(p: tsk, RUSAGE_SELF, ru: &sq_usage); |
| 158 | put_task_struct(t: tsk); |
| 159 | sq_pid = sq->task_pid; |
| 160 | sq_cpu = sq->sq_cpu; |
| 161 | sq_total_time = (sq_usage.ru_stime.tv_sec * 1000000 |
| 162 | + sq_usage.ru_stime.tv_usec); |
| 163 | sq_work_time = sq->work_time; |
| 164 | } else { |
| 165 | rcu_read_unlock(); |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | seq_printf(m, fmt: "SqThread:\t%d\n" , sq_pid); |
| 170 | seq_printf(m, fmt: "SqThreadCpu:\t%d\n" , sq_cpu); |
| 171 | seq_printf(m, fmt: "SqTotalTime:\t%llu\n" , sq_total_time); |
| 172 | seq_printf(m, fmt: "SqWorkTime:\t%llu\n" , sq_work_time); |
| 173 | seq_printf(m, fmt: "UserFiles:\t%u\n" , ctx->file_table.data.nr); |
| 174 | for (i = 0; i < ctx->file_table.data.nr; i++) { |
| 175 | struct file *f = NULL; |
| 176 | |
| 177 | if (ctx->file_table.data.nodes[i]) |
| 178 | f = io_slot_file(node: ctx->file_table.data.nodes[i]); |
| 179 | if (f) { |
| 180 | seq_printf(m, fmt: "%5u: " , i); |
| 181 | seq_file_path(m, f, " \t\n\\" ); |
| 182 | seq_puts(m, s: "\n" ); |
| 183 | } |
| 184 | } |
| 185 | seq_printf(m, fmt: "UserBufs:\t%u\n" , ctx->buf_table.nr); |
| 186 | for (i = 0; i < ctx->buf_table.nr; i++) { |
| 187 | struct io_mapped_ubuf *buf = NULL; |
| 188 | |
| 189 | if (ctx->buf_table.nodes[i]) |
| 190 | buf = ctx->buf_table.nodes[i]->buf; |
| 191 | if (buf) |
| 192 | seq_printf(m, fmt: "%5u: 0x%llx/%u\n" , i, buf->ubuf, buf->len); |
| 193 | else |
| 194 | seq_printf(m, fmt: "%5u: <none>\n" , i); |
| 195 | } |
| 196 | |
| 197 | seq_puts(m, s: "PollList:\n" ); |
| 198 | for (i = 0; i < (1U << ctx->cancel_table.hash_bits); i++) { |
| 199 | struct io_hash_bucket *hb = &ctx->cancel_table.hbs[i]; |
| 200 | struct io_kiocb *req; |
| 201 | |
| 202 | hlist_for_each_entry(req, &hb->list, hash_node) |
| 203 | seq_printf(m, fmt: " op=%d, task_works=%d\n" , req->opcode, |
| 204 | task_work_pending(task: req->tctx->task)); |
| 205 | } |
| 206 | |
| 207 | seq_puts(m, s: "CqOverflowList:\n" ); |
| 208 | spin_lock(lock: &ctx->completion_lock); |
| 209 | list_for_each_entry(ocqe, &ctx->cq_overflow_list, list) { |
| 210 | struct io_uring_cqe *cqe = &ocqe->cqe; |
| 211 | |
| 212 | seq_printf(m, fmt: " user_data=%llu, res=%d, flags=%x\n" , |
| 213 | cqe->user_data, cqe->res, cqe->flags); |
| 214 | |
| 215 | } |
| 216 | spin_unlock(lock: &ctx->completion_lock); |
| 217 | napi_show_fdinfo(ctx, m); |
| 218 | } |
| 219 | |
| 220 | /* |
| 221 | * Caller holds a reference to the file already, we don't need to do |
| 222 | * anything else to get an extra reference. |
| 223 | */ |
| 224 | __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *file) |
| 225 | { |
| 226 | struct io_ring_ctx *ctx = file->private_data; |
| 227 | |
| 228 | /* |
| 229 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, |
| 230 | * since fdinfo case grabs it in the opposite direction of normal use |
| 231 | * cases. |
| 232 | */ |
| 233 | if (mutex_trylock(lock: &ctx->uring_lock)) { |
| 234 | __io_uring_show_fdinfo(ctx, m); |
| 235 | mutex_unlock(lock: &ctx->uring_lock); |
| 236 | } |
| 237 | } |
| 238 | |