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/nospec.h>
9 #include <linux/io_uring.h>
10
11 #include <uapi/linux/io_uring.h>
12
13 #include "filetable.h"
14 #include "sqpoll.h"
15 #include "fdinfo.h"
16 #include "cancel.h"
17 #include "rsrc.h"
18 #include "opdef.h"
19
20 #ifdef CONFIG_NET_RX_BUSY_POLL
common_tracking_show_fdinfo(struct io_ring_ctx * ctx,struct seq_file * m,const char * tracking_strategy)21 static __cold void common_tracking_show_fdinfo(struct io_ring_ctx *ctx,
22 struct seq_file *m,
23 const char *tracking_strategy)
24 {
25 seq_puts(m, "NAPI:\tenabled\n");
26 seq_printf(m, "napi tracking:\t%s\n", tracking_strategy);
27 seq_printf(m, "napi_busy_poll_dt:\t%llu\n", ctx->napi_busy_poll_dt);
28 if (ctx->napi_prefer_busy_poll)
29 seq_puts(m, "napi_prefer_busy_poll:\ttrue\n");
30 else
31 seq_puts(m, "napi_prefer_busy_poll:\tfalse\n");
32 }
33
napi_show_fdinfo(struct io_ring_ctx * ctx,struct seq_file * m)34 static __cold void napi_show_fdinfo(struct io_ring_ctx *ctx,
35 struct seq_file *m)
36 {
37 unsigned int mode = READ_ONCE(ctx->napi_track_mode);
38
39 switch (mode) {
40 case IO_URING_NAPI_TRACKING_INACTIVE:
41 seq_puts(m, "NAPI:\tdisabled\n");
42 break;
43 case IO_URING_NAPI_TRACKING_DYNAMIC:
44 common_tracking_show_fdinfo(ctx, m, "dynamic");
45 break;
46 case IO_URING_NAPI_TRACKING_STATIC:
47 common_tracking_show_fdinfo(ctx, m, "static");
48 break;
49 default:
50 seq_printf(m, "NAPI:\tunknown mode (%u)\n", mode);
51 }
52 }
53 #else
napi_show_fdinfo(struct io_ring_ctx * ctx,struct seq_file * m)54 static inline void napi_show_fdinfo(struct io_ring_ctx *ctx,
55 struct seq_file *m)
56 {
57 }
58 #endif
59
__io_uring_show_fdinfo(struct io_ring_ctx * ctx,struct seq_file * m)60 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
61 {
62 struct io_overflow_cqe *ocqe;
63 struct io_rings *r = ctx->rings;
64 unsigned int sq_mask = ctx->sq_entries - 1, cq_mask = ctx->cq_entries - 1;
65 unsigned int sq_head = READ_ONCE(r->sq.head);
66 unsigned int sq_tail = READ_ONCE(r->sq.tail);
67 unsigned int cq_head = READ_ONCE(r->cq.head);
68 unsigned int cq_tail = READ_ONCE(r->cq.tail);
69 unsigned int sq_shift = 0;
70 unsigned int cq_entries, sq_entries;
71 int sq_pid = -1, sq_cpu = -1;
72 u64 sq_total_time = 0, sq_work_time = 0;
73 unsigned int i;
74
75 if (ctx->flags & IORING_SETUP_SQE128)
76 sq_shift = 1;
77
78 /*
79 * we may get imprecise sqe and cqe info if uring is actively running
80 * since we get cached_sq_head and cached_cq_tail without uring_lock
81 * and sq_tail and cq_head are changed by userspace. But it's ok since
82 * we usually use these info when it is stuck.
83 */
84 seq_printf(m, "SqMask:\t0x%x\n", sq_mask);
85 seq_printf(m, "SqHead:\t%u\n", sq_head);
86 seq_printf(m, "SqTail:\t%u\n", sq_tail);
87 seq_printf(m, "CachedSqHead:\t%u\n", data_race(ctx->cached_sq_head));
88 seq_printf(m, "CqMask:\t0x%x\n", cq_mask);
89 seq_printf(m, "CqHead:\t%u\n", cq_head);
90 seq_printf(m, "CqTail:\t%u\n", cq_tail);
91 seq_printf(m, "CachedCqTail:\t%u\n", data_race(ctx->cached_cq_tail));
92 seq_printf(m, "SQEs:\t%u\n", sq_tail - sq_head);
93 sq_entries = min(sq_tail - sq_head, ctx->sq_entries);
94 for (i = 0; i < sq_entries; i++) {
95 unsigned int entry = i + sq_head;
96 struct io_uring_sqe *sqe;
97 unsigned int sq_idx;
98 bool sqe128 = false;
99 u8 opcode;
100
101 if (ctx->flags & IORING_SETUP_NO_SQARRAY)
102 sq_idx = entry & sq_mask;
103 else
104 sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]);
105 if (sq_idx > sq_mask)
106 continue;
107
108 sqe = &ctx->sq_sqes[sq_idx << sq_shift];
109 opcode = READ_ONCE(sqe->opcode);
110 if (opcode >= IORING_OP_LAST)
111 continue;
112 opcode = array_index_nospec(opcode, IORING_OP_LAST);
113 if (sq_shift) {
114 sqe128 = true;
115 } else if (io_issue_defs[opcode].is_128) {
116 if (!(ctx->flags & IORING_SETUP_SQE_MIXED)) {
117 seq_printf(m,
118 "%5u: invalid sqe, 128B entry on non-mixed sq\n",
119 sq_idx);
120 break;
121 }
122 if ((++sq_head & sq_mask) == 0) {
123 seq_printf(m,
124 "%5u: corrupted sqe, wrapping 128B entry\n",
125 sq_idx);
126 break;
127 }
128 sqe128 = true;
129 }
130 seq_printf(m, "%5u: opcode:%s, fd:%d, flags:%x, off:%llu, "
131 "addr:0x%llx, rw_flags:0x%x, buf_index:%d "
132 "user_data:%llu",
133 sq_idx, io_uring_get_opcode(opcode), sqe->fd,
134 sqe->flags, (unsigned long long) sqe->off,
135 (unsigned long long) sqe->addr, sqe->rw_flags,
136 sqe->buf_index, sqe->user_data);
137 if (sqe128) {
138 u64 *sqeb = (void *) (sqe + 1);
139 int size = sizeof(struct io_uring_sqe) / sizeof(u64);
140 int j;
141
142 for (j = 0; j < size; j++) {
143 seq_printf(m, ", e%d:0x%llx", j,
144 (unsigned long long) *sqeb);
145 sqeb++;
146 }
147 }
148 seq_printf(m, "\n");
149 cond_resched();
150 }
151 seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head);
152 cq_entries = min(cq_tail - cq_head, ctx->cq_entries);
153 for (i = 0; i < cq_entries; i++) {
154 struct io_uring_cqe *cqe;
155 bool cqe32 = false;
156
157 cqe = &r->cqes[(cq_head & cq_mask)];
158 if (cqe->flags & IORING_CQE_F_32 || ctx->flags & IORING_SETUP_CQE32)
159 cqe32 = true;
160 seq_printf(m, "%5u: user_data:%llu, res:%d, flags:%x",
161 cq_head & cq_mask, cqe->user_data, cqe->res,
162 cqe->flags);
163 if (cqe32)
164 seq_printf(m, ", extra1:%llu, extra2:%llu",
165 cqe->big_cqe[0], cqe->big_cqe[1]);
166 seq_printf(m, "\n");
167 cq_head++;
168 if (cqe32) {
169 cq_head++;
170 i++;
171 }
172 cond_resched();
173 }
174
175 if (ctx->flags & IORING_SETUP_SQPOLL) {
176 struct io_sq_data *sq = ctx->sq_data;
177 struct task_struct *tsk;
178
179 rcu_read_lock();
180 tsk = rcu_dereference(sq->thread);
181 /*
182 * sq->thread might be NULL if we raced with the sqpoll
183 * thread termination.
184 */
185 if (tsk) {
186 u64 usec;
187
188 get_task_struct(tsk);
189 rcu_read_unlock();
190 usec = io_sq_cpu_usec(tsk);
191 put_task_struct(tsk);
192 sq_pid = sq->task_pid;
193 sq_cpu = sq->sq_cpu;
194 sq_total_time = usec;
195 sq_work_time = sq->work_time;
196 } else {
197 rcu_read_unlock();
198 }
199 }
200
201 seq_printf(m, "SqThread:\t%d\n", sq_pid);
202 seq_printf(m, "SqThreadCpu:\t%d\n", sq_cpu);
203 seq_printf(m, "SqTotalTime:\t%llu\n", sq_total_time);
204 seq_printf(m, "SqWorkTime:\t%llu\n", sq_work_time);
205 seq_printf(m, "UserFiles:\t%u\n", ctx->file_table.data.nr);
206 for (i = 0; i < ctx->file_table.data.nr; i++) {
207 struct file *f = NULL;
208
209 if (ctx->file_table.data.nodes[i])
210 f = io_slot_file(ctx->file_table.data.nodes[i]);
211 if (f) {
212 seq_printf(m, "%5u: ", i);
213 seq_file_path(m, f, " \t\n\\");
214 seq_puts(m, "\n");
215 }
216 }
217 seq_printf(m, "UserBufs:\t%u\n", ctx->buf_table.nr);
218 for (i = 0; i < ctx->buf_table.nr; i++) {
219 struct io_mapped_ubuf *buf = NULL;
220
221 if (ctx->buf_table.nodes[i])
222 buf = ctx->buf_table.nodes[i]->buf;
223 if (buf)
224 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, buf->len);
225 else
226 seq_printf(m, "%5u: <none>\n", i);
227 }
228
229 seq_puts(m, "PollList:\n");
230 for (i = 0; i < (1U << ctx->cancel_table.hash_bits); i++) {
231 struct io_hash_bucket *hb = &ctx->cancel_table.hbs[i];
232 struct io_kiocb *req;
233
234 hlist_for_each_entry(req, &hb->list, hash_node)
235 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
236 task_work_pending(req->tctx->task));
237 }
238
239 seq_puts(m, "CqOverflowList:\n");
240 spin_lock(&ctx->completion_lock);
241 list_for_each_entry(ocqe, &ctx->cq_overflow_list, list) {
242 struct io_uring_cqe *cqe = &ocqe->cqe;
243
244 seq_printf(m, " user_data=%llu, res=%d, flags=%x\n",
245 cqe->user_data, cqe->res, cqe->flags);
246
247 }
248 spin_unlock(&ctx->completion_lock);
249 napi_show_fdinfo(ctx, m);
250 }
251
252 /*
253 * Caller holds a reference to the file already, we don't need to do
254 * anything else to get an extra reference.
255 */
io_uring_show_fdinfo(struct seq_file * m,struct file * file)256 __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *file)
257 {
258 struct io_ring_ctx *ctx = file->private_data;
259
260 /*
261 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
262 * since fdinfo case grabs it in the opposite direction of normal use
263 * cases.
264 */
265 if (mutex_trylock(&ctx->uring_lock)) {
266 __io_uring_show_fdinfo(ctx, m);
267 mutex_unlock(&ctx->uring_lock);
268 }
269 }
270