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 "io_uring.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, "NAPI:\tenabled\n");
24 	seq_printf(m, "napi tracking:\t%s\n", tracking_strategy);
25 	seq_printf(m, "napi_busy_poll_dt:\t%llu\n", ctx->napi_busy_poll_dt);
26 	if (ctx->napi_prefer_busy_poll)
27 		seq_puts(m, "napi_prefer_busy_poll:\ttrue\n");
28 	else
29 		seq_puts(m, "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, "NAPI:\tdisabled\n");
40 		break;
41 	case IO_URING_NAPI_TRACKING_DYNAMIC:
42 		common_tracking_show_fdinfo(ctx, m, "dynamic");
43 		break;
44 	case IO_URING_NAPI_TRACKING_STATIC:
45 		common_tracking_show_fdinfo(ctx, m, "static");
46 		break;
47 	default:
48 		seq_printf(m, "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 cq_shift = 0;
69 	unsigned int sq_shift = 0;
70 	unsigned int sq_entries, cq_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_CQE32)
76 		cq_shift = 1;
77 	if (ctx->flags & IORING_SETUP_SQE128)
78 		sq_shift = 1;
79 
80 	/*
81 	 * we may get imprecise sqe and cqe info if uring is actively running
82 	 * since we get cached_sq_head and cached_cq_tail without uring_lock
83 	 * and sq_tail and cq_head are changed by userspace. But it's ok since
84 	 * we usually use these info when it is stuck.
85 	 */
86 	seq_printf(m, "SqMask:\t0x%x\n", sq_mask);
87 	seq_printf(m, "SqHead:\t%u\n", sq_head);
88 	seq_printf(m, "SqTail:\t%u\n", sq_tail);
89 	seq_printf(m, "CachedSqHead:\t%u\n", data_race(ctx->cached_sq_head));
90 	seq_printf(m, "CqMask:\t0x%x\n", cq_mask);
91 	seq_printf(m, "CqHead:\t%u\n", cq_head);
92 	seq_printf(m, "CqTail:\t%u\n", cq_tail);
93 	seq_printf(m, "CachedCqTail:\t%u\n", data_race(ctx->cached_cq_tail));
94 	seq_printf(m, "SQEs:\t%u\n", sq_tail - sq_head);
95 	sq_entries = min(sq_tail - sq_head, ctx->sq_entries);
96 	for (i = 0; i < sq_entries; i++) {
97 		unsigned int entry = i + sq_head;
98 		struct io_uring_sqe *sqe;
99 		unsigned int sq_idx;
100 
101 		if (ctx->flags & IORING_SETUP_NO_SQARRAY)
102 			break;
103 		sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]);
104 		if (sq_idx > sq_mask)
105 			continue;
106 		sqe = &ctx->sq_sqes[sq_idx << sq_shift];
107 		seq_printf(m, "%5u: opcode:%s, fd:%d, flags:%x, off:%llu, "
108 			      "addr:0x%llx, rw_flags:0x%x, buf_index:%d "
109 			      "user_data:%llu",
110 			   sq_idx, io_uring_get_opcode(sqe->opcode), sqe->fd,
111 			   sqe->flags, (unsigned long long) sqe->off,
112 			   (unsigned long long) sqe->addr, sqe->rw_flags,
113 			   sqe->buf_index, sqe->user_data);
114 		if (sq_shift) {
115 			u64 *sqeb = (void *) (sqe + 1);
116 			int size = sizeof(struct io_uring_sqe) / sizeof(u64);
117 			int j;
118 
119 			for (j = 0; j < size; j++) {
120 				seq_printf(m, ", e%d:0x%llx", j,
121 						(unsigned long long) *sqeb);
122 				sqeb++;
123 			}
124 		}
125 		seq_printf(m, "\n");
126 	}
127 	seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head);
128 	cq_entries = min(cq_tail - cq_head, ctx->cq_entries);
129 	for (i = 0; i < cq_entries; i++) {
130 		unsigned int entry = i + cq_head;
131 		struct io_uring_cqe *cqe = &r->cqes[(entry & cq_mask) << cq_shift];
132 
133 		seq_printf(m, "%5u: user_data:%llu, res:%d, flag:%x",
134 			   entry & cq_mask, cqe->user_data, cqe->res,
135 			   cqe->flags);
136 		if (cq_shift)
137 			seq_printf(m, ", extra1:%llu, extra2:%llu\n",
138 					cqe->big_cqe[0], cqe->big_cqe[1]);
139 		seq_printf(m, "\n");
140 	}
141 
142 	if (ctx->flags & IORING_SETUP_SQPOLL) {
143 		struct io_sq_data *sq = ctx->sq_data;
144 
145 		/*
146 		 * sq->thread might be NULL if we raced with the sqpoll
147 		 * thread termination.
148 		 */
149 		if (sq->thread) {
150 			sq_pid = sq->task_pid;
151 			sq_cpu = sq->sq_cpu;
152 			getrusage(sq->thread, RUSAGE_SELF, &sq_usage);
153 			sq_total_time = (sq_usage.ru_stime.tv_sec * 1000000
154 					 + sq_usage.ru_stime.tv_usec);
155 			sq_work_time = sq->work_time;
156 		}
157 	}
158 
159 	seq_printf(m, "SqThread:\t%d\n", sq_pid);
160 	seq_printf(m, "SqThreadCpu:\t%d\n", sq_cpu);
161 	seq_printf(m, "SqTotalTime:\t%llu\n", sq_total_time);
162 	seq_printf(m, "SqWorkTime:\t%llu\n", sq_work_time);
163 	seq_printf(m, "UserFiles:\t%u\n", ctx->file_table.data.nr);
164 	for (i = 0; i < ctx->file_table.data.nr; i++) {
165 		struct file *f = NULL;
166 
167 		if (ctx->file_table.data.nodes[i])
168 			f = io_slot_file(ctx->file_table.data.nodes[i]);
169 		if (f) {
170 			seq_printf(m, "%5u: ", i);
171 			seq_file_path(m, f, " \t\n\\");
172 			seq_puts(m, "\n");
173 		}
174 	}
175 	seq_printf(m, "UserBufs:\t%u\n", ctx->buf_table.nr);
176 	for (i = 0; i < ctx->buf_table.nr; i++) {
177 		struct io_mapped_ubuf *buf = NULL;
178 
179 		if (ctx->buf_table.nodes[i])
180 			buf = ctx->buf_table.nodes[i]->buf;
181 		if (buf)
182 			seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, buf->len);
183 		else
184 			seq_printf(m, "%5u: <none>\n", i);
185 	}
186 
187 	seq_puts(m, "PollList:\n");
188 	for (i = 0; i < (1U << ctx->cancel_table.hash_bits); i++) {
189 		struct io_hash_bucket *hb = &ctx->cancel_table.hbs[i];
190 		struct io_kiocb *req;
191 
192 		hlist_for_each_entry(req, &hb->list, hash_node)
193 			seq_printf(m, "  op=%d, task_works=%d\n", req->opcode,
194 					task_work_pending(req->tctx->task));
195 	}
196 
197 	seq_puts(m, "CqOverflowList:\n");
198 	spin_lock(&ctx->completion_lock);
199 	list_for_each_entry(ocqe, &ctx->cq_overflow_list, list) {
200 		struct io_uring_cqe *cqe = &ocqe->cqe;
201 
202 		seq_printf(m, "  user_data=%llu, res=%d, flags=%x\n",
203 			   cqe->user_data, cqe->res, cqe->flags);
204 
205 	}
206 	spin_unlock(&ctx->completion_lock);
207 	napi_show_fdinfo(ctx, m);
208 }
209 
210 /*
211  * Caller holds a reference to the file already, we don't need to do
212  * anything else to get an extra reference.
213  */
214 __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *file)
215 {
216 	struct io_ring_ctx *ctx = file->private_data;
217 
218 	/*
219 	 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
220 	 * since fdinfo case grabs it in the opposite direction of normal use
221 	 * cases.
222 	 */
223 	if (mutex_trylock(&ctx->uring_lock)) {
224 		__io_uring_show_fdinfo(ctx, m);
225 		mutex_unlock(&ctx->uring_lock);
226 	}
227 }
228