xref: /openbmc/linux/block/blk-timeout.c (revision 8bd1369b)
1 /*
2  * Functions related to generic timeout handling of requests.
3  */
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/blkdev.h>
7 #include <linux/fault-inject.h>
8 
9 #include "blk.h"
10 #include "blk-mq.h"
11 
12 #ifdef CONFIG_FAIL_IO_TIMEOUT
13 
14 static DECLARE_FAULT_ATTR(fail_io_timeout);
15 
16 static int __init setup_fail_io_timeout(char *str)
17 {
18 	return setup_fault_attr(&fail_io_timeout, str);
19 }
20 __setup("fail_io_timeout=", setup_fail_io_timeout);
21 
22 int blk_should_fake_timeout(struct request_queue *q)
23 {
24 	if (!test_bit(QUEUE_FLAG_FAIL_IO, &q->queue_flags))
25 		return 0;
26 
27 	return should_fail(&fail_io_timeout, 1);
28 }
29 
30 static int __init fail_io_timeout_debugfs(void)
31 {
32 	struct dentry *dir = fault_create_debugfs_attr("fail_io_timeout",
33 						NULL, &fail_io_timeout);
34 
35 	return PTR_ERR_OR_ZERO(dir);
36 }
37 
38 late_initcall(fail_io_timeout_debugfs);
39 
40 ssize_t part_timeout_show(struct device *dev, struct device_attribute *attr,
41 			  char *buf)
42 {
43 	struct gendisk *disk = dev_to_disk(dev);
44 	int set = test_bit(QUEUE_FLAG_FAIL_IO, &disk->queue->queue_flags);
45 
46 	return sprintf(buf, "%d\n", set != 0);
47 }
48 
49 ssize_t part_timeout_store(struct device *dev, struct device_attribute *attr,
50 			   const char *buf, size_t count)
51 {
52 	struct gendisk *disk = dev_to_disk(dev);
53 	int val;
54 
55 	if (count) {
56 		struct request_queue *q = disk->queue;
57 		char *p = (char *) buf;
58 
59 		val = simple_strtoul(p, &p, 10);
60 		if (val)
61 			blk_queue_flag_set(QUEUE_FLAG_FAIL_IO, q);
62 		else
63 			blk_queue_flag_clear(QUEUE_FLAG_FAIL_IO, q);
64 	}
65 
66 	return count;
67 }
68 
69 #endif /* CONFIG_FAIL_IO_TIMEOUT */
70 
71 /*
72  * blk_delete_timer - Delete/cancel timer for a given function.
73  * @req:	request that we are canceling timer for
74  *
75  */
76 void blk_delete_timer(struct request *req)
77 {
78 	list_del_init(&req->timeout_list);
79 }
80 
81 static void blk_rq_timed_out(struct request *req)
82 {
83 	struct request_queue *q = req->q;
84 	enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
85 
86 	if (q->rq_timed_out_fn)
87 		ret = q->rq_timed_out_fn(req);
88 	switch (ret) {
89 	case BLK_EH_RESET_TIMER:
90 		blk_add_timer(req);
91 		blk_clear_rq_complete(req);
92 		break;
93 	case BLK_EH_DONE:
94 		/*
95 		 * LLD handles this for now but in the future
96 		 * we can send a request msg to abort the command
97 		 * and we can move more of the generic scsi eh code to
98 		 * the blk layer.
99 		 */
100 		break;
101 	default:
102 		printk(KERN_ERR "block: bad eh return: %d\n", ret);
103 		break;
104 	}
105 }
106 
107 static void blk_rq_check_expired(struct request *rq, unsigned long *next_timeout,
108 			  unsigned int *next_set)
109 {
110 	const unsigned long deadline = blk_rq_deadline(rq);
111 
112 	if (time_after_eq(jiffies, deadline)) {
113 		list_del_init(&rq->timeout_list);
114 
115 		/*
116 		 * Check if we raced with end io completion
117 		 */
118 		if (!blk_mark_rq_complete(rq))
119 			blk_rq_timed_out(rq);
120 	} else if (!*next_set || time_after(*next_timeout, deadline)) {
121 		*next_timeout = deadline;
122 		*next_set = 1;
123 	}
124 }
125 
126 void blk_timeout_work(struct work_struct *work)
127 {
128 	struct request_queue *q =
129 		container_of(work, struct request_queue, timeout_work);
130 	unsigned long flags, next = 0;
131 	struct request *rq, *tmp;
132 	int next_set = 0;
133 
134 	spin_lock_irqsave(q->queue_lock, flags);
135 
136 	list_for_each_entry_safe(rq, tmp, &q->timeout_list, timeout_list)
137 		blk_rq_check_expired(rq, &next, &next_set);
138 
139 	if (next_set)
140 		mod_timer(&q->timeout, round_jiffies_up(next));
141 
142 	spin_unlock_irqrestore(q->queue_lock, flags);
143 }
144 
145 /**
146  * blk_abort_request -- Request request recovery for the specified command
147  * @req:	pointer to the request of interest
148  *
149  * This function requests that the block layer start recovery for the
150  * request by deleting the timer and calling the q's timeout function.
151  * LLDDs who implement their own error recovery MAY ignore the timeout
152  * event if they generated blk_abort_req. Must hold queue lock.
153  */
154 void blk_abort_request(struct request *req)
155 {
156 	if (req->q->mq_ops) {
157 		/*
158 		 * All we need to ensure is that timeout scan takes place
159 		 * immediately and that scan sees the new timeout value.
160 		 * No need for fancy synchronizations.
161 		 */
162 		blk_rq_set_deadline(req, jiffies);
163 		kblockd_schedule_work(&req->q->timeout_work);
164 	} else {
165 		if (blk_mark_rq_complete(req))
166 			return;
167 		blk_delete_timer(req);
168 		blk_rq_timed_out(req);
169 	}
170 }
171 EXPORT_SYMBOL_GPL(blk_abort_request);
172 
173 unsigned long blk_rq_timeout(unsigned long timeout)
174 {
175 	unsigned long maxt;
176 
177 	maxt = round_jiffies_up(jiffies + BLK_MAX_TIMEOUT);
178 	if (time_after(timeout, maxt))
179 		timeout = maxt;
180 
181 	return timeout;
182 }
183 
184 /**
185  * blk_add_timer - Start timeout timer for a single request
186  * @req:	request that is about to start running.
187  *
188  * Notes:
189  *    Each request has its own timer, and as it is added to the queue, we
190  *    set up the timer. When the request completes, we cancel the timer.
191  */
192 void blk_add_timer(struct request *req)
193 {
194 	struct request_queue *q = req->q;
195 	unsigned long expiry;
196 
197 	if (!q->mq_ops)
198 		lockdep_assert_held(q->queue_lock);
199 
200 	/* blk-mq has its own handler, so we don't need ->rq_timed_out_fn */
201 	if (!q->mq_ops && !q->rq_timed_out_fn)
202 		return;
203 
204 	BUG_ON(!list_empty(&req->timeout_list));
205 
206 	/*
207 	 * Some LLDs, like scsi, peek at the timeout to prevent a
208 	 * command from being retried forever.
209 	 */
210 	if (!req->timeout)
211 		req->timeout = q->rq_timeout;
212 
213 	req->rq_flags &= ~RQF_TIMED_OUT;
214 	blk_rq_set_deadline(req, jiffies + req->timeout);
215 
216 	/*
217 	 * Only the non-mq case needs to add the request to a protected list.
218 	 * For the mq case we simply scan the tag map.
219 	 */
220 	if (!q->mq_ops)
221 		list_add_tail(&req->timeout_list, &req->q->timeout_list);
222 
223 	/*
224 	 * If the timer isn't already pending or this timeout is earlier
225 	 * than an existing one, modify the timer. Round up to next nearest
226 	 * second.
227 	 */
228 	expiry = blk_rq_timeout(round_jiffies_up(blk_rq_deadline(req)));
229 
230 	if (!timer_pending(&q->timeout) ||
231 	    time_before(expiry, q->timeout.expires)) {
232 		unsigned long diff = q->timeout.expires - expiry;
233 
234 		/*
235 		 * Due to added timer slack to group timers, the timer
236 		 * will often be a little in front of what we asked for.
237 		 * So apply some tolerance here too, otherwise we keep
238 		 * modifying the timer because expires for value X
239 		 * will be X + something.
240 		 */
241 		if (!timer_pending(&q->timeout) || (diff >= HZ / 2))
242 			mod_timer(&q->timeout, expiry);
243 	}
244 
245 }
246