xref: /openbmc/linux/block/blk-timeout.c (revision 94c7b6fc)
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 		spin_lock_irq(q->queue_lock);
61 		if (val)
62 			queue_flag_set(QUEUE_FLAG_FAIL_IO, q);
63 		else
64 			queue_flag_clear(QUEUE_FLAG_FAIL_IO, q);
65 		spin_unlock_irq(q->queue_lock);
66 	}
67 
68 	return count;
69 }
70 
71 #endif /* CONFIG_FAIL_IO_TIMEOUT */
72 
73 /*
74  * blk_delete_timer - Delete/cancel timer for a given function.
75  * @req:	request that we are canceling timer for
76  *
77  */
78 void blk_delete_timer(struct request *req)
79 {
80 	list_del_init(&req->timeout_list);
81 }
82 
83 static void blk_rq_timed_out(struct request *req)
84 {
85 	struct request_queue *q = req->q;
86 	enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
87 
88 	if (q->rq_timed_out_fn)
89 		ret = q->rq_timed_out_fn(req);
90 	switch (ret) {
91 	case BLK_EH_HANDLED:
92 		/* Can we use req->errors here? */
93 		if (q->mq_ops)
94 			__blk_mq_complete_request(req);
95 		else
96 			__blk_complete_request(req);
97 		break;
98 	case BLK_EH_RESET_TIMER:
99 		blk_add_timer(req);
100 		blk_clear_rq_complete(req);
101 		break;
102 	case BLK_EH_NOT_HANDLED:
103 		/*
104 		 * LLD handles this for now but in the future
105 		 * we can send a request msg to abort the command
106 		 * and we can move more of the generic scsi eh code to
107 		 * the blk layer.
108 		 */
109 		break;
110 	default:
111 		printk(KERN_ERR "block: bad eh return: %d\n", ret);
112 		break;
113 	}
114 }
115 
116 void blk_rq_check_expired(struct request *rq, unsigned long *next_timeout,
117 			  unsigned int *next_set)
118 {
119 	if (time_after_eq(jiffies, rq->deadline)) {
120 		list_del_init(&rq->timeout_list);
121 
122 		/*
123 		 * Check if we raced with end io completion
124 		 */
125 		if (!blk_mark_rq_complete(rq))
126 			blk_rq_timed_out(rq);
127 	} else if (!*next_set || time_after(*next_timeout, rq->deadline)) {
128 		*next_timeout = rq->deadline;
129 		*next_set = 1;
130 	}
131 }
132 
133 void blk_rq_timed_out_timer(unsigned long data)
134 {
135 	struct request_queue *q = (struct request_queue *) data;
136 	unsigned long flags, next = 0;
137 	struct request *rq, *tmp;
138 	int next_set = 0;
139 
140 	spin_lock_irqsave(q->queue_lock, flags);
141 
142 	list_for_each_entry_safe(rq, tmp, &q->timeout_list, timeout_list)
143 		blk_rq_check_expired(rq, &next, &next_set);
144 
145 	if (next_set)
146 		mod_timer(&q->timeout, round_jiffies_up(next));
147 
148 	spin_unlock_irqrestore(q->queue_lock, flags);
149 }
150 
151 /**
152  * blk_abort_request -- Request request recovery for the specified command
153  * @req:	pointer to the request of interest
154  *
155  * This function requests that the block layer start recovery for the
156  * request by deleting the timer and calling the q's timeout function.
157  * LLDDs who implement their own error recovery MAY ignore the timeout
158  * event if they generated blk_abort_req. Must hold queue lock.
159  */
160 void blk_abort_request(struct request *req)
161 {
162 	if (blk_mark_rq_complete(req))
163 		return;
164 	blk_delete_timer(req);
165 	blk_rq_timed_out(req);
166 }
167 EXPORT_SYMBOL_GPL(blk_abort_request);
168 
169 unsigned long blk_rq_timeout(unsigned long timeout)
170 {
171 	unsigned long maxt;
172 
173 	maxt = round_jiffies_up(jiffies + BLK_MAX_TIMEOUT);
174 	if (time_after(timeout, maxt))
175 		timeout = maxt;
176 
177 	return timeout;
178 }
179 
180 /**
181  * blk_add_timer - Start timeout timer for a single request
182  * @req:	request that is about to start running.
183  *
184  * Notes:
185  *    Each request has its own timer, and as it is added to the queue, we
186  *    set up the timer. When the request completes, we cancel the timer.
187  */
188 void blk_add_timer(struct request *req)
189 {
190 	struct request_queue *q = req->q;
191 	unsigned long expiry;
192 
193 	if (!q->rq_timed_out_fn)
194 		return;
195 
196 	BUG_ON(!list_empty(&req->timeout_list));
197 
198 	/*
199 	 * Some LLDs, like scsi, peek at the timeout to prevent a
200 	 * command from being retried forever.
201 	 */
202 	if (!req->timeout)
203 		req->timeout = q->rq_timeout;
204 
205 	req->deadline = jiffies + req->timeout;
206 	if (!q->mq_ops)
207 		list_add_tail(&req->timeout_list, &req->q->timeout_list);
208 
209 	/*
210 	 * If the timer isn't already pending or this timeout is earlier
211 	 * than an existing one, modify the timer. Round up to next nearest
212 	 * second.
213 	 */
214 	expiry = blk_rq_timeout(round_jiffies_up(req->deadline));
215 
216 	if (!timer_pending(&q->timeout) ||
217 	    time_before(expiry, q->timeout.expires)) {
218 		unsigned long diff = q->timeout.expires - expiry;
219 
220 		/*
221 		 * Due to added timer slack to group timers, the timer
222 		 * will often be a little in front of what we asked for.
223 		 * So apply some tolerance here too, otherwise we keep
224 		 * modifying the timer because expires for value X
225 		 * will be X + something.
226 		 */
227 		if (!timer_pending(&q->timeout) || (diff >= HZ / 2))
228 			mod_timer(&q->timeout, expiry);
229 	}
230 
231 }
232