blob: abcfa3fa9992957006461c202c54380018a17787 [file] [log] [blame]
Mike Marshall1182fca2015-07-17 10:38:15 -04001/*
2 * (C) 2001 Clemson University and The University of Chicago
3 * (C) 2011 Omnibond Systems
4 *
5 * Changes by Acxiom Corporation to implement generic service_operation()
6 * function, Copyright Acxiom Corporation, 2005.
7 *
8 * See COPYING in top-level directory.
9 */
10
11/*
12 * In-kernel waitqueue operations.
13 */
14
15#include "protocol.h"
Mike Marshall575e9462015-12-04 12:56:14 -050016#include "orangefs-kernel.h"
17#include "orangefs-bufmap.h"
Mike Marshall1182fca2015-07-17 10:38:15 -040018
Al Viro05b39a82016-02-13 11:04:19 -050019static int wait_for_matching_downcall(struct orangefs_kernel_op_s *, long, bool);
Al Virod2d87a32016-02-13 10:15:22 -050020static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *);
Al Viroade3d782016-01-21 22:58:58 -050021
Mike Marshall1182fca2015-07-17 10:38:15 -040022/*
23 * What we do in this function is to walk the list of operations that are
24 * present in the request queue and mark them as purged.
25 * NOTE: This is called from the device close after client-core has
26 * guaranteed that no new operations could appear on the list since the
27 * client-core is anyway going to exit.
28 */
29void purge_waiting_ops(void)
30{
Yi Liu8bb8aef2015-11-24 15:12:14 -050031 struct orangefs_kernel_op_s *op;
Mike Marshall1182fca2015-07-17 10:38:15 -040032
Yi Liu8bb8aef2015-11-24 15:12:14 -050033 spin_lock(&orangefs_request_list_lock);
34 list_for_each_entry(op, &orangefs_request_list, list) {
Mike Marshall1182fca2015-07-17 10:38:15 -040035 gossip_debug(GOSSIP_WAIT_DEBUG,
36 "pvfs2-client-core: purging op tag %llu %s\n",
37 llu(op->tag),
38 get_opname_string(op));
Mike Marshall1182fca2015-07-17 10:38:15 -040039 set_op_state_purged(op);
Mike Marshall9d9e7ba2016-03-03 13:46:48 -050040 gossip_debug(GOSSIP_DEV_DEBUG,
41 "%s: op:%s: op_state:%d: process:%s:\n",
42 __func__,
43 get_opname_string(op),
44 op->op_state,
45 current->comm);
Mike Marshall1182fca2015-07-17 10:38:15 -040046 }
Yi Liu8bb8aef2015-11-24 15:12:14 -050047 spin_unlock(&orangefs_request_list_lock);
Mike Marshall1182fca2015-07-17 10:38:15 -040048}
49
50/*
Yi Liu8bb8aef2015-11-24 15:12:14 -050051 * submits a ORANGEFS operation and waits for it to complete
Mike Marshall1182fca2015-07-17 10:38:15 -040052 *
53 * Note op->downcall.status will contain the status of the operation (in
54 * errno format), whether provided by pvfs2-client or a result of failure to
55 * service the operation. If the caller wishes to distinguish, then
56 * op->state can be checked to see if it was serviced or not.
57 *
58 * Returns contents of op->downcall.status for convenience
59 */
Yi Liu8bb8aef2015-11-24 15:12:14 -050060int service_operation(struct orangefs_kernel_op_s *op,
Mike Marshall1182fca2015-07-17 10:38:15 -040061 const char *op_name,
62 int flags)
63{
Al Viro05b39a82016-02-13 11:04:19 -050064 long timeout = MAX_SCHEDULE_TIMEOUT;
Mike Marshall1182fca2015-07-17 10:38:15 -040065 int ret = 0;
66
Mike Marshallce6c4142015-12-14 14:54:46 -050067 DEFINE_WAIT(wait_entry);
Mike Marshall1182fca2015-07-17 10:38:15 -040068
69 op->upcall.tgid = current->tgid;
70 op->upcall.pid = current->pid;
71
72retry_servicing:
73 op->downcall.status = 0;
74 gossip_debug(GOSSIP_WAIT_DEBUG,
Mike Marshall52534872016-02-16 17:09:09 -050075 "%s: %s op:%p: process:%s: pid:%d:\n",
76 __func__,
Mike Marshall1182fca2015-07-17 10:38:15 -040077 op_name,
Mike Marshall52534872016-02-16 17:09:09 -050078 op,
Mike Marshall1182fca2015-07-17 10:38:15 -040079 current->comm,
80 current->pid);
81
Mike Marshalladcf34a2016-02-24 16:54:27 -050082 /*
83 * If ORANGEFS_OP_NO_MUTEX was set in flags, we need to avoid
Mike Marshallca9f5182016-02-26 10:21:12 -050084 * acquiring the request_mutex because we're servicing a
Mike Marshalladcf34a2016-02-24 16:54:27 -050085 * high priority remount operation and the request_mutex is
86 * already taken.
87 */
88 if (!(flags & ORANGEFS_OP_NO_MUTEX)) {
Al Viroc72f15b2016-02-13 10:49:24 -050089 if (flags & ORANGEFS_OP_INTERRUPTIBLE)
Martin Brandenburg1d503612016-08-16 11:38:14 -040090 ret = mutex_lock_interruptible(&orangefs_request_mutex);
Al Viroc72f15b2016-02-13 10:49:24 -050091 else
Martin Brandenburg1d503612016-08-16 11:38:14 -040092 ret = mutex_lock_killable(&orangefs_request_mutex);
Mike Marshall1182fca2015-07-17 10:38:15 -040093 /*
94 * check to see if we were interrupted while waiting for
Mike Marshalladcf34a2016-02-24 16:54:27 -050095 * mutex
Mike Marshall1182fca2015-07-17 10:38:15 -040096 */
97 if (ret < 0) {
Mike Marshall1182fca2015-07-17 10:38:15 -040098 op->downcall.status = ret;
99 gossip_debug(GOSSIP_WAIT_DEBUG,
Mike Marshallca9f5182016-02-26 10:21:12 -0500100 "%s: service_operation interrupted.\n",
101 __func__);
Mike Marshall1182fca2015-07-17 10:38:15 -0400102 return ret;
103 }
104 }
105
Al Viro98815ad2016-02-13 10:38:23 -0500106 /* queue up the operation */
107 spin_lock(&orangefs_request_list_lock);
108 spin_lock(&op->lock);
109 set_op_state_waiting(op);
Mike Marshall9d9e7ba2016-03-03 13:46:48 -0500110 gossip_debug(GOSSIP_DEV_DEBUG,
111 "%s: op:%s: op_state:%d: process:%s:\n",
112 __func__,
113 get_opname_string(op),
114 op->op_state,
115 current->comm);
Mike Marshalladcf34a2016-02-24 16:54:27 -0500116 /* add high priority remount op to the front of the line. */
Al Viro98815ad2016-02-13 10:38:23 -0500117 if (flags & ORANGEFS_OP_PRIORITY)
118 list_add(&op->list, &orangefs_request_list);
119 else
120 list_add_tail(&op->list, &orangefs_request_list);
121 spin_unlock(&op->lock);
122 wake_up_interruptible(&orangefs_request_list_waitq);
123 if (!__is_daemon_in_service()) {
Mike Marshall1182fca2015-07-17 10:38:15 -0400124 gossip_debug(GOSSIP_WAIT_DEBUG,
Al Viro98815ad2016-02-13 10:38:23 -0500125 "%s:client core is NOT in service.\n",
126 __func__);
Al Viro05b39a82016-02-13 11:04:19 -0500127 timeout = op_timeout_secs * HZ;
Mike Marshall1182fca2015-07-17 10:38:15 -0400128 }
Al Viro98815ad2016-02-13 10:38:23 -0500129 spin_unlock(&orangefs_request_list_lock);
Mike Marshall1182fca2015-07-17 10:38:15 -0400130
Mike Marshalladcf34a2016-02-24 16:54:27 -0500131 if (!(flags & ORANGEFS_OP_NO_MUTEX))
Martin Brandenburg1d503612016-08-16 11:38:14 -0400132 mutex_unlock(&orangefs_request_mutex);
Mike Marshall1182fca2015-07-17 10:38:15 -0400133
Al Viro05b39a82016-02-13 11:04:19 -0500134 ret = wait_for_matching_downcall(op, timeout,
135 flags & ORANGEFS_OP_INTERRUPTIBLE);
Mike Marshall52534872016-02-16 17:09:09 -0500136
137 gossip_debug(GOSSIP_WAIT_DEBUG,
138 "%s: wait_for_matching_downcall returned %d for %p\n",
139 __func__,
140 ret,
141 op);
142
Mike Marshallca9f5182016-02-26 10:21:12 -0500143 /* got matching downcall; make sure status is in errno format */
Al Viro05b39a82016-02-13 11:04:19 -0500144 if (!ret) {
Al Virod2d87a32016-02-13 10:15:22 -0500145 spin_unlock(&op->lock);
Mike Marshall1182fca2015-07-17 10:38:15 -0400146 op->downcall.status =
Yi Liu8bb8aef2015-11-24 15:12:14 -0500147 orangefs_normalize_to_errno(op->downcall.status);
Mike Marshall1182fca2015-07-17 10:38:15 -0400148 ret = op->downcall.status;
Al Viro05b39a82016-02-13 11:04:19 -0500149 goto out;
Mike Marshall1182fca2015-07-17 10:38:15 -0400150 }
151
Al Viro05b39a82016-02-13 11:04:19 -0500152 /* failed to get matching downcall */
153 if (ret == -ETIMEDOUT) {
Mike Marshalladcf34a2016-02-24 16:54:27 -0500154 gossip_err("%s: %s -- wait timed out; aborting attempt.\n",
155 __func__,
Al Viro05b39a82016-02-13 11:04:19 -0500156 op_name);
157 }
Mike Marshalladcf34a2016-02-24 16:54:27 -0500158
159 /*
Mike Marshallca9f5182016-02-26 10:21:12 -0500160 * remove a waiting op from the request list or
161 * remove an in-progress op from the in-progress list.
Mike Marshalladcf34a2016-02-24 16:54:27 -0500162 */
Al Viro05b39a82016-02-13 11:04:19 -0500163 orangefs_clean_up_interrupted_operation(op);
Mike Marshalladcf34a2016-02-24 16:54:27 -0500164
Al Viro05b39a82016-02-13 11:04:19 -0500165 op->downcall.status = ret;
Mike Marshall1182fca2015-07-17 10:38:15 -0400166 /* retry if operation has not been serviced and if requested */
Al Viro05b39a82016-02-13 11:04:19 -0500167 if (ret == -EAGAIN) {
168 op->attempts++;
169 timeout = op_timeout_secs * HZ;
Mike Marshall1182fca2015-07-17 10:38:15 -0400170 gossip_debug(GOSSIP_WAIT_DEBUG,
Yi Liu8bb8aef2015-11-24 15:12:14 -0500171 "orangefs: tag %llu (%s)"
Mike Marshall1182fca2015-07-17 10:38:15 -0400172 " -- operation to be retried (%d attempt)\n",
173 llu(op->tag),
174 op_name,
Al Viro05b39a82016-02-13 11:04:19 -0500175 op->attempts);
Mike Marshall1182fca2015-07-17 10:38:15 -0400176
Mike Marshalladcf34a2016-02-24 16:54:27 -0500177 /*
178 * io ops (ops that use the shared memory buffer) have
179 * to be returned to their caller for a retry. Other ops
180 * can just be recycled here.
181 */
Mike Marshall1182fca2015-07-17 10:38:15 -0400182 if (!op->uses_shared_memory)
Mike Marshall1182fca2015-07-17 10:38:15 -0400183 goto retry_servicing;
Mike Marshall1182fca2015-07-17 10:38:15 -0400184 }
185
Al Viro05b39a82016-02-13 11:04:19 -0500186out:
Mike Marshall1182fca2015-07-17 10:38:15 -0400187 gossip_debug(GOSSIP_WAIT_DEBUG,
Mike Marshall9d9e7ba2016-03-03 13:46:48 -0500188 "%s: %s returning: %d for %p.\n",
189 __func__,
Mike Marshall1182fca2015-07-17 10:38:15 -0400190 op_name,
191 ret,
192 op);
193 return ret;
194}
195
Mike Marshallca9f5182016-02-26 10:21:12 -0500196/* This can get called on an I/O op if it had a bad service_operation. */
Al Viro78699e22016-02-11 23:07:19 -0500197bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op)
198{
199 u64 tag = op->tag;
200 if (!op_state_in_progress(op))
201 return false;
202
203 op->slot_to_free = op->upcall.req.io.buf_index;
204 memset(&op->upcall, 0, sizeof(op->upcall));
205 memset(&op->downcall, 0, sizeof(op->downcall));
206 op->upcall.type = ORANGEFS_VFS_OP_CANCEL;
207 op->upcall.req.cancel.op_tag = tag;
208 op->downcall.type = ORANGEFS_VFS_OP_INVALID;
209 op->downcall.status = -1;
210 orangefs_new_tag(op);
211
212 spin_lock(&orangefs_request_list_lock);
213 /* orangefs_request_list_lock is enough of a barrier here */
214 if (!__is_daemon_in_service()) {
215 spin_unlock(&orangefs_request_list_lock);
216 return false;
217 }
Al Viro98815ad2016-02-13 10:38:23 -0500218 spin_lock(&op->lock);
219 set_op_state_waiting(op);
Mike Marshall9d9e7ba2016-03-03 13:46:48 -0500220 gossip_debug(GOSSIP_DEV_DEBUG,
221 "%s: op:%s: op_state:%d: process:%s:\n",
222 __func__,
223 get_opname_string(op),
224 op->op_state,
225 current->comm);
Al Viro98815ad2016-02-13 10:38:23 -0500226 list_add(&op->list, &orangefs_request_list);
227 spin_unlock(&op->lock);
Al Viro78699e22016-02-11 23:07:19 -0500228 spin_unlock(&orangefs_request_list_lock);
229
Mike Marshallca9f5182016-02-26 10:21:12 -0500230 gossip_debug(GOSSIP_WAIT_DEBUG,
Al Viro78699e22016-02-11 23:07:19 -0500231 "Attempting ORANGEFS operation cancellation of tag %llu\n",
232 llu(tag));
233 return true;
234}
235
Mike Marshallca9f5182016-02-26 10:21:12 -0500236/*
237 * Change an op to the "given up" state and remove it from its list.
238 */
239static void
240 orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
Mike Marshall1182fca2015-07-17 10:38:15 -0400241{
242 /*
243 * handle interrupted cases depending on what state we were in when
Mike Marshallca9f5182016-02-26 10:21:12 -0500244 * the interruption is detected.
Mike Marshall1182fca2015-07-17 10:38:15 -0400245 *
Al Viroeab9b382016-01-23 13:09:05 -0500246 * Called with op->lock held.
Mike Marshall1182fca2015-07-17 10:38:15 -0400247 */
Mike Marshallca9f5182016-02-26 10:21:12 -0500248
249 /*
250 * List manipulation code elsewhere will ignore ops that
251 * have been given up upon.
252 */
Al Viroed42fe02016-01-22 19:47:47 -0500253 op->op_state |= OP_VFS_STATE_GIVEN_UP;
Mike Marshallca9f5182016-02-26 10:21:12 -0500254
Al Viro05a50a52016-02-18 18:59:44 -0500255 if (list_empty(&op->list)) {
256 /* caught copying to/from daemon */
257 BUG_ON(op_state_serviced(op));
258 spin_unlock(&op->lock);
259 wait_for_completion(&op->waitq);
260 } else if (op_state_waiting(op)) {
Mike Marshall1182fca2015-07-17 10:38:15 -0400261 /*
262 * upcall hasn't been read; remove op from upcall request
263 * list.
264 */
265 spin_unlock(&op->lock);
Al Viroed42fe02016-01-22 19:47:47 -0500266 spin_lock(&orangefs_request_list_lock);
Al Viro05a50a52016-02-18 18:59:44 -0500267 list_del_init(&op->list);
Al Viroed42fe02016-01-22 19:47:47 -0500268 spin_unlock(&orangefs_request_list_lock);
Mike Marshall1182fca2015-07-17 10:38:15 -0400269 gossip_debug(GOSSIP_WAIT_DEBUG,
270 "Interrupted: Removed op %p from request_list\n",
271 op);
272 } else if (op_state_in_progress(op)) {
273 /* op must be removed from the in progress htable */
274 spin_unlock(&op->lock);
Martin Brandenburg1d503612016-08-16 11:38:14 -0400275 spin_lock(&orangefs_htable_ops_in_progress_lock);
Al Viro05a50a52016-02-18 18:59:44 -0500276 list_del_init(&op->list);
Martin Brandenburg1d503612016-08-16 11:38:14 -0400277 spin_unlock(&orangefs_htable_ops_in_progress_lock);
Mike Marshall1182fca2015-07-17 10:38:15 -0400278 gossip_debug(GOSSIP_WAIT_DEBUG,
279 "Interrupted: Removed op %p"
280 " from htable_ops_in_progress\n",
281 op);
Al Viro05a50a52016-02-18 18:59:44 -0500282 } else {
Mike Marshall1182fca2015-07-17 10:38:15 -0400283 spin_unlock(&op->lock);
284 gossip_err("interrupted operation is in a weird state 0x%x\n",
285 op->op_state);
Mike Marshall1182fca2015-07-17 10:38:15 -0400286 }
Al Virod2d87a32016-02-13 10:15:22 -0500287 reinit_completion(&op->waitq);
Mike Marshall1182fca2015-07-17 10:38:15 -0400288}
289
290/*
Mike Marshallca9f5182016-02-26 10:21:12 -0500291 * Sleeps on waitqueue waiting for matching downcall.
292 * If client-core finishes servicing, then we are good to go.
Mike Marshall1182fca2015-07-17 10:38:15 -0400293 * else if client-core exits, we get woken up here, and retry with a timeout
294 *
Mike Marshallca9f5182016-02-26 10:21:12 -0500295 * When this call returns to the caller, the specified op will no
296 * longer be in either the in_progress hash table or on the request list.
Mike Marshall1182fca2015-07-17 10:38:15 -0400297 *
298 * Returns 0 on success and -errno on failure
299 * Errors are:
300 * EAGAIN in case we want the caller to requeue and try again..
301 * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
302 * operation since client-core seems to be exiting too often
303 * or if we were interrupted.
Al Virod2d87a32016-02-13 10:15:22 -0500304 *
305 * Returns with op->lock taken.
Mike Marshall1182fca2015-07-17 10:38:15 -0400306 */
Al Viroc72f15b2016-02-13 10:49:24 -0500307static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op,
Al Viro05b39a82016-02-13 11:04:19 -0500308 long timeout,
Al Viroc72f15b2016-02-13 10:49:24 -0500309 bool interruptible)
Mike Marshall1182fca2015-07-17 10:38:15 -0400310{
Al Viro05b39a82016-02-13 11:04:19 -0500311 long n;
Al Viroc72f15b2016-02-13 10:49:24 -0500312
Mike Marshallca9f5182016-02-26 10:21:12 -0500313 /*
314 * There's a "schedule_timeout" inside of these wait
315 * primitives, during which the op is out of the hands of the
316 * user process that needs something done and is being
317 * manipulated by the client-core process.
318 */
Al Viroc72f15b2016-02-13 10:49:24 -0500319 if (interruptible)
Mike Marshalladcf34a2016-02-24 16:54:27 -0500320 n = wait_for_completion_interruptible_timeout(&op->waitq,
321 timeout);
Al Viroc72f15b2016-02-13 10:49:24 -0500322 else
323 n = wait_for_completion_killable_timeout(&op->waitq, timeout);
324
Mike Marshall1182fca2015-07-17 10:38:15 -0400325 spin_lock(&op->lock);
Mike Marshall1182fca2015-07-17 10:38:15 -0400326
Al Virod2d87a32016-02-13 10:15:22 -0500327 if (op_state_serviced(op))
328 return 0;
329
330 if (unlikely(n < 0)) {
331 gossip_debug(GOSSIP_WAIT_DEBUG,
Mike Marshall9d9e7ba2016-03-03 13:46:48 -0500332 "%s: operation interrupted, tag %llu, %p\n",
Al Virod2d87a32016-02-13 10:15:22 -0500333 __func__,
334 llu(op->tag),
335 op);
336 return -EINTR;
337 }
Al Virod2d87a32016-02-13 10:15:22 -0500338 if (op_state_purged(op)) {
339 gossip_debug(GOSSIP_WAIT_DEBUG,
Mike Marshall9d9e7ba2016-03-03 13:46:48 -0500340 "%s: operation purged, tag %llu, %p, %d\n",
Al Virod2d87a32016-02-13 10:15:22 -0500341 __func__,
342 llu(op->tag),
343 op,
344 op->attempts);
345 return (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
346 -EAGAIN :
347 -EIO;
348 }
349 /* must have timed out, then... */
350 gossip_debug(GOSSIP_WAIT_DEBUG,
Mike Marshall9d9e7ba2016-03-03 13:46:48 -0500351 "%s: operation timed out, tag %llu, %p, %d)\n",
Al Virod2d87a32016-02-13 10:15:22 -0500352 __func__,
353 llu(op->tag),
354 op,
355 op->attempts);
356 return -ETIMEDOUT;
Mike Marshall1182fca2015-07-17 10:38:15 -0400357}