1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * (C) 2001 Clemson University and The University of Chicago
4 * (C) 2011 Omnibond Systems
5 *
6 * Changes by Acxiom Corporation to implement generic service_operation()
7 * function, Copyright Acxiom Corporation, 2005.
8 *
9 * See COPYING in top-level directory.
10 */
11
12 /*
13 * In-kernel waitqueue operations.
14 */
15
16 #include "protocol.h"
17 #include "orangefs-kernel.h"
18 #include "orangefs-bufmap.h"
19
20 static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op,
21 long timeout,
22 int flags)
23 __acquires(op->lock);
24 static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
25 __releases(op->lock);
26
27 /*
28 * What we do in this function is to walk the list of operations that are
29 * present in the request queue and mark them as purged.
30 * NOTE: This is called from the device close after client-core has
31 * guaranteed that no new operations could appear on the list since the
32 * client-core is anyway going to exit.
33 */
34 void purge_waiting_ops(void)
35 {
36 struct orangefs_kernel_op_s *op, *tmp;
37
38 spin_lock(&orangefs_request_list_lock);
39 list_for_each_entry_safe(op, tmp, &orangefs_request_list, list) {
40 gossip_debug(GOSSIP_WAIT_DEBUG,
41 "pvfs2-client-core: purging op tag %llu %s\n",
42 llu(op->tag),
43 get_opname_string(op));
44 set_op_state_purged(op);
45 gossip_debug(GOSSIP_DEV_DEBUG,
46 "%s: op:%s: op_state:%d: process:%s:\n",
47 __func__,
48 get_opname_string(op),
49 op->op_state,
50 current->comm);
51 }
52 spin_unlock(&orangefs_request_list_lock);
53 }
54
55 /*
56 * submits a ORANGEFS operation and waits for it to complete
57 *
58 * Note op->downcall.status will contain the status of the operation (in
59 * errno format), whether provided by pvfs2-client or a result of failure to
60 * service the operation. If the caller wishes to distinguish, then
61 * op->state can be checked to see if it was serviced or not.
62 *
63 * Returns contents of op->downcall.status for convenience
64 */
65 int service_operation(struct orangefs_kernel_op_s *op,
66 const char *op_name,
67 int flags)
68 {
69 long timeout = MAX_SCHEDULE_TIMEOUT;
70 int ret = 0;
71
72 DEFINE_WAIT(wait_entry);
73
74 op->upcall.tgid = current->tgid;
75 op->upcall.pid = current->pid;
76
77 retry_servicing:
78 op->downcall.status = 0;
79 gossip_debug(GOSSIP_WAIT_DEBUG,
80 "%s: %s op:%p: process:%s: pid:%d:\n",
81 __func__,
82 op_name,
83 op,
84 current->comm,
85 current->pid);
86
87 /*
88 * If ORANGEFS_OP_NO_MUTEX was set in flags, we need to avoid
89 * acquiring the request_mutex because we're servicing a
90 * high priority remount operation and the request_mutex is
91 * already taken.
92 */
93 if (!(flags & ORANGEFS_OP_NO_MUTEX)) {
94 if (flags & ORANGEFS_OP_INTERRUPTIBLE)
95 ret = mutex_lock_interruptible(&orangefs_request_mutex);
96 else
97 ret = mutex_lock_killable(&orangefs_request_mutex);
98 /*
99 * check to see if we were interrupted while waiting for
100 * mutex
101 */
102 if (ret < 0) {
103 op->downcall.status = ret;
104 gossip_debug(GOSSIP_WAIT_DEBUG,
105 "%s: service_operation interrupted.\n",
106 __func__);
107 return ret;
108 }
109 }
110
111 /* queue up the operation */
112 spin_lock(&orangefs_request_list_lock);
113 spin_lock(&op->lock);
114 set_op_state_waiting(op);
115 gossip_debug(GOSSIP_DEV_DEBUG,
116 "%s: op:%s: op_state:%d: process:%s:\n",
117 __func__,
118 get_opname_string(op),
119 op->op_state,
120 current->comm);
121 /* add high priority remount op to the front of the line. */
122 if (flags & ORANGEFS_OP_PRIORITY)
123 list_add(&op->list, &orangefs_request_list);
124 else
125 list_add_tail(&op->list, &orangefs_request_list);
126 spin_unlock(&op->lock);
127 wake_up_interruptible(&orangefs_request_list_waitq);
128 if (!__is_daemon_in_service()) {
129 gossip_debug(GOSSIP_WAIT_DEBUG,
130 "%s:client core is NOT in service.\n",
131 __func__);
132 /*
133 * Don't wait for the userspace component to return if
134 * the filesystem is being umounted anyway.
135 */
136 if (op->upcall.type == ORANGEFS_VFS_OP_FS_UMOUNT)
137 timeout = 0;
138 else
139 timeout = op_timeout_secs * HZ;
140 }
141 spin_unlock(&orangefs_request_list_lock);
142
143 if (!(flags & ORANGEFS_OP_NO_MUTEX))
144 mutex_unlock(&orangefs_request_mutex);
145
146 ret = wait_for_matching_downcall(op, timeout, flags);
147 gossip_debug(GOSSIP_WAIT_DEBUG,
148 "%s: wait_for_matching_downcall returned %d for %p\n",
149 __func__,
150 ret,
151 op);
152
153 /* got matching downcall; make sure status is in errno format */
154 if (!ret) {
155 spin_unlock(&op->lock);
156 op->downcall.status =
157 orangefs_normalize_to_errno(op->downcall.status);
158 ret = op->downcall.status;
159 goto out;
160 }
161
162 /* failed to get matching downcall */
163 if (ret == -ETIMEDOUT) {
164 gossip_err("%s: %s -- wait timed out; aborting attempt.\n",
165 __func__,
166 op_name);
167 }
168
169 /*
170 * remove a waiting op from the request list or
171 * remove an in-progress op from the in-progress list.
172 */
173 orangefs_clean_up_interrupted_operation(op);
174
175 op->downcall.status = ret;
176 /* retry if operation has not been serviced and if requested */
177 if (ret == -EAGAIN) {
178 op->attempts++;
179 timeout = op_timeout_secs * HZ;
180 gossip_debug(GOSSIP_WAIT_DEBUG,
181 "orangefs: tag %llu (%s)"
182 " -- operation to be retried (%d attempt)\n",
183 llu(op->tag),
184 op_name,
185 op->attempts);
186
187 /*
188 * io ops (ops that use the shared memory buffer) have
189 * to be returned to their caller for a retry. Other ops
190 * can just be recycled here.
191 */
192 if (!op->uses_shared_memory)
193 goto retry_servicing;
194 }
195
196 out:
197 gossip_debug(GOSSIP_WAIT_DEBUG,
198 "%s: %s returning: %d for %p.\n",
199 __func__,
200 op_name,
201 ret,
202 op);
203 return ret;
204 }
205
206 /* This can get called on an I/O op if it had a bad service_operation. */
207 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op)
208 {
209 u64 tag = op->tag;
210 if (!op_state_in_progress(op))
211 return false;
212
213 op->slot_to_free = op->upcall.req.io.buf_index;
214 memset(&op->upcall, 0, sizeof(op->upcall));
215 memset(&op->downcall, 0, sizeof(op->downcall));
216 op->upcall.type = ORANGEFS_VFS_OP_CANCEL;
217 op->upcall.req.cancel.op_tag = tag;
218 op->downcall.type = ORANGEFS_VFS_OP_INVALID;
219 op->downcall.status = -1;
220 orangefs_new_tag(op);
221
222 spin_lock(&orangefs_request_list_lock);
223 /* orangefs_request_list_lock is enough of a barrier here */
224 if (!__is_daemon_in_service()) {
225 spin_unlock(&orangefs_request_list_lock);
226 return false;
227 }
228 spin_lock(&op->lock);
229 set_op_state_waiting(op);
230 gossip_debug(GOSSIP_DEV_DEBUG,
231 "%s: op:%s: op_state:%d: process:%s:\n",
232 __func__,
233 get_opname_string(op),
234 op->op_state,
235 current->comm);
236 list_add(&op->list, &orangefs_request_list);
237 spin_unlock(&op->lock);
238 spin_unlock(&orangefs_request_list_lock);
239
240 gossip_debug(GOSSIP_WAIT_DEBUG,
241 "Attempting ORANGEFS operation cancellation of tag %llu\n",
242 llu(tag));
243 return true;
244 }
245
246 /*
247 * Change an op to the "given up" state and remove it from its list.
248 */
249 static void
250 orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
251 __releases(op->lock)
252 {
253 /*
254 * handle interrupted cases depending on what state we were in when
255 * the interruption is detected.
256 *
257 * Called with op->lock held.
258 */
259
260 /*
261 * List manipulation code elsewhere will ignore ops that
262 * have been given up upon.
263 */
264 op->op_state |= OP_VFS_STATE_GIVEN_UP;
265
266 if (list_empty(&op->list)) {
267 /* caught copying to/from daemon */
268 BUG_ON(op_state_serviced(op));
269 spin_unlock(&op->lock);
270 wait_for_completion(&op->waitq);
271 } else if (op_state_waiting(op)) {
272 /*
273 * upcall hasn't been read; remove op from upcall request
274 * list.
275 */
276 spin_unlock(&op->lock);
277 spin_lock(&orangefs_request_list_lock);
278 list_del_init(&op->list);
279 spin_unlock(&orangefs_request_list_lock);
280 gossip_debug(GOSSIP_WAIT_DEBUG,
281 "Interrupted: Removed op %p from request_list\n",
282 op);
283 } else if (op_state_in_progress(op)) {
284 /* op must be removed from the in progress htable */
285 spin_unlock(&op->lock);
286 spin_lock(&orangefs_htable_ops_in_progress_lock);
287 list_del_init(&op->list);
288 spin_unlock(&orangefs_htable_ops_in_progress_lock);
289 gossip_debug(GOSSIP_WAIT_DEBUG,
290 "Interrupted: Removed op %p"
291 " from htable_ops_in_progress\n",
292 op);
293 } else {
294 spin_unlock(&op->lock);
295 gossip_err("interrupted operation is in a weird state 0x%x\n",
296 op->op_state);
297 }
298 reinit_completion(&op->waitq);
299 }
300
301 /*
302 * Sleeps on waitqueue waiting for matching downcall.
303 * If client-core finishes servicing, then we are good to go.
304 * else if client-core exits, we get woken up here, and retry with a timeout
305 *
306 * When this call returns to the caller, the specified op will no
307 * longer be in either the in_progress hash table or on the request list.
308 *
309 * Returns 0 on success and -errno on failure
310 * Errors are:
311 * EAGAIN in case we want the caller to requeue and try again..
312 * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
313 * operation since client-core seems to be exiting too often
314 * or if we were interrupted.
315 *
316 * Returns with op->lock taken.
317 */
318 static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op,
319 long timeout,
320 int flags)
321 __acquires(op->lock)
322 {
323 long n;
324 int writeback = flags & ORANGEFS_OP_WRITEBACK,
325 interruptible = flags & ORANGEFS_OP_INTERRUPTIBLE;
326
327 /*
328 * There's a "schedule_timeout" inside of these wait
329 * primitives, during which the op is out of the hands of the
330 * user process that needs something done and is being
331 * manipulated by the client-core process.
332 */
333 if (writeback)
334 n = wait_for_completion_io_timeout(&op->waitq, timeout);
335 else if (!writeback && interruptible)
336 n = wait_for_completion_interruptible_timeout(&op->waitq,
337 timeout);
338 else /* !writeback && !interruptible but compiler complains */
339 n = wait_for_completion_killable_timeout(&op->waitq, timeout);
340
341 spin_lock(&op->lock);
342
343 if (op_state_serviced(op))
344 return 0;
345
346 if (unlikely(n < 0)) {
347 gossip_debug(GOSSIP_WAIT_DEBUG,
348 "%s: operation interrupted, tag %llu, %p\n",
349 __func__,
350 llu(op->tag),
351 op);
352 return -EINTR;
353 }
354 if (op_state_purged(op)) {
355 gossip_debug(GOSSIP_WAIT_DEBUG,
356 "%s: operation purged, tag %llu, %p, %d\n",
357 __func__,
358 llu(op->tag),
359 op,
360 op->attempts);
361 return (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
362 -EAGAIN :
363 -EIO;
364 }
365 /* must have timed out, then... */
366 gossip_debug(GOSSIP_WAIT_DEBUG,
367 "%s: operation timed out, tag %llu, %p, %d)\n",
368 __func__,
369 llu(op->tag),
370 op,
371 op->attempts);
372 return -ETIMEDOUT;
373 }