blob: 7dc0f6299a52bbd8a0b236c7fa29d9de738eab72 [file] [log] [blame]
Sage Weilf24e9982009-10-06 11:31:10 -07001#include "ceph_debug.h"
2
3#include <linux/err.h>
4#include <linux/highmem.h>
5#include <linux/mm.h>
6#include <linux/pagemap.h>
7#include <linux/slab.h>
8#include <linux/uaccess.h>
9
10#include "super.h"
11#include "osd_client.h"
12#include "messenger.h"
13#include "decode.h"
14
15const static struct ceph_connection_operations osd_con_ops;
16
17static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
18
19/*
20 * Implement client access to distributed object storage cluster.
21 *
22 * All data objects are stored within a cluster/cloud of OSDs, or
23 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
24 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
25 * remote daemons serving up and coordinating consistent and safe
26 * access to storage.
27 *
28 * Cluster membership and the mapping of data objects onto storage devices
29 * are described by the osd map.
30 *
31 * We keep track of pending OSD requests (read, write), resubmit
32 * requests to different OSDs when the cluster topology/data layout
33 * change, or retry the affected requests when the communications
34 * channel with an OSD is reset.
35 */
36
37/*
38 * calculate the mapping of a file extent onto an object, and fill out the
39 * request accordingly. shorten extent as necessary if it crosses an
40 * object boundary.
41 *
42 * fill osd op in request message.
43 */
44static void calc_layout(struct ceph_osd_client *osdc,
45 struct ceph_vino vino, struct ceph_file_layout *layout,
46 u64 off, u64 *plen,
47 struct ceph_osd_request *req)
48{
49 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
50 struct ceph_osd_op *op = (void *)(reqhead + 1);
51 u64 orig_len = *plen;
52 u64 objoff, objlen; /* extent in object */
53 u64 bno;
54
55 reqhead->snapid = cpu_to_le64(vino.snap);
56
57 /* object extent? */
58 ceph_calc_file_object_mapping(layout, off, plen, &bno,
59 &objoff, &objlen);
60 if (*plen < orig_len)
61 dout(" skipping last %llu, final file extent %llu~%llu\n",
62 orig_len - *plen, off, *plen);
63
64 sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
65 req->r_oid_len = strlen(req->r_oid);
66
67 op->extent.offset = cpu_to_le64(objoff);
68 op->extent.length = cpu_to_le64(objlen);
69 req->r_num_pages = calc_pages_for(off, *plen);
70
71 dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
72 req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
73}
74
75
76/*
77 * requests
78 */
79void ceph_osdc_put_request(struct ceph_osd_request *req)
80{
81 dout("osdc put_request %p %d -> %d\n", req, atomic_read(&req->r_ref),
82 atomic_read(&req->r_ref)-1);
83 BUG_ON(atomic_read(&req->r_ref) <= 0);
84 if (atomic_dec_and_test(&req->r_ref)) {
85 if (req->r_request)
86 ceph_msg_put(req->r_request);
87 if (req->r_reply)
88 ceph_msg_put(req->r_reply);
89 if (req->r_own_pages)
90 ceph_release_page_vector(req->r_pages,
91 req->r_num_pages);
92 ceph_put_snap_context(req->r_snapc);
93 if (req->r_mempool)
94 mempool_free(req, req->r_osdc->req_mempool);
95 else
96 kfree(req);
97 }
98}
99
100/*
101 * build new request AND message, calculate layout, and adjust file
102 * extent as needed.
103 *
104 * if the file was recently truncated, we include information about its
105 * old and new size so that the object can be updated appropriately. (we
106 * avoid synchronously deleting truncated objects because it's slow.)
107 *
108 * if @do_sync, include a 'startsync' command so that the osd will flush
109 * data quickly.
110 */
111struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
112 struct ceph_file_layout *layout,
113 struct ceph_vino vino,
114 u64 off, u64 *plen,
115 int opcode, int flags,
116 struct ceph_snap_context *snapc,
117 int do_sync,
118 u32 truncate_seq,
119 u64 truncate_size,
120 struct timespec *mtime,
121 bool use_mempool, int num_reply)
122{
123 struct ceph_osd_request *req;
124 struct ceph_msg *msg;
125 struct ceph_osd_request_head *head;
126 struct ceph_osd_op *op;
127 void *p;
128 int do_trunc = truncate_seq && (off + *plen > truncate_size);
129 int num_op = 1 + do_sync + do_trunc;
130 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
131 int err, i;
132 u64 prevofs;
133
134 if (use_mempool) {
135 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
136 memset(req, 0, sizeof(*req));
137 } else {
138 req = kzalloc(sizeof(*req), GFP_NOFS);
139 }
140 if (req == NULL)
141 return ERR_PTR(-ENOMEM);
142
143 err = ceph_msgpool_resv(&osdc->msgpool_op_reply, num_reply);
144 if (err) {
145 ceph_osdc_put_request(req);
146 return ERR_PTR(-ENOMEM);
147 }
148
149 req->r_osdc = osdc;
150 req->r_mempool = use_mempool;
151 atomic_set(&req->r_ref, 1);
152 init_completion(&req->r_completion);
153 init_completion(&req->r_safe_completion);
154 INIT_LIST_HEAD(&req->r_unsafe_item);
155 req->r_flags = flags;
156
157 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
158
159 /* create message; allow space for oid */
160 msg_size += 40;
161 if (snapc)
162 msg_size += sizeof(u64) * snapc->num_snaps;
163 if (use_mempool)
Sage Weil8f3bc052009-10-14 17:36:07 -0700164 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
Sage Weilf24e9982009-10-06 11:31:10 -0700165 else
166 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
167 if (IS_ERR(msg)) {
168 ceph_msgpool_resv(&osdc->msgpool_op_reply, num_reply);
169 ceph_osdc_put_request(req);
170 return ERR_PTR(PTR_ERR(msg));
171 }
172 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
173 memset(msg->front.iov_base, 0, msg->front.iov_len);
174 head = msg->front.iov_base;
175 op = (void *)(head + 1);
176 p = (void *)(op + num_op);
177
178 req->r_request = msg;
179 req->r_snapc = ceph_get_snap_context(snapc);
180
181 head->client_inc = cpu_to_le32(1); /* always, for now. */
182 head->flags = cpu_to_le32(flags);
183 if (flags & CEPH_OSD_FLAG_WRITE)
184 ceph_encode_timespec(&head->mtime, mtime);
185 head->num_ops = cpu_to_le16(num_op);
186 op->op = cpu_to_le16(opcode);
187
188 /* calculate max write size */
189 calc_layout(osdc, vino, layout, off, plen, req);
190 req->r_file_layout = *layout; /* keep a copy */
191
192 if (flags & CEPH_OSD_FLAG_WRITE) {
193 req->r_request->hdr.data_off = cpu_to_le16(off);
194 req->r_request->hdr.data_len = cpu_to_le32(*plen);
195 op->payload_len = cpu_to_le32(*plen);
196 }
197
198 /* fill in oid */
199 head->object_len = cpu_to_le32(req->r_oid_len);
200 memcpy(p, req->r_oid, req->r_oid_len);
201 p += req->r_oid_len;
202
203 /* additional ops */
204 if (do_trunc) {
205 op++;
206 op->op = cpu_to_le16(opcode == CEPH_OSD_OP_READ ?
207 CEPH_OSD_OP_MASKTRUNC : CEPH_OSD_OP_SETTRUNC);
208 op->trunc.truncate_seq = cpu_to_le32(truncate_seq);
209 prevofs = le64_to_cpu((op-1)->extent.offset);
210 op->trunc.truncate_size = cpu_to_le64(truncate_size -
211 (off-prevofs));
212 }
213 if (do_sync) {
214 op++;
215 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
216 }
217 if (snapc) {
218 head->snap_seq = cpu_to_le64(snapc->seq);
219 head->num_snaps = cpu_to_le32(snapc->num_snaps);
220 for (i = 0; i < snapc->num_snaps; i++) {
221 put_unaligned_le64(snapc->snaps[i], p);
222 p += sizeof(u64);
223 }
224 }
225
226 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
227 return req;
228}
229
230/*
231 * We keep osd requests in an rbtree, sorted by ->r_tid.
232 */
233static void __insert_request(struct ceph_osd_client *osdc,
234 struct ceph_osd_request *new)
235{
236 struct rb_node **p = &osdc->requests.rb_node;
237 struct rb_node *parent = NULL;
238 struct ceph_osd_request *req = NULL;
239
240 while (*p) {
241 parent = *p;
242 req = rb_entry(parent, struct ceph_osd_request, r_node);
243 if (new->r_tid < req->r_tid)
244 p = &(*p)->rb_left;
245 else if (new->r_tid > req->r_tid)
246 p = &(*p)->rb_right;
247 else
248 BUG();
249 }
250
251 rb_link_node(&new->r_node, parent, p);
252 rb_insert_color(&new->r_node, &osdc->requests);
253}
254
255static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
256 u64 tid)
257{
258 struct ceph_osd_request *req;
259 struct rb_node *n = osdc->requests.rb_node;
260
261 while (n) {
262 req = rb_entry(n, struct ceph_osd_request, r_node);
263 if (tid < req->r_tid)
264 n = n->rb_left;
265 else if (tid > req->r_tid)
266 n = n->rb_right;
267 else
268 return req;
269 }
270 return NULL;
271}
272
273static struct ceph_osd_request *
274__lookup_request_ge(struct ceph_osd_client *osdc,
275 u64 tid)
276{
277 struct ceph_osd_request *req;
278 struct rb_node *n = osdc->requests.rb_node;
279
280 while (n) {
281 req = rb_entry(n, struct ceph_osd_request, r_node);
282 if (tid < req->r_tid) {
283 if (!n->rb_left)
284 return req;
285 n = n->rb_left;
286 } else if (tid > req->r_tid) {
287 n = n->rb_right;
288 } else {
289 return req;
290 }
291 }
292 return NULL;
293}
294
295
296/*
Sage Weil81b024e2009-10-09 10:29:18 -0700297 * If the osd connection drops, we need to resubmit all requests.
Sage Weilf24e9982009-10-06 11:31:10 -0700298 */
299static void osd_reset(struct ceph_connection *con)
300{
301 struct ceph_osd *osd = con->private;
302 struct ceph_osd_client *osdc;
303
304 if (!osd)
305 return;
306 dout("osd_reset osd%d\n", osd->o_osd);
307 osdc = osd->o_osdc;
308 osd->o_incarnation++;
309 down_read(&osdc->map_sem);
310 kick_requests(osdc, osd);
311 up_read(&osdc->map_sem);
312}
313
314/*
315 * Track open sessions with osds.
316 */
317static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
318{
319 struct ceph_osd *osd;
320
321 osd = kzalloc(sizeof(*osd), GFP_NOFS);
322 if (!osd)
323 return NULL;
324
325 atomic_set(&osd->o_ref, 1);
326 osd->o_osdc = osdc;
327 INIT_LIST_HEAD(&osd->o_requests);
328 osd->o_incarnation = 1;
329
330 ceph_con_init(osdc->client->msgr, &osd->o_con);
331 osd->o_con.private = osd;
332 osd->o_con.ops = &osd_con_ops;
333 osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
334 return osd;
335}
336
337static struct ceph_osd *get_osd(struct ceph_osd *osd)
338{
339 if (atomic_inc_not_zero(&osd->o_ref)) {
340 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
341 atomic_read(&osd->o_ref));
342 return osd;
343 } else {
344 dout("get_osd %p FAIL\n", osd);
345 return NULL;
346 }
347}
348
349static void put_osd(struct ceph_osd *osd)
350{
351 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
352 atomic_read(&osd->o_ref) - 1);
353 if (atomic_dec_and_test(&osd->o_ref)) {
354 ceph_con_shutdown(&osd->o_con);
355 kfree(osd);
356 }
357}
358
359/*
360 * remove an osd from our map
361 */
362static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
363{
364 dout("remove_osd %p\n", osd);
365 BUG_ON(!list_empty(&osd->o_requests));
366 rb_erase(&osd->o_node, &osdc->osds);
367 ceph_con_close(&osd->o_con);
368 put_osd(osd);
369}
370
371/*
372 * reset osd connect
373 */
374static int reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
375{
376 int ret = 0;
377
378 dout("reset_osd %p osd%d\n", osd, osd->o_osd);
379 if (list_empty(&osd->o_requests)) {
380 remove_osd(osdc, osd);
381 } else {
382 ceph_con_close(&osd->o_con);
383 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
384 osd->o_incarnation++;
385 }
386 return ret;
387}
388
389static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
390{
391 struct rb_node **p = &osdc->osds.rb_node;
392 struct rb_node *parent = NULL;
393 struct ceph_osd *osd = NULL;
394
395 while (*p) {
396 parent = *p;
397 osd = rb_entry(parent, struct ceph_osd, o_node);
398 if (new->o_osd < osd->o_osd)
399 p = &(*p)->rb_left;
400 else if (new->o_osd > osd->o_osd)
401 p = &(*p)->rb_right;
402 else
403 BUG();
404 }
405
406 rb_link_node(&new->o_node, parent, p);
407 rb_insert_color(&new->o_node, &osdc->osds);
408}
409
410static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
411{
412 struct ceph_osd *osd;
413 struct rb_node *n = osdc->osds.rb_node;
414
415 while (n) {
416 osd = rb_entry(n, struct ceph_osd, o_node);
417 if (o < osd->o_osd)
418 n = n->rb_left;
419 else if (o > osd->o_osd)
420 n = n->rb_right;
421 else
422 return osd;
423 }
424 return NULL;
425}
426
427
428/*
429 * Register request, assign tid. If this is the first request, set up
430 * the timeout event.
431 */
432static void register_request(struct ceph_osd_client *osdc,
433 struct ceph_osd_request *req)
434{
435 struct ceph_osd_request_head *head = req->r_request->front.iov_base;
436
437 mutex_lock(&osdc->request_mutex);
438 req->r_tid = ++osdc->last_tid;
439 head->tid = cpu_to_le64(req->r_tid);
440
441 dout("register_request %p tid %lld\n", req, req->r_tid);
442 __insert_request(osdc, req);
443 ceph_osdc_get_request(req);
444 osdc->num_requests++;
445
446 req->r_timeout_stamp =
Sage Weil6b805182009-10-27 11:50:50 -0700447 jiffies + osdc->client->mount_args->osd_timeout*HZ;
Sage Weilf24e9982009-10-06 11:31:10 -0700448
449 if (osdc->num_requests == 1) {
450 osdc->timeout_tid = req->r_tid;
451 dout(" timeout on tid %llu at %lu\n", req->r_tid,
452 req->r_timeout_stamp);
453 schedule_delayed_work(&osdc->timeout_work,
454 round_jiffies_relative(req->r_timeout_stamp - jiffies));
455 }
456 mutex_unlock(&osdc->request_mutex);
457}
458
459/*
460 * called under osdc->request_mutex
461 */
462static void __unregister_request(struct ceph_osd_client *osdc,
463 struct ceph_osd_request *req)
464{
465 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
466 rb_erase(&req->r_node, &osdc->requests);
467 osdc->num_requests--;
468
Sage Weil0ba64782009-10-08 16:57:16 -0700469 if (req->r_osd) {
470 /* make sure the original request isn't in flight. */
471 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
472
473 list_del_init(&req->r_osd_item);
474 if (list_empty(&req->r_osd->o_requests))
475 remove_osd(osdc, req->r_osd);
476 req->r_osd = NULL;
477 }
Sage Weilf24e9982009-10-06 11:31:10 -0700478
479 ceph_osdc_put_request(req);
480
481 if (req->r_tid == osdc->timeout_tid) {
482 if (osdc->num_requests == 0) {
483 dout("no requests, canceling timeout\n");
484 osdc->timeout_tid = 0;
485 cancel_delayed_work(&osdc->timeout_work);
486 } else {
487 req = rb_entry(rb_first(&osdc->requests),
488 struct ceph_osd_request, r_node);
489 osdc->timeout_tid = req->r_tid;
490 dout("rescheduled timeout on tid %llu at %lu\n",
491 req->r_tid, req->r_timeout_stamp);
492 schedule_delayed_work(&osdc->timeout_work,
493 round_jiffies_relative(req->r_timeout_stamp -
494 jiffies));
495 }
496 }
497}
498
499/*
500 * Cancel a previously queued request message
501 */
502static void __cancel_request(struct ceph_osd_request *req)
503{
504 if (req->r_sent) {
505 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
506 req->r_sent = 0;
507 }
508}
509
510/*
511 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
512 * (as needed), and set the request r_osd appropriately. If there is
513 * no up osd, set r_osd to NULL.
514 *
515 * Return 0 if unchanged, 1 if changed, or negative on error.
516 *
517 * Caller should hold map_sem for read and request_mutex.
518 */
519static int __map_osds(struct ceph_osd_client *osdc,
520 struct ceph_osd_request *req)
521{
522 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
523 union ceph_pg pgid;
524 int o = -1;
525 int err;
526 struct ceph_osd *newosd = NULL;
527
528 dout("map_osds %p tid %lld\n", req, req->r_tid);
529 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
530 &req->r_file_layout, osdc->osdmap);
531 if (err)
532 return err;
533 pgid.pg64 = le64_to_cpu(reqhead->layout.ol_pgid);
534 o = ceph_calc_pg_primary(osdc->osdmap, pgid);
535
536 if ((req->r_osd && req->r_osd->o_osd == o &&
537 req->r_sent >= req->r_osd->o_incarnation) ||
538 (req->r_osd == NULL && o == -1))
539 return 0; /* no change */
540
541 dout("map_osds tid %llu pgid %llx pool %d osd%d (was osd%d)\n",
542 req->r_tid, pgid.pg64, pgid.pg.pool, o,
543 req->r_osd ? req->r_osd->o_osd : -1);
544
545 if (req->r_osd) {
546 __cancel_request(req);
547 list_del_init(&req->r_osd_item);
548 if (list_empty(&req->r_osd->o_requests)) {
549 /* try to re-use r_osd if possible */
550 newosd = get_osd(req->r_osd);
551 remove_osd(osdc, newosd);
552 }
553 req->r_osd = NULL;
554 }
555
556 req->r_osd = __lookup_osd(osdc, o);
557 if (!req->r_osd && o >= 0) {
558 if (newosd) {
559 req->r_osd = newosd;
560 newosd = NULL;
561 } else {
562 err = -ENOMEM;
563 req->r_osd = create_osd(osdc);
564 if (!req->r_osd)
565 goto out;
566 }
567
568 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
569 req->r_osd->o_osd = o;
570 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
571 __insert_osd(osdc, req->r_osd);
572
573 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
574 }
575
576 if (req->r_osd)
577 list_add(&req->r_osd_item, &req->r_osd->o_requests);
578 err = 1; /* osd changed */
579
580out:
581 if (newosd)
582 put_osd(newosd);
583 return err;
584}
585
586/*
587 * caller should hold map_sem (for read) and request_mutex
588 */
589static int __send_request(struct ceph_osd_client *osdc,
590 struct ceph_osd_request *req)
591{
592 struct ceph_osd_request_head *reqhead;
593 int err;
594
595 err = __map_osds(osdc, req);
596 if (err < 0)
597 return err;
598 if (req->r_osd == NULL) {
599 dout("send_request %p no up osds in pg\n", req);
600 ceph_monc_request_next_osdmap(&osdc->client->monc);
601 return 0;
602 }
603
604 dout("send_request %p tid %llu to osd%d flags %d\n",
605 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
606
607 reqhead = req->r_request->front.iov_base;
608 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
609 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
610 reqhead->reassert_version = req->r_reassert_version;
611
Sage Weil6b805182009-10-27 11:50:50 -0700612 req->r_timeout_stamp = jiffies+osdc->client->mount_args->osd_timeout*HZ;
Sage Weilf24e9982009-10-06 11:31:10 -0700613
614 ceph_msg_get(req->r_request); /* send consumes a ref */
615 ceph_con_send(&req->r_osd->o_con, req->r_request);
616 req->r_sent = req->r_osd->o_incarnation;
617 return 0;
618}
619
620/*
621 * Timeout callback, called every N seconds when 1 or more osd
622 * requests has been active for more than N seconds. When this
623 * happens, we ping all OSDs with requests who have timed out to
624 * ensure any communications channel reset is detected. Reset the
625 * request timeouts another N seconds in the future as we go.
626 * Reschedule the timeout event another N seconds in future (unless
627 * there are no open requests).
628 */
629static void handle_timeout(struct work_struct *work)
630{
631 struct ceph_osd_client *osdc =
632 container_of(work, struct ceph_osd_client, timeout_work.work);
633 struct ceph_osd_request *req;
634 struct ceph_osd *osd;
Sage Weil6b805182009-10-27 11:50:50 -0700635 unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
Sage Weilf24e9982009-10-06 11:31:10 -0700636 unsigned long next_timeout = timeout + jiffies;
637 struct rb_node *p;
638
639 dout("timeout\n");
640 down_read(&osdc->map_sem);
641
642 ceph_monc_request_next_osdmap(&osdc->client->monc);
643
644 mutex_lock(&osdc->request_mutex);
645 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
646 req = rb_entry(p, struct ceph_osd_request, r_node);
647
648 if (req->r_resend) {
649 int err;
650
651 dout("osdc resending prev failed %lld\n", req->r_tid);
652 err = __send_request(osdc, req);
653 if (err)
654 dout("osdc failed again on %lld\n", req->r_tid);
655 else
656 req->r_resend = false;
657 continue;
658 }
659 }
660 for (p = rb_first(&osdc->osds); p; p = rb_next(p)) {
661 osd = rb_entry(p, struct ceph_osd, o_node);
662 if (list_empty(&osd->o_requests))
663 continue;
664 req = list_first_entry(&osd->o_requests,
665 struct ceph_osd_request, r_osd_item);
666 if (time_before(jiffies, req->r_timeout_stamp))
667 continue;
668
669 dout(" tid %llu (at least) timed out on osd%d\n",
670 req->r_tid, osd->o_osd);
671 req->r_timeout_stamp = next_timeout;
672 ceph_con_keepalive(&osd->o_con);
673 }
674
675 if (osdc->timeout_tid)
676 schedule_delayed_work(&osdc->timeout_work,
677 round_jiffies_relative(timeout));
678
679 mutex_unlock(&osdc->request_mutex);
680
681 up_read(&osdc->map_sem);
682}
683
684/*
685 * handle osd op reply. either call the callback if it is specified,
686 * or do the completion to wake up the waiting thread.
687 */
688static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
689{
690 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
691 struct ceph_osd_request *req;
692 u64 tid;
693 int numops, object_len, flags;
694
695 if (msg->front.iov_len < sizeof(*rhead))
696 goto bad;
697 tid = le64_to_cpu(rhead->tid);
698 numops = le32_to_cpu(rhead->num_ops);
699 object_len = le32_to_cpu(rhead->object_len);
700 if (msg->front.iov_len != sizeof(*rhead) + object_len +
701 numops * sizeof(struct ceph_osd_op))
702 goto bad;
703 dout("handle_reply %p tid %llu\n", msg, tid);
704
705 /* lookup */
706 mutex_lock(&osdc->request_mutex);
707 req = __lookup_request(osdc, tid);
708 if (req == NULL) {
709 dout("handle_reply tid %llu dne\n", tid);
710 mutex_unlock(&osdc->request_mutex);
711 return;
712 }
713 ceph_osdc_get_request(req);
714 flags = le32_to_cpu(rhead->flags);
715
716 if (req->r_reply) {
717 /*
718 * once we see the message has been received, we don't
719 * need a ref (which is only needed for revoking
720 * pages)
721 */
722 ceph_msg_put(req->r_reply);
723 req->r_reply = NULL;
724 }
725
726 if (!req->r_got_reply) {
727 unsigned bytes;
728
729 req->r_result = le32_to_cpu(rhead->result);
730 bytes = le32_to_cpu(msg->hdr.data_len);
731 dout("handle_reply result %d bytes %d\n", req->r_result,
732 bytes);
733 if (req->r_result == 0)
734 req->r_result = bytes;
735
736 /* in case this is a write and we need to replay, */
737 req->r_reassert_version = rhead->reassert_version;
738
739 req->r_got_reply = 1;
740 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
741 dout("handle_reply tid %llu dup ack\n", tid);
742 goto done;
743 }
744
745 dout("handle_reply tid %llu flags %d\n", tid, flags);
746
747 /* either this is a read, or we got the safe response */
748 if ((flags & CEPH_OSD_FLAG_ONDISK) ||
749 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
750 __unregister_request(osdc, req);
751
752 mutex_unlock(&osdc->request_mutex);
753
754 if (req->r_callback)
755 req->r_callback(req, msg);
756 else
757 complete(&req->r_completion);
758
759 if (flags & CEPH_OSD_FLAG_ONDISK) {
760 if (req->r_safe_callback)
761 req->r_safe_callback(req, msg);
762 complete(&req->r_safe_completion); /* fsync waiter */
763 }
764
765done:
766 ceph_osdc_put_request(req);
767 return;
768
769bad:
770 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
771 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
772 (int)sizeof(*rhead));
773}
774
775
776/*
777 * Resubmit osd requests whose osd or osd address has changed. Request
778 * a new osd map if osds are down, or we are otherwise unable to determine
779 * how to direct a request.
780 *
781 * Close connections to down osds.
782 *
783 * If @who is specified, resubmit requests for that specific osd.
784 *
785 * Caller should hold map_sem for read and request_mutex.
786 */
787static void kick_requests(struct ceph_osd_client *osdc,
788 struct ceph_osd *kickosd)
789{
790 struct ceph_osd_request *req;
791 struct rb_node *p, *n;
792 int needmap = 0;
793 int err;
794
795 dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
796 mutex_lock(&osdc->request_mutex);
797 if (!kickosd) {
798 for (p = rb_first(&osdc->osds); p; p = n) {
799 struct ceph_osd *osd =
800 rb_entry(p, struct ceph_osd, o_node);
801
802 n = rb_next(p);
803 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
804 !ceph_entity_addr_equal(&osd->o_con.peer_addr,
805 ceph_osd_addr(osdc->osdmap,
806 osd->o_osd)))
807 reset_osd(osdc, osd);
808 }
809 }
810
811 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
812 req = rb_entry(p, struct ceph_osd_request, r_node);
813
814 if (req->r_resend) {
815 dout(" r_resend set on tid %llu\n", req->r_tid);
Sage Weil266673d2009-10-09 10:31:32 -0700816 __cancel_request(req);
Sage Weilf24e9982009-10-06 11:31:10 -0700817 goto kick;
818 }
Sage Weil266673d2009-10-09 10:31:32 -0700819 if (req->r_osd && kickosd == req->r_osd) {
820 __cancel_request(req);
Sage Weilf24e9982009-10-06 11:31:10 -0700821 goto kick;
Sage Weil266673d2009-10-09 10:31:32 -0700822 }
Sage Weilf24e9982009-10-06 11:31:10 -0700823
824 err = __map_osds(osdc, req);
825 if (err == 0)
826 continue; /* no change */
827 if (err < 0) {
828 /*
829 * FIXME: really, we should set the request
830 * error and fail if this isn't a 'nofail'
831 * request, but that's a fair bit more
832 * complicated to do. So retry!
833 */
834 dout(" setting r_resend on %llu\n", req->r_tid);
835 req->r_resend = true;
836 continue;
837 }
838 if (req->r_osd == NULL) {
839 dout("tid %llu maps to no valid osd\n", req->r_tid);
840 needmap++; /* request a newer map */
841 continue;
842 }
843
844kick:
Sage Weilc1ea8822009-10-08 16:55:47 -0700845 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
846 req->r_osd->o_osd);
Sage Weilf24e9982009-10-06 11:31:10 -0700847 req->r_flags |= CEPH_OSD_FLAG_RETRY;
848 err = __send_request(osdc, req);
849 if (err) {
850 dout(" setting r_resend on %llu\n", req->r_tid);
851 req->r_resend = true;
852 }
853 }
854 mutex_unlock(&osdc->request_mutex);
855
856 if (needmap) {
857 dout("%d requests for down osds, need new map\n", needmap);
858 ceph_monc_request_next_osdmap(&osdc->client->monc);
859 }
860}
861
862/*
863 * Process updated osd map.
864 *
865 * The message contains any number of incremental and full maps, normally
866 * indicating some sort of topology change in the cluster. Kick requests
867 * off to different OSDs as needed.
868 */
869void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
870{
871 void *p, *end, *next;
872 u32 nr_maps, maplen;
873 u32 epoch;
874 struct ceph_osdmap *newmap = NULL, *oldmap;
875 int err;
876 struct ceph_fsid fsid;
877
878 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
879 p = msg->front.iov_base;
880 end = p + msg->front.iov_len;
881
882 /* verify fsid */
883 ceph_decode_need(&p, end, sizeof(fsid), bad);
884 ceph_decode_copy(&p, &fsid, sizeof(fsid));
885 if (ceph_fsid_compare(&fsid, &osdc->client->monc.monmap->fsid)) {
886 pr_err("got osdmap with wrong fsid, ignoring\n");
887 return;
888 }
889
890 down_write(&osdc->map_sem);
891
892 /* incremental maps */
893 ceph_decode_32_safe(&p, end, nr_maps, bad);
894 dout(" %d inc maps\n", nr_maps);
895 while (nr_maps > 0) {
896 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
Sage Weilc89136e2009-10-14 09:59:09 -0700897 epoch = ceph_decode_32(&p);
898 maplen = ceph_decode_32(&p);
Sage Weilf24e9982009-10-06 11:31:10 -0700899 ceph_decode_need(&p, end, maplen, bad);
900 next = p + maplen;
901 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
902 dout("applying incremental map %u len %d\n",
903 epoch, maplen);
904 newmap = osdmap_apply_incremental(&p, next,
905 osdc->osdmap,
906 osdc->client->msgr);
907 if (IS_ERR(newmap)) {
908 err = PTR_ERR(newmap);
909 goto bad;
910 }
911 if (newmap != osdc->osdmap) {
912 ceph_osdmap_destroy(osdc->osdmap);
913 osdc->osdmap = newmap;
914 }
915 } else {
916 dout("ignoring incremental map %u len %d\n",
917 epoch, maplen);
918 }
919 p = next;
920 nr_maps--;
921 }
922 if (newmap)
923 goto done;
924
925 /* full maps */
926 ceph_decode_32_safe(&p, end, nr_maps, bad);
927 dout(" %d full maps\n", nr_maps);
928 while (nr_maps) {
929 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
Sage Weilc89136e2009-10-14 09:59:09 -0700930 epoch = ceph_decode_32(&p);
931 maplen = ceph_decode_32(&p);
Sage Weilf24e9982009-10-06 11:31:10 -0700932 ceph_decode_need(&p, end, maplen, bad);
933 if (nr_maps > 1) {
934 dout("skipping non-latest full map %u len %d\n",
935 epoch, maplen);
936 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
937 dout("skipping full map %u len %d, "
938 "older than our %u\n", epoch, maplen,
939 osdc->osdmap->epoch);
940 } else {
941 dout("taking full map %u len %d\n", epoch, maplen);
942 newmap = osdmap_decode(&p, p+maplen);
943 if (IS_ERR(newmap)) {
944 err = PTR_ERR(newmap);
945 goto bad;
946 }
947 oldmap = osdc->osdmap;
948 osdc->osdmap = newmap;
949 if (oldmap)
950 ceph_osdmap_destroy(oldmap);
951 }
952 p += maplen;
953 nr_maps--;
954 }
955
956done:
957 downgrade_write(&osdc->map_sem);
958 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
959 if (newmap)
960 kick_requests(osdc, NULL);
961 up_read(&osdc->map_sem);
962 return;
963
964bad:
965 pr_err("osdc handle_map corrupt msg\n");
966 up_write(&osdc->map_sem);
967 return;
968}
969
970
971/*
972 * A read request prepares specific pages that data is to be read into.
973 * When a message is being read off the wire, we call prepare_pages to
974 * find those pages.
975 * 0 = success, -1 failure.
976 */
977static int prepare_pages(struct ceph_connection *con, struct ceph_msg *m,
978 int want)
979{
980 struct ceph_osd *osd = con->private;
981 struct ceph_osd_client *osdc;
982 struct ceph_osd_reply_head *rhead = m->front.iov_base;
983 struct ceph_osd_request *req;
984 u64 tid;
985 int ret = -1;
986 int type = le16_to_cpu(m->hdr.type);
987
988 if (!osd)
989 return -1;
990 osdc = osd->o_osdc;
991
992 dout("prepare_pages on msg %p want %d\n", m, want);
993 if (unlikely(type != CEPH_MSG_OSD_OPREPLY))
994 return -1; /* hmm! */
995
996 tid = le64_to_cpu(rhead->tid);
997 mutex_lock(&osdc->request_mutex);
998 req = __lookup_request(osdc, tid);
999 if (!req) {
1000 dout("prepare_pages unknown tid %llu\n", tid);
1001 goto out;
1002 }
1003 dout("prepare_pages tid %llu has %d pages, want %d\n",
1004 tid, req->r_num_pages, want);
1005 if (likely(req->r_num_pages >= want && !req->r_prepared_pages)) {
1006 m->pages = req->r_pages;
1007 m->nr_pages = req->r_num_pages;
1008 req->r_reply = m; /* only for duration of read over socket */
1009 ceph_msg_get(m);
1010 req->r_prepared_pages = 1;
1011 ret = 0; /* success */
1012 }
1013out:
1014 mutex_unlock(&osdc->request_mutex);
1015 return ret;
1016}
1017
1018/*
1019 * Register request, send initial attempt.
1020 */
1021int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1022 struct ceph_osd_request *req,
1023 bool nofail)
1024{
Sage Weilc1ea8822009-10-08 16:55:47 -07001025 int rc = 0;
Sage Weilf24e9982009-10-06 11:31:10 -07001026
1027 req->r_request->pages = req->r_pages;
1028 req->r_request->nr_pages = req->r_num_pages;
1029
1030 register_request(osdc, req);
1031
1032 down_read(&osdc->map_sem);
1033 mutex_lock(&osdc->request_mutex);
Sage Weilc1ea8822009-10-08 16:55:47 -07001034 /*
1035 * a racing kick_requests() may have sent the message for us
1036 * while we dropped request_mutex above, so only send now if
1037 * the request still han't been touched yet.
1038 */
1039 if (req->r_sent == 0) {
1040 rc = __send_request(osdc, req);
1041 if (rc) {
1042 if (nofail) {
1043 dout("osdc_start_request failed send, "
1044 " marking %lld\n", req->r_tid);
1045 req->r_resend = true;
1046 rc = 0;
1047 } else {
1048 __unregister_request(osdc, req);
1049 }
Sage Weilf24e9982009-10-06 11:31:10 -07001050 }
1051 }
1052 mutex_unlock(&osdc->request_mutex);
1053 up_read(&osdc->map_sem);
1054 return rc;
1055}
1056
1057/*
1058 * wait for a request to complete
1059 */
1060int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1061 struct ceph_osd_request *req)
1062{
1063 int rc;
1064
1065 rc = wait_for_completion_interruptible(&req->r_completion);
1066 if (rc < 0) {
1067 mutex_lock(&osdc->request_mutex);
1068 __cancel_request(req);
1069 mutex_unlock(&osdc->request_mutex);
1070 dout("wait_request tid %llu timed out\n", req->r_tid);
1071 return rc;
1072 }
1073
1074 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1075 return req->r_result;
1076}
1077
1078/*
1079 * sync - wait for all in-flight requests to flush. avoid starvation.
1080 */
1081void ceph_osdc_sync(struct ceph_osd_client *osdc)
1082{
1083 struct ceph_osd_request *req;
1084 u64 last_tid, next_tid = 0;
1085
1086 mutex_lock(&osdc->request_mutex);
1087 last_tid = osdc->last_tid;
1088 while (1) {
1089 req = __lookup_request_ge(osdc, next_tid);
1090 if (!req)
1091 break;
1092 if (req->r_tid > last_tid)
1093 break;
1094
1095 next_tid = req->r_tid + 1;
1096 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1097 continue;
1098
1099 ceph_osdc_get_request(req);
1100 mutex_unlock(&osdc->request_mutex);
1101 dout("sync waiting on tid %llu (last is %llu)\n",
1102 req->r_tid, last_tid);
1103 wait_for_completion(&req->r_safe_completion);
1104 mutex_lock(&osdc->request_mutex);
1105 ceph_osdc_put_request(req);
1106 }
1107 mutex_unlock(&osdc->request_mutex);
1108 dout("sync done (thru tid %llu)\n", last_tid);
1109}
1110
1111/*
1112 * init, shutdown
1113 */
1114int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1115{
1116 int err;
1117
1118 dout("init\n");
1119 osdc->client = client;
1120 osdc->osdmap = NULL;
1121 init_rwsem(&osdc->map_sem);
1122 init_completion(&osdc->map_waiters);
1123 osdc->last_requested_map = 0;
1124 mutex_init(&osdc->request_mutex);
1125 osdc->timeout_tid = 0;
1126 osdc->last_tid = 0;
1127 osdc->osds = RB_ROOT;
1128 osdc->requests = RB_ROOT;
1129 osdc->num_requests = 0;
1130 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1131
1132 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1133 sizeof(struct ceph_osd_request));
1134 if (!osdc->req_mempool)
1135 return -ENOMEM;
1136
1137 err = ceph_msgpool_init(&osdc->msgpool_op, 4096, 10, true);
1138 if (err < 0)
1139 return -ENOMEM;
1140 err = ceph_msgpool_init(&osdc->msgpool_op_reply, 512, 0, false);
1141 if (err < 0)
1142 return -ENOMEM;
1143
1144 return 0;
1145}
1146
1147void ceph_osdc_stop(struct ceph_osd_client *osdc)
1148{
1149 cancel_delayed_work_sync(&osdc->timeout_work);
1150 if (osdc->osdmap) {
1151 ceph_osdmap_destroy(osdc->osdmap);
1152 osdc->osdmap = NULL;
1153 }
1154 mempool_destroy(osdc->req_mempool);
1155 ceph_msgpool_destroy(&osdc->msgpool_op);
1156 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1157}
1158
1159/*
1160 * Read some contiguous pages. If we cross a stripe boundary, shorten
1161 * *plen. Return number of bytes read, or error.
1162 */
1163int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1164 struct ceph_vino vino, struct ceph_file_layout *layout,
1165 u64 off, u64 *plen,
1166 u32 truncate_seq, u64 truncate_size,
1167 struct page **pages, int num_pages)
1168{
1169 struct ceph_osd_request *req;
1170 int rc = 0;
1171
1172 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1173 vino.snap, off, *plen);
1174 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1175 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1176 NULL, 0, truncate_seq, truncate_size, NULL,
1177 false, 1);
1178 if (IS_ERR(req))
1179 return PTR_ERR(req);
1180
1181 /* it may be a short read due to an object boundary */
1182 req->r_pages = pages;
1183 num_pages = calc_pages_for(off, *plen);
1184 req->r_num_pages = num_pages;
1185
1186 dout("readpages final extent is %llu~%llu (%d pages)\n",
1187 off, *plen, req->r_num_pages);
1188
1189 rc = ceph_osdc_start_request(osdc, req, false);
1190 if (!rc)
1191 rc = ceph_osdc_wait_request(osdc, req);
1192
1193 ceph_osdc_put_request(req);
1194 dout("readpages result %d\n", rc);
1195 return rc;
1196}
1197
1198/*
1199 * do a synchronous write on N pages
1200 */
1201int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1202 struct ceph_file_layout *layout,
1203 struct ceph_snap_context *snapc,
1204 u64 off, u64 len,
1205 u32 truncate_seq, u64 truncate_size,
1206 struct timespec *mtime,
1207 struct page **pages, int num_pages,
1208 int flags, int do_sync, bool nofail)
1209{
1210 struct ceph_osd_request *req;
1211 int rc = 0;
1212
1213 BUG_ON(vino.snap != CEPH_NOSNAP);
1214 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1215 CEPH_OSD_OP_WRITE,
1216 flags | CEPH_OSD_FLAG_ONDISK |
1217 CEPH_OSD_FLAG_WRITE,
1218 snapc, do_sync,
1219 truncate_seq, truncate_size, mtime,
1220 nofail, 1);
1221 if (IS_ERR(req))
1222 return PTR_ERR(req);
1223
1224 /* it may be a short write due to an object boundary */
1225 req->r_pages = pages;
1226 req->r_num_pages = calc_pages_for(off, len);
1227 dout("writepages %llu~%llu (%d pages)\n", off, len,
1228 req->r_num_pages);
1229
1230 rc = ceph_osdc_start_request(osdc, req, nofail);
1231 if (!rc)
1232 rc = ceph_osdc_wait_request(osdc, req);
1233
1234 ceph_osdc_put_request(req);
1235 if (rc == 0)
1236 rc = len;
1237 dout("writepages result %d\n", rc);
1238 return rc;
1239}
1240
1241/*
1242 * handle incoming message
1243 */
1244static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1245{
1246 struct ceph_osd *osd = con->private;
1247 struct ceph_osd_client *osdc = osd->o_osdc;
1248 int type = le16_to_cpu(msg->hdr.type);
1249
1250 if (!osd)
1251 return;
1252
1253 switch (type) {
1254 case CEPH_MSG_OSD_MAP:
1255 ceph_osdc_handle_map(osdc, msg);
1256 break;
1257 case CEPH_MSG_OSD_OPREPLY:
1258 handle_reply(osdc, msg);
1259 break;
1260
1261 default:
1262 pr_err("received unknown message type %d %s\n", type,
1263 ceph_msg_type_name(type));
1264 }
1265 ceph_msg_put(msg);
1266}
1267
1268static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1269 struct ceph_msg_header *hdr)
1270{
1271 struct ceph_osd *osd = con->private;
1272 struct ceph_osd_client *osdc = osd->o_osdc;
1273 int type = le16_to_cpu(hdr->type);
Sage Weil8f3bc052009-10-14 17:36:07 -07001274 int front = le32_to_cpu(hdr->front_len);
Sage Weilf24e9982009-10-06 11:31:10 -07001275
1276 switch (type) {
1277 case CEPH_MSG_OSD_OPREPLY:
Sage Weil8f3bc052009-10-14 17:36:07 -07001278 return ceph_msgpool_get(&osdc->msgpool_op_reply, front);
Sage Weilf24e9982009-10-06 11:31:10 -07001279 }
1280 return ceph_alloc_msg(con, hdr);
1281}
1282
1283/*
1284 * Wrappers to refcount containing ceph_osd struct
1285 */
1286static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1287{
1288 struct ceph_osd *osd = con->private;
1289 if (get_osd(osd))
1290 return con;
1291 return NULL;
1292}
1293
1294static void put_osd_con(struct ceph_connection *con)
1295{
1296 struct ceph_osd *osd = con->private;
1297 put_osd(osd);
1298}
1299
1300const static struct ceph_connection_operations osd_con_ops = {
1301 .get = get_osd_con,
1302 .put = put_osd_con,
1303 .dispatch = dispatch,
1304 .alloc_msg = alloc_msg,
Sage Weil81b024e2009-10-09 10:29:18 -07001305 .fault = osd_reset,
Sage Weilf24e9982009-10-06 11:31:10 -07001306 .alloc_middle = ceph_alloc_middle,
1307 .prepare_pages = prepare_pages,
1308};