Merge git://git.kernel.org/pub/scm/linux/kernel/git/joern/logfs
[pandora-kernel.git] / fs / ceph / osd_client.c
1 #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 #include "auth.h"
15
16 #define OSD_OP_FRONT_LEN        4096
17 #define OSD_OPREPLY_FRONT_LEN   512
18
19 const static struct ceph_connection_operations osd_con_ops;
20 static int __kick_requests(struct ceph_osd_client *osdc,
21                           struct ceph_osd *kickosd);
22
23 static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
24
25 /*
26  * Implement client access to distributed object storage cluster.
27  *
28  * All data objects are stored within a cluster/cloud of OSDs, or
29  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
30  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
31  * remote daemons serving up and coordinating consistent and safe
32  * access to storage.
33  *
34  * Cluster membership and the mapping of data objects onto storage devices
35  * are described by the osd map.
36  *
37  * We keep track of pending OSD requests (read, write), resubmit
38  * requests to different OSDs when the cluster topology/data layout
39  * change, or retry the affected requests when the communications
40  * channel with an OSD is reset.
41  */
42
43 /*
44  * calculate the mapping of a file extent onto an object, and fill out the
45  * request accordingly.  shorten extent as necessary if it crosses an
46  * object boundary.
47  *
48  * fill osd op in request message.
49  */
50 static void calc_layout(struct ceph_osd_client *osdc,
51                         struct ceph_vino vino, struct ceph_file_layout *layout,
52                         u64 off, u64 *plen,
53                         struct ceph_osd_request *req)
54 {
55         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
56         struct ceph_osd_op *op = (void *)(reqhead + 1);
57         u64 orig_len = *plen;
58         u64 objoff, objlen;    /* extent in object */
59         u64 bno;
60
61         reqhead->snapid = cpu_to_le64(vino.snap);
62
63         /* object extent? */
64         ceph_calc_file_object_mapping(layout, off, plen, &bno,
65                                       &objoff, &objlen);
66         if (*plen < orig_len)
67                 dout(" skipping last %llu, final file extent %llu~%llu\n",
68                      orig_len - *plen, off, *plen);
69
70         sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
71         req->r_oid_len = strlen(req->r_oid);
72
73         op->extent.offset = cpu_to_le64(objoff);
74         op->extent.length = cpu_to_le64(objlen);
75         req->r_num_pages = calc_pages_for(off, *plen);
76
77         dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
78              req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
79 }
80
81 /*
82  * requests
83  */
84 void ceph_osdc_release_request(struct kref *kref)
85 {
86         struct ceph_osd_request *req = container_of(kref,
87                                                     struct ceph_osd_request,
88                                                     r_kref);
89
90         if (req->r_request)
91                 ceph_msg_put(req->r_request);
92         if (req->r_reply)
93                 ceph_msg_put(req->r_reply);
94         if (req->r_con_filling_msg) {
95                 dout("release_request revoking pages %p from con %p\n",
96                      req->r_pages, req->r_con_filling_msg);
97                 ceph_con_revoke_message(req->r_con_filling_msg,
98                                       req->r_reply);
99                 ceph_con_put(req->r_con_filling_msg);
100         }
101         if (req->r_own_pages)
102                 ceph_release_page_vector(req->r_pages,
103                                          req->r_num_pages);
104         ceph_put_snap_context(req->r_snapc);
105         if (req->r_mempool)
106                 mempool_free(req, req->r_osdc->req_mempool);
107         else
108                 kfree(req);
109 }
110
111 /*
112  * build new request AND message, calculate layout, and adjust file
113  * extent as needed.
114  *
115  * if the file was recently truncated, we include information about its
116  * old and new size so that the object can be updated appropriately.  (we
117  * avoid synchronously deleting truncated objects because it's slow.)
118  *
119  * if @do_sync, include a 'startsync' command so that the osd will flush
120  * data quickly.
121  */
122 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
123                                                struct ceph_file_layout *layout,
124                                                struct ceph_vino vino,
125                                                u64 off, u64 *plen,
126                                                int opcode, int flags,
127                                                struct ceph_snap_context *snapc,
128                                                int do_sync,
129                                                u32 truncate_seq,
130                                                u64 truncate_size,
131                                                struct timespec *mtime,
132                                                bool use_mempool, int num_reply)
133 {
134         struct ceph_osd_request *req;
135         struct ceph_msg *msg;
136         struct ceph_osd_request_head *head;
137         struct ceph_osd_op *op;
138         void *p;
139         int num_op = 1 + do_sync;
140         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
141         int i;
142
143         if (use_mempool) {
144                 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
145                 memset(req, 0, sizeof(*req));
146         } else {
147                 req = kzalloc(sizeof(*req), GFP_NOFS);
148         }
149         if (req == NULL)
150                 return ERR_PTR(-ENOMEM);
151
152         req->r_osdc = osdc;
153         req->r_mempool = use_mempool;
154         kref_init(&req->r_kref);
155         init_completion(&req->r_completion);
156         init_completion(&req->r_safe_completion);
157         INIT_LIST_HEAD(&req->r_unsafe_item);
158         req->r_flags = flags;
159
160         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
161
162         /* create reply message */
163         if (use_mempool)
164                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
165         else
166                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
167                                    OSD_OPREPLY_FRONT_LEN, 0, 0, NULL);
168         if (IS_ERR(msg)) {
169                 ceph_osdc_put_request(req);
170                 return ERR_PTR(PTR_ERR(msg));
171         }
172         req->r_reply = msg;
173
174         /* create request message; allow space for oid */
175         msg_size += 40;
176         if (snapc)
177                 msg_size += sizeof(u64) * snapc->num_snaps;
178         if (use_mempool)
179                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
180         else
181                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
182         if (IS_ERR(msg)) {
183                 ceph_osdc_put_request(req);
184                 return ERR_PTR(PTR_ERR(msg));
185         }
186         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
187         memset(msg->front.iov_base, 0, msg->front.iov_len);
188         head = msg->front.iov_base;
189         op = (void *)(head + 1);
190         p = (void *)(op + num_op);
191
192         req->r_request = msg;
193         req->r_snapc = ceph_get_snap_context(snapc);
194
195         head->client_inc = cpu_to_le32(1); /* always, for now. */
196         head->flags = cpu_to_le32(flags);
197         if (flags & CEPH_OSD_FLAG_WRITE)
198                 ceph_encode_timespec(&head->mtime, mtime);
199         head->num_ops = cpu_to_le16(num_op);
200         op->op = cpu_to_le16(opcode);
201
202         /* calculate max write size */
203         calc_layout(osdc, vino, layout, off, plen, req);
204         req->r_file_layout = *layout;  /* keep a copy */
205
206         if (flags & CEPH_OSD_FLAG_WRITE) {
207                 req->r_request->hdr.data_off = cpu_to_le16(off);
208                 req->r_request->hdr.data_len = cpu_to_le32(*plen);
209                 op->payload_len = cpu_to_le32(*plen);
210         }
211         op->extent.truncate_size = cpu_to_le64(truncate_size);
212         op->extent.truncate_seq = cpu_to_le32(truncate_seq);
213
214         /* fill in oid */
215         head->object_len = cpu_to_le32(req->r_oid_len);
216         memcpy(p, req->r_oid, req->r_oid_len);
217         p += req->r_oid_len;
218
219         if (do_sync) {
220                 op++;
221                 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
222         }
223         if (snapc) {
224                 head->snap_seq = cpu_to_le64(snapc->seq);
225                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
226                 for (i = 0; i < snapc->num_snaps; i++) {
227                         put_unaligned_le64(snapc->snaps[i], p);
228                         p += sizeof(u64);
229                 }
230         }
231
232         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
233         msg_size = p - msg->front.iov_base;
234         msg->front.iov_len = msg_size;
235         msg->hdr.front_len = cpu_to_le32(msg_size);
236         return req;
237 }
238
239 /*
240  * We keep osd requests in an rbtree, sorted by ->r_tid.
241  */
242 static void __insert_request(struct ceph_osd_client *osdc,
243                              struct ceph_osd_request *new)
244 {
245         struct rb_node **p = &osdc->requests.rb_node;
246         struct rb_node *parent = NULL;
247         struct ceph_osd_request *req = NULL;
248
249         while (*p) {
250                 parent = *p;
251                 req = rb_entry(parent, struct ceph_osd_request, r_node);
252                 if (new->r_tid < req->r_tid)
253                         p = &(*p)->rb_left;
254                 else if (new->r_tid > req->r_tid)
255                         p = &(*p)->rb_right;
256                 else
257                         BUG();
258         }
259
260         rb_link_node(&new->r_node, parent, p);
261         rb_insert_color(&new->r_node, &osdc->requests);
262 }
263
264 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
265                                                  u64 tid)
266 {
267         struct ceph_osd_request *req;
268         struct rb_node *n = osdc->requests.rb_node;
269
270         while (n) {
271                 req = rb_entry(n, struct ceph_osd_request, r_node);
272                 if (tid < req->r_tid)
273                         n = n->rb_left;
274                 else if (tid > req->r_tid)
275                         n = n->rb_right;
276                 else
277                         return req;
278         }
279         return NULL;
280 }
281
282 static struct ceph_osd_request *
283 __lookup_request_ge(struct ceph_osd_client *osdc,
284                     u64 tid)
285 {
286         struct ceph_osd_request *req;
287         struct rb_node *n = osdc->requests.rb_node;
288
289         while (n) {
290                 req = rb_entry(n, struct ceph_osd_request, r_node);
291                 if (tid < req->r_tid) {
292                         if (!n->rb_left)
293                                 return req;
294                         n = n->rb_left;
295                 } else if (tid > req->r_tid) {
296                         n = n->rb_right;
297                 } else {
298                         return req;
299                 }
300         }
301         return NULL;
302 }
303
304
305 /*
306  * If the osd connection drops, we need to resubmit all requests.
307  */
308 static void osd_reset(struct ceph_connection *con)
309 {
310         struct ceph_osd *osd = con->private;
311         struct ceph_osd_client *osdc;
312
313         if (!osd)
314                 return;
315         dout("osd_reset osd%d\n", osd->o_osd);
316         osdc = osd->o_osdc;
317         down_read(&osdc->map_sem);
318         kick_requests(osdc, osd);
319         up_read(&osdc->map_sem);
320 }
321
322 /*
323  * Track open sessions with osds.
324  */
325 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
326 {
327         struct ceph_osd *osd;
328
329         osd = kzalloc(sizeof(*osd), GFP_NOFS);
330         if (!osd)
331                 return NULL;
332
333         atomic_set(&osd->o_ref, 1);
334         osd->o_osdc = osdc;
335         INIT_LIST_HEAD(&osd->o_requests);
336         INIT_LIST_HEAD(&osd->o_osd_lru);
337         osd->o_incarnation = 1;
338
339         ceph_con_init(osdc->client->msgr, &osd->o_con);
340         osd->o_con.private = osd;
341         osd->o_con.ops = &osd_con_ops;
342         osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
343
344         INIT_LIST_HEAD(&osd->o_keepalive_item);
345         return osd;
346 }
347
348 static struct ceph_osd *get_osd(struct ceph_osd *osd)
349 {
350         if (atomic_inc_not_zero(&osd->o_ref)) {
351                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
352                      atomic_read(&osd->o_ref));
353                 return osd;
354         } else {
355                 dout("get_osd %p FAIL\n", osd);
356                 return NULL;
357         }
358 }
359
360 static void put_osd(struct ceph_osd *osd)
361 {
362         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
363              atomic_read(&osd->o_ref) - 1);
364         if (atomic_dec_and_test(&osd->o_ref))
365                 kfree(osd);
366 }
367
368 /*
369  * remove an osd from our map
370  */
371 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
372 {
373         dout("__remove_osd %p\n", osd);
374         BUG_ON(!list_empty(&osd->o_requests));
375         rb_erase(&osd->o_node, &osdc->osds);
376         list_del_init(&osd->o_osd_lru);
377         ceph_con_close(&osd->o_con);
378         put_osd(osd);
379 }
380
381 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
382                               struct ceph_osd *osd)
383 {
384         dout("__move_osd_to_lru %p\n", osd);
385         BUG_ON(!list_empty(&osd->o_osd_lru));
386         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
387         osd->lru_ttl = jiffies + osdc->client->mount_args->osd_idle_ttl * HZ;
388 }
389
390 static void __remove_osd_from_lru(struct ceph_osd *osd)
391 {
392         dout("__remove_osd_from_lru %p\n", osd);
393         if (!list_empty(&osd->o_osd_lru))
394                 list_del_init(&osd->o_osd_lru);
395 }
396
397 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
398 {
399         struct ceph_osd *osd, *nosd;
400
401         dout("__remove_old_osds %p\n", osdc);
402         mutex_lock(&osdc->request_mutex);
403         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
404                 if (!remove_all && time_before(jiffies, osd->lru_ttl))
405                         break;
406                 __remove_osd(osdc, osd);
407         }
408         mutex_unlock(&osdc->request_mutex);
409 }
410
411 /*
412  * reset osd connect
413  */
414 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
415 {
416         struct ceph_osd_request *req;
417         int ret = 0;
418
419         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
420         if (list_empty(&osd->o_requests)) {
421                 __remove_osd(osdc, osd);
422         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
423                           &osd->o_con.peer_addr,
424                           sizeof(osd->o_con.peer_addr)) == 0 &&
425                    !ceph_con_opened(&osd->o_con)) {
426                 dout(" osd addr hasn't changed and connection never opened,"
427                      " letting msgr retry");
428                 /* touch each r_stamp for handle_timeout()'s benfit */
429                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
430                         req->r_stamp = jiffies;
431                 ret = -EAGAIN;
432         } else {
433                 ceph_con_close(&osd->o_con);
434                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
435                 osd->o_incarnation++;
436         }
437         return ret;
438 }
439
440 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
441 {
442         struct rb_node **p = &osdc->osds.rb_node;
443         struct rb_node *parent = NULL;
444         struct ceph_osd *osd = NULL;
445
446         while (*p) {
447                 parent = *p;
448                 osd = rb_entry(parent, struct ceph_osd, o_node);
449                 if (new->o_osd < osd->o_osd)
450                         p = &(*p)->rb_left;
451                 else if (new->o_osd > osd->o_osd)
452                         p = &(*p)->rb_right;
453                 else
454                         BUG();
455         }
456
457         rb_link_node(&new->o_node, parent, p);
458         rb_insert_color(&new->o_node, &osdc->osds);
459 }
460
461 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
462 {
463         struct ceph_osd *osd;
464         struct rb_node *n = osdc->osds.rb_node;
465
466         while (n) {
467                 osd = rb_entry(n, struct ceph_osd, o_node);
468                 if (o < osd->o_osd)
469                         n = n->rb_left;
470                 else if (o > osd->o_osd)
471                         n = n->rb_right;
472                 else
473                         return osd;
474         }
475         return NULL;
476 }
477
478 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
479 {
480         schedule_delayed_work(&osdc->timeout_work,
481                         osdc->client->mount_args->osd_keepalive_timeout * HZ);
482 }
483
484 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
485 {
486         cancel_delayed_work(&osdc->timeout_work);
487 }
488
489 /*
490  * Register request, assign tid.  If this is the first request, set up
491  * the timeout event.
492  */
493 static void register_request(struct ceph_osd_client *osdc,
494                              struct ceph_osd_request *req)
495 {
496         mutex_lock(&osdc->request_mutex);
497         req->r_tid = ++osdc->last_tid;
498         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
499         INIT_LIST_HEAD(&req->r_req_lru_item);
500
501         dout("register_request %p tid %lld\n", req, req->r_tid);
502         __insert_request(osdc, req);
503         ceph_osdc_get_request(req);
504         osdc->num_requests++;
505
506         if (osdc->num_requests == 1) {
507                 dout(" first request, scheduling timeout\n");
508                 __schedule_osd_timeout(osdc);
509         }
510         mutex_unlock(&osdc->request_mutex);
511 }
512
513 /*
514  * called under osdc->request_mutex
515  */
516 static void __unregister_request(struct ceph_osd_client *osdc,
517                                  struct ceph_osd_request *req)
518 {
519         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
520         rb_erase(&req->r_node, &osdc->requests);
521         osdc->num_requests--;
522
523         if (req->r_osd) {
524                 /* make sure the original request isn't in flight. */
525                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
526
527                 list_del_init(&req->r_osd_item);
528                 if (list_empty(&req->r_osd->o_requests))
529                         __move_osd_to_lru(osdc, req->r_osd);
530                 req->r_osd = NULL;
531         }
532
533         ceph_osdc_put_request(req);
534
535         list_del_init(&req->r_req_lru_item);
536         if (osdc->num_requests == 0) {
537                 dout(" no requests, canceling timeout\n");
538                 __cancel_osd_timeout(osdc);
539         }
540 }
541
542 /*
543  * Cancel a previously queued request message
544  */
545 static void __cancel_request(struct ceph_osd_request *req)
546 {
547         if (req->r_sent) {
548                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
549                 req->r_sent = 0;
550         }
551         list_del_init(&req->r_req_lru_item);
552 }
553
554 /*
555  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
556  * (as needed), and set the request r_osd appropriately.  If there is
557  * no up osd, set r_osd to NULL.
558  *
559  * Return 0 if unchanged, 1 if changed, or negative on error.
560  *
561  * Caller should hold map_sem for read and request_mutex.
562  */
563 static int __map_osds(struct ceph_osd_client *osdc,
564                       struct ceph_osd_request *req)
565 {
566         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
567         struct ceph_pg pgid;
568         int o = -1;
569         int err;
570
571         dout("map_osds %p tid %lld\n", req, req->r_tid);
572         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
573                                       &req->r_file_layout, osdc->osdmap);
574         if (err)
575                 return err;
576         pgid = reqhead->layout.ol_pgid;
577         req->r_pgid = pgid;
578
579         o = ceph_calc_pg_primary(osdc->osdmap, pgid);
580
581         if ((req->r_osd && req->r_osd->o_osd == o &&
582              req->r_sent >= req->r_osd->o_incarnation) ||
583             (req->r_osd == NULL && o == -1))
584                 return 0;  /* no change */
585
586         dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
587              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
588              req->r_osd ? req->r_osd->o_osd : -1);
589
590         if (req->r_osd) {
591                 __cancel_request(req);
592                 list_del_init(&req->r_osd_item);
593                 req->r_osd = NULL;
594         }
595
596         req->r_osd = __lookup_osd(osdc, o);
597         if (!req->r_osd && o >= 0) {
598                 err = -ENOMEM;
599                 req->r_osd = create_osd(osdc);
600                 if (!req->r_osd)
601                         goto out;
602
603                 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
604                 req->r_osd->o_osd = o;
605                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
606                 __insert_osd(osdc, req->r_osd);
607
608                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
609         }
610
611         if (req->r_osd) {
612                 __remove_osd_from_lru(req->r_osd);
613                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
614         }
615         err = 1;   /* osd changed */
616
617 out:
618         return err;
619 }
620
621 /*
622  * caller should hold map_sem (for read) and request_mutex
623  */
624 static int __send_request(struct ceph_osd_client *osdc,
625                           struct ceph_osd_request *req)
626 {
627         struct ceph_osd_request_head *reqhead;
628         int err;
629
630         err = __map_osds(osdc, req);
631         if (err < 0)
632                 return err;
633         if (req->r_osd == NULL) {
634                 dout("send_request %p no up osds in pg\n", req);
635                 ceph_monc_request_next_osdmap(&osdc->client->monc);
636                 return 0;
637         }
638
639         dout("send_request %p tid %llu to osd%d flags %d\n",
640              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
641
642         reqhead = req->r_request->front.iov_base;
643         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
644         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
645         reqhead->reassert_version = req->r_reassert_version;
646
647         req->r_stamp = jiffies;
648         list_move_tail(&osdc->req_lru, &req->r_req_lru_item);
649
650         ceph_msg_get(req->r_request); /* send consumes a ref */
651         ceph_con_send(&req->r_osd->o_con, req->r_request);
652         req->r_sent = req->r_osd->o_incarnation;
653         return 0;
654 }
655
656 /*
657  * Timeout callback, called every N seconds when 1 or more osd
658  * requests has been active for more than N seconds.  When this
659  * happens, we ping all OSDs with requests who have timed out to
660  * ensure any communications channel reset is detected.  Reset the
661  * request timeouts another N seconds in the future as we go.
662  * Reschedule the timeout event another N seconds in future (unless
663  * there are no open requests).
664  */
665 static void handle_timeout(struct work_struct *work)
666 {
667         struct ceph_osd_client *osdc =
668                 container_of(work, struct ceph_osd_client, timeout_work.work);
669         struct ceph_osd_request *req, *last_req = NULL;
670         struct ceph_osd *osd;
671         unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
672         unsigned long keepalive =
673                 osdc->client->mount_args->osd_keepalive_timeout * HZ;
674         unsigned long last_stamp = 0;
675         struct rb_node *p;
676         struct list_head slow_osds;
677
678         dout("timeout\n");
679         down_read(&osdc->map_sem);
680
681         ceph_monc_request_next_osdmap(&osdc->client->monc);
682
683         mutex_lock(&osdc->request_mutex);
684         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
685                 req = rb_entry(p, struct ceph_osd_request, r_node);
686
687                 if (req->r_resend) {
688                         int err;
689
690                         dout("osdc resending prev failed %lld\n", req->r_tid);
691                         err = __send_request(osdc, req);
692                         if (err)
693                                 dout("osdc failed again on %lld\n", req->r_tid);
694                         else
695                                 req->r_resend = false;
696                         continue;
697                 }
698         }
699
700         /*
701          * reset osds that appear to be _really_ unresponsive.  this
702          * is a failsafe measure.. we really shouldn't be getting to
703          * this point if the system is working properly.  the monitors
704          * should mark the osd as failed and we should find out about
705          * it from an updated osd map.
706          */
707         while (!list_empty(&osdc->req_lru)) {
708                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
709                                  r_req_lru_item);
710
711                 if (time_before(jiffies, req->r_stamp + timeout))
712                         break;
713
714                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
715                 last_req = req;
716                 last_stamp = req->r_stamp;
717
718                 osd = req->r_osd;
719                 BUG_ON(!osd);
720                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
721                            req->r_tid, osd->o_osd);
722                 __kick_requests(osdc, osd);
723         }
724
725         /*
726          * ping osds that are a bit slow.  this ensures that if there
727          * is a break in the TCP connection we will notice, and reopen
728          * a connection with that osd (from the fault callback).
729          */
730         INIT_LIST_HEAD(&slow_osds);
731         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
732                 if (time_before(jiffies, req->r_stamp + keepalive))
733                         break;
734
735                 osd = req->r_osd;
736                 BUG_ON(!osd);
737                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
738                      req->r_tid, osd->o_osd);
739                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
740         }
741         while (!list_empty(&slow_osds)) {
742                 osd = list_entry(slow_osds.next, struct ceph_osd,
743                                  o_keepalive_item);
744                 list_del_init(&osd->o_keepalive_item);
745                 ceph_con_keepalive(&osd->o_con);
746         }
747
748         __schedule_osd_timeout(osdc);
749         mutex_unlock(&osdc->request_mutex);
750
751         up_read(&osdc->map_sem);
752 }
753
754 static void handle_osds_timeout(struct work_struct *work)
755 {
756         struct ceph_osd_client *osdc =
757                 container_of(work, struct ceph_osd_client,
758                              osds_timeout_work.work);
759         unsigned long delay =
760                 osdc->client->mount_args->osd_idle_ttl * HZ >> 2;
761
762         dout("osds timeout\n");
763         down_read(&osdc->map_sem);
764         remove_old_osds(osdc, 0);
765         up_read(&osdc->map_sem);
766
767         schedule_delayed_work(&osdc->osds_timeout_work,
768                               round_jiffies_relative(delay));
769 }
770
771 /*
772  * handle osd op reply.  either call the callback if it is specified,
773  * or do the completion to wake up the waiting thread.
774  */
775 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
776                          struct ceph_connection *con)
777 {
778         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
779         struct ceph_osd_request *req;
780         u64 tid;
781         int numops, object_len, flags;
782
783         tid = le64_to_cpu(msg->hdr.tid);
784         if (msg->front.iov_len < sizeof(*rhead))
785                 goto bad;
786         numops = le32_to_cpu(rhead->num_ops);
787         object_len = le32_to_cpu(rhead->object_len);
788         if (msg->front.iov_len != sizeof(*rhead) + object_len +
789             numops * sizeof(struct ceph_osd_op))
790                 goto bad;
791         dout("handle_reply %p tid %llu\n", msg, tid);
792
793         /* lookup */
794         mutex_lock(&osdc->request_mutex);
795         req = __lookup_request(osdc, tid);
796         if (req == NULL) {
797                 dout("handle_reply tid %llu dne\n", tid);
798                 mutex_unlock(&osdc->request_mutex);
799                 return;
800         }
801         ceph_osdc_get_request(req);
802         flags = le32_to_cpu(rhead->flags);
803
804         /*
805          * if this connection filled our message, drop our reference now, to
806          * avoid a (safe but slower) revoke later.
807          */
808         if (req->r_con_filling_msg == con && req->r_reply == msg) {
809                 dout(" dropping con_filling_msg ref %p\n", con);
810                 req->r_con_filling_msg = NULL;
811                 ceph_con_put(con);
812         }
813
814         if (!req->r_got_reply) {
815                 unsigned bytes;
816
817                 req->r_result = le32_to_cpu(rhead->result);
818                 bytes = le32_to_cpu(msg->hdr.data_len);
819                 dout("handle_reply result %d bytes %d\n", req->r_result,
820                      bytes);
821                 if (req->r_result == 0)
822                         req->r_result = bytes;
823
824                 /* in case this is a write and we need to replay, */
825                 req->r_reassert_version = rhead->reassert_version;
826
827                 req->r_got_reply = 1;
828         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
829                 dout("handle_reply tid %llu dup ack\n", tid);
830                 mutex_unlock(&osdc->request_mutex);
831                 goto done;
832         }
833
834         dout("handle_reply tid %llu flags %d\n", tid, flags);
835
836         /* either this is a read, or we got the safe response */
837         if ((flags & CEPH_OSD_FLAG_ONDISK) ||
838             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
839                 __unregister_request(osdc, req);
840
841         mutex_unlock(&osdc->request_mutex);
842
843         if (req->r_callback)
844                 req->r_callback(req, msg);
845         else
846                 complete(&req->r_completion);
847
848         if (flags & CEPH_OSD_FLAG_ONDISK) {
849                 if (req->r_safe_callback)
850                         req->r_safe_callback(req, msg);
851                 complete(&req->r_safe_completion);  /* fsync waiter */
852         }
853
854 done:
855         ceph_osdc_put_request(req);
856         return;
857
858 bad:
859         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
860                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
861                (int)sizeof(*rhead));
862         ceph_msg_dump(msg);
863 }
864
865
866 static int __kick_requests(struct ceph_osd_client *osdc,
867                           struct ceph_osd *kickosd)
868 {
869         struct ceph_osd_request *req;
870         struct rb_node *p, *n;
871         int needmap = 0;
872         int err;
873
874         dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
875         if (kickosd) {
876                 err = __reset_osd(osdc, kickosd);
877                 if (err == -EAGAIN)
878                         return 1;
879         } else {
880                 for (p = rb_first(&osdc->osds); p; p = n) {
881                         struct ceph_osd *osd =
882                                 rb_entry(p, struct ceph_osd, o_node);
883
884                         n = rb_next(p);
885                         if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
886                             memcmp(&osd->o_con.peer_addr,
887                                    ceph_osd_addr(osdc->osdmap,
888                                                  osd->o_osd),
889                                    sizeof(struct ceph_entity_addr)) != 0)
890                                 __reset_osd(osdc, osd);
891                 }
892         }
893
894         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
895                 req = rb_entry(p, struct ceph_osd_request, r_node);
896
897                 if (req->r_resend) {
898                         dout(" r_resend set on tid %llu\n", req->r_tid);
899                         __cancel_request(req);
900                         goto kick;
901                 }
902                 if (req->r_osd && kickosd == req->r_osd) {
903                         __cancel_request(req);
904                         goto kick;
905                 }
906
907                 err = __map_osds(osdc, req);
908                 if (err == 0)
909                         continue;  /* no change */
910                 if (err < 0) {
911                         /*
912                          * FIXME: really, we should set the request
913                          * error and fail if this isn't a 'nofail'
914                          * request, but that's a fair bit more
915                          * complicated to do.  So retry!
916                          */
917                         dout(" setting r_resend on %llu\n", req->r_tid);
918                         req->r_resend = true;
919                         continue;
920                 }
921                 if (req->r_osd == NULL) {
922                         dout("tid %llu maps to no valid osd\n", req->r_tid);
923                         needmap++;  /* request a newer map */
924                         continue;
925                 }
926
927 kick:
928                 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
929                      req->r_osd ? req->r_osd->o_osd : -1);
930                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
931                 err = __send_request(osdc, req);
932                 if (err) {
933                         dout(" setting r_resend on %llu\n", req->r_tid);
934                         req->r_resend = true;
935                 }
936         }
937
938         return needmap;
939 }
940
941 /*
942  * Resubmit osd requests whose osd or osd address has changed.  Request
943  * a new osd map if osds are down, or we are otherwise unable to determine
944  * how to direct a request.
945  *
946  * Close connections to down osds.
947  *
948  * If @who is specified, resubmit requests for that specific osd.
949  *
950  * Caller should hold map_sem for read and request_mutex.
951  */
952 static void kick_requests(struct ceph_osd_client *osdc,
953                           struct ceph_osd *kickosd)
954 {
955         int needmap;
956
957         mutex_lock(&osdc->request_mutex);
958         needmap = __kick_requests(osdc, kickosd);
959         mutex_unlock(&osdc->request_mutex);
960
961         if (needmap) {
962                 dout("%d requests for down osds, need new map\n", needmap);
963                 ceph_monc_request_next_osdmap(&osdc->client->monc);
964         }
965
966 }
967 /*
968  * Process updated osd map.
969  *
970  * The message contains any number of incremental and full maps, normally
971  * indicating some sort of topology change in the cluster.  Kick requests
972  * off to different OSDs as needed.
973  */
974 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
975 {
976         void *p, *end, *next;
977         u32 nr_maps, maplen;
978         u32 epoch;
979         struct ceph_osdmap *newmap = NULL, *oldmap;
980         int err;
981         struct ceph_fsid fsid;
982
983         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
984         p = msg->front.iov_base;
985         end = p + msg->front.iov_len;
986
987         /* verify fsid */
988         ceph_decode_need(&p, end, sizeof(fsid), bad);
989         ceph_decode_copy(&p, &fsid, sizeof(fsid));
990         if (ceph_check_fsid(osdc->client, &fsid) < 0)
991                 return;
992
993         down_write(&osdc->map_sem);
994
995         /* incremental maps */
996         ceph_decode_32_safe(&p, end, nr_maps, bad);
997         dout(" %d inc maps\n", nr_maps);
998         while (nr_maps > 0) {
999                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1000                 epoch = ceph_decode_32(&p);
1001                 maplen = ceph_decode_32(&p);
1002                 ceph_decode_need(&p, end, maplen, bad);
1003                 next = p + maplen;
1004                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1005                         dout("applying incremental map %u len %d\n",
1006                              epoch, maplen);
1007                         newmap = osdmap_apply_incremental(&p, next,
1008                                                           osdc->osdmap,
1009                                                           osdc->client->msgr);
1010                         if (IS_ERR(newmap)) {
1011                                 err = PTR_ERR(newmap);
1012                                 goto bad;
1013                         }
1014                         BUG_ON(!newmap);
1015                         if (newmap != osdc->osdmap) {
1016                                 ceph_osdmap_destroy(osdc->osdmap);
1017                                 osdc->osdmap = newmap;
1018                         }
1019                 } else {
1020                         dout("ignoring incremental map %u len %d\n",
1021                              epoch, maplen);
1022                 }
1023                 p = next;
1024                 nr_maps--;
1025         }
1026         if (newmap)
1027                 goto done;
1028
1029         /* full maps */
1030         ceph_decode_32_safe(&p, end, nr_maps, bad);
1031         dout(" %d full maps\n", nr_maps);
1032         while (nr_maps) {
1033                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1034                 epoch = ceph_decode_32(&p);
1035                 maplen = ceph_decode_32(&p);
1036                 ceph_decode_need(&p, end, maplen, bad);
1037                 if (nr_maps > 1) {
1038                         dout("skipping non-latest full map %u len %d\n",
1039                              epoch, maplen);
1040                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1041                         dout("skipping full map %u len %d, "
1042                              "older than our %u\n", epoch, maplen,
1043                              osdc->osdmap->epoch);
1044                 } else {
1045                         dout("taking full map %u len %d\n", epoch, maplen);
1046                         newmap = osdmap_decode(&p, p+maplen);
1047                         if (IS_ERR(newmap)) {
1048                                 err = PTR_ERR(newmap);
1049                                 goto bad;
1050                         }
1051                         BUG_ON(!newmap);
1052                         oldmap = osdc->osdmap;
1053                         osdc->osdmap = newmap;
1054                         if (oldmap)
1055                                 ceph_osdmap_destroy(oldmap);
1056                 }
1057                 p += maplen;
1058                 nr_maps--;
1059         }
1060
1061 done:
1062         downgrade_write(&osdc->map_sem);
1063         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1064         if (newmap)
1065                 kick_requests(osdc, NULL);
1066         up_read(&osdc->map_sem);
1067         return;
1068
1069 bad:
1070         pr_err("osdc handle_map corrupt msg\n");
1071         ceph_msg_dump(msg);
1072         up_write(&osdc->map_sem);
1073         return;
1074 }
1075
1076
1077 /*
1078  * A read request prepares specific pages that data is to be read into.
1079  * When a message is being read off the wire, we call prepare_pages to
1080  * find those pages.
1081  *  0 = success, -1 failure.
1082  */
1083 static int __prepare_pages(struct ceph_connection *con,
1084                          struct ceph_msg_header *hdr,
1085                          struct ceph_osd_request *req,
1086                          u64 tid,
1087                          struct ceph_msg *m)
1088 {
1089         struct ceph_osd *osd = con->private;
1090         struct ceph_osd_client *osdc;
1091         int ret = -1;
1092         int data_len = le32_to_cpu(hdr->data_len);
1093         unsigned data_off = le16_to_cpu(hdr->data_off);
1094
1095         int want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1096
1097         if (!osd)
1098                 return -1;
1099
1100         osdc = osd->o_osdc;
1101
1102         dout("__prepare_pages on msg %p tid %llu, has %d pages, want %d\n", m,
1103              tid, req->r_num_pages, want);
1104         if (unlikely(req->r_num_pages < want))
1105                 goto out;
1106         m->pages = req->r_pages;
1107         m->nr_pages = req->r_num_pages;
1108         ret = 0; /* success */
1109 out:
1110         BUG_ON(ret < 0 || m->nr_pages < want);
1111
1112         return ret;
1113 }
1114
1115 /*
1116  * Register request, send initial attempt.
1117  */
1118 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1119                             struct ceph_osd_request *req,
1120                             bool nofail)
1121 {
1122         int rc = 0;
1123
1124         req->r_request->pages = req->r_pages;
1125         req->r_request->nr_pages = req->r_num_pages;
1126
1127         register_request(osdc, req);
1128
1129         down_read(&osdc->map_sem);
1130         mutex_lock(&osdc->request_mutex);
1131         /*
1132          * a racing kick_requests() may have sent the message for us
1133          * while we dropped request_mutex above, so only send now if
1134          * the request still han't been touched yet.
1135          */
1136         if (req->r_sent == 0) {
1137                 rc = __send_request(osdc, req);
1138                 if (rc) {
1139                         if (nofail) {
1140                                 dout("osdc_start_request failed send, "
1141                                      " marking %lld\n", req->r_tid);
1142                                 req->r_resend = true;
1143                                 rc = 0;
1144                         } else {
1145                                 __unregister_request(osdc, req);
1146                         }
1147                 }
1148         }
1149         mutex_unlock(&osdc->request_mutex);
1150         up_read(&osdc->map_sem);
1151         return rc;
1152 }
1153
1154 /*
1155  * wait for a request to complete
1156  */
1157 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1158                            struct ceph_osd_request *req)
1159 {
1160         int rc;
1161
1162         rc = wait_for_completion_interruptible(&req->r_completion);
1163         if (rc < 0) {
1164                 mutex_lock(&osdc->request_mutex);
1165                 __cancel_request(req);
1166                 __unregister_request(osdc, req);
1167                 mutex_unlock(&osdc->request_mutex);
1168                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1169                 return rc;
1170         }
1171
1172         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1173         return req->r_result;
1174 }
1175
1176 /*
1177  * sync - wait for all in-flight requests to flush.  avoid starvation.
1178  */
1179 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1180 {
1181         struct ceph_osd_request *req;
1182         u64 last_tid, next_tid = 0;
1183
1184         mutex_lock(&osdc->request_mutex);
1185         last_tid = osdc->last_tid;
1186         while (1) {
1187                 req = __lookup_request_ge(osdc, next_tid);
1188                 if (!req)
1189                         break;
1190                 if (req->r_tid > last_tid)
1191                         break;
1192
1193                 next_tid = req->r_tid + 1;
1194                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1195                         continue;
1196
1197                 ceph_osdc_get_request(req);
1198                 mutex_unlock(&osdc->request_mutex);
1199                 dout("sync waiting on tid %llu (last is %llu)\n",
1200                      req->r_tid, last_tid);
1201                 wait_for_completion(&req->r_safe_completion);
1202                 mutex_lock(&osdc->request_mutex);
1203                 ceph_osdc_put_request(req);
1204         }
1205         mutex_unlock(&osdc->request_mutex);
1206         dout("sync done (thru tid %llu)\n", last_tid);
1207 }
1208
1209 /*
1210  * init, shutdown
1211  */
1212 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1213 {
1214         int err;
1215
1216         dout("init\n");
1217         osdc->client = client;
1218         osdc->osdmap = NULL;
1219         init_rwsem(&osdc->map_sem);
1220         init_completion(&osdc->map_waiters);
1221         osdc->last_requested_map = 0;
1222         mutex_init(&osdc->request_mutex);
1223         osdc->last_tid = 0;
1224         osdc->osds = RB_ROOT;
1225         INIT_LIST_HEAD(&osdc->osd_lru);
1226         osdc->requests = RB_ROOT;
1227         INIT_LIST_HEAD(&osdc->req_lru);
1228         osdc->num_requests = 0;
1229         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1230         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1231
1232         schedule_delayed_work(&osdc->osds_timeout_work,
1233            round_jiffies_relative(osdc->client->mount_args->osd_idle_ttl * HZ));
1234
1235         err = -ENOMEM;
1236         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1237                                         sizeof(struct ceph_osd_request));
1238         if (!osdc->req_mempool)
1239                 goto out;
1240
1241         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true);
1242         if (err < 0)
1243                 goto out_mempool;
1244         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1245                                 OSD_OPREPLY_FRONT_LEN, 10, true);
1246         if (err < 0)
1247                 goto out_msgpool;
1248         return 0;
1249
1250 out_msgpool:
1251         ceph_msgpool_destroy(&osdc->msgpool_op);
1252 out_mempool:
1253         mempool_destroy(osdc->req_mempool);
1254 out:
1255         return err;
1256 }
1257
1258 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1259 {
1260         cancel_delayed_work_sync(&osdc->timeout_work);
1261         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1262         if (osdc->osdmap) {
1263                 ceph_osdmap_destroy(osdc->osdmap);
1264                 osdc->osdmap = NULL;
1265         }
1266         remove_old_osds(osdc, 1);
1267         mempool_destroy(osdc->req_mempool);
1268         ceph_msgpool_destroy(&osdc->msgpool_op);
1269         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1270 }
1271
1272 /*
1273  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1274  * *plen.  Return number of bytes read, or error.
1275  */
1276 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1277                         struct ceph_vino vino, struct ceph_file_layout *layout,
1278                         u64 off, u64 *plen,
1279                         u32 truncate_seq, u64 truncate_size,
1280                         struct page **pages, int num_pages)
1281 {
1282         struct ceph_osd_request *req;
1283         int rc = 0;
1284
1285         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1286              vino.snap, off, *plen);
1287         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1288                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1289                                     NULL, 0, truncate_seq, truncate_size, NULL,
1290                                     false, 1);
1291         if (IS_ERR(req))
1292                 return PTR_ERR(req);
1293
1294         /* it may be a short read due to an object boundary */
1295         req->r_pages = pages;
1296         num_pages = calc_pages_for(off, *plen);
1297         req->r_num_pages = num_pages;
1298
1299         dout("readpages  final extent is %llu~%llu (%d pages)\n",
1300              off, *plen, req->r_num_pages);
1301
1302         rc = ceph_osdc_start_request(osdc, req, false);
1303         if (!rc)
1304                 rc = ceph_osdc_wait_request(osdc, req);
1305
1306         ceph_osdc_put_request(req);
1307         dout("readpages result %d\n", rc);
1308         return rc;
1309 }
1310
1311 /*
1312  * do a synchronous write on N pages
1313  */
1314 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1315                          struct ceph_file_layout *layout,
1316                          struct ceph_snap_context *snapc,
1317                          u64 off, u64 len,
1318                          u32 truncate_seq, u64 truncate_size,
1319                          struct timespec *mtime,
1320                          struct page **pages, int num_pages,
1321                          int flags, int do_sync, bool nofail)
1322 {
1323         struct ceph_osd_request *req;
1324         int rc = 0;
1325
1326         BUG_ON(vino.snap != CEPH_NOSNAP);
1327         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1328                                     CEPH_OSD_OP_WRITE,
1329                                     flags | CEPH_OSD_FLAG_ONDISK |
1330                                             CEPH_OSD_FLAG_WRITE,
1331                                     snapc, do_sync,
1332                                     truncate_seq, truncate_size, mtime,
1333                                     nofail, 1);
1334         if (IS_ERR(req))
1335                 return PTR_ERR(req);
1336
1337         /* it may be a short write due to an object boundary */
1338         req->r_pages = pages;
1339         req->r_num_pages = calc_pages_for(off, len);
1340         dout("writepages %llu~%llu (%d pages)\n", off, len,
1341              req->r_num_pages);
1342
1343         rc = ceph_osdc_start_request(osdc, req, nofail);
1344         if (!rc)
1345                 rc = ceph_osdc_wait_request(osdc, req);
1346
1347         ceph_osdc_put_request(req);
1348         if (rc == 0)
1349                 rc = len;
1350         dout("writepages result %d\n", rc);
1351         return rc;
1352 }
1353
1354 /*
1355  * handle incoming message
1356  */
1357 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1358 {
1359         struct ceph_osd *osd = con->private;
1360         struct ceph_osd_client *osdc;
1361         int type = le16_to_cpu(msg->hdr.type);
1362
1363         if (!osd)
1364                 return;
1365         osdc = osd->o_osdc;
1366
1367         switch (type) {
1368         case CEPH_MSG_OSD_MAP:
1369                 ceph_osdc_handle_map(osdc, msg);
1370                 break;
1371         case CEPH_MSG_OSD_OPREPLY:
1372                 handle_reply(osdc, msg, con);
1373                 break;
1374
1375         default:
1376                 pr_err("received unknown message type %d %s\n", type,
1377                        ceph_msg_type_name(type));
1378         }
1379         ceph_msg_put(msg);
1380 }
1381
1382 /*
1383  * lookup and return message for incoming reply
1384  */
1385 static struct ceph_msg *get_reply(struct ceph_connection *con,
1386                                   struct ceph_msg_header *hdr,
1387                                   int *skip)
1388 {
1389         struct ceph_osd *osd = con->private;
1390         struct ceph_osd_client *osdc = osd->o_osdc;
1391         struct ceph_msg *m;
1392         struct ceph_osd_request *req;
1393         int front = le32_to_cpu(hdr->front_len);
1394         int data_len = le32_to_cpu(hdr->data_len);
1395         u64 tid;
1396         int err;
1397
1398         tid = le64_to_cpu(hdr->tid);
1399         mutex_lock(&osdc->request_mutex);
1400         req = __lookup_request(osdc, tid);
1401         if (!req) {
1402                 *skip = 1;
1403                 m = NULL;
1404                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
1405                         osd->o_osd);
1406                 goto out;
1407         }
1408
1409         if (req->r_con_filling_msg) {
1410                 dout("get_reply revoking msg %p from old con %p\n",
1411                      req->r_reply, req->r_con_filling_msg);
1412                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
1413                 ceph_con_put(req->r_con_filling_msg);
1414         }
1415
1416         if (front > req->r_reply->front.iov_len) {
1417                 pr_warning("get_reply front %d > preallocated %d\n",
1418                            front, (int)req->r_reply->front.iov_len);
1419                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, 0, 0, NULL);
1420                 if (IS_ERR(m))
1421                         goto out;
1422                 ceph_msg_put(req->r_reply);
1423                 req->r_reply = m;
1424         }
1425         m = ceph_msg_get(req->r_reply);
1426
1427         if (data_len > 0) {
1428                 err = __prepare_pages(con, hdr, req, tid, m);
1429                 if (err < 0) {
1430                         *skip = 1;
1431                         ceph_msg_put(m);
1432                         m = ERR_PTR(err);
1433                 }
1434         }
1435         *skip = 0;
1436         req->r_con_filling_msg = ceph_con_get(con);
1437         dout("get_reply tid %lld %p\n", tid, m);
1438
1439 out:
1440         mutex_unlock(&osdc->request_mutex);
1441         return m;
1442
1443 }
1444
1445 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1446                                   struct ceph_msg_header *hdr,
1447                                   int *skip)
1448 {
1449         struct ceph_osd *osd = con->private;
1450         int type = le16_to_cpu(hdr->type);
1451         int front = le32_to_cpu(hdr->front_len);
1452
1453         switch (type) {
1454         case CEPH_MSG_OSD_MAP:
1455                 return ceph_msg_new(type, front, 0, 0, NULL);
1456         case CEPH_MSG_OSD_OPREPLY:
1457                 return get_reply(con, hdr, skip);
1458         default:
1459                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
1460                         osd->o_osd);
1461                 *skip = 1;
1462                 return NULL;
1463         }
1464 }
1465
1466 /*
1467  * Wrappers to refcount containing ceph_osd struct
1468  */
1469 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1470 {
1471         struct ceph_osd *osd = con->private;
1472         if (get_osd(osd))
1473                 return con;
1474         return NULL;
1475 }
1476
1477 static void put_osd_con(struct ceph_connection *con)
1478 {
1479         struct ceph_osd *osd = con->private;
1480         put_osd(osd);
1481 }
1482
1483 /*
1484  * authentication
1485  */
1486 static int get_authorizer(struct ceph_connection *con,
1487                           void **buf, int *len, int *proto,
1488                           void **reply_buf, int *reply_len, int force_new)
1489 {
1490         struct ceph_osd *o = con->private;
1491         struct ceph_osd_client *osdc = o->o_osdc;
1492         struct ceph_auth_client *ac = osdc->client->monc.auth;
1493         int ret = 0;
1494
1495         if (force_new && o->o_authorizer) {
1496                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1497                 o->o_authorizer = NULL;
1498         }
1499         if (o->o_authorizer == NULL) {
1500                 ret = ac->ops->create_authorizer(
1501                         ac, CEPH_ENTITY_TYPE_OSD,
1502                         &o->o_authorizer,
1503                         &o->o_authorizer_buf,
1504                         &o->o_authorizer_buf_len,
1505                         &o->o_authorizer_reply_buf,
1506                         &o->o_authorizer_reply_buf_len);
1507                 if (ret)
1508                 return ret;
1509         }
1510
1511         *proto = ac->protocol;
1512         *buf = o->o_authorizer_buf;
1513         *len = o->o_authorizer_buf_len;
1514         *reply_buf = o->o_authorizer_reply_buf;
1515         *reply_len = o->o_authorizer_reply_buf_len;
1516         return 0;
1517 }
1518
1519
1520 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1521 {
1522         struct ceph_osd *o = con->private;
1523         struct ceph_osd_client *osdc = o->o_osdc;
1524         struct ceph_auth_client *ac = osdc->client->monc.auth;
1525
1526         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1527 }
1528
1529 static int invalidate_authorizer(struct ceph_connection *con)
1530 {
1531         struct ceph_osd *o = con->private;
1532         struct ceph_osd_client *osdc = o->o_osdc;
1533         struct ceph_auth_client *ac = osdc->client->monc.auth;
1534
1535         if (ac->ops->invalidate_authorizer)
1536                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
1537
1538         return ceph_monc_validate_auth(&osdc->client->monc);
1539 }
1540
1541 const static struct ceph_connection_operations osd_con_ops = {
1542         .get = get_osd_con,
1543         .put = put_osd_con,
1544         .dispatch = dispatch,
1545         .get_authorizer = get_authorizer,
1546         .verify_authorizer_reply = verify_authorizer_reply,
1547         .invalidate_authorizer = invalidate_authorizer,
1548         .alloc_msg = alloc_msg,
1549         .fault = osd_reset,
1550 };