Merge git://bedivere.hansenpartnership.com/git/scsi-rc-fixes-2.6
[linux-2.6.git] / net / ceph / osd_client.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/module.h>
4 #include <linux/err.h>
5 #include <linux/highmem.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/slab.h>
9 #include <linux/uaccess.h>
10 #ifdef CONFIG_BLOCK
11 #include <linux/bio.h>
12 #endif
13
14 #include <linux/ceph/libceph.h>
15 #include <linux/ceph/osd_client.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/auth.h>
19 #include <linux/ceph/pagelist.h>
20
21 #define OSD_OP_FRONT_LEN        4096
22 #define OSD_OPREPLY_FRONT_LEN   512
23
24 static const struct ceph_connection_operations osd_con_ops;
25
26 static void send_queued(struct ceph_osd_client *osdc);
27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
28 static void __register_request(struct ceph_osd_client *osdc,
29                                struct ceph_osd_request *req);
30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
31                                         struct ceph_osd_request *req);
32 static int __send_request(struct ceph_osd_client *osdc,
33                           struct ceph_osd_request *req);
34
35 static int op_needs_trail(int op)
36 {
37         switch (op) {
38         case CEPH_OSD_OP_GETXATTR:
39         case CEPH_OSD_OP_SETXATTR:
40         case CEPH_OSD_OP_CMPXATTR:
41         case CEPH_OSD_OP_CALL:
42         case CEPH_OSD_OP_NOTIFY:
43                 return 1;
44         default:
45                 return 0;
46         }
47 }
48
49 static int op_has_extent(int op)
50 {
51         return (op == CEPH_OSD_OP_READ ||
52                 op == CEPH_OSD_OP_WRITE);
53 }
54
55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
56                         struct ceph_file_layout *layout,
57                         u64 snapid,
58                         u64 off, u64 *plen, u64 *bno,
59                         struct ceph_osd_request *req,
60                         struct ceph_osd_req_op *op)
61 {
62         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
63         u64 orig_len = *plen;
64         u64 objoff, objlen;    /* extent in object */
65
66         reqhead->snapid = cpu_to_le64(snapid);
67
68         /* object extent? */
69         ceph_calc_file_object_mapping(layout, off, plen, bno,
70                                       &objoff, &objlen);
71         if (*plen < orig_len)
72                 dout(" skipping last %llu, final file extent %llu~%llu\n",
73                      orig_len - *plen, off, *plen);
74
75         if (op_has_extent(op->op)) {
76                 op->extent.offset = objoff;
77                 op->extent.length = objlen;
78         }
79         req->r_num_pages = calc_pages_for(off, *plen);
80         req->r_page_alignment = off & ~PAGE_MASK;
81         if (op->op == CEPH_OSD_OP_WRITE)
82                 op->payload_len = *plen;
83
84         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
85              *bno, objoff, objlen, req->r_num_pages);
86
87 }
88 EXPORT_SYMBOL(ceph_calc_raw_layout);
89
90 /*
91  * Implement client access to distributed object storage cluster.
92  *
93  * All data objects are stored within a cluster/cloud of OSDs, or
94  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
95  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
96  * remote daemons serving up and coordinating consistent and safe
97  * access to storage.
98  *
99  * Cluster membership and the mapping of data objects onto storage devices
100  * are described by the osd map.
101  *
102  * We keep track of pending OSD requests (read, write), resubmit
103  * requests to different OSDs when the cluster topology/data layout
104  * change, or retry the affected requests when the communications
105  * channel with an OSD is reset.
106  */
107
108 /*
109  * calculate the mapping of a file extent onto an object, and fill out the
110  * request accordingly.  shorten extent as necessary if it crosses an
111  * object boundary.
112  *
113  * fill osd op in request message.
114  */
115 static void calc_layout(struct ceph_osd_client *osdc,
116                         struct ceph_vino vino,
117                         struct ceph_file_layout *layout,
118                         u64 off, u64 *plen,
119                         struct ceph_osd_request *req,
120                         struct ceph_osd_req_op *op)
121 {
122         u64 bno;
123
124         ceph_calc_raw_layout(osdc, layout, vino.snap, off,
125                              plen, &bno, req, op);
126
127         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
128         req->r_oid_len = strlen(req->r_oid);
129 }
130
131 /*
132  * requests
133  */
134 void ceph_osdc_release_request(struct kref *kref)
135 {
136         struct ceph_osd_request *req = container_of(kref,
137                                                     struct ceph_osd_request,
138                                                     r_kref);
139
140         if (req->r_request)
141                 ceph_msg_put(req->r_request);
142         if (req->r_reply)
143                 ceph_msg_put(req->r_reply);
144         if (req->r_con_filling_msg) {
145                 dout("release_request revoking pages %p from con %p\n",
146                      req->r_pages, req->r_con_filling_msg);
147                 ceph_con_revoke_message(req->r_con_filling_msg,
148                                       req->r_reply);
149                 ceph_con_put(req->r_con_filling_msg);
150         }
151         if (req->r_own_pages)
152                 ceph_release_page_vector(req->r_pages,
153                                          req->r_num_pages);
154 #ifdef CONFIG_BLOCK
155         if (req->r_bio)
156                 bio_put(req->r_bio);
157 #endif
158         ceph_put_snap_context(req->r_snapc);
159         if (req->r_trail) {
160                 ceph_pagelist_release(req->r_trail);
161                 kfree(req->r_trail);
162         }
163         if (req->r_mempool)
164                 mempool_free(req, req->r_osdc->req_mempool);
165         else
166                 kfree(req);
167 }
168 EXPORT_SYMBOL(ceph_osdc_release_request);
169
170 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
171 {
172         int i = 0;
173
174         if (needs_trail)
175                 *needs_trail = 0;
176         while (ops[i].op) {
177                 if (needs_trail && op_needs_trail(ops[i].op))
178                         *needs_trail = 1;
179                 i++;
180         }
181
182         return i;
183 }
184
185 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
186                                                int flags,
187                                                struct ceph_snap_context *snapc,
188                                                struct ceph_osd_req_op *ops,
189                                                bool use_mempool,
190                                                gfp_t gfp_flags,
191                                                struct page **pages,
192                                                struct bio *bio)
193 {
194         struct ceph_osd_request *req;
195         struct ceph_msg *msg;
196         int needs_trail;
197         int num_op = get_num_ops(ops, &needs_trail);
198         size_t msg_size = sizeof(struct ceph_osd_request_head);
199
200         msg_size += num_op*sizeof(struct ceph_osd_op);
201
202         if (use_mempool) {
203                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
204                 memset(req, 0, sizeof(*req));
205         } else {
206                 req = kzalloc(sizeof(*req), gfp_flags);
207         }
208         if (req == NULL)
209                 return NULL;
210
211         req->r_osdc = osdc;
212         req->r_mempool = use_mempool;
213
214         kref_init(&req->r_kref);
215         init_completion(&req->r_completion);
216         init_completion(&req->r_safe_completion);
217         INIT_LIST_HEAD(&req->r_unsafe_item);
218         INIT_LIST_HEAD(&req->r_linger_item);
219         INIT_LIST_HEAD(&req->r_linger_osd);
220         req->r_flags = flags;
221
222         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
223
224         /* create reply message */
225         if (use_mempool)
226                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
227         else
228                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
229                                    OSD_OPREPLY_FRONT_LEN, gfp_flags);
230         if (!msg) {
231                 ceph_osdc_put_request(req);
232                 return NULL;
233         }
234         req->r_reply = msg;
235
236         /* allocate space for the trailing data */
237         if (needs_trail) {
238                 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
239                 if (!req->r_trail) {
240                         ceph_osdc_put_request(req);
241                         return NULL;
242                 }
243                 ceph_pagelist_init(req->r_trail);
244         }
245         /* create request message; allow space for oid */
246         msg_size += 40;
247         if (snapc)
248                 msg_size += sizeof(u64) * snapc->num_snaps;
249         if (use_mempool)
250                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
251         else
252                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags);
253         if (!msg) {
254                 ceph_osdc_put_request(req);
255                 return NULL;
256         }
257
258         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
259         memset(msg->front.iov_base, 0, msg->front.iov_len);
260
261         req->r_request = msg;
262         req->r_pages = pages;
263 #ifdef CONFIG_BLOCK
264         if (bio) {
265                 req->r_bio = bio;
266                 bio_get(req->r_bio);
267         }
268 #endif
269
270         return req;
271 }
272 EXPORT_SYMBOL(ceph_osdc_alloc_request);
273
274 static void osd_req_encode_op(struct ceph_osd_request *req,
275                               struct ceph_osd_op *dst,
276                               struct ceph_osd_req_op *src)
277 {
278         dst->op = cpu_to_le16(src->op);
279
280         switch (dst->op) {
281         case CEPH_OSD_OP_READ:
282         case CEPH_OSD_OP_WRITE:
283                 dst->extent.offset =
284                         cpu_to_le64(src->extent.offset);
285                 dst->extent.length =
286                         cpu_to_le64(src->extent.length);
287                 dst->extent.truncate_size =
288                         cpu_to_le64(src->extent.truncate_size);
289                 dst->extent.truncate_seq =
290                         cpu_to_le32(src->extent.truncate_seq);
291                 break;
292
293         case CEPH_OSD_OP_GETXATTR:
294         case CEPH_OSD_OP_SETXATTR:
295         case CEPH_OSD_OP_CMPXATTR:
296                 BUG_ON(!req->r_trail);
297
298                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
299                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
300                 dst->xattr.cmp_op = src->xattr.cmp_op;
301                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
302                 ceph_pagelist_append(req->r_trail, src->xattr.name,
303                                      src->xattr.name_len);
304                 ceph_pagelist_append(req->r_trail, src->xattr.val,
305                                      src->xattr.value_len);
306                 break;
307         case CEPH_OSD_OP_CALL:
308                 BUG_ON(!req->r_trail);
309
310                 dst->cls.class_len = src->cls.class_len;
311                 dst->cls.method_len = src->cls.method_len;
312                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
313
314                 ceph_pagelist_append(req->r_trail, src->cls.class_name,
315                                      src->cls.class_len);
316                 ceph_pagelist_append(req->r_trail, src->cls.method_name,
317                                      src->cls.method_len);
318                 ceph_pagelist_append(req->r_trail, src->cls.indata,
319                                      src->cls.indata_len);
320                 break;
321         case CEPH_OSD_OP_ROLLBACK:
322                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
323                 break;
324         case CEPH_OSD_OP_STARTSYNC:
325                 break;
326         case CEPH_OSD_OP_NOTIFY:
327                 {
328                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
329                         __le32 timeout = cpu_to_le32(src->watch.timeout);
330
331                         BUG_ON(!req->r_trail);
332
333                         ceph_pagelist_append(req->r_trail,
334                                                 &prot_ver, sizeof(prot_ver));
335                         ceph_pagelist_append(req->r_trail,
336                                                 &timeout, sizeof(timeout));
337                 }
338         case CEPH_OSD_OP_NOTIFY_ACK:
339         case CEPH_OSD_OP_WATCH:
340                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
341                 dst->watch.ver = cpu_to_le64(src->watch.ver);
342                 dst->watch.flag = src->watch.flag;
343                 break;
344         default:
345                 pr_err("unrecognized osd opcode %d\n", dst->op);
346                 WARN_ON(1);
347                 break;
348         }
349         dst->payload_len = cpu_to_le32(src->payload_len);
350 }
351
352 /*
353  * build new request AND message
354  *
355  */
356 void ceph_osdc_build_request(struct ceph_osd_request *req,
357                              u64 off, u64 *plen,
358                              struct ceph_osd_req_op *src_ops,
359                              struct ceph_snap_context *snapc,
360                              struct timespec *mtime,
361                              const char *oid,
362                              int oid_len)
363 {
364         struct ceph_msg *msg = req->r_request;
365         struct ceph_osd_request_head *head;
366         struct ceph_osd_req_op *src_op;
367         struct ceph_osd_op *op;
368         void *p;
369         int num_op = get_num_ops(src_ops, NULL);
370         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
371         int flags = req->r_flags;
372         u64 data_len = 0;
373         int i;
374
375         head = msg->front.iov_base;
376         op = (void *)(head + 1);
377         p = (void *)(op + num_op);
378
379         req->r_snapc = ceph_get_snap_context(snapc);
380
381         head->client_inc = cpu_to_le32(1); /* always, for now. */
382         head->flags = cpu_to_le32(flags);
383         if (flags & CEPH_OSD_FLAG_WRITE)
384                 ceph_encode_timespec(&head->mtime, mtime);
385         head->num_ops = cpu_to_le16(num_op);
386
387
388         /* fill in oid */
389         head->object_len = cpu_to_le32(oid_len);
390         memcpy(p, oid, oid_len);
391         p += oid_len;
392
393         src_op = src_ops;
394         while (src_op->op) {
395                 osd_req_encode_op(req, op, src_op);
396                 src_op++;
397                 op++;
398         }
399
400         if (req->r_trail)
401                 data_len += req->r_trail->length;
402
403         if (snapc) {
404                 head->snap_seq = cpu_to_le64(snapc->seq);
405                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
406                 for (i = 0; i < snapc->num_snaps; i++) {
407                         put_unaligned_le64(snapc->snaps[i], p);
408                         p += sizeof(u64);
409                 }
410         }
411
412         if (flags & CEPH_OSD_FLAG_WRITE) {
413                 req->r_request->hdr.data_off = cpu_to_le16(off);
414                 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
415         } else if (data_len) {
416                 req->r_request->hdr.data_off = 0;
417                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
418         }
419
420         req->r_request->page_alignment = req->r_page_alignment;
421
422         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
423         msg_size = p - msg->front.iov_base;
424         msg->front.iov_len = msg_size;
425         msg->hdr.front_len = cpu_to_le32(msg_size);
426         return;
427 }
428 EXPORT_SYMBOL(ceph_osdc_build_request);
429
430 /*
431  * build new request AND message, calculate layout, and adjust file
432  * extent as needed.
433  *
434  * if the file was recently truncated, we include information about its
435  * old and new size so that the object can be updated appropriately.  (we
436  * avoid synchronously deleting truncated objects because it's slow.)
437  *
438  * if @do_sync, include a 'startsync' command so that the osd will flush
439  * data quickly.
440  */
441 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
442                                                struct ceph_file_layout *layout,
443                                                struct ceph_vino vino,
444                                                u64 off, u64 *plen,
445                                                int opcode, int flags,
446                                                struct ceph_snap_context *snapc,
447                                                int do_sync,
448                                                u32 truncate_seq,
449                                                u64 truncate_size,
450                                                struct timespec *mtime,
451                                                bool use_mempool, int num_reply,
452                                                int page_align)
453 {
454         struct ceph_osd_req_op ops[3];
455         struct ceph_osd_request *req;
456
457         ops[0].op = opcode;
458         ops[0].extent.truncate_seq = truncate_seq;
459         ops[0].extent.truncate_size = truncate_size;
460         ops[0].payload_len = 0;
461
462         if (do_sync) {
463                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
464                 ops[1].payload_len = 0;
465                 ops[2].op = 0;
466         } else
467                 ops[1].op = 0;
468
469         req = ceph_osdc_alloc_request(osdc, flags,
470                                          snapc, ops,
471                                          use_mempool,
472                                          GFP_NOFS, NULL, NULL);
473         if (!req)
474                 return NULL;
475
476         /* calculate max write size */
477         calc_layout(osdc, vino, layout, off, plen, req, ops);
478         req->r_file_layout = *layout;  /* keep a copy */
479
480         /* in case it differs from natural (file) alignment that
481            calc_layout filled in for us */
482         req->r_num_pages = calc_pages_for(page_align, *plen);
483         req->r_page_alignment = page_align;
484
485         ceph_osdc_build_request(req, off, plen, ops,
486                                 snapc,
487                                 mtime,
488                                 req->r_oid, req->r_oid_len);
489
490         return req;
491 }
492 EXPORT_SYMBOL(ceph_osdc_new_request);
493
494 /*
495  * We keep osd requests in an rbtree, sorted by ->r_tid.
496  */
497 static void __insert_request(struct ceph_osd_client *osdc,
498                              struct ceph_osd_request *new)
499 {
500         struct rb_node **p = &osdc->requests.rb_node;
501         struct rb_node *parent = NULL;
502         struct ceph_osd_request *req = NULL;
503
504         while (*p) {
505                 parent = *p;
506                 req = rb_entry(parent, struct ceph_osd_request, r_node);
507                 if (new->r_tid < req->r_tid)
508                         p = &(*p)->rb_left;
509                 else if (new->r_tid > req->r_tid)
510                         p = &(*p)->rb_right;
511                 else
512                         BUG();
513         }
514
515         rb_link_node(&new->r_node, parent, p);
516         rb_insert_color(&new->r_node, &osdc->requests);
517 }
518
519 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
520                                                  u64 tid)
521 {
522         struct ceph_osd_request *req;
523         struct rb_node *n = osdc->requests.rb_node;
524
525         while (n) {
526                 req = rb_entry(n, struct ceph_osd_request, r_node);
527                 if (tid < req->r_tid)
528                         n = n->rb_left;
529                 else if (tid > req->r_tid)
530                         n = n->rb_right;
531                 else
532                         return req;
533         }
534         return NULL;
535 }
536
537 static struct ceph_osd_request *
538 __lookup_request_ge(struct ceph_osd_client *osdc,
539                     u64 tid)
540 {
541         struct ceph_osd_request *req;
542         struct rb_node *n = osdc->requests.rb_node;
543
544         while (n) {
545                 req = rb_entry(n, struct ceph_osd_request, r_node);
546                 if (tid < req->r_tid) {
547                         if (!n->rb_left)
548                                 return req;
549                         n = n->rb_left;
550                 } else if (tid > req->r_tid) {
551                         n = n->rb_right;
552                 } else {
553                         return req;
554                 }
555         }
556         return NULL;
557 }
558
559 /*
560  * Resubmit requests pending on the given osd.
561  */
562 static void __kick_osd_requests(struct ceph_osd_client *osdc,
563                                 struct ceph_osd *osd)
564 {
565         struct ceph_osd_request *req, *nreq;
566         int err;
567
568         dout("__kick_osd_requests osd%d\n", osd->o_osd);
569         err = __reset_osd(osdc, osd);
570         if (err == -EAGAIN)
571                 return;
572
573         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
574                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
575                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
576                      osd->o_osd);
577                 if (!req->r_linger)
578                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
579         }
580
581         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
582                                  r_linger_osd) {
583                 /*
584                  * reregister request prior to unregistering linger so
585                  * that r_osd is preserved.
586                  */
587                 BUG_ON(!list_empty(&req->r_req_lru_item));
588                 __register_request(osdc, req);
589                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
590                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
591                 __unregister_linger_request(osdc, req);
592                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
593                      osd->o_osd);
594         }
595 }
596
597 static void kick_osd_requests(struct ceph_osd_client *osdc,
598                               struct ceph_osd *kickosd)
599 {
600         mutex_lock(&osdc->request_mutex);
601         __kick_osd_requests(osdc, kickosd);
602         mutex_unlock(&osdc->request_mutex);
603 }
604
605 /*
606  * If the osd connection drops, we need to resubmit all requests.
607  */
608 static void osd_reset(struct ceph_connection *con)
609 {
610         struct ceph_osd *osd = con->private;
611         struct ceph_osd_client *osdc;
612
613         if (!osd)
614                 return;
615         dout("osd_reset osd%d\n", osd->o_osd);
616         osdc = osd->o_osdc;
617         down_read(&osdc->map_sem);
618         kick_osd_requests(osdc, osd);
619         send_queued(osdc);
620         up_read(&osdc->map_sem);
621 }
622
623 /*
624  * Track open sessions with osds.
625  */
626 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
627 {
628         struct ceph_osd *osd;
629
630         osd = kzalloc(sizeof(*osd), GFP_NOFS);
631         if (!osd)
632                 return NULL;
633
634         atomic_set(&osd->o_ref, 1);
635         osd->o_osdc = osdc;
636         INIT_LIST_HEAD(&osd->o_requests);
637         INIT_LIST_HEAD(&osd->o_linger_requests);
638         INIT_LIST_HEAD(&osd->o_osd_lru);
639         osd->o_incarnation = 1;
640
641         ceph_con_init(osdc->client->msgr, &osd->o_con);
642         osd->o_con.private = osd;
643         osd->o_con.ops = &osd_con_ops;
644         osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
645
646         INIT_LIST_HEAD(&osd->o_keepalive_item);
647         return osd;
648 }
649
650 static struct ceph_osd *get_osd(struct ceph_osd *osd)
651 {
652         if (atomic_inc_not_zero(&osd->o_ref)) {
653                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
654                      atomic_read(&osd->o_ref));
655                 return osd;
656         } else {
657                 dout("get_osd %p FAIL\n", osd);
658                 return NULL;
659         }
660 }
661
662 static void put_osd(struct ceph_osd *osd)
663 {
664         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
665              atomic_read(&osd->o_ref) - 1);
666         if (atomic_dec_and_test(&osd->o_ref)) {
667                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
668
669                 if (osd->o_authorizer)
670                         ac->ops->destroy_authorizer(ac, osd->o_authorizer);
671                 kfree(osd);
672         }
673 }
674
675 /*
676  * remove an osd from our map
677  */
678 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
679 {
680         dout("__remove_osd %p\n", osd);
681         BUG_ON(!list_empty(&osd->o_requests));
682         rb_erase(&osd->o_node, &osdc->osds);
683         list_del_init(&osd->o_osd_lru);
684         ceph_con_close(&osd->o_con);
685         put_osd(osd);
686 }
687
688 static void remove_all_osds(struct ceph_osd_client *osdc)
689 {
690         dout("__remove_old_osds %p\n", osdc);
691         mutex_lock(&osdc->request_mutex);
692         while (!RB_EMPTY_ROOT(&osdc->osds)) {
693                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
694                                                 struct ceph_osd, o_node);
695                 __remove_osd(osdc, osd);
696         }
697         mutex_unlock(&osdc->request_mutex);
698 }
699
700 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
701                               struct ceph_osd *osd)
702 {
703         dout("__move_osd_to_lru %p\n", osd);
704         BUG_ON(!list_empty(&osd->o_osd_lru));
705         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
706         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
707 }
708
709 static void __remove_osd_from_lru(struct ceph_osd *osd)
710 {
711         dout("__remove_osd_from_lru %p\n", osd);
712         if (!list_empty(&osd->o_osd_lru))
713                 list_del_init(&osd->o_osd_lru);
714 }
715
716 static void remove_old_osds(struct ceph_osd_client *osdc)
717 {
718         struct ceph_osd *osd, *nosd;
719
720         dout("__remove_old_osds %p\n", osdc);
721         mutex_lock(&osdc->request_mutex);
722         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
723                 if (time_before(jiffies, osd->lru_ttl))
724                         break;
725                 __remove_osd(osdc, osd);
726         }
727         mutex_unlock(&osdc->request_mutex);
728 }
729
730 /*
731  * reset osd connect
732  */
733 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
734 {
735         struct ceph_osd_request *req;
736         int ret = 0;
737
738         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
739         if (list_empty(&osd->o_requests) &&
740             list_empty(&osd->o_linger_requests)) {
741                 __remove_osd(osdc, osd);
742         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
743                           &osd->o_con.peer_addr,
744                           sizeof(osd->o_con.peer_addr)) == 0 &&
745                    !ceph_con_opened(&osd->o_con)) {
746                 dout(" osd addr hasn't changed and connection never opened,"
747                      " letting msgr retry");
748                 /* touch each r_stamp for handle_timeout()'s benfit */
749                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
750                         req->r_stamp = jiffies;
751                 ret = -EAGAIN;
752         } else {
753                 ceph_con_close(&osd->o_con);
754                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
755                 osd->o_incarnation++;
756         }
757         return ret;
758 }
759
760 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
761 {
762         struct rb_node **p = &osdc->osds.rb_node;
763         struct rb_node *parent = NULL;
764         struct ceph_osd *osd = NULL;
765
766         dout("__insert_osd %p osd%d\n", new, new->o_osd);
767         while (*p) {
768                 parent = *p;
769                 osd = rb_entry(parent, struct ceph_osd, o_node);
770                 if (new->o_osd < osd->o_osd)
771                         p = &(*p)->rb_left;
772                 else if (new->o_osd > osd->o_osd)
773                         p = &(*p)->rb_right;
774                 else
775                         BUG();
776         }
777
778         rb_link_node(&new->o_node, parent, p);
779         rb_insert_color(&new->o_node, &osdc->osds);
780 }
781
782 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
783 {
784         struct ceph_osd *osd;
785         struct rb_node *n = osdc->osds.rb_node;
786
787         while (n) {
788                 osd = rb_entry(n, struct ceph_osd, o_node);
789                 if (o < osd->o_osd)
790                         n = n->rb_left;
791                 else if (o > osd->o_osd)
792                         n = n->rb_right;
793                 else
794                         return osd;
795         }
796         return NULL;
797 }
798
799 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
800 {
801         schedule_delayed_work(&osdc->timeout_work,
802                         osdc->client->options->osd_keepalive_timeout * HZ);
803 }
804
805 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
806 {
807         cancel_delayed_work(&osdc->timeout_work);
808 }
809
810 /*
811  * Register request, assign tid.  If this is the first request, set up
812  * the timeout event.
813  */
814 static void __register_request(struct ceph_osd_client *osdc,
815                                struct ceph_osd_request *req)
816 {
817         req->r_tid = ++osdc->last_tid;
818         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
819         INIT_LIST_HEAD(&req->r_req_lru_item);
820
821         dout("__register_request %p tid %lld\n", req, req->r_tid);
822         __insert_request(osdc, req);
823         ceph_osdc_get_request(req);
824         osdc->num_requests++;
825
826         if (osdc->num_requests == 1) {
827                 dout(" first request, scheduling timeout\n");
828                 __schedule_osd_timeout(osdc);
829         }
830 }
831
832 static void register_request(struct ceph_osd_client *osdc,
833                              struct ceph_osd_request *req)
834 {
835         mutex_lock(&osdc->request_mutex);
836         __register_request(osdc, req);
837         mutex_unlock(&osdc->request_mutex);
838 }
839
840 /*
841  * called under osdc->request_mutex
842  */
843 static void __unregister_request(struct ceph_osd_client *osdc,
844                                  struct ceph_osd_request *req)
845 {
846         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
847         rb_erase(&req->r_node, &osdc->requests);
848         osdc->num_requests--;
849
850         if (req->r_osd) {
851                 /* make sure the original request isn't in flight. */
852                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
853
854                 list_del_init(&req->r_osd_item);
855                 if (list_empty(&req->r_osd->o_requests) &&
856                     list_empty(&req->r_osd->o_linger_requests)) {
857                         dout("moving osd to %p lru\n", req->r_osd);
858                         __move_osd_to_lru(osdc, req->r_osd);
859                 }
860                 if (list_empty(&req->r_linger_item))
861                         req->r_osd = NULL;
862         }
863
864         ceph_osdc_put_request(req);
865
866         list_del_init(&req->r_req_lru_item);
867         if (osdc->num_requests == 0) {
868                 dout(" no requests, canceling timeout\n");
869                 __cancel_osd_timeout(osdc);
870         }
871 }
872
873 /*
874  * Cancel a previously queued request message
875  */
876 static void __cancel_request(struct ceph_osd_request *req)
877 {
878         if (req->r_sent && req->r_osd) {
879                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
880                 req->r_sent = 0;
881         }
882 }
883
884 static void __register_linger_request(struct ceph_osd_client *osdc,
885                                     struct ceph_osd_request *req)
886 {
887         dout("__register_linger_request %p\n", req);
888         list_add_tail(&req->r_linger_item, &osdc->req_linger);
889         list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests);
890 }
891
892 static void __unregister_linger_request(struct ceph_osd_client *osdc,
893                                         struct ceph_osd_request *req)
894 {
895         dout("__unregister_linger_request %p\n", req);
896         if (req->r_osd) {
897                 list_del_init(&req->r_linger_item);
898                 list_del_init(&req->r_linger_osd);
899
900                 if (list_empty(&req->r_osd->o_requests) &&
901                     list_empty(&req->r_osd->o_linger_requests)) {
902                         dout("moving osd to %p lru\n", req->r_osd);
903                         __move_osd_to_lru(osdc, req->r_osd);
904                 }
905                 if (list_empty(&req->r_osd_item))
906                         req->r_osd = NULL;
907         }
908 }
909
910 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
911                                          struct ceph_osd_request *req)
912 {
913         mutex_lock(&osdc->request_mutex);
914         if (req->r_linger) {
915                 __unregister_linger_request(osdc, req);
916                 ceph_osdc_put_request(req);
917         }
918         mutex_unlock(&osdc->request_mutex);
919 }
920 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
921
922 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
923                                   struct ceph_osd_request *req)
924 {
925         if (!req->r_linger) {
926                 dout("set_request_linger %p\n", req);
927                 req->r_linger = 1;
928                 /*
929                  * caller is now responsible for calling
930                  * unregister_linger_request
931                  */
932                 ceph_osdc_get_request(req);
933         }
934 }
935 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
936
937 /*
938  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
939  * (as needed), and set the request r_osd appropriately.  If there is
940  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
941  * (unsent, homeless) or leave on in-flight lru.
942  *
943  * Return 0 if unchanged, 1 if changed, or negative on error.
944  *
945  * Caller should hold map_sem for read and request_mutex.
946  */
947 static int __map_request(struct ceph_osd_client *osdc,
948                          struct ceph_osd_request *req)
949 {
950         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
951         struct ceph_pg pgid;
952         int acting[CEPH_PG_MAX_SIZE];
953         int o = -1, num = 0;
954         int err;
955
956         dout("map_request %p tid %lld\n", req, req->r_tid);
957         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
958                                       &req->r_file_layout, osdc->osdmap);
959         if (err) {
960                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
961                 return err;
962         }
963         pgid = reqhead->layout.ol_pgid;
964         req->r_pgid = pgid;
965
966         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
967         if (err > 0) {
968                 o = acting[0];
969                 num = err;
970         }
971
972         if ((req->r_osd && req->r_osd->o_osd == o &&
973              req->r_sent >= req->r_osd->o_incarnation &&
974              req->r_num_pg_osds == num &&
975              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
976             (req->r_osd == NULL && o == -1))
977                 return 0;  /* no change */
978
979         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
980              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
981              req->r_osd ? req->r_osd->o_osd : -1);
982
983         /* record full pg acting set */
984         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
985         req->r_num_pg_osds = num;
986
987         if (req->r_osd) {
988                 __cancel_request(req);
989                 list_del_init(&req->r_osd_item);
990                 req->r_osd = NULL;
991         }
992
993         req->r_osd = __lookup_osd(osdc, o);
994         if (!req->r_osd && o >= 0) {
995                 err = -ENOMEM;
996                 req->r_osd = create_osd(osdc);
997                 if (!req->r_osd) {
998                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
999                         goto out;
1000                 }
1001
1002                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1003                 req->r_osd->o_osd = o;
1004                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
1005                 __insert_osd(osdc, req->r_osd);
1006
1007                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
1008         }
1009
1010         if (req->r_osd) {
1011                 __remove_osd_from_lru(req->r_osd);
1012                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1013                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1014         } else {
1015                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1016         }
1017         err = 1;   /* osd or pg changed */
1018
1019 out:
1020         return err;
1021 }
1022
1023 /*
1024  * caller should hold map_sem (for read) and request_mutex
1025  */
1026 static int __send_request(struct ceph_osd_client *osdc,
1027                           struct ceph_osd_request *req)
1028 {
1029         struct ceph_osd_request_head *reqhead;
1030
1031         dout("send_request %p tid %llu to osd%d flags %d\n",
1032              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1033
1034         reqhead = req->r_request->front.iov_base;
1035         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1036         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1037         reqhead->reassert_version = req->r_reassert_version;
1038
1039         req->r_stamp = jiffies;
1040         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1041
1042         ceph_msg_get(req->r_request); /* send consumes a ref */
1043         ceph_con_send(&req->r_osd->o_con, req->r_request);
1044         req->r_sent = req->r_osd->o_incarnation;
1045         return 0;
1046 }
1047
1048 /*
1049  * Send any requests in the queue (req_unsent).
1050  */
1051 static void send_queued(struct ceph_osd_client *osdc)
1052 {
1053         struct ceph_osd_request *req, *tmp;
1054
1055         dout("send_queued\n");
1056         mutex_lock(&osdc->request_mutex);
1057         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1058                 __send_request(osdc, req);
1059         }
1060         mutex_unlock(&osdc->request_mutex);
1061 }
1062
1063 /*
1064  * Timeout callback, called every N seconds when 1 or more osd
1065  * requests has been active for more than N seconds.  When this
1066  * happens, we ping all OSDs with requests who have timed out to
1067  * ensure any communications channel reset is detected.  Reset the
1068  * request timeouts another N seconds in the future as we go.
1069  * Reschedule the timeout event another N seconds in future (unless
1070  * there are no open requests).
1071  */
1072 static void handle_timeout(struct work_struct *work)
1073 {
1074         struct ceph_osd_client *osdc =
1075                 container_of(work, struct ceph_osd_client, timeout_work.work);
1076         struct ceph_osd_request *req, *last_req = NULL;
1077         struct ceph_osd *osd;
1078         unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1079         unsigned long keepalive =
1080                 osdc->client->options->osd_keepalive_timeout * HZ;
1081         unsigned long last_stamp = 0;
1082         struct list_head slow_osds;
1083         dout("timeout\n");
1084         down_read(&osdc->map_sem);
1085
1086         ceph_monc_request_next_osdmap(&osdc->client->monc);
1087
1088         mutex_lock(&osdc->request_mutex);
1089
1090         /*
1091          * reset osds that appear to be _really_ unresponsive.  this
1092          * is a failsafe measure.. we really shouldn't be getting to
1093          * this point if the system is working properly.  the monitors
1094          * should mark the osd as failed and we should find out about
1095          * it from an updated osd map.
1096          */
1097         while (timeout && !list_empty(&osdc->req_lru)) {
1098                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1099                                  r_req_lru_item);
1100
1101                 /* hasn't been long enough since we sent it? */
1102                 if (time_before(jiffies, req->r_stamp + timeout))
1103                         break;
1104
1105                 /* hasn't been long enough since it was acked? */
1106                 if (req->r_request->ack_stamp == 0 ||
1107                     time_before(jiffies, req->r_request->ack_stamp + timeout))
1108                         break;
1109
1110                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
1111                 last_req = req;
1112                 last_stamp = req->r_stamp;
1113
1114                 osd = req->r_osd;
1115                 BUG_ON(!osd);
1116                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1117                            req->r_tid, osd->o_osd);
1118                 __kick_osd_requests(osdc, osd);
1119         }
1120
1121         /*
1122          * ping osds that are a bit slow.  this ensures that if there
1123          * is a break in the TCP connection we will notice, and reopen
1124          * a connection with that osd (from the fault callback).
1125          */
1126         INIT_LIST_HEAD(&slow_osds);
1127         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1128                 if (time_before(jiffies, req->r_stamp + keepalive))
1129                         break;
1130
1131                 osd = req->r_osd;
1132                 BUG_ON(!osd);
1133                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1134                      req->r_tid, osd->o_osd);
1135                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1136         }
1137         while (!list_empty(&slow_osds)) {
1138                 osd = list_entry(slow_osds.next, struct ceph_osd,
1139                                  o_keepalive_item);
1140                 list_del_init(&osd->o_keepalive_item);
1141                 ceph_con_keepalive(&osd->o_con);
1142         }
1143
1144         __schedule_osd_timeout(osdc);
1145         mutex_unlock(&osdc->request_mutex);
1146         send_queued(osdc);
1147         up_read(&osdc->map_sem);
1148 }
1149
1150 static void handle_osds_timeout(struct work_struct *work)
1151 {
1152         struct ceph_osd_client *osdc =
1153                 container_of(work, struct ceph_osd_client,
1154                              osds_timeout_work.work);
1155         unsigned long delay =
1156                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1157
1158         dout("osds timeout\n");
1159         down_read(&osdc->map_sem);
1160         remove_old_osds(osdc);
1161         up_read(&osdc->map_sem);
1162
1163         schedule_delayed_work(&osdc->osds_timeout_work,
1164                               round_jiffies_relative(delay));
1165 }
1166
1167 static void complete_request(struct ceph_osd_request *req)
1168 {
1169         if (req->r_safe_callback)
1170                 req->r_safe_callback(req, NULL);
1171         complete_all(&req->r_safe_completion);  /* fsync waiter */
1172 }
1173
1174 /*
1175  * handle osd op reply.  either call the callback if it is specified,
1176  * or do the completion to wake up the waiting thread.
1177  */
1178 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1179                          struct ceph_connection *con)
1180 {
1181         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1182         struct ceph_osd_request *req;
1183         u64 tid;
1184         int numops, object_len, flags;
1185         s32 result;
1186
1187         tid = le64_to_cpu(msg->hdr.tid);
1188         if (msg->front.iov_len < sizeof(*rhead))
1189                 goto bad;
1190         numops = le32_to_cpu(rhead->num_ops);
1191         object_len = le32_to_cpu(rhead->object_len);
1192         result = le32_to_cpu(rhead->result);
1193         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1194             numops * sizeof(struct ceph_osd_op))
1195                 goto bad;
1196         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1197         /* lookup */
1198         mutex_lock(&osdc->request_mutex);
1199         req = __lookup_request(osdc, tid);
1200         if (req == NULL) {
1201                 dout("handle_reply tid %llu dne\n", tid);
1202                 mutex_unlock(&osdc->request_mutex);
1203                 return;
1204         }
1205         ceph_osdc_get_request(req);
1206         flags = le32_to_cpu(rhead->flags);
1207
1208         /*
1209          * if this connection filled our message, drop our reference now, to
1210          * avoid a (safe but slower) revoke later.
1211          */
1212         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1213                 dout(" dropping con_filling_msg ref %p\n", con);
1214                 req->r_con_filling_msg = NULL;
1215                 ceph_con_put(con);
1216         }
1217
1218         if (!req->r_got_reply) {
1219                 unsigned bytes;
1220
1221                 req->r_result = le32_to_cpu(rhead->result);
1222                 bytes = le32_to_cpu(msg->hdr.data_len);
1223                 dout("handle_reply result %d bytes %d\n", req->r_result,
1224                      bytes);
1225                 if (req->r_result == 0)
1226                         req->r_result = bytes;
1227
1228                 /* in case this is a write and we need to replay, */
1229                 req->r_reassert_version = rhead->reassert_version;
1230
1231                 req->r_got_reply = 1;
1232         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1233                 dout("handle_reply tid %llu dup ack\n", tid);
1234                 mutex_unlock(&osdc->request_mutex);
1235                 goto done;
1236         }
1237
1238         dout("handle_reply tid %llu flags %d\n", tid, flags);
1239
1240         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1241                 __register_linger_request(osdc, req);
1242
1243         /* either this is a read, or we got the safe response */
1244         if (result < 0 ||
1245             (flags & CEPH_OSD_FLAG_ONDISK) ||
1246             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1247                 __unregister_request(osdc, req);
1248
1249         mutex_unlock(&osdc->request_mutex);
1250
1251         if (req->r_callback)
1252                 req->r_callback(req, msg);
1253         else
1254                 complete_all(&req->r_completion);
1255
1256         if (flags & CEPH_OSD_FLAG_ONDISK)
1257                 complete_request(req);
1258
1259 done:
1260         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1261         ceph_osdc_put_request(req);
1262         return;
1263
1264 bad:
1265         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1266                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1267                (int)sizeof(*rhead));
1268         ceph_msg_dump(msg);
1269 }
1270
1271 static void reset_changed_osds(struct ceph_osd_client *osdc)
1272 {
1273         struct rb_node *p, *n;
1274
1275         for (p = rb_first(&osdc->osds); p; p = n) {
1276                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1277
1278                 n = rb_next(p);
1279                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1280                     memcmp(&osd->o_con.peer_addr,
1281                            ceph_osd_addr(osdc->osdmap,
1282                                          osd->o_osd),
1283                            sizeof(struct ceph_entity_addr)) != 0)
1284                         __reset_osd(osdc, osd);
1285         }
1286 }
1287
1288 /*
1289  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1290  * no osd, request a new map.
1291  *
1292  * Caller should hold map_sem for read and request_mutex.
1293  */
1294 static void kick_requests(struct ceph_osd_client *osdc)
1295 {
1296         struct ceph_osd_request *req, *nreq;
1297         struct rb_node *p;
1298         int needmap = 0;
1299         int err;
1300
1301         dout("kick_requests\n");
1302         mutex_lock(&osdc->request_mutex);
1303         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1304                 req = rb_entry(p, struct ceph_osd_request, r_node);
1305                 err = __map_request(osdc, req);
1306                 if (err < 0)
1307                         continue;  /* error */
1308                 if (req->r_osd == NULL) {
1309                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1310                         needmap++;  /* request a newer map */
1311                 } else if (err > 0) {
1312                         dout("%p tid %llu requeued on osd%d\n", req, req->r_tid,
1313                              req->r_osd ? req->r_osd->o_osd : -1);
1314                         if (!req->r_linger)
1315                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1316                 }
1317         }
1318
1319         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1320                                  r_linger_item) {
1321                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1322
1323                 err = __map_request(osdc, req);
1324                 if (err == 0)
1325                         continue;  /* no change and no osd was specified */
1326                 if (err < 0)
1327                         continue;  /* hrm! */
1328                 if (req->r_osd == NULL) {
1329                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1330                         needmap++;  /* request a newer map */
1331                         continue;
1332                 }
1333
1334                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1335                      req->r_osd ? req->r_osd->o_osd : -1);
1336                 __unregister_linger_request(osdc, req);
1337                 __register_request(osdc, req);
1338         }
1339         mutex_unlock(&osdc->request_mutex);
1340
1341         if (needmap) {
1342                 dout("%d requests for down osds, need new map\n", needmap);
1343                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1344         }
1345 }
1346
1347
1348 /*
1349  * Process updated osd map.
1350  *
1351  * The message contains any number of incremental and full maps, normally
1352  * indicating some sort of topology change in the cluster.  Kick requests
1353  * off to different OSDs as needed.
1354  */
1355 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1356 {
1357         void *p, *end, *next;
1358         u32 nr_maps, maplen;
1359         u32 epoch;
1360         struct ceph_osdmap *newmap = NULL, *oldmap;
1361         int err;
1362         struct ceph_fsid fsid;
1363
1364         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1365         p = msg->front.iov_base;
1366         end = p + msg->front.iov_len;
1367
1368         /* verify fsid */
1369         ceph_decode_need(&p, end, sizeof(fsid), bad);
1370         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1371         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1372                 return;
1373
1374         down_write(&osdc->map_sem);
1375
1376         /* incremental maps */
1377         ceph_decode_32_safe(&p, end, nr_maps, bad);
1378         dout(" %d inc maps\n", nr_maps);
1379         while (nr_maps > 0) {
1380                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1381                 epoch = ceph_decode_32(&p);
1382                 maplen = ceph_decode_32(&p);
1383                 ceph_decode_need(&p, end, maplen, bad);
1384                 next = p + maplen;
1385                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1386                         dout("applying incremental map %u len %d\n",
1387                              epoch, maplen);
1388                         newmap = osdmap_apply_incremental(&p, next,
1389                                                           osdc->osdmap,
1390                                                           osdc->client->msgr);
1391                         if (IS_ERR(newmap)) {
1392                                 err = PTR_ERR(newmap);
1393                                 goto bad;
1394                         }
1395                         BUG_ON(!newmap);
1396                         if (newmap != osdc->osdmap) {
1397                                 ceph_osdmap_destroy(osdc->osdmap);
1398                                 osdc->osdmap = newmap;
1399                         }
1400                         kick_requests(osdc);
1401                         reset_changed_osds(osdc);
1402                 } else {
1403                         dout("ignoring incremental map %u len %d\n",
1404                              epoch, maplen);
1405                 }
1406                 p = next;
1407                 nr_maps--;
1408         }
1409         if (newmap)
1410                 goto done;
1411
1412         /* full maps */
1413         ceph_decode_32_safe(&p, end, nr_maps, bad);
1414         dout(" %d full maps\n", nr_maps);
1415         while (nr_maps) {
1416                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1417                 epoch = ceph_decode_32(&p);
1418                 maplen = ceph_decode_32(&p);
1419                 ceph_decode_need(&p, end, maplen, bad);
1420                 if (nr_maps > 1) {
1421                         dout("skipping non-latest full map %u len %d\n",
1422                              epoch, maplen);
1423                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1424                         dout("skipping full map %u len %d, "
1425                              "older than our %u\n", epoch, maplen,
1426                              osdc->osdmap->epoch);
1427                 } else {
1428                         dout("taking full map %u len %d\n", epoch, maplen);
1429                         newmap = osdmap_decode(&p, p+maplen);
1430                         if (IS_ERR(newmap)) {
1431                                 err = PTR_ERR(newmap);
1432                                 goto bad;
1433                         }
1434                         BUG_ON(!newmap);
1435                         oldmap = osdc->osdmap;
1436                         osdc->osdmap = newmap;
1437                         if (oldmap)
1438                                 ceph_osdmap_destroy(oldmap);
1439                         kick_requests(osdc);
1440                 }
1441                 p += maplen;
1442                 nr_maps--;
1443         }
1444
1445 done:
1446         downgrade_write(&osdc->map_sem);
1447         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1448
1449         /*
1450          * subscribe to subsequent osdmap updates if full to ensure
1451          * we find out when we are no longer full and stop returning
1452          * ENOSPC.
1453          */
1454         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1455                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1456
1457         send_queued(osdc);
1458         up_read(&osdc->map_sem);
1459         wake_up_all(&osdc->client->auth_wq);
1460         return;
1461
1462 bad:
1463         pr_err("osdc handle_map corrupt msg\n");
1464         ceph_msg_dump(msg);
1465         up_write(&osdc->map_sem);
1466         return;
1467 }
1468
1469 /*
1470  * watch/notify callback event infrastructure
1471  *
1472  * These callbacks are used both for watch and notify operations.
1473  */
1474 static void __release_event(struct kref *kref)
1475 {
1476         struct ceph_osd_event *event =
1477                 container_of(kref, struct ceph_osd_event, kref);
1478
1479         dout("__release_event %p\n", event);
1480         kfree(event);
1481 }
1482
1483 static void get_event(struct ceph_osd_event *event)
1484 {
1485         kref_get(&event->kref);
1486 }
1487
1488 void ceph_osdc_put_event(struct ceph_osd_event *event)
1489 {
1490         kref_put(&event->kref, __release_event);
1491 }
1492 EXPORT_SYMBOL(ceph_osdc_put_event);
1493
1494 static void __insert_event(struct ceph_osd_client *osdc,
1495                              struct ceph_osd_event *new)
1496 {
1497         struct rb_node **p = &osdc->event_tree.rb_node;
1498         struct rb_node *parent = NULL;
1499         struct ceph_osd_event *event = NULL;
1500
1501         while (*p) {
1502                 parent = *p;
1503                 event = rb_entry(parent, struct ceph_osd_event, node);
1504                 if (new->cookie < event->cookie)
1505                         p = &(*p)->rb_left;
1506                 else if (new->cookie > event->cookie)
1507                         p = &(*p)->rb_right;
1508                 else
1509                         BUG();
1510         }
1511
1512         rb_link_node(&new->node, parent, p);
1513         rb_insert_color(&new->node, &osdc->event_tree);
1514 }
1515
1516 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1517                                                 u64 cookie)
1518 {
1519         struct rb_node **p = &osdc->event_tree.rb_node;
1520         struct rb_node *parent = NULL;
1521         struct ceph_osd_event *event = NULL;
1522
1523         while (*p) {
1524                 parent = *p;
1525                 event = rb_entry(parent, struct ceph_osd_event, node);
1526                 if (cookie < event->cookie)
1527                         p = &(*p)->rb_left;
1528                 else if (cookie > event->cookie)
1529                         p = &(*p)->rb_right;
1530                 else
1531                         return event;
1532         }
1533         return NULL;
1534 }
1535
1536 static void __remove_event(struct ceph_osd_event *event)
1537 {
1538         struct ceph_osd_client *osdc = event->osdc;
1539
1540         if (!RB_EMPTY_NODE(&event->node)) {
1541                 dout("__remove_event removed %p\n", event);
1542                 rb_erase(&event->node, &osdc->event_tree);
1543                 ceph_osdc_put_event(event);
1544         } else {
1545                 dout("__remove_event didn't remove %p\n", event);
1546         }
1547 }
1548
1549 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1550                            void (*event_cb)(u64, u64, u8, void *),
1551                            int one_shot, void *data,
1552                            struct ceph_osd_event **pevent)
1553 {
1554         struct ceph_osd_event *event;
1555
1556         event = kmalloc(sizeof(*event), GFP_NOIO);
1557         if (!event)
1558                 return -ENOMEM;
1559
1560         dout("create_event %p\n", event);
1561         event->cb = event_cb;
1562         event->one_shot = one_shot;
1563         event->data = data;
1564         event->osdc = osdc;
1565         INIT_LIST_HEAD(&event->osd_node);
1566         kref_init(&event->kref);   /* one ref for us */
1567         kref_get(&event->kref);    /* one ref for the caller */
1568         init_completion(&event->completion);
1569
1570         spin_lock(&osdc->event_lock);
1571         event->cookie = ++osdc->event_count;
1572         __insert_event(osdc, event);
1573         spin_unlock(&osdc->event_lock);
1574
1575         *pevent = event;
1576         return 0;
1577 }
1578 EXPORT_SYMBOL(ceph_osdc_create_event);
1579
1580 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1581 {
1582         struct ceph_osd_client *osdc = event->osdc;
1583
1584         dout("cancel_event %p\n", event);
1585         spin_lock(&osdc->event_lock);
1586         __remove_event(event);
1587         spin_unlock(&osdc->event_lock);
1588         ceph_osdc_put_event(event); /* caller's */
1589 }
1590 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1591
1592
1593 static void do_event_work(struct work_struct *work)
1594 {
1595         struct ceph_osd_event_work *event_work =
1596                 container_of(work, struct ceph_osd_event_work, work);
1597         struct ceph_osd_event *event = event_work->event;
1598         u64 ver = event_work->ver;
1599         u64 notify_id = event_work->notify_id;
1600         u8 opcode = event_work->opcode;
1601
1602         dout("do_event_work completing %p\n", event);
1603         event->cb(ver, notify_id, opcode, event->data);
1604         complete(&event->completion);
1605         dout("do_event_work completed %p\n", event);
1606         ceph_osdc_put_event(event);
1607         kfree(event_work);
1608 }
1609
1610
1611 /*
1612  * Process osd watch notifications
1613  */
1614 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1615 {
1616         void *p, *end;
1617         u8 proto_ver;
1618         u64 cookie, ver, notify_id;
1619         u8 opcode;
1620         struct ceph_osd_event *event;
1621         struct ceph_osd_event_work *event_work;
1622
1623         p = msg->front.iov_base;
1624         end = p + msg->front.iov_len;
1625
1626         ceph_decode_8_safe(&p, end, proto_ver, bad);
1627         ceph_decode_8_safe(&p, end, opcode, bad);
1628         ceph_decode_64_safe(&p, end, cookie, bad);
1629         ceph_decode_64_safe(&p, end, ver, bad);
1630         ceph_decode_64_safe(&p, end, notify_id, bad);
1631
1632         spin_lock(&osdc->event_lock);
1633         event = __find_event(osdc, cookie);
1634         if (event) {
1635                 get_event(event);
1636                 if (event->one_shot)
1637                         __remove_event(event);
1638         }
1639         spin_unlock(&osdc->event_lock);
1640         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1641              cookie, ver, event);
1642         if (event) {
1643                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1644                 if (!event_work) {
1645                         dout("ERROR: could not allocate event_work\n");
1646                         goto done_err;
1647                 }
1648                 INIT_WORK(&event_work->work, do_event_work);
1649                 event_work->event = event;
1650                 event_work->ver = ver;
1651                 event_work->notify_id = notify_id;
1652                 event_work->opcode = opcode;
1653                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1654                         dout("WARNING: failed to queue notify event work\n");
1655                         goto done_err;
1656                 }
1657         }
1658
1659         return;
1660
1661 done_err:
1662         complete(&event->completion);
1663         ceph_osdc_put_event(event);
1664         return;
1665
1666 bad:
1667         pr_err("osdc handle_watch_notify corrupt msg\n");
1668         return;
1669 }
1670
1671 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1672 {
1673         int err;
1674
1675         dout("wait_event %p\n", event);
1676         err = wait_for_completion_interruptible_timeout(&event->completion,
1677                                                         timeout * HZ);
1678         ceph_osdc_put_event(event);
1679         if (err > 0)
1680                 err = 0;
1681         dout("wait_event %p returns %d\n", event, err);
1682         return err;
1683 }
1684 EXPORT_SYMBOL(ceph_osdc_wait_event);
1685
1686 /*
1687  * Register request, send initial attempt.
1688  */
1689 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1690                             struct ceph_osd_request *req,
1691                             bool nofail)
1692 {
1693         int rc = 0;
1694
1695         req->r_request->pages = req->r_pages;
1696         req->r_request->nr_pages = req->r_num_pages;
1697 #ifdef CONFIG_BLOCK
1698         req->r_request->bio = req->r_bio;
1699 #endif
1700         req->r_request->trail = req->r_trail;
1701
1702         register_request(osdc, req);
1703
1704         down_read(&osdc->map_sem);
1705         mutex_lock(&osdc->request_mutex);
1706         /*
1707          * a racing kick_requests() may have sent the message for us
1708          * while we dropped request_mutex above, so only send now if
1709          * the request still han't been touched yet.
1710          */
1711         if (req->r_sent == 0) {
1712                 rc = __map_request(osdc, req);
1713                 if (rc < 0) {
1714                         if (nofail) {
1715                                 dout("osdc_start_request failed map, "
1716                                      " will retry %lld\n", req->r_tid);
1717                                 rc = 0;
1718                         }
1719                         goto out_unlock;
1720                 }
1721                 if (req->r_osd == NULL) {
1722                         dout("send_request %p no up osds in pg\n", req);
1723                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1724                 } else {
1725                         rc = __send_request(osdc, req);
1726                         if (rc) {
1727                                 if (nofail) {
1728                                         dout("osdc_start_request failed send, "
1729                                              " will retry %lld\n", req->r_tid);
1730                                         rc = 0;
1731                                 } else {
1732                                         __unregister_request(osdc, req);
1733                                 }
1734                         }
1735                 }
1736         }
1737
1738 out_unlock:
1739         mutex_unlock(&osdc->request_mutex);
1740         up_read(&osdc->map_sem);
1741         return rc;
1742 }
1743 EXPORT_SYMBOL(ceph_osdc_start_request);
1744
1745 /*
1746  * wait for a request to complete
1747  */
1748 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1749                            struct ceph_osd_request *req)
1750 {
1751         int rc;
1752
1753         rc = wait_for_completion_interruptible(&req->r_completion);
1754         if (rc < 0) {
1755                 mutex_lock(&osdc->request_mutex);
1756                 __cancel_request(req);
1757                 __unregister_request(osdc, req);
1758                 mutex_unlock(&osdc->request_mutex);
1759                 complete_request(req);
1760                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1761                 return rc;
1762         }
1763
1764         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1765         return req->r_result;
1766 }
1767 EXPORT_SYMBOL(ceph_osdc_wait_request);
1768
1769 /*
1770  * sync - wait for all in-flight requests to flush.  avoid starvation.
1771  */
1772 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1773 {
1774         struct ceph_osd_request *req;
1775         u64 last_tid, next_tid = 0;
1776
1777         mutex_lock(&osdc->request_mutex);
1778         last_tid = osdc->last_tid;
1779         while (1) {
1780                 req = __lookup_request_ge(osdc, next_tid);
1781                 if (!req)
1782                         break;
1783                 if (req->r_tid > last_tid)
1784                         break;
1785
1786                 next_tid = req->r_tid + 1;
1787                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1788                         continue;
1789
1790                 ceph_osdc_get_request(req);
1791                 mutex_unlock(&osdc->request_mutex);
1792                 dout("sync waiting on tid %llu (last is %llu)\n",
1793                      req->r_tid, last_tid);
1794                 wait_for_completion(&req->r_safe_completion);
1795                 mutex_lock(&osdc->request_mutex);
1796                 ceph_osdc_put_request(req);
1797         }
1798         mutex_unlock(&osdc->request_mutex);
1799         dout("sync done (thru tid %llu)\n", last_tid);
1800 }
1801 EXPORT_SYMBOL(ceph_osdc_sync);
1802
1803 /*
1804  * init, shutdown
1805  */
1806 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1807 {
1808         int err;
1809
1810         dout("init\n");
1811         osdc->client = client;
1812         osdc->osdmap = NULL;
1813         init_rwsem(&osdc->map_sem);
1814         init_completion(&osdc->map_waiters);
1815         osdc->last_requested_map = 0;
1816         mutex_init(&osdc->request_mutex);
1817         osdc->last_tid = 0;
1818         osdc->osds = RB_ROOT;
1819         INIT_LIST_HEAD(&osdc->osd_lru);
1820         osdc->requests = RB_ROOT;
1821         INIT_LIST_HEAD(&osdc->req_lru);
1822         INIT_LIST_HEAD(&osdc->req_unsent);
1823         INIT_LIST_HEAD(&osdc->req_notarget);
1824         INIT_LIST_HEAD(&osdc->req_linger);
1825         osdc->num_requests = 0;
1826         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1827         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1828         spin_lock_init(&osdc->event_lock);
1829         osdc->event_tree = RB_ROOT;
1830         osdc->event_count = 0;
1831
1832         schedule_delayed_work(&osdc->osds_timeout_work,
1833            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1834
1835         err = -ENOMEM;
1836         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1837                                         sizeof(struct ceph_osd_request));
1838         if (!osdc->req_mempool)
1839                 goto out;
1840
1841         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1842                                 "osd_op");
1843         if (err < 0)
1844                 goto out_mempool;
1845         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1846                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1847                                 "osd_op_reply");
1848         if (err < 0)
1849                 goto out_msgpool;
1850
1851         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1852         if (IS_ERR(osdc->notify_wq)) {
1853                 err = PTR_ERR(osdc->notify_wq);
1854                 osdc->notify_wq = NULL;
1855                 goto out_msgpool;
1856         }
1857         return 0;
1858
1859 out_msgpool:
1860         ceph_msgpool_destroy(&osdc->msgpool_op);
1861 out_mempool:
1862         mempool_destroy(osdc->req_mempool);
1863 out:
1864         return err;
1865 }
1866 EXPORT_SYMBOL(ceph_osdc_init);
1867
1868 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1869 {
1870         flush_workqueue(osdc->notify_wq);
1871         destroy_workqueue(osdc->notify_wq);
1872         cancel_delayed_work_sync(&osdc->timeout_work);
1873         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1874         if (osdc->osdmap) {
1875                 ceph_osdmap_destroy(osdc->osdmap);
1876                 osdc->osdmap = NULL;
1877         }
1878         remove_all_osds(osdc);
1879         mempool_destroy(osdc->req_mempool);
1880         ceph_msgpool_destroy(&osdc->msgpool_op);
1881         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1882 }
1883 EXPORT_SYMBOL(ceph_osdc_stop);
1884
1885 /*
1886  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1887  * *plen.  Return number of bytes read, or error.
1888  */
1889 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1890                         struct ceph_vino vino, struct ceph_file_layout *layout,
1891                         u64 off, u64 *plen,
1892                         u32 truncate_seq, u64 truncate_size,
1893                         struct page **pages, int num_pages, int page_align)
1894 {
1895         struct ceph_osd_request *req;
1896         int rc = 0;
1897
1898         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1899              vino.snap, off, *plen);
1900         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1901                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1902                                     NULL, 0, truncate_seq, truncate_size, NULL,
1903                                     false, 1, page_align);
1904         if (!req)
1905                 return -ENOMEM;
1906
1907         /* it may be a short read due to an object boundary */
1908         req->r_pages = pages;
1909
1910         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1911              off, *plen, req->r_num_pages, page_align);
1912
1913         rc = ceph_osdc_start_request(osdc, req, false);
1914         if (!rc)
1915                 rc = ceph_osdc_wait_request(osdc, req);
1916
1917         ceph_osdc_put_request(req);
1918         dout("readpages result %d\n", rc);
1919         return rc;
1920 }
1921 EXPORT_SYMBOL(ceph_osdc_readpages);
1922
1923 /*
1924  * do a synchronous write on N pages
1925  */
1926 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1927                          struct ceph_file_layout *layout,
1928                          struct ceph_snap_context *snapc,
1929                          u64 off, u64 len,
1930                          u32 truncate_seq, u64 truncate_size,
1931                          struct timespec *mtime,
1932                          struct page **pages, int num_pages,
1933                          int flags, int do_sync, bool nofail)
1934 {
1935         struct ceph_osd_request *req;
1936         int rc = 0;
1937         int page_align = off & ~PAGE_MASK;
1938
1939         BUG_ON(vino.snap != CEPH_NOSNAP);
1940         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1941                                     CEPH_OSD_OP_WRITE,
1942                                     flags | CEPH_OSD_FLAG_ONDISK |
1943                                             CEPH_OSD_FLAG_WRITE,
1944                                     snapc, do_sync,
1945                                     truncate_seq, truncate_size, mtime,
1946                                     nofail, 1, page_align);
1947         if (!req)
1948                 return -ENOMEM;
1949
1950         /* it may be a short write due to an object boundary */
1951         req->r_pages = pages;
1952         dout("writepages %llu~%llu (%d pages)\n", off, len,
1953              req->r_num_pages);
1954
1955         rc = ceph_osdc_start_request(osdc, req, nofail);
1956         if (!rc)
1957                 rc = ceph_osdc_wait_request(osdc, req);
1958
1959         ceph_osdc_put_request(req);
1960         if (rc == 0)
1961                 rc = len;
1962         dout("writepages result %d\n", rc);
1963         return rc;
1964 }
1965 EXPORT_SYMBOL(ceph_osdc_writepages);
1966
1967 /*
1968  * handle incoming message
1969  */
1970 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1971 {
1972         struct ceph_osd *osd = con->private;
1973         struct ceph_osd_client *osdc;
1974         int type = le16_to_cpu(msg->hdr.type);
1975
1976         if (!osd)
1977                 goto out;
1978         osdc = osd->o_osdc;
1979
1980         switch (type) {
1981         case CEPH_MSG_OSD_MAP:
1982                 ceph_osdc_handle_map(osdc, msg);
1983                 break;
1984         case CEPH_MSG_OSD_OPREPLY:
1985                 handle_reply(osdc, msg, con);
1986                 break;
1987         case CEPH_MSG_WATCH_NOTIFY:
1988                 handle_watch_notify(osdc, msg);
1989                 break;
1990
1991         default:
1992                 pr_err("received unknown message type %d %s\n", type,
1993                        ceph_msg_type_name(type));
1994         }
1995 out:
1996         ceph_msg_put(msg);
1997 }
1998
1999 /*
2000  * lookup and return message for incoming reply.  set up reply message
2001  * pages.
2002  */
2003 static struct ceph_msg *get_reply(struct ceph_connection *con,
2004                                   struct ceph_msg_header *hdr,
2005                                   int *skip)
2006 {
2007         struct ceph_osd *osd = con->private;
2008         struct ceph_osd_client *osdc = osd->o_osdc;
2009         struct ceph_msg *m;
2010         struct ceph_osd_request *req;
2011         int front = le32_to_cpu(hdr->front_len);
2012         int data_len = le32_to_cpu(hdr->data_len);
2013         u64 tid;
2014
2015         tid = le64_to_cpu(hdr->tid);
2016         mutex_lock(&osdc->request_mutex);
2017         req = __lookup_request(osdc, tid);
2018         if (!req) {
2019                 *skip = 1;
2020                 m = NULL;
2021                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
2022                         osd->o_osd);
2023                 goto out;
2024         }
2025
2026         if (req->r_con_filling_msg) {
2027                 dout("get_reply revoking msg %p from old con %p\n",
2028                      req->r_reply, req->r_con_filling_msg);
2029                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
2030                 ceph_con_put(req->r_con_filling_msg);
2031                 req->r_con_filling_msg = NULL;
2032         }
2033
2034         if (front > req->r_reply->front.iov_len) {
2035                 pr_warning("get_reply front %d > preallocated %d\n",
2036                            front, (int)req->r_reply->front.iov_len);
2037                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS);
2038                 if (!m)
2039                         goto out;
2040                 ceph_msg_put(req->r_reply);
2041                 req->r_reply = m;
2042         }
2043         m = ceph_msg_get(req->r_reply);
2044
2045         if (data_len > 0) {
2046                 int want = calc_pages_for(req->r_page_alignment, data_len);
2047
2048                 if (unlikely(req->r_num_pages < want)) {
2049                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2050                                    " had only %d pages ready\n", tid, data_len,
2051                                    want, req->r_num_pages);
2052                         *skip = 1;
2053                         ceph_msg_put(m);
2054                         m = NULL;
2055                         goto out;
2056                 }
2057                 m->pages = req->r_pages;
2058                 m->nr_pages = req->r_num_pages;
2059                 m->page_alignment = req->r_page_alignment;
2060 #ifdef CONFIG_BLOCK
2061                 m->bio = req->r_bio;
2062 #endif
2063         }
2064         *skip = 0;
2065         req->r_con_filling_msg = ceph_con_get(con);
2066         dout("get_reply tid %lld %p\n", tid, m);
2067
2068 out:
2069         mutex_unlock(&osdc->request_mutex);
2070         return m;
2071
2072 }
2073
2074 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2075                                   struct ceph_msg_header *hdr,
2076                                   int *skip)
2077 {
2078         struct ceph_osd *osd = con->private;
2079         int type = le16_to_cpu(hdr->type);
2080         int front = le32_to_cpu(hdr->front_len);
2081
2082         switch (type) {
2083         case CEPH_MSG_OSD_MAP:
2084         case CEPH_MSG_WATCH_NOTIFY:
2085                 return ceph_msg_new(type, front, GFP_NOFS);
2086         case CEPH_MSG_OSD_OPREPLY:
2087                 return get_reply(con, hdr, skip);
2088         default:
2089                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2090                         osd->o_osd);
2091                 *skip = 1;
2092                 return NULL;
2093         }
2094 }
2095
2096 /*
2097  * Wrappers to refcount containing ceph_osd struct
2098  */
2099 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2100 {
2101         struct ceph_osd *osd = con->private;
2102         if (get_osd(osd))
2103                 return con;
2104         return NULL;
2105 }
2106
2107 static void put_osd_con(struct ceph_connection *con)
2108 {
2109         struct ceph_osd *osd = con->private;
2110         put_osd(osd);
2111 }
2112
2113 /*
2114  * authentication
2115  */
2116 static int get_authorizer(struct ceph_connection *con,
2117                           void **buf, int *len, int *proto,
2118                           void **reply_buf, int *reply_len, int force_new)
2119 {
2120         struct ceph_osd *o = con->private;
2121         struct ceph_osd_client *osdc = o->o_osdc;
2122         struct ceph_auth_client *ac = osdc->client->monc.auth;
2123         int ret = 0;
2124
2125         if (force_new && o->o_authorizer) {
2126                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
2127                 o->o_authorizer = NULL;
2128         }
2129         if (o->o_authorizer == NULL) {
2130                 ret = ac->ops->create_authorizer(
2131                         ac, CEPH_ENTITY_TYPE_OSD,
2132                         &o->o_authorizer,
2133                         &o->o_authorizer_buf,
2134                         &o->o_authorizer_buf_len,
2135                         &o->o_authorizer_reply_buf,
2136                         &o->o_authorizer_reply_buf_len);
2137                 if (ret)
2138                         return ret;
2139         }
2140
2141         *proto = ac->protocol;
2142         *buf = o->o_authorizer_buf;
2143         *len = o->o_authorizer_buf_len;
2144         *reply_buf = o->o_authorizer_reply_buf;
2145         *reply_len = o->o_authorizer_reply_buf_len;
2146         return 0;
2147 }
2148
2149
2150 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2151 {
2152         struct ceph_osd *o = con->private;
2153         struct ceph_osd_client *osdc = o->o_osdc;
2154         struct ceph_auth_client *ac = osdc->client->monc.auth;
2155
2156         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
2157 }
2158
2159 static int invalidate_authorizer(struct ceph_connection *con)
2160 {
2161         struct ceph_osd *o = con->private;
2162         struct ceph_osd_client *osdc = o->o_osdc;
2163         struct ceph_auth_client *ac = osdc->client->monc.auth;
2164
2165         if (ac->ops->invalidate_authorizer)
2166                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2167
2168         return ceph_monc_validate_auth(&osdc->client->monc);
2169 }
2170
2171 static const struct ceph_connection_operations osd_con_ops = {
2172         .get = get_osd_con,
2173         .put = put_osd_con,
2174         .dispatch = dispatch,
2175         .get_authorizer = get_authorizer,
2176         .verify_authorizer_reply = verify_authorizer_reply,
2177         .invalidate_authorizer = invalidate_authorizer,
2178         .alloc_msg = alloc_msg,
2179         .fault = osd_reset,
2180 };