]> git.karo-electronics.de Git - karo-tx-linux.git/blob - net/ceph/osd_client.c
3ee4a1da02cd0a8b5812635c45f511d51bae7727
[karo-tx-linux.git] / net / ceph / osd_client.c
1
2 #include <linux/ceph/ceph_debug.h>
3
4 #include <linux/module.h>
5 #include <linux/err.h>
6 #include <linux/highmem.h>
7 #include <linux/mm.h>
8 #include <linux/pagemap.h>
9 #include <linux/slab.h>
10 #include <linux/uaccess.h>
11 #ifdef CONFIG_BLOCK
12 #include <linux/bio.h>
13 #endif
14
15 #include <linux/ceph/libceph.h>
16 #include <linux/ceph/osd_client.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/pagelist.h>
21
22 #define OSD_OP_FRONT_LEN        4096
23 #define OSD_OPREPLY_FRONT_LEN   512
24
25 static struct kmem_cache        *ceph_osd_request_cache;
26
27 static const struct ceph_connection_operations osd_con_ops;
28
29 static void __send_queued(struct ceph_osd_client *osdc);
30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
31 static void __register_request(struct ceph_osd_client *osdc,
32                                struct ceph_osd_request *req);
33 static void __unregister_linger_request(struct ceph_osd_client *osdc,
34                                         struct ceph_osd_request *req);
35 static void __send_request(struct ceph_osd_client *osdc,
36                            struct ceph_osd_request *req);
37
38 /*
39  * Implement client access to distributed object storage cluster.
40  *
41  * All data objects are stored within a cluster/cloud of OSDs, or
42  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
43  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
44  * remote daemons serving up and coordinating consistent and safe
45  * access to storage.
46  *
47  * Cluster membership and the mapping of data objects onto storage devices
48  * are described by the osd map.
49  *
50  * We keep track of pending OSD requests (read, write), resubmit
51  * requests to different OSDs when the cluster topology/data layout
52  * change, or retry the affected requests when the communications
53  * channel with an OSD is reset.
54  */
55
56 /*
57  * calculate the mapping of a file extent onto an object, and fill out the
58  * request accordingly.  shorten extent as necessary if it crosses an
59  * object boundary.
60  *
61  * fill osd op in request message.
62  */
63 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
64                         u64 *objnum, u64 *objoff, u64 *objlen)
65 {
66         u64 orig_len = *plen;
67         int r;
68
69         /* object extent? */
70         r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
71                                           objoff, objlen);
72         if (r < 0)
73                 return r;
74         if (*objlen < orig_len) {
75                 *plen = *objlen;
76                 dout(" skipping last %llu, final file extent %llu~%llu\n",
77                      orig_len - *plen, off, *plen);
78         }
79
80         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
81
82         return 0;
83 }
84
85 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
86 {
87         memset(osd_data, 0, sizeof (*osd_data));
88         osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
89 }
90
91 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
92                         struct page **pages, u64 length, u32 alignment,
93                         bool pages_from_pool, bool own_pages)
94 {
95         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
96         osd_data->pages = pages;
97         osd_data->length = length;
98         osd_data->alignment = alignment;
99         osd_data->pages_from_pool = pages_from_pool;
100         osd_data->own_pages = own_pages;
101 }
102
103 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
104                         struct ceph_pagelist *pagelist)
105 {
106         osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
107         osd_data->pagelist = pagelist;
108 }
109
110 #ifdef CONFIG_BLOCK
111 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
112                         struct bio *bio, size_t bio_length)
113 {
114         osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
115         osd_data->bio = bio;
116         osd_data->bio_length = bio_length;
117 }
118 #endif /* CONFIG_BLOCK */
119
120 #define osd_req_op_data(oreq, whch, typ, fld)   \
121         ({                                              \
122                 BUG_ON(whch >= (oreq)->r_num_ops);      \
123                 &(oreq)->r_ops[whch].typ.fld;           \
124         })
125
126 static struct ceph_osd_data *
127 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
128 {
129         BUG_ON(which >= osd_req->r_num_ops);
130
131         return &osd_req->r_ops[which].raw_data_in;
132 }
133
134 struct ceph_osd_data *
135 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
136                         unsigned int which)
137 {
138         return osd_req_op_data(osd_req, which, extent, osd_data);
139 }
140 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
141
142 struct ceph_osd_data *
143 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
144                         unsigned int which)
145 {
146         return osd_req_op_data(osd_req, which, cls, response_data);
147 }
148 EXPORT_SYMBOL(osd_req_op_cls_response_data);    /* ??? */
149
150 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
151                         unsigned int which, struct page **pages,
152                         u64 length, u32 alignment,
153                         bool pages_from_pool, bool own_pages)
154 {
155         struct ceph_osd_data *osd_data;
156
157         osd_data = osd_req_op_raw_data_in(osd_req, which);
158         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
159                                 pages_from_pool, own_pages);
160 }
161 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
162
163 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
164                         unsigned int which, struct page **pages,
165                         u64 length, u32 alignment,
166                         bool pages_from_pool, bool own_pages)
167 {
168         struct ceph_osd_data *osd_data;
169
170         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
171         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
172                                 pages_from_pool, own_pages);
173 }
174 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
175
176 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
177                         unsigned int which, struct ceph_pagelist *pagelist)
178 {
179         struct ceph_osd_data *osd_data;
180
181         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
182         ceph_osd_data_pagelist_init(osd_data, pagelist);
183 }
184 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
185
186 #ifdef CONFIG_BLOCK
187 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
188                         unsigned int which, struct bio *bio, size_t bio_length)
189 {
190         struct ceph_osd_data *osd_data;
191
192         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
193         ceph_osd_data_bio_init(osd_data, bio, bio_length);
194 }
195 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
196 #endif /* CONFIG_BLOCK */
197
198 static void osd_req_op_cls_request_info_pagelist(
199                         struct ceph_osd_request *osd_req,
200                         unsigned int which, struct ceph_pagelist *pagelist)
201 {
202         struct ceph_osd_data *osd_data;
203
204         osd_data = osd_req_op_data(osd_req, which, cls, request_info);
205         ceph_osd_data_pagelist_init(osd_data, pagelist);
206 }
207
208 void osd_req_op_cls_request_data_pagelist(
209                         struct ceph_osd_request *osd_req,
210                         unsigned int which, struct ceph_pagelist *pagelist)
211 {
212         struct ceph_osd_data *osd_data;
213
214         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
215         ceph_osd_data_pagelist_init(osd_data, pagelist);
216 }
217 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
218
219 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
220                         unsigned int which, struct page **pages, u64 length,
221                         u32 alignment, bool pages_from_pool, bool own_pages)
222 {
223         struct ceph_osd_data *osd_data;
224
225         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
226         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
227                                 pages_from_pool, own_pages);
228 }
229 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
230
231 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
232                         unsigned int which, struct page **pages, u64 length,
233                         u32 alignment, bool pages_from_pool, bool own_pages)
234 {
235         struct ceph_osd_data *osd_data;
236
237         osd_data = osd_req_op_data(osd_req, which, cls, response_data);
238         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
239                                 pages_from_pool, own_pages);
240 }
241 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
242
243 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
244 {
245         switch (osd_data->type) {
246         case CEPH_OSD_DATA_TYPE_NONE:
247                 return 0;
248         case CEPH_OSD_DATA_TYPE_PAGES:
249                 return osd_data->length;
250         case CEPH_OSD_DATA_TYPE_PAGELIST:
251                 return (u64)osd_data->pagelist->length;
252 #ifdef CONFIG_BLOCK
253         case CEPH_OSD_DATA_TYPE_BIO:
254                 return (u64)osd_data->bio_length;
255 #endif /* CONFIG_BLOCK */
256         default:
257                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
258                 return 0;
259         }
260 }
261
262 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
263 {
264         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
265                 int num_pages;
266
267                 num_pages = calc_pages_for((u64)osd_data->alignment,
268                                                 (u64)osd_data->length);
269                 ceph_release_page_vector(osd_data->pages, num_pages);
270         }
271         ceph_osd_data_init(osd_data);
272 }
273
274 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
275                         unsigned int which)
276 {
277         struct ceph_osd_req_op *op;
278
279         BUG_ON(which >= osd_req->r_num_ops);
280         op = &osd_req->r_ops[which];
281
282         switch (op->op) {
283         case CEPH_OSD_OP_READ:
284         case CEPH_OSD_OP_WRITE:
285                 ceph_osd_data_release(&op->extent.osd_data);
286                 break;
287         case CEPH_OSD_OP_CALL:
288                 ceph_osd_data_release(&op->cls.request_info);
289                 ceph_osd_data_release(&op->cls.request_data);
290                 ceph_osd_data_release(&op->cls.response_data);
291                 break;
292         default:
293                 break;
294         }
295 }
296
297 /*
298  * requests
299  */
300 void ceph_osdc_release_request(struct kref *kref)
301 {
302         struct ceph_osd_request *req;
303         unsigned int which;
304
305         req = container_of(kref, struct ceph_osd_request, r_kref);
306         if (req->r_request)
307                 ceph_msg_put(req->r_request);
308         if (req->r_reply) {
309                 ceph_msg_revoke_incoming(req->r_reply);
310                 ceph_msg_put(req->r_reply);
311         }
312
313         for (which = 0; which < req->r_num_ops; which++)
314                 osd_req_op_data_release(req, which);
315
316         ceph_put_snap_context(req->r_snapc);
317         if (req->r_mempool)
318                 mempool_free(req, req->r_osdc->req_mempool);
319         else
320                 kmem_cache_free(ceph_osd_request_cache, req);
321
322 }
323 EXPORT_SYMBOL(ceph_osdc_release_request);
324
325 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
326                                                struct ceph_snap_context *snapc,
327                                                unsigned int num_ops,
328                                                bool use_mempool,
329                                                gfp_t gfp_flags)
330 {
331         struct ceph_osd_request *req;
332         struct ceph_msg *msg;
333         size_t msg_size;
334
335         BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
336         BUG_ON(num_ops > CEPH_OSD_MAX_OP);
337
338         msg_size = 4 + 4 + 8 + 8 + 4+8;
339         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
340         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
341         msg_size += 4 + MAX_OBJ_NAME_SIZE;
342         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
343         msg_size += 8;  /* snapid */
344         msg_size += 8;  /* snap_seq */
345         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
346         msg_size += 4;
347
348         if (use_mempool) {
349                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
350                 memset(req, 0, sizeof(*req));
351         } else {
352                 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
353         }
354         if (req == NULL)
355                 return NULL;
356
357         req->r_osdc = osdc;
358         req->r_mempool = use_mempool;
359         req->r_num_ops = num_ops;
360
361         kref_init(&req->r_kref);
362         init_completion(&req->r_completion);
363         init_completion(&req->r_safe_completion);
364         RB_CLEAR_NODE(&req->r_node);
365         INIT_LIST_HEAD(&req->r_unsafe_item);
366         INIT_LIST_HEAD(&req->r_linger_item);
367         INIT_LIST_HEAD(&req->r_linger_osd);
368         INIT_LIST_HEAD(&req->r_req_lru_item);
369         INIT_LIST_HEAD(&req->r_osd_item);
370
371         /* create reply message */
372         if (use_mempool)
373                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
374         else
375                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
376                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
377         if (!msg) {
378                 ceph_osdc_put_request(req);
379                 return NULL;
380         }
381         req->r_reply = msg;
382
383         /* create request message; allow space for oid */
384         if (use_mempool)
385                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
386         else
387                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
388         if (!msg) {
389                 ceph_osdc_put_request(req);
390                 return NULL;
391         }
392
393         memset(msg->front.iov_base, 0, msg->front.iov_len);
394
395         req->r_request = msg;
396
397         return req;
398 }
399 EXPORT_SYMBOL(ceph_osdc_alloc_request);
400
401 static bool osd_req_opcode_valid(u16 opcode)
402 {
403         switch (opcode) {
404         case CEPH_OSD_OP_READ:
405         case CEPH_OSD_OP_STAT:
406         case CEPH_OSD_OP_MAPEXT:
407         case CEPH_OSD_OP_MASKTRUNC:
408         case CEPH_OSD_OP_SPARSE_READ:
409         case CEPH_OSD_OP_NOTIFY:
410         case CEPH_OSD_OP_NOTIFY_ACK:
411         case CEPH_OSD_OP_ASSERT_VER:
412         case CEPH_OSD_OP_WRITE:
413         case CEPH_OSD_OP_WRITEFULL:
414         case CEPH_OSD_OP_TRUNCATE:
415         case CEPH_OSD_OP_ZERO:
416         case CEPH_OSD_OP_DELETE:
417         case CEPH_OSD_OP_APPEND:
418         case CEPH_OSD_OP_STARTSYNC:
419         case CEPH_OSD_OP_SETTRUNC:
420         case CEPH_OSD_OP_TRIMTRUNC:
421         case CEPH_OSD_OP_TMAPUP:
422         case CEPH_OSD_OP_TMAPPUT:
423         case CEPH_OSD_OP_TMAPGET:
424         case CEPH_OSD_OP_CREATE:
425         case CEPH_OSD_OP_ROLLBACK:
426         case CEPH_OSD_OP_WATCH:
427         case CEPH_OSD_OP_OMAPGETKEYS:
428         case CEPH_OSD_OP_OMAPGETVALS:
429         case CEPH_OSD_OP_OMAPGETHEADER:
430         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
431         case CEPH_OSD_OP_OMAPSETVALS:
432         case CEPH_OSD_OP_OMAPSETHEADER:
433         case CEPH_OSD_OP_OMAPCLEAR:
434         case CEPH_OSD_OP_OMAPRMKEYS:
435         case CEPH_OSD_OP_OMAP_CMP:
436         case CEPH_OSD_OP_CLONERANGE:
437         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
438         case CEPH_OSD_OP_SRC_CMPXATTR:
439         case CEPH_OSD_OP_GETXATTR:
440         case CEPH_OSD_OP_GETXATTRS:
441         case CEPH_OSD_OP_CMPXATTR:
442         case CEPH_OSD_OP_SETXATTR:
443         case CEPH_OSD_OP_SETXATTRS:
444         case CEPH_OSD_OP_RESETXATTRS:
445         case CEPH_OSD_OP_RMXATTR:
446         case CEPH_OSD_OP_PULL:
447         case CEPH_OSD_OP_PUSH:
448         case CEPH_OSD_OP_BALANCEREADS:
449         case CEPH_OSD_OP_UNBALANCEREADS:
450         case CEPH_OSD_OP_SCRUB:
451         case CEPH_OSD_OP_SCRUB_RESERVE:
452         case CEPH_OSD_OP_SCRUB_UNRESERVE:
453         case CEPH_OSD_OP_SCRUB_STOP:
454         case CEPH_OSD_OP_SCRUB_MAP:
455         case CEPH_OSD_OP_WRLOCK:
456         case CEPH_OSD_OP_WRUNLOCK:
457         case CEPH_OSD_OP_RDLOCK:
458         case CEPH_OSD_OP_RDUNLOCK:
459         case CEPH_OSD_OP_UPLOCK:
460         case CEPH_OSD_OP_DNLOCK:
461         case CEPH_OSD_OP_CALL:
462         case CEPH_OSD_OP_PGLS:
463         case CEPH_OSD_OP_PGLS_FILTER:
464                 return true;
465         default:
466                 return false;
467         }
468 }
469
470 /*
471  * This is an osd op init function for opcodes that have no data or
472  * other information associated with them.  It also serves as a
473  * common init routine for all the other init functions, below.
474  */
475 static struct ceph_osd_req_op *
476 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
477                                 u16 opcode)
478 {
479         struct ceph_osd_req_op *op;
480
481         BUG_ON(which >= osd_req->r_num_ops);
482         BUG_ON(!osd_req_opcode_valid(opcode));
483
484         op = &osd_req->r_ops[which];
485         memset(op, 0, sizeof (*op));
486         op->op = opcode;
487
488         return op;
489 }
490
491 void osd_req_op_init(struct ceph_osd_request *osd_req,
492                                 unsigned int which, u16 opcode)
493 {
494         (void)_osd_req_op_init(osd_req, which, opcode);
495 }
496 EXPORT_SYMBOL(osd_req_op_init);
497
498 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
499                                 unsigned int which, u16 opcode,
500                                 u64 offset, u64 length,
501                                 u64 truncate_size, u32 truncate_seq)
502 {
503         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
504         size_t payload_len = 0;
505
506         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
507                opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
508                opcode != CEPH_OSD_OP_TRUNCATE);
509
510         op->extent.offset = offset;
511         op->extent.length = length;
512         op->extent.truncate_size = truncate_size;
513         op->extent.truncate_seq = truncate_seq;
514         if (opcode == CEPH_OSD_OP_WRITE)
515                 payload_len += length;
516
517         op->payload_len = payload_len;
518 }
519 EXPORT_SYMBOL(osd_req_op_extent_init);
520
521 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
522                                 unsigned int which, u64 length)
523 {
524         struct ceph_osd_req_op *op;
525         u64 previous;
526
527         BUG_ON(which >= osd_req->r_num_ops);
528         op = &osd_req->r_ops[which];
529         previous = op->extent.length;
530
531         if (length == previous)
532                 return;         /* Nothing to do */
533         BUG_ON(length > previous);
534
535         op->extent.length = length;
536         op->payload_len -= previous - length;
537 }
538 EXPORT_SYMBOL(osd_req_op_extent_update);
539
540 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
541                         u16 opcode, const char *class, const char *method)
542 {
543         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
544         struct ceph_pagelist *pagelist;
545         size_t payload_len = 0;
546         size_t size;
547
548         BUG_ON(opcode != CEPH_OSD_OP_CALL);
549
550         pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
551         BUG_ON(!pagelist);
552         ceph_pagelist_init(pagelist);
553
554         op->cls.class_name = class;
555         size = strlen(class);
556         BUG_ON(size > (size_t) U8_MAX);
557         op->cls.class_len = size;
558         ceph_pagelist_append(pagelist, class, size);
559         payload_len += size;
560
561         op->cls.method_name = method;
562         size = strlen(method);
563         BUG_ON(size > (size_t) U8_MAX);
564         op->cls.method_len = size;
565         ceph_pagelist_append(pagelist, method, size);
566         payload_len += size;
567
568         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
569
570         op->cls.argc = 0;       /* currently unused */
571
572         op->payload_len = payload_len;
573 }
574 EXPORT_SYMBOL(osd_req_op_cls_init);
575
576 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
577                                 unsigned int which, u16 opcode,
578                                 u64 cookie, u64 version, int flag)
579 {
580         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
581
582         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
583
584         op->watch.cookie = cookie;
585         op->watch.ver = version;
586         if (opcode == CEPH_OSD_OP_WATCH && flag)
587                 op->watch.flag = (u8)1;
588 }
589 EXPORT_SYMBOL(osd_req_op_watch_init);
590
591 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
592                                 struct ceph_osd_data *osd_data)
593 {
594         u64 length = ceph_osd_data_length(osd_data);
595
596         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
597                 BUG_ON(length > (u64) SIZE_MAX);
598                 if (length)
599                         ceph_msg_data_add_pages(msg, osd_data->pages,
600                                         length, osd_data->alignment);
601         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
602                 BUG_ON(!length);
603                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
604 #ifdef CONFIG_BLOCK
605         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
606                 ceph_msg_data_add_bio(msg, osd_data->bio, length);
607 #endif
608         } else {
609                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
610         }
611 }
612
613 static u64 osd_req_encode_op(struct ceph_osd_request *req,
614                               struct ceph_osd_op *dst, unsigned int which)
615 {
616         struct ceph_osd_req_op *src;
617         struct ceph_osd_data *osd_data;
618         u64 request_data_len = 0;
619         u64 data_length;
620
621         BUG_ON(which >= req->r_num_ops);
622         src = &req->r_ops[which];
623         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
624                 pr_err("unrecognized osd opcode %d\n", src->op);
625
626                 return 0;
627         }
628
629         switch (src->op) {
630         case CEPH_OSD_OP_STAT:
631                 osd_data = &src->raw_data_in;
632                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
633                 break;
634         case CEPH_OSD_OP_READ:
635         case CEPH_OSD_OP_WRITE:
636         case CEPH_OSD_OP_ZERO:
637         case CEPH_OSD_OP_DELETE:
638         case CEPH_OSD_OP_TRUNCATE:
639                 if (src->op == CEPH_OSD_OP_WRITE)
640                         request_data_len = src->extent.length;
641                 dst->extent.offset = cpu_to_le64(src->extent.offset);
642                 dst->extent.length = cpu_to_le64(src->extent.length);
643                 dst->extent.truncate_size =
644                         cpu_to_le64(src->extent.truncate_size);
645                 dst->extent.truncate_seq =
646                         cpu_to_le32(src->extent.truncate_seq);
647                 osd_data = &src->extent.osd_data;
648                 if (src->op == CEPH_OSD_OP_WRITE)
649                         ceph_osdc_msg_data_add(req->r_request, osd_data);
650                 else
651                         ceph_osdc_msg_data_add(req->r_reply, osd_data);
652                 break;
653         case CEPH_OSD_OP_CALL:
654                 dst->cls.class_len = src->cls.class_len;
655                 dst->cls.method_len = src->cls.method_len;
656                 osd_data = &src->cls.request_info;
657                 ceph_osdc_msg_data_add(req->r_request, osd_data);
658                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
659                 request_data_len = osd_data->pagelist->length;
660
661                 osd_data = &src->cls.request_data;
662                 data_length = ceph_osd_data_length(osd_data);
663                 if (data_length) {
664                         BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
665                         dst->cls.indata_len = cpu_to_le32(data_length);
666                         ceph_osdc_msg_data_add(req->r_request, osd_data);
667                         src->payload_len += data_length;
668                         request_data_len += data_length;
669                 }
670                 osd_data = &src->cls.response_data;
671                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
672                 break;
673         case CEPH_OSD_OP_STARTSYNC:
674                 break;
675         case CEPH_OSD_OP_NOTIFY_ACK:
676         case CEPH_OSD_OP_WATCH:
677                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
678                 dst->watch.ver = cpu_to_le64(src->watch.ver);
679                 dst->watch.flag = src->watch.flag;
680                 break;
681         default:
682                 pr_err("unsupported osd opcode %s\n",
683                         ceph_osd_op_name(src->op));
684                 WARN_ON(1);
685
686                 return 0;
687         }
688         dst->op = cpu_to_le16(src->op);
689         dst->payload_len = cpu_to_le32(src->payload_len);
690
691         return request_data_len;
692 }
693
694 /*
695  * build new request AND message, calculate layout, and adjust file
696  * extent as needed.
697  *
698  * if the file was recently truncated, we include information about its
699  * old and new size so that the object can be updated appropriately.  (we
700  * avoid synchronously deleting truncated objects because it's slow.)
701  *
702  * if @do_sync, include a 'startsync' command so that the osd will flush
703  * data quickly.
704  */
705 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
706                                                struct ceph_file_layout *layout,
707                                                struct ceph_vino vino,
708                                                u64 off, u64 *plen, int num_ops,
709                                                int opcode, int flags,
710                                                struct ceph_snap_context *snapc,
711                                                u32 truncate_seq,
712                                                u64 truncate_size,
713                                                bool use_mempool)
714 {
715         struct ceph_osd_request *req;
716         u64 objnum = 0;
717         u64 objoff = 0;
718         u64 objlen = 0;
719         u32 object_size;
720         u64 object_base;
721         int r;
722
723         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
724                opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
725                opcode != CEPH_OSD_OP_TRUNCATE);
726
727         req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
728                                         GFP_NOFS);
729         if (!req)
730                 return ERR_PTR(-ENOMEM);
731
732         req->r_flags = flags;
733
734         /* calculate max write size */
735         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
736         if (r < 0) {
737                 ceph_osdc_put_request(req);
738                 return ERR_PTR(r);
739         }
740
741         object_size = le32_to_cpu(layout->fl_object_size);
742         object_base = off - objoff;
743         if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
744                 if (truncate_size <= object_base) {
745                         truncate_size = 0;
746                 } else {
747                         truncate_size -= object_base;
748                         if (truncate_size > object_size)
749                                 truncate_size = object_size;
750                 }
751         }
752
753         osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
754                                 truncate_size, truncate_seq);
755
756         /*
757          * A second op in the ops array means the caller wants to
758          * also issue a include a 'startsync' command so that the
759          * osd will flush data quickly.
760          */
761         if (num_ops > 1)
762                 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
763
764         req->r_file_layout = *layout;  /* keep a copy */
765
766         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx",
767                 vino.ino, objnum);
768         req->r_oid_len = strlen(req->r_oid);
769
770         return req;
771 }
772 EXPORT_SYMBOL(ceph_osdc_new_request);
773
774 /*
775  * We keep osd requests in an rbtree, sorted by ->r_tid.
776  */
777 static void __insert_request(struct ceph_osd_client *osdc,
778                              struct ceph_osd_request *new)
779 {
780         struct rb_node **p = &osdc->requests.rb_node;
781         struct rb_node *parent = NULL;
782         struct ceph_osd_request *req = NULL;
783
784         while (*p) {
785                 parent = *p;
786                 req = rb_entry(parent, struct ceph_osd_request, r_node);
787                 if (new->r_tid < req->r_tid)
788                         p = &(*p)->rb_left;
789                 else if (new->r_tid > req->r_tid)
790                         p = &(*p)->rb_right;
791                 else
792                         BUG();
793         }
794
795         rb_link_node(&new->r_node, parent, p);
796         rb_insert_color(&new->r_node, &osdc->requests);
797 }
798
799 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
800                                                  u64 tid)
801 {
802         struct ceph_osd_request *req;
803         struct rb_node *n = osdc->requests.rb_node;
804
805         while (n) {
806                 req = rb_entry(n, struct ceph_osd_request, r_node);
807                 if (tid < req->r_tid)
808                         n = n->rb_left;
809                 else if (tid > req->r_tid)
810                         n = n->rb_right;
811                 else
812                         return req;
813         }
814         return NULL;
815 }
816
817 static struct ceph_osd_request *
818 __lookup_request_ge(struct ceph_osd_client *osdc,
819                     u64 tid)
820 {
821         struct ceph_osd_request *req;
822         struct rb_node *n = osdc->requests.rb_node;
823
824         while (n) {
825                 req = rb_entry(n, struct ceph_osd_request, r_node);
826                 if (tid < req->r_tid) {
827                         if (!n->rb_left)
828                                 return req;
829                         n = n->rb_left;
830                 } else if (tid > req->r_tid) {
831                         n = n->rb_right;
832                 } else {
833                         return req;
834                 }
835         }
836         return NULL;
837 }
838
839 /*
840  * Resubmit requests pending on the given osd.
841  */
842 static void __kick_osd_requests(struct ceph_osd_client *osdc,
843                                 struct ceph_osd *osd)
844 {
845         struct ceph_osd_request *req, *nreq;
846         LIST_HEAD(resend);
847         int err;
848
849         dout("__kick_osd_requests osd%d\n", osd->o_osd);
850         err = __reset_osd(osdc, osd);
851         if (err)
852                 return;
853         /*
854          * Build up a list of requests to resend by traversing the
855          * osd's list of requests.  Requests for a given object are
856          * sent in tid order, and that is also the order they're
857          * kept on this list.  Therefore all requests that are in
858          * flight will be found first, followed by all requests that
859          * have not yet been sent.  And to resend requests while
860          * preserving this order we will want to put any sent
861          * requests back on the front of the osd client's unsent
862          * list.
863          *
864          * So we build a separate ordered list of already-sent
865          * requests for the affected osd and splice it onto the
866          * front of the osd client's unsent list.  Once we've seen a
867          * request that has not yet been sent we're done.  Those
868          * requests are already sitting right where they belong.
869          */
870         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
871                 if (!req->r_sent)
872                         break;
873                 list_move_tail(&req->r_req_lru_item, &resend);
874                 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
875                      osd->o_osd);
876                 if (!req->r_linger)
877                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
878         }
879         list_splice(&resend, &osdc->req_unsent);
880
881         /*
882          * Linger requests are re-registered before sending, which
883          * sets up a new tid for each.  We add them to the unsent
884          * list at the end to keep things in tid order.
885          */
886         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
887                                  r_linger_osd) {
888                 /*
889                  * reregister request prior to unregistering linger so
890                  * that r_osd is preserved.
891                  */
892                 BUG_ON(!list_empty(&req->r_req_lru_item));
893                 __register_request(osdc, req);
894                 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
895                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
896                 __unregister_linger_request(osdc, req);
897                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
898                      osd->o_osd);
899         }
900 }
901
902 /*
903  * If the osd connection drops, we need to resubmit all requests.
904  */
905 static void osd_reset(struct ceph_connection *con)
906 {
907         struct ceph_osd *osd = con->private;
908         struct ceph_osd_client *osdc;
909
910         if (!osd)
911                 return;
912         dout("osd_reset osd%d\n", osd->o_osd);
913         osdc = osd->o_osdc;
914         down_read(&osdc->map_sem);
915         mutex_lock(&osdc->request_mutex);
916         __kick_osd_requests(osdc, osd);
917         __send_queued(osdc);
918         mutex_unlock(&osdc->request_mutex);
919         up_read(&osdc->map_sem);
920 }
921
922 /*
923  * Track open sessions with osds.
924  */
925 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
926 {
927         struct ceph_osd *osd;
928
929         osd = kzalloc(sizeof(*osd), GFP_NOFS);
930         if (!osd)
931                 return NULL;
932
933         atomic_set(&osd->o_ref, 1);
934         osd->o_osdc = osdc;
935         osd->o_osd = onum;
936         RB_CLEAR_NODE(&osd->o_node);
937         INIT_LIST_HEAD(&osd->o_requests);
938         INIT_LIST_HEAD(&osd->o_linger_requests);
939         INIT_LIST_HEAD(&osd->o_osd_lru);
940         osd->o_incarnation = 1;
941
942         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
943
944         INIT_LIST_HEAD(&osd->o_keepalive_item);
945         return osd;
946 }
947
948 static struct ceph_osd *get_osd(struct ceph_osd *osd)
949 {
950         if (atomic_inc_not_zero(&osd->o_ref)) {
951                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
952                      atomic_read(&osd->o_ref));
953                 return osd;
954         } else {
955                 dout("get_osd %p FAIL\n", osd);
956                 return NULL;
957         }
958 }
959
960 static void put_osd(struct ceph_osd *osd)
961 {
962         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
963              atomic_read(&osd->o_ref) - 1);
964         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
965                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
966
967                 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
968                 kfree(osd);
969         }
970 }
971
972 /*
973  * remove an osd from our map
974  */
975 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
976 {
977         dout("__remove_osd %p\n", osd);
978         BUG_ON(!list_empty(&osd->o_requests));
979         rb_erase(&osd->o_node, &osdc->osds);
980         list_del_init(&osd->o_osd_lru);
981         ceph_con_close(&osd->o_con);
982         put_osd(osd);
983 }
984
985 static void remove_all_osds(struct ceph_osd_client *osdc)
986 {
987         dout("%s %p\n", __func__, osdc);
988         mutex_lock(&osdc->request_mutex);
989         while (!RB_EMPTY_ROOT(&osdc->osds)) {
990                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
991                                                 struct ceph_osd, o_node);
992                 __remove_osd(osdc, osd);
993         }
994         mutex_unlock(&osdc->request_mutex);
995 }
996
997 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
998                               struct ceph_osd *osd)
999 {
1000         dout("__move_osd_to_lru %p\n", osd);
1001         BUG_ON(!list_empty(&osd->o_osd_lru));
1002         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1003         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1004 }
1005
1006 static void __remove_osd_from_lru(struct ceph_osd *osd)
1007 {
1008         dout("__remove_osd_from_lru %p\n", osd);
1009         if (!list_empty(&osd->o_osd_lru))
1010                 list_del_init(&osd->o_osd_lru);
1011 }
1012
1013 static void remove_old_osds(struct ceph_osd_client *osdc)
1014 {
1015         struct ceph_osd *osd, *nosd;
1016
1017         dout("__remove_old_osds %p\n", osdc);
1018         mutex_lock(&osdc->request_mutex);
1019         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1020                 if (time_before(jiffies, osd->lru_ttl))
1021                         break;
1022                 __remove_osd(osdc, osd);
1023         }
1024         mutex_unlock(&osdc->request_mutex);
1025 }
1026
1027 /*
1028  * reset osd connect
1029  */
1030 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1031 {
1032         struct ceph_entity_addr *peer_addr;
1033
1034         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1035         if (list_empty(&osd->o_requests) &&
1036             list_empty(&osd->o_linger_requests)) {
1037                 __remove_osd(osdc, osd);
1038
1039                 return -ENODEV;
1040         }
1041
1042         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1043         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1044                         !ceph_con_opened(&osd->o_con)) {
1045                 struct ceph_osd_request *req;
1046
1047                 dout(" osd addr hasn't changed and connection never opened,"
1048                      " letting msgr retry");
1049                 /* touch each r_stamp for handle_timeout()'s benfit */
1050                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1051                         req->r_stamp = jiffies;
1052
1053                 return -EAGAIN;
1054         }
1055
1056         ceph_con_close(&osd->o_con);
1057         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1058         osd->o_incarnation++;
1059
1060         return 0;
1061 }
1062
1063 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1064 {
1065         struct rb_node **p = &osdc->osds.rb_node;
1066         struct rb_node *parent = NULL;
1067         struct ceph_osd *osd = NULL;
1068
1069         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1070         while (*p) {
1071                 parent = *p;
1072                 osd = rb_entry(parent, struct ceph_osd, o_node);
1073                 if (new->o_osd < osd->o_osd)
1074                         p = &(*p)->rb_left;
1075                 else if (new->o_osd > osd->o_osd)
1076                         p = &(*p)->rb_right;
1077                 else
1078                         BUG();
1079         }
1080
1081         rb_link_node(&new->o_node, parent, p);
1082         rb_insert_color(&new->o_node, &osdc->osds);
1083 }
1084
1085 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1086 {
1087         struct ceph_osd *osd;
1088         struct rb_node *n = osdc->osds.rb_node;
1089
1090         while (n) {
1091                 osd = rb_entry(n, struct ceph_osd, o_node);
1092                 if (o < osd->o_osd)
1093                         n = n->rb_left;
1094                 else if (o > osd->o_osd)
1095                         n = n->rb_right;
1096                 else
1097                         return osd;
1098         }
1099         return NULL;
1100 }
1101
1102 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1103 {
1104         schedule_delayed_work(&osdc->timeout_work,
1105                         osdc->client->options->osd_keepalive_timeout * HZ);
1106 }
1107
1108 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1109 {
1110         cancel_delayed_work(&osdc->timeout_work);
1111 }
1112
1113 /*
1114  * Register request, assign tid.  If this is the first request, set up
1115  * the timeout event.
1116  */
1117 static void __register_request(struct ceph_osd_client *osdc,
1118                                struct ceph_osd_request *req)
1119 {
1120         req->r_tid = ++osdc->last_tid;
1121         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1122         dout("__register_request %p tid %lld\n", req, req->r_tid);
1123         __insert_request(osdc, req);
1124         ceph_osdc_get_request(req);
1125         osdc->num_requests++;
1126         if (osdc->num_requests == 1) {
1127                 dout(" first request, scheduling timeout\n");
1128                 __schedule_osd_timeout(osdc);
1129         }
1130 }
1131
1132 /*
1133  * called under osdc->request_mutex
1134  */
1135 static void __unregister_request(struct ceph_osd_client *osdc,
1136                                  struct ceph_osd_request *req)
1137 {
1138         if (RB_EMPTY_NODE(&req->r_node)) {
1139                 dout("__unregister_request %p tid %lld not registered\n",
1140                         req, req->r_tid);
1141                 return;
1142         }
1143
1144         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1145         rb_erase(&req->r_node, &osdc->requests);
1146         osdc->num_requests--;
1147
1148         if (req->r_osd) {
1149                 /* make sure the original request isn't in flight. */
1150                 ceph_msg_revoke(req->r_request);
1151
1152                 list_del_init(&req->r_osd_item);
1153                 if (list_empty(&req->r_osd->o_requests) &&
1154                     list_empty(&req->r_osd->o_linger_requests)) {
1155                         dout("moving osd to %p lru\n", req->r_osd);
1156                         __move_osd_to_lru(osdc, req->r_osd);
1157                 }
1158                 if (list_empty(&req->r_linger_item))
1159                         req->r_osd = NULL;
1160         }
1161
1162         list_del_init(&req->r_req_lru_item);
1163         ceph_osdc_put_request(req);
1164
1165         if (osdc->num_requests == 0) {
1166                 dout(" no requests, canceling timeout\n");
1167                 __cancel_osd_timeout(osdc);
1168         }
1169 }
1170
1171 /*
1172  * Cancel a previously queued request message
1173  */
1174 static void __cancel_request(struct ceph_osd_request *req)
1175 {
1176         if (req->r_sent && req->r_osd) {
1177                 ceph_msg_revoke(req->r_request);
1178                 req->r_sent = 0;
1179         }
1180 }
1181
1182 static void __register_linger_request(struct ceph_osd_client *osdc,
1183                                     struct ceph_osd_request *req)
1184 {
1185         dout("__register_linger_request %p\n", req);
1186         ceph_osdc_get_request(req);
1187         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1188         if (req->r_osd)
1189                 list_add_tail(&req->r_linger_osd,
1190                               &req->r_osd->o_linger_requests);
1191 }
1192
1193 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1194                                         struct ceph_osd_request *req)
1195 {
1196         dout("__unregister_linger_request %p\n", req);
1197         list_del_init(&req->r_linger_item);
1198         if (req->r_osd) {
1199                 list_del_init(&req->r_linger_osd);
1200
1201                 if (list_empty(&req->r_osd->o_requests) &&
1202                     list_empty(&req->r_osd->o_linger_requests)) {
1203                         dout("moving osd to %p lru\n", req->r_osd);
1204                         __move_osd_to_lru(osdc, req->r_osd);
1205                 }
1206                 if (list_empty(&req->r_osd_item))
1207                         req->r_osd = NULL;
1208         }
1209         ceph_osdc_put_request(req);
1210 }
1211
1212 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1213                                          struct ceph_osd_request *req)
1214 {
1215         mutex_lock(&osdc->request_mutex);
1216         if (req->r_linger) {
1217                 req->r_linger = 0;
1218                 __unregister_linger_request(osdc, req);
1219         }
1220         mutex_unlock(&osdc->request_mutex);
1221 }
1222 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1223
1224 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1225                                   struct ceph_osd_request *req)
1226 {
1227         if (!req->r_linger) {
1228                 dout("set_request_linger %p\n", req);
1229                 req->r_linger = 1;
1230         }
1231 }
1232 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1233
1234 /*
1235  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1236  * (as needed), and set the request r_osd appropriately.  If there is
1237  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1238  * (unsent, homeless) or leave on in-flight lru.
1239  *
1240  * Return 0 if unchanged, 1 if changed, or negative on error.
1241  *
1242  * Caller should hold map_sem for read and request_mutex.
1243  */
1244 static int __map_request(struct ceph_osd_client *osdc,
1245                          struct ceph_osd_request *req, int force_resend)
1246 {
1247         struct ceph_pg pgid;
1248         int acting[CEPH_PG_MAX_SIZE];
1249         int o = -1, num = 0;
1250         int err;
1251
1252         dout("map_request %p tid %lld\n", req, req->r_tid);
1253         err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap,
1254                                 ceph_file_layout_pg_pool(req->r_file_layout));
1255         if (err) {
1256                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1257                 return err;
1258         }
1259         req->r_pgid = pgid;
1260
1261         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
1262         if (err > 0) {
1263                 o = acting[0];
1264                 num = err;
1265         }
1266
1267         if ((!force_resend &&
1268              req->r_osd && req->r_osd->o_osd == o &&
1269              req->r_sent >= req->r_osd->o_incarnation &&
1270              req->r_num_pg_osds == num &&
1271              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1272             (req->r_osd == NULL && o == -1))
1273                 return 0;  /* no change */
1274
1275         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1276              req->r_tid, pgid.pool, pgid.seed, o,
1277              req->r_osd ? req->r_osd->o_osd : -1);
1278
1279         /* record full pg acting set */
1280         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1281         req->r_num_pg_osds = num;
1282
1283         if (req->r_osd) {
1284                 __cancel_request(req);
1285                 list_del_init(&req->r_osd_item);
1286                 req->r_osd = NULL;
1287         }
1288
1289         req->r_osd = __lookup_osd(osdc, o);
1290         if (!req->r_osd && o >= 0) {
1291                 err = -ENOMEM;
1292                 req->r_osd = create_osd(osdc, o);
1293                 if (!req->r_osd) {
1294                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1295                         goto out;
1296                 }
1297
1298                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1299                 __insert_osd(osdc, req->r_osd);
1300
1301                 ceph_con_open(&req->r_osd->o_con,
1302                               CEPH_ENTITY_TYPE_OSD, o,
1303                               &osdc->osdmap->osd_addr[o]);
1304         }
1305
1306         if (req->r_osd) {
1307                 __remove_osd_from_lru(req->r_osd);
1308                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1309                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1310         } else {
1311                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1312         }
1313         err = 1;   /* osd or pg changed */
1314
1315 out:
1316         return err;
1317 }
1318
1319 /*
1320  * caller should hold map_sem (for read) and request_mutex
1321  */
1322 static void __send_request(struct ceph_osd_client *osdc,
1323                            struct ceph_osd_request *req)
1324 {
1325         void *p;
1326
1327         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1328              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1329              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1330
1331         /* fill in message content that changes each time we send it */
1332         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1333         put_unaligned_le32(req->r_flags, req->r_request_flags);
1334         put_unaligned_le64(req->r_pgid.pool, req->r_request_pool);
1335         p = req->r_request_pgid;
1336         ceph_encode_64(&p, req->r_pgid.pool);
1337         ceph_encode_32(&p, req->r_pgid.seed);
1338         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1339         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1340                sizeof(req->r_reassert_version));
1341
1342         req->r_stamp = jiffies;
1343         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1344
1345         ceph_msg_get(req->r_request); /* send consumes a ref */
1346
1347         req->r_sent = req->r_osd->o_incarnation;
1348
1349         ceph_con_send(&req->r_osd->o_con, req->r_request);
1350 }
1351
1352 /*
1353  * Send any requests in the queue (req_unsent).
1354  */
1355 static void __send_queued(struct ceph_osd_client *osdc)
1356 {
1357         struct ceph_osd_request *req, *tmp;
1358
1359         dout("__send_queued\n");
1360         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1361                 __send_request(osdc, req);
1362 }
1363
1364 /*
1365  * Timeout callback, called every N seconds when 1 or more osd
1366  * requests has been active for more than N seconds.  When this
1367  * happens, we ping all OSDs with requests who have timed out to
1368  * ensure any communications channel reset is detected.  Reset the
1369  * request timeouts another N seconds in the future as we go.
1370  * Reschedule the timeout event another N seconds in future (unless
1371  * there are no open requests).
1372  */
1373 static void handle_timeout(struct work_struct *work)
1374 {
1375         struct ceph_osd_client *osdc =
1376                 container_of(work, struct ceph_osd_client, timeout_work.work);
1377         struct ceph_osd_request *req;
1378         struct ceph_osd *osd;
1379         unsigned long keepalive =
1380                 osdc->client->options->osd_keepalive_timeout * HZ;
1381         struct list_head slow_osds;
1382         dout("timeout\n");
1383         down_read(&osdc->map_sem);
1384
1385         ceph_monc_request_next_osdmap(&osdc->client->monc);
1386
1387         mutex_lock(&osdc->request_mutex);
1388
1389         /*
1390          * ping osds that are a bit slow.  this ensures that if there
1391          * is a break in the TCP connection we will notice, and reopen
1392          * a connection with that osd (from the fault callback).
1393          */
1394         INIT_LIST_HEAD(&slow_osds);
1395         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1396                 if (time_before(jiffies, req->r_stamp + keepalive))
1397                         break;
1398
1399                 osd = req->r_osd;
1400                 BUG_ON(!osd);
1401                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1402                      req->r_tid, osd->o_osd);
1403                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1404         }
1405         while (!list_empty(&slow_osds)) {
1406                 osd = list_entry(slow_osds.next, struct ceph_osd,
1407                                  o_keepalive_item);
1408                 list_del_init(&osd->o_keepalive_item);
1409                 ceph_con_keepalive(&osd->o_con);
1410         }
1411
1412         __schedule_osd_timeout(osdc);
1413         __send_queued(osdc);
1414         mutex_unlock(&osdc->request_mutex);
1415         up_read(&osdc->map_sem);
1416 }
1417
1418 static void handle_osds_timeout(struct work_struct *work)
1419 {
1420         struct ceph_osd_client *osdc =
1421                 container_of(work, struct ceph_osd_client,
1422                              osds_timeout_work.work);
1423         unsigned long delay =
1424                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1425
1426         dout("osds timeout\n");
1427         down_read(&osdc->map_sem);
1428         remove_old_osds(osdc);
1429         up_read(&osdc->map_sem);
1430
1431         schedule_delayed_work(&osdc->osds_timeout_work,
1432                               round_jiffies_relative(delay));
1433 }
1434
1435 static void complete_request(struct ceph_osd_request *req)
1436 {
1437         complete_all(&req->r_safe_completion);  /* fsync waiter */
1438 }
1439
1440 /*
1441  * handle osd op reply.  either call the callback if it is specified,
1442  * or do the completion to wake up the waiting thread.
1443  */
1444 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1445                          struct ceph_connection *con)
1446 {
1447         void *p, *end;
1448         struct ceph_osd_request *req;
1449         u64 tid;
1450         int object_len;
1451         unsigned int numops;
1452         int payload_len, flags;
1453         s32 result;
1454         s32 retry_attempt;
1455         struct ceph_pg pg;
1456         int err;
1457         u32 reassert_epoch;
1458         u64 reassert_version;
1459         u32 osdmap_epoch;
1460         int already_completed;
1461         u32 bytes;
1462         unsigned int i;
1463
1464         tid = le64_to_cpu(msg->hdr.tid);
1465         dout("handle_reply %p tid %llu\n", msg, tid);
1466
1467         p = msg->front.iov_base;
1468         end = p + msg->front.iov_len;
1469
1470         ceph_decode_need(&p, end, 4, bad);
1471         object_len = ceph_decode_32(&p);
1472         ceph_decode_need(&p, end, object_len, bad);
1473         p += object_len;
1474
1475         err = ceph_decode_pgid(&p, end, &pg);
1476         if (err)
1477                 goto bad;
1478
1479         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1480         flags = ceph_decode_64(&p);
1481         result = ceph_decode_32(&p);
1482         reassert_epoch = ceph_decode_32(&p);
1483         reassert_version = ceph_decode_64(&p);
1484         osdmap_epoch = ceph_decode_32(&p);
1485
1486         /* lookup */
1487         mutex_lock(&osdc->request_mutex);
1488         req = __lookup_request(osdc, tid);
1489         if (req == NULL) {
1490                 dout("handle_reply tid %llu dne\n", tid);
1491                 goto bad_mutex;
1492         }
1493         ceph_osdc_get_request(req);
1494
1495         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1496              req, result);
1497
1498         ceph_decode_need(&p, end, 4, bad_put);
1499         numops = ceph_decode_32(&p);
1500         if (numops > CEPH_OSD_MAX_OP)
1501                 goto bad_put;
1502         if (numops != req->r_num_ops)
1503                 goto bad_put;
1504         payload_len = 0;
1505         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1506         for (i = 0; i < numops; i++) {
1507                 struct ceph_osd_op *op = p;
1508                 int len;
1509
1510                 len = le32_to_cpu(op->payload_len);
1511                 req->r_reply_op_len[i] = len;
1512                 dout(" op %d has %d bytes\n", i, len);
1513                 payload_len += len;
1514                 p += sizeof(*op);
1515         }
1516         bytes = le32_to_cpu(msg->hdr.data_len);
1517         if (payload_len != bytes) {
1518                 pr_warning("sum of op payload lens %d != data_len %d",
1519                            payload_len, bytes);
1520                 goto bad_put;
1521         }
1522
1523         ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1524         retry_attempt = ceph_decode_32(&p);
1525         for (i = 0; i < numops; i++)
1526                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1527
1528         already_completed = req->r_got_reply;
1529
1530         if (!req->r_got_reply) {
1531
1532                 req->r_result = result;
1533                 dout("handle_reply result %d bytes %d\n", req->r_result,
1534                      bytes);
1535                 if (req->r_result == 0)
1536                         req->r_result = bytes;
1537
1538                 /* in case this is a write and we need to replay, */
1539                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1540                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1541
1542                 req->r_got_reply = 1;
1543         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1544                 dout("handle_reply tid %llu dup ack\n", tid);
1545                 mutex_unlock(&osdc->request_mutex);
1546                 goto done;
1547         }
1548
1549         dout("handle_reply tid %llu flags %d\n", tid, flags);
1550
1551         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1552                 __register_linger_request(osdc, req);
1553
1554         /* either this is a read, or we got the safe response */
1555         if (result < 0 ||
1556             (flags & CEPH_OSD_FLAG_ONDISK) ||
1557             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1558                 __unregister_request(osdc, req);
1559
1560         mutex_unlock(&osdc->request_mutex);
1561
1562         if (!already_completed) {
1563                 if (req->r_unsafe_callback &&
1564                     result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1565                         req->r_unsafe_callback(req, true);
1566                 if (req->r_callback)
1567                         req->r_callback(req, msg);
1568                 else
1569                         complete_all(&req->r_completion);
1570         }
1571
1572         if (flags & CEPH_OSD_FLAG_ONDISK) {
1573                 if (req->r_unsafe_callback && already_completed)
1574                         req->r_unsafe_callback(req, false);
1575                 complete_request(req);
1576         }
1577
1578 done:
1579         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1580         ceph_osdc_put_request(req);
1581         return;
1582
1583 bad_put:
1584         ceph_osdc_put_request(req);
1585 bad_mutex:
1586         mutex_unlock(&osdc->request_mutex);
1587 bad:
1588         pr_err("corrupt osd_op_reply got %d %d\n",
1589                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1590         ceph_msg_dump(msg);
1591 }
1592
1593 static void reset_changed_osds(struct ceph_osd_client *osdc)
1594 {
1595         struct rb_node *p, *n;
1596
1597         for (p = rb_first(&osdc->osds); p; p = n) {
1598                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1599
1600                 n = rb_next(p);
1601                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1602                     memcmp(&osd->o_con.peer_addr,
1603                            ceph_osd_addr(osdc->osdmap,
1604                                          osd->o_osd),
1605                            sizeof(struct ceph_entity_addr)) != 0)
1606                         __reset_osd(osdc, osd);
1607         }
1608 }
1609
1610 /*
1611  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1612  * no osd, request a new map.
1613  *
1614  * Caller should hold map_sem for read.
1615  */
1616 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1617 {
1618         struct ceph_osd_request *req, *nreq;
1619         struct rb_node *p;
1620         int needmap = 0;
1621         int err;
1622
1623         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1624         mutex_lock(&osdc->request_mutex);
1625         for (p = rb_first(&osdc->requests); p; ) {
1626                 req = rb_entry(p, struct ceph_osd_request, r_node);
1627                 p = rb_next(p);
1628
1629                 /*
1630                  * For linger requests that have not yet been
1631                  * registered, move them to the linger list; they'll
1632                  * be sent to the osd in the loop below.  Unregister
1633                  * the request before re-registering it as a linger
1634                  * request to ensure the __map_request() below
1635                  * will decide it needs to be sent.
1636                  */
1637                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1638                         dout("%p tid %llu restart on osd%d\n",
1639                              req, req->r_tid,
1640                              req->r_osd ? req->r_osd->o_osd : -1);
1641                         ceph_osdc_get_request(req);
1642                         __unregister_request(osdc, req);
1643                         __register_linger_request(osdc, req);
1644                         ceph_osdc_put_request(req);
1645                         continue;
1646                 }
1647
1648                 err = __map_request(osdc, req, force_resend);
1649                 if (err < 0)
1650                         continue;  /* error */
1651                 if (req->r_osd == NULL) {
1652                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1653                         needmap++;  /* request a newer map */
1654                 } else if (err > 0) {
1655                         if (!req->r_linger) {
1656                                 dout("%p tid %llu requeued on osd%d\n", req,
1657                                      req->r_tid,
1658                                      req->r_osd ? req->r_osd->o_osd : -1);
1659                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1660                         }
1661                 }
1662         }
1663
1664         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1665                                  r_linger_item) {
1666                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1667
1668                 err = __map_request(osdc, req, force_resend);
1669                 dout("__map_request returned %d\n", err);
1670                 if (err == 0)
1671                         continue;  /* no change and no osd was specified */
1672                 if (err < 0)
1673                         continue;  /* hrm! */
1674                 if (req->r_osd == NULL) {
1675                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1676                         needmap++;  /* request a newer map */
1677                         continue;
1678                 }
1679
1680                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1681                      req->r_osd ? req->r_osd->o_osd : -1);
1682                 __register_request(osdc, req);
1683                 __unregister_linger_request(osdc, req);
1684         }
1685         reset_changed_osds(osdc);
1686         mutex_unlock(&osdc->request_mutex);
1687
1688         if (needmap) {
1689                 dout("%d requests for down osds, need new map\n", needmap);
1690                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1691         }
1692 }
1693
1694
1695 /*
1696  * Process updated osd map.
1697  *
1698  * The message contains any number of incremental and full maps, normally
1699  * indicating some sort of topology change in the cluster.  Kick requests
1700  * off to different OSDs as needed.
1701  */
1702 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1703 {
1704         void *p, *end, *next;
1705         u32 nr_maps, maplen;
1706         u32 epoch;
1707         struct ceph_osdmap *newmap = NULL, *oldmap;
1708         int err;
1709         struct ceph_fsid fsid;
1710
1711         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1712         p = msg->front.iov_base;
1713         end = p + msg->front.iov_len;
1714
1715         /* verify fsid */
1716         ceph_decode_need(&p, end, sizeof(fsid), bad);
1717         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1718         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1719                 return;
1720
1721         down_write(&osdc->map_sem);
1722
1723         /* incremental maps */
1724         ceph_decode_32_safe(&p, end, nr_maps, bad);
1725         dout(" %d inc maps\n", nr_maps);
1726         while (nr_maps > 0) {
1727                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1728                 epoch = ceph_decode_32(&p);
1729                 maplen = ceph_decode_32(&p);
1730                 ceph_decode_need(&p, end, maplen, bad);
1731                 next = p + maplen;
1732                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1733                         dout("applying incremental map %u len %d\n",
1734                              epoch, maplen);
1735                         newmap = osdmap_apply_incremental(&p, next,
1736                                                           osdc->osdmap,
1737                                                           &osdc->client->msgr);
1738                         if (IS_ERR(newmap)) {
1739                                 err = PTR_ERR(newmap);
1740                                 goto bad;
1741                         }
1742                         BUG_ON(!newmap);
1743                         if (newmap != osdc->osdmap) {
1744                                 ceph_osdmap_destroy(osdc->osdmap);
1745                                 osdc->osdmap = newmap;
1746                         }
1747                         kick_requests(osdc, 0);
1748                 } else {
1749                         dout("ignoring incremental map %u len %d\n",
1750                              epoch, maplen);
1751                 }
1752                 p = next;
1753                 nr_maps--;
1754         }
1755         if (newmap)
1756                 goto done;
1757
1758         /* full maps */
1759         ceph_decode_32_safe(&p, end, nr_maps, bad);
1760         dout(" %d full maps\n", nr_maps);
1761         while (nr_maps) {
1762                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1763                 epoch = ceph_decode_32(&p);
1764                 maplen = ceph_decode_32(&p);
1765                 ceph_decode_need(&p, end, maplen, bad);
1766                 if (nr_maps > 1) {
1767                         dout("skipping non-latest full map %u len %d\n",
1768                              epoch, maplen);
1769                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1770                         dout("skipping full map %u len %d, "
1771                              "older than our %u\n", epoch, maplen,
1772                              osdc->osdmap->epoch);
1773                 } else {
1774                         int skipped_map = 0;
1775
1776                         dout("taking full map %u len %d\n", epoch, maplen);
1777                         newmap = osdmap_decode(&p, p+maplen);
1778                         if (IS_ERR(newmap)) {
1779                                 err = PTR_ERR(newmap);
1780                                 goto bad;
1781                         }
1782                         BUG_ON(!newmap);
1783                         oldmap = osdc->osdmap;
1784                         osdc->osdmap = newmap;
1785                         if (oldmap) {
1786                                 if (oldmap->epoch + 1 < newmap->epoch)
1787                                         skipped_map = 1;
1788                                 ceph_osdmap_destroy(oldmap);
1789                         }
1790                         kick_requests(osdc, skipped_map);
1791                 }
1792                 p += maplen;
1793                 nr_maps--;
1794         }
1795
1796 done:
1797         downgrade_write(&osdc->map_sem);
1798         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1799
1800         /*
1801          * subscribe to subsequent osdmap updates if full to ensure
1802          * we find out when we are no longer full and stop returning
1803          * ENOSPC.
1804          */
1805         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1806                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1807
1808         mutex_lock(&osdc->request_mutex);
1809         __send_queued(osdc);
1810         mutex_unlock(&osdc->request_mutex);
1811         up_read(&osdc->map_sem);
1812         wake_up_all(&osdc->client->auth_wq);
1813         return;
1814
1815 bad:
1816         pr_err("osdc handle_map corrupt msg\n");
1817         ceph_msg_dump(msg);
1818         up_write(&osdc->map_sem);
1819         return;
1820 }
1821
1822 /*
1823  * watch/notify callback event infrastructure
1824  *
1825  * These callbacks are used both for watch and notify operations.
1826  */
1827 static void __release_event(struct kref *kref)
1828 {
1829         struct ceph_osd_event *event =
1830                 container_of(kref, struct ceph_osd_event, kref);
1831
1832         dout("__release_event %p\n", event);
1833         kfree(event);
1834 }
1835
1836 static void get_event(struct ceph_osd_event *event)
1837 {
1838         kref_get(&event->kref);
1839 }
1840
1841 void ceph_osdc_put_event(struct ceph_osd_event *event)
1842 {
1843         kref_put(&event->kref, __release_event);
1844 }
1845 EXPORT_SYMBOL(ceph_osdc_put_event);
1846
1847 static void __insert_event(struct ceph_osd_client *osdc,
1848                              struct ceph_osd_event *new)
1849 {
1850         struct rb_node **p = &osdc->event_tree.rb_node;
1851         struct rb_node *parent = NULL;
1852         struct ceph_osd_event *event = NULL;
1853
1854         while (*p) {
1855                 parent = *p;
1856                 event = rb_entry(parent, struct ceph_osd_event, node);
1857                 if (new->cookie < event->cookie)
1858                         p = &(*p)->rb_left;
1859                 else if (new->cookie > event->cookie)
1860                         p = &(*p)->rb_right;
1861                 else
1862                         BUG();
1863         }
1864
1865         rb_link_node(&new->node, parent, p);
1866         rb_insert_color(&new->node, &osdc->event_tree);
1867 }
1868
1869 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1870                                                 u64 cookie)
1871 {
1872         struct rb_node **p = &osdc->event_tree.rb_node;
1873         struct rb_node *parent = NULL;
1874         struct ceph_osd_event *event = NULL;
1875
1876         while (*p) {
1877                 parent = *p;
1878                 event = rb_entry(parent, struct ceph_osd_event, node);
1879                 if (cookie < event->cookie)
1880                         p = &(*p)->rb_left;
1881                 else if (cookie > event->cookie)
1882                         p = &(*p)->rb_right;
1883                 else
1884                         return event;
1885         }
1886         return NULL;
1887 }
1888
1889 static void __remove_event(struct ceph_osd_event *event)
1890 {
1891         struct ceph_osd_client *osdc = event->osdc;
1892
1893         if (!RB_EMPTY_NODE(&event->node)) {
1894                 dout("__remove_event removed %p\n", event);
1895                 rb_erase(&event->node, &osdc->event_tree);
1896                 ceph_osdc_put_event(event);
1897         } else {
1898                 dout("__remove_event didn't remove %p\n", event);
1899         }
1900 }
1901
1902 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1903                            void (*event_cb)(u64, u64, u8, void *),
1904                            void *data, struct ceph_osd_event **pevent)
1905 {
1906         struct ceph_osd_event *event;
1907
1908         event = kmalloc(sizeof(*event), GFP_NOIO);
1909         if (!event)
1910                 return -ENOMEM;
1911
1912         dout("create_event %p\n", event);
1913         event->cb = event_cb;
1914         event->one_shot = 0;
1915         event->data = data;
1916         event->osdc = osdc;
1917         INIT_LIST_HEAD(&event->osd_node);
1918         RB_CLEAR_NODE(&event->node);
1919         kref_init(&event->kref);   /* one ref for us */
1920         kref_get(&event->kref);    /* one ref for the caller */
1921
1922         spin_lock(&osdc->event_lock);
1923         event->cookie = ++osdc->event_count;
1924         __insert_event(osdc, event);
1925         spin_unlock(&osdc->event_lock);
1926
1927         *pevent = event;
1928         return 0;
1929 }
1930 EXPORT_SYMBOL(ceph_osdc_create_event);
1931
1932 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1933 {
1934         struct ceph_osd_client *osdc = event->osdc;
1935
1936         dout("cancel_event %p\n", event);
1937         spin_lock(&osdc->event_lock);
1938         __remove_event(event);
1939         spin_unlock(&osdc->event_lock);
1940         ceph_osdc_put_event(event); /* caller's */
1941 }
1942 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1943
1944
1945 static void do_event_work(struct work_struct *work)
1946 {
1947         struct ceph_osd_event_work *event_work =
1948                 container_of(work, struct ceph_osd_event_work, work);
1949         struct ceph_osd_event *event = event_work->event;
1950         u64 ver = event_work->ver;
1951         u64 notify_id = event_work->notify_id;
1952         u8 opcode = event_work->opcode;
1953
1954         dout("do_event_work completing %p\n", event);
1955         event->cb(ver, notify_id, opcode, event->data);
1956         dout("do_event_work completed %p\n", event);
1957         ceph_osdc_put_event(event);
1958         kfree(event_work);
1959 }
1960
1961
1962 /*
1963  * Process osd watch notifications
1964  */
1965 static void handle_watch_notify(struct ceph_osd_client *osdc,
1966                                 struct ceph_msg *msg)
1967 {
1968         void *p, *end;
1969         u8 proto_ver;
1970         u64 cookie, ver, notify_id;
1971         u8 opcode;
1972         struct ceph_osd_event *event;
1973         struct ceph_osd_event_work *event_work;
1974
1975         p = msg->front.iov_base;
1976         end = p + msg->front.iov_len;
1977
1978         ceph_decode_8_safe(&p, end, proto_ver, bad);
1979         ceph_decode_8_safe(&p, end, opcode, bad);
1980         ceph_decode_64_safe(&p, end, cookie, bad);
1981         ceph_decode_64_safe(&p, end, ver, bad);
1982         ceph_decode_64_safe(&p, end, notify_id, bad);
1983
1984         spin_lock(&osdc->event_lock);
1985         event = __find_event(osdc, cookie);
1986         if (event) {
1987                 BUG_ON(event->one_shot);
1988                 get_event(event);
1989         }
1990         spin_unlock(&osdc->event_lock);
1991         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1992              cookie, ver, event);
1993         if (event) {
1994                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1995                 if (!event_work) {
1996                         dout("ERROR: could not allocate event_work\n");
1997                         goto done_err;
1998                 }
1999                 INIT_WORK(&event_work->work, do_event_work);
2000                 event_work->event = event;
2001                 event_work->ver = ver;
2002                 event_work->notify_id = notify_id;
2003                 event_work->opcode = opcode;
2004                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
2005                         dout("WARNING: failed to queue notify event work\n");
2006                         goto done_err;
2007                 }
2008         }
2009
2010         return;
2011
2012 done_err:
2013         ceph_osdc_put_event(event);
2014         return;
2015
2016 bad:
2017         pr_err("osdc handle_watch_notify corrupt msg\n");
2018         return;
2019 }
2020
2021 /*
2022  * build new request AND message
2023  *
2024  */
2025 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2026                                 struct ceph_snap_context *snapc, u64 snap_id,
2027                                 struct timespec *mtime)
2028 {
2029         struct ceph_msg *msg = req->r_request;
2030         void *p;
2031         size_t msg_size;
2032         int flags = req->r_flags;
2033         u64 data_len;
2034         unsigned int i;
2035
2036         req->r_snapid = snap_id;
2037         req->r_snapc = ceph_get_snap_context(snapc);
2038
2039         /* encode request */
2040         msg->hdr.version = cpu_to_le16(4);
2041
2042         p = msg->front.iov_base;
2043         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2044         req->r_request_osdmap_epoch = p;
2045         p += 4;
2046         req->r_request_flags = p;
2047         p += 4;
2048         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2049                 ceph_encode_timespec(p, mtime);
2050         p += sizeof(struct ceph_timespec);
2051         req->r_request_reassert_version = p;
2052         p += sizeof(struct ceph_eversion); /* will get filled in */
2053
2054         /* oloc */
2055         ceph_encode_8(&p, 4);
2056         ceph_encode_8(&p, 4);
2057         ceph_encode_32(&p, 8 + 4 + 4);
2058         req->r_request_pool = p;
2059         p += 8;
2060         ceph_encode_32(&p, -1);  /* preferred */
2061         ceph_encode_32(&p, 0);   /* key len */
2062
2063         ceph_encode_8(&p, 1);
2064         req->r_request_pgid = p;
2065         p += 8 + 4;
2066         ceph_encode_32(&p, -1);  /* preferred */
2067
2068         /* oid */
2069         ceph_encode_32(&p, req->r_oid_len);
2070         memcpy(p, req->r_oid, req->r_oid_len);
2071         dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len);
2072         p += req->r_oid_len;
2073
2074         /* ops--can imply data */
2075         ceph_encode_16(&p, (u16)req->r_num_ops);
2076         data_len = 0;
2077         for (i = 0; i < req->r_num_ops; i++) {
2078                 data_len += osd_req_encode_op(req, p, i);
2079                 p += sizeof(struct ceph_osd_op);
2080         }
2081
2082         /* snaps */
2083         ceph_encode_64(&p, req->r_snapid);
2084         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2085         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2086         if (req->r_snapc) {
2087                 for (i = 0; i < snapc->num_snaps; i++) {
2088                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2089                 }
2090         }
2091
2092         req->r_request_attempts = p;
2093         p += 4;
2094
2095         /* data */
2096         if (flags & CEPH_OSD_FLAG_WRITE) {
2097                 u16 data_off;
2098
2099                 /*
2100                  * The header "data_off" is a hint to the receiver
2101                  * allowing it to align received data into its
2102                  * buffers such that there's no need to re-copy
2103                  * it before writing it to disk (direct I/O).
2104                  */
2105                 data_off = (u16) (off & 0xffff);
2106                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2107         }
2108         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2109
2110         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2111         msg_size = p - msg->front.iov_base;
2112         msg->front.iov_len = msg_size;
2113         msg->hdr.front_len = cpu_to_le32(msg_size);
2114
2115         dout("build_request msg_size was %d\n", (int)msg_size);
2116 }
2117 EXPORT_SYMBOL(ceph_osdc_build_request);
2118
2119 /*
2120  * Register request, send initial attempt.
2121  */
2122 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2123                             struct ceph_osd_request *req,
2124                             bool nofail)
2125 {
2126         int rc = 0;
2127
2128         down_read(&osdc->map_sem);
2129         mutex_lock(&osdc->request_mutex);
2130         __register_request(osdc, req);
2131         req->r_sent = 0;
2132         req->r_got_reply = 0;
2133         rc = __map_request(osdc, req, 0);
2134         if (rc < 0) {
2135                 if (nofail) {
2136                         dout("osdc_start_request failed map, "
2137                                 " will retry %lld\n", req->r_tid);
2138                         rc = 0;
2139                 } else {
2140                         __unregister_request(osdc, req);
2141                 }
2142                 goto out_unlock;
2143         }
2144         if (req->r_osd == NULL) {
2145                 dout("send_request %p no up osds in pg\n", req);
2146                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2147         } else {
2148                 __send_queued(osdc);
2149         }
2150         rc = 0;
2151 out_unlock:
2152         mutex_unlock(&osdc->request_mutex);
2153         up_read(&osdc->map_sem);
2154         return rc;
2155 }
2156 EXPORT_SYMBOL(ceph_osdc_start_request);
2157
2158 /*
2159  * wait for a request to complete
2160  */
2161 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2162                            struct ceph_osd_request *req)
2163 {
2164         int rc;
2165
2166         rc = wait_for_completion_interruptible(&req->r_completion);
2167         if (rc < 0) {
2168                 mutex_lock(&osdc->request_mutex);
2169                 __cancel_request(req);
2170                 __unregister_request(osdc, req);
2171                 mutex_unlock(&osdc->request_mutex);
2172                 complete_request(req);
2173                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
2174                 return rc;
2175         }
2176
2177         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
2178         return req->r_result;
2179 }
2180 EXPORT_SYMBOL(ceph_osdc_wait_request);
2181
2182 /*
2183  * sync - wait for all in-flight requests to flush.  avoid starvation.
2184  */
2185 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2186 {
2187         struct ceph_osd_request *req;
2188         u64 last_tid, next_tid = 0;
2189
2190         mutex_lock(&osdc->request_mutex);
2191         last_tid = osdc->last_tid;
2192         while (1) {
2193                 req = __lookup_request_ge(osdc, next_tid);
2194                 if (!req)
2195                         break;
2196                 if (req->r_tid > last_tid)
2197                         break;
2198
2199                 next_tid = req->r_tid + 1;
2200                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2201                         continue;
2202
2203                 ceph_osdc_get_request(req);
2204                 mutex_unlock(&osdc->request_mutex);
2205                 dout("sync waiting on tid %llu (last is %llu)\n",
2206                      req->r_tid, last_tid);
2207                 wait_for_completion(&req->r_safe_completion);
2208                 mutex_lock(&osdc->request_mutex);
2209                 ceph_osdc_put_request(req);
2210         }
2211         mutex_unlock(&osdc->request_mutex);
2212         dout("sync done (thru tid %llu)\n", last_tid);
2213 }
2214 EXPORT_SYMBOL(ceph_osdc_sync);
2215
2216 /*
2217  * init, shutdown
2218  */
2219 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2220 {
2221         int err;
2222
2223         dout("init\n");
2224         osdc->client = client;
2225         osdc->osdmap = NULL;
2226         init_rwsem(&osdc->map_sem);
2227         init_completion(&osdc->map_waiters);
2228         osdc->last_requested_map = 0;
2229         mutex_init(&osdc->request_mutex);
2230         osdc->last_tid = 0;
2231         osdc->osds = RB_ROOT;
2232         INIT_LIST_HEAD(&osdc->osd_lru);
2233         osdc->requests = RB_ROOT;
2234         INIT_LIST_HEAD(&osdc->req_lru);
2235         INIT_LIST_HEAD(&osdc->req_unsent);
2236         INIT_LIST_HEAD(&osdc->req_notarget);
2237         INIT_LIST_HEAD(&osdc->req_linger);
2238         osdc->num_requests = 0;
2239         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2240         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2241         spin_lock_init(&osdc->event_lock);
2242         osdc->event_tree = RB_ROOT;
2243         osdc->event_count = 0;
2244
2245         schedule_delayed_work(&osdc->osds_timeout_work,
2246            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2247
2248         err = -ENOMEM;
2249         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2250                                         sizeof(struct ceph_osd_request));
2251         if (!osdc->req_mempool)
2252                 goto out;
2253
2254         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2255                                 OSD_OP_FRONT_LEN, 10, true,
2256                                 "osd_op");
2257         if (err < 0)
2258                 goto out_mempool;
2259         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2260                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2261                                 "osd_op_reply");
2262         if (err < 0)
2263                 goto out_msgpool;
2264
2265         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2266         if (IS_ERR(osdc->notify_wq)) {
2267                 err = PTR_ERR(osdc->notify_wq);
2268                 osdc->notify_wq = NULL;
2269                 goto out_msgpool;
2270         }
2271         return 0;
2272
2273 out_msgpool:
2274         ceph_msgpool_destroy(&osdc->msgpool_op);
2275 out_mempool:
2276         mempool_destroy(osdc->req_mempool);
2277 out:
2278         return err;
2279 }
2280
2281 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2282 {
2283         flush_workqueue(osdc->notify_wq);
2284         destroy_workqueue(osdc->notify_wq);
2285         cancel_delayed_work_sync(&osdc->timeout_work);
2286         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2287         if (osdc->osdmap) {
2288                 ceph_osdmap_destroy(osdc->osdmap);
2289                 osdc->osdmap = NULL;
2290         }
2291         remove_all_osds(osdc);
2292         mempool_destroy(osdc->req_mempool);
2293         ceph_msgpool_destroy(&osdc->msgpool_op);
2294         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2295 }
2296
2297 /*
2298  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2299  * *plen.  Return number of bytes read, or error.
2300  */
2301 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2302                         struct ceph_vino vino, struct ceph_file_layout *layout,
2303                         u64 off, u64 *plen,
2304                         u32 truncate_seq, u64 truncate_size,
2305                         struct page **pages, int num_pages, int page_align)
2306 {
2307         struct ceph_osd_request *req;
2308         int rc = 0;
2309
2310         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2311              vino.snap, off, *plen);
2312         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2313                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2314                                     NULL, truncate_seq, truncate_size,
2315                                     false);
2316         if (IS_ERR(req))
2317                 return PTR_ERR(req);
2318
2319         /* it may be a short read due to an object boundary */
2320
2321         osd_req_op_extent_osd_data_pages(req, 0,
2322                                 pages, *plen, page_align, false, false);
2323
2324         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2325              off, *plen, *plen, page_align);
2326
2327         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2328
2329         rc = ceph_osdc_start_request(osdc, req, false);
2330         if (!rc)
2331                 rc = ceph_osdc_wait_request(osdc, req);
2332
2333         ceph_osdc_put_request(req);
2334         dout("readpages result %d\n", rc);
2335         return rc;
2336 }
2337 EXPORT_SYMBOL(ceph_osdc_readpages);
2338
2339 /*
2340  * do a synchronous write on N pages
2341  */
2342 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2343                          struct ceph_file_layout *layout,
2344                          struct ceph_snap_context *snapc,
2345                          u64 off, u64 len,
2346                          u32 truncate_seq, u64 truncate_size,
2347                          struct timespec *mtime,
2348                          struct page **pages, int num_pages)
2349 {
2350         struct ceph_osd_request *req;
2351         int rc = 0;
2352         int page_align = off & ~PAGE_MASK;
2353
2354         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2355         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2356                                     CEPH_OSD_OP_WRITE,
2357                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2358                                     snapc, truncate_seq, truncate_size,
2359                                     true);
2360         if (IS_ERR(req))
2361                 return PTR_ERR(req);
2362
2363         /* it may be a short write due to an object boundary */
2364         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2365                                 false, false);
2366         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2367
2368         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2369
2370         rc = ceph_osdc_start_request(osdc, req, true);
2371         if (!rc)
2372                 rc = ceph_osdc_wait_request(osdc, req);
2373
2374         ceph_osdc_put_request(req);
2375         if (rc == 0)
2376                 rc = len;
2377         dout("writepages result %d\n", rc);
2378         return rc;
2379 }
2380 EXPORT_SYMBOL(ceph_osdc_writepages);
2381
2382 int ceph_osdc_setup(void)
2383 {
2384         BUG_ON(ceph_osd_request_cache);
2385         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2386                                         sizeof (struct ceph_osd_request),
2387                                         __alignof__(struct ceph_osd_request),
2388                                         0, NULL);
2389
2390         return ceph_osd_request_cache ? 0 : -ENOMEM;
2391 }
2392 EXPORT_SYMBOL(ceph_osdc_setup);
2393
2394 void ceph_osdc_cleanup(void)
2395 {
2396         BUG_ON(!ceph_osd_request_cache);
2397         kmem_cache_destroy(ceph_osd_request_cache);
2398         ceph_osd_request_cache = NULL;
2399 }
2400 EXPORT_SYMBOL(ceph_osdc_cleanup);
2401
2402 /*
2403  * handle incoming message
2404  */
2405 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2406 {
2407         struct ceph_osd *osd = con->private;
2408         struct ceph_osd_client *osdc;
2409         int type = le16_to_cpu(msg->hdr.type);
2410
2411         if (!osd)
2412                 goto out;
2413         osdc = osd->o_osdc;
2414
2415         switch (type) {
2416         case CEPH_MSG_OSD_MAP:
2417                 ceph_osdc_handle_map(osdc, msg);
2418                 break;
2419         case CEPH_MSG_OSD_OPREPLY:
2420                 handle_reply(osdc, msg, con);
2421                 break;
2422         case CEPH_MSG_WATCH_NOTIFY:
2423                 handle_watch_notify(osdc, msg);
2424                 break;
2425
2426         default:
2427                 pr_err("received unknown message type %d %s\n", type,
2428                        ceph_msg_type_name(type));
2429         }
2430 out:
2431         ceph_msg_put(msg);
2432 }
2433
2434 /*
2435  * lookup and return message for incoming reply.  set up reply message
2436  * pages.
2437  */
2438 static struct ceph_msg *get_reply(struct ceph_connection *con,
2439                                   struct ceph_msg_header *hdr,
2440                                   int *skip)
2441 {
2442         struct ceph_osd *osd = con->private;
2443         struct ceph_osd_client *osdc = osd->o_osdc;
2444         struct ceph_msg *m;
2445         struct ceph_osd_request *req;
2446         int front = le32_to_cpu(hdr->front_len);
2447         int data_len = le32_to_cpu(hdr->data_len);
2448         u64 tid;
2449
2450         tid = le64_to_cpu(hdr->tid);
2451         mutex_lock(&osdc->request_mutex);
2452         req = __lookup_request(osdc, tid);
2453         if (!req) {
2454                 *skip = 1;
2455                 m = NULL;
2456                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2457                      osd->o_osd);
2458                 goto out;
2459         }
2460
2461         if (req->r_reply->con)
2462                 dout("%s revoking msg %p from old con %p\n", __func__,
2463                      req->r_reply, req->r_reply->con);
2464         ceph_msg_revoke_incoming(req->r_reply);
2465
2466         if (front > req->r_reply->front.iov_len) {
2467                 pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n",
2468                            front, (int)req->r_reply->front.iov_len,
2469                            (unsigned int)con->peer_name.type,
2470                            le64_to_cpu(con->peer_name.num));
2471                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2472                 if (!m)
2473                         goto out;
2474                 ceph_msg_put(req->r_reply);
2475                 req->r_reply = m;
2476         }
2477         m = ceph_msg_get(req->r_reply);
2478
2479         if (data_len > 0) {
2480                 struct ceph_osd_data *osd_data;
2481
2482                 /*
2483                  * XXX This is assuming there is only one op containing
2484                  * XXX page data.  Probably OK for reads, but this
2485                  * XXX ought to be done more generally.
2486                  */
2487                 osd_data = osd_req_op_extent_osd_data(req, 0);
2488                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2489                         if (osd_data->pages &&
2490                                 unlikely(osd_data->length < data_len)) {
2491
2492                                 pr_warning("tid %lld reply has %d bytes "
2493                                         "we had only %llu bytes ready\n",
2494                                         tid, data_len, osd_data->length);
2495                                 *skip = 1;
2496                                 ceph_msg_put(m);
2497                                 m = NULL;
2498                                 goto out;
2499                         }
2500                 }
2501         }
2502         *skip = 0;
2503         dout("get_reply tid %lld %p\n", tid, m);
2504
2505 out:
2506         mutex_unlock(&osdc->request_mutex);
2507         return m;
2508
2509 }
2510
2511 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2512                                   struct ceph_msg_header *hdr,
2513                                   int *skip)
2514 {
2515         struct ceph_osd *osd = con->private;
2516         int type = le16_to_cpu(hdr->type);
2517         int front = le32_to_cpu(hdr->front_len);
2518
2519         *skip = 0;
2520         switch (type) {
2521         case CEPH_MSG_OSD_MAP:
2522         case CEPH_MSG_WATCH_NOTIFY:
2523                 return ceph_msg_new(type, front, GFP_NOFS, false);
2524         case CEPH_MSG_OSD_OPREPLY:
2525                 return get_reply(con, hdr, skip);
2526         default:
2527                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2528                         osd->o_osd);
2529                 *skip = 1;
2530                 return NULL;
2531         }
2532 }
2533
2534 /*
2535  * Wrappers to refcount containing ceph_osd struct
2536  */
2537 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2538 {
2539         struct ceph_osd *osd = con->private;
2540         if (get_osd(osd))
2541                 return con;
2542         return NULL;
2543 }
2544
2545 static void put_osd_con(struct ceph_connection *con)
2546 {
2547         struct ceph_osd *osd = con->private;
2548         put_osd(osd);
2549 }
2550
2551 /*
2552  * authentication
2553  */
2554 /*
2555  * Note: returned pointer is the address of a structure that's
2556  * managed separately.  Caller must *not* attempt to free it.
2557  */
2558 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2559                                         int *proto, int force_new)
2560 {
2561         struct ceph_osd *o = con->private;
2562         struct ceph_osd_client *osdc = o->o_osdc;
2563         struct ceph_auth_client *ac = osdc->client->monc.auth;
2564         struct ceph_auth_handshake *auth = &o->o_auth;
2565
2566         if (force_new && auth->authorizer) {
2567                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2568                 auth->authorizer = NULL;
2569         }
2570         if (!auth->authorizer) {
2571                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2572                                                       auth);
2573                 if (ret)
2574                         return ERR_PTR(ret);
2575         } else {
2576                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2577                                                      auth);
2578                 if (ret)
2579                         return ERR_PTR(ret);
2580         }
2581         *proto = ac->protocol;
2582
2583         return auth;
2584 }
2585
2586
2587 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2588 {
2589         struct ceph_osd *o = con->private;
2590         struct ceph_osd_client *osdc = o->o_osdc;
2591         struct ceph_auth_client *ac = osdc->client->monc.auth;
2592
2593         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2594 }
2595
2596 static int invalidate_authorizer(struct ceph_connection *con)
2597 {
2598         struct ceph_osd *o = con->private;
2599         struct ceph_osd_client *osdc = o->o_osdc;
2600         struct ceph_auth_client *ac = osdc->client->monc.auth;
2601
2602         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2603         return ceph_monc_validate_auth(&osdc->client->monc);
2604 }
2605
2606 static const struct ceph_connection_operations osd_con_ops = {
2607         .get = get_osd_con,
2608         .put = put_osd_con,
2609         .dispatch = dispatch,
2610         .get_authorizer = get_authorizer,
2611         .verify_authorizer_reply = verify_authorizer_reply,
2612         .invalidate_authorizer = invalidate_authorizer,
2613         .alloc_msg = alloc_msg,
2614         .fault = osd_reset,
2615 };