projects / linux-2.6-block.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Merge branch 'for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/j.anaszewski...
[linux-2.6-block.git] / net / ceph / osd_client.c
... / ...
CommitLineData
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
25static struct kmem_cache *ceph_osd_request_cache;
26
27static const struct ceph_connection_operations osd_con_ops;
28
29static void __send_queued(struct ceph_osd_client *osdc);
30static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
31static void __register_request(struct ceph_osd_client *osdc,
32 struct ceph_osd_request *req);
33static void __unregister_request(struct ceph_osd_client *osdc,
34 struct ceph_osd_request *req);
35static void __unregister_linger_request(struct ceph_osd_client *osdc,
36 struct ceph_osd_request *req);
37static void __enqueue_request(struct ceph_osd_request *req);
38static void __send_request(struct ceph_osd_client *osdc,
39 struct ceph_osd_request *req);
40
41/*
42 * Implement client access to distributed object storage cluster.
43 *
44 * All data objects are stored within a cluster/cloud of OSDs, or
45 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
46 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
47 * remote daemons serving up and coordinating consistent and safe
48 * access to storage.
49 *
50 * Cluster membership and the mapping of data objects onto storage devices
51 * are described by the osd map.
52 *
53 * We keep track of pending OSD requests (read, write), resubmit
54 * requests to different OSDs when the cluster topology/data layout
55 * change, or retry the affected requests when the communications
56 * channel with an OSD is reset.
57 */
58
59/*
60 * calculate the mapping of a file extent onto an object, and fill out the
61 * request accordingly. shorten extent as necessary if it crosses an
62 * object boundary.
63 *
64 * fill osd op in request message.
65 */
66static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
67 u64 *objnum, u64 *objoff, u64 *objlen)
68{
69 u64 orig_len = *plen;
70 int r;
71
72 /* object extent? */
73 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
74 objoff, objlen);
75 if (r < 0)
76 return r;
77 if (*objlen < orig_len) {
78 *plen = *objlen;
79 dout(" skipping last %llu, final file extent %llu~%llu\n",
80 orig_len - *plen, off, *plen);
81 }
82
83 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
84
85 return 0;
86}
87
88static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
89{
90 memset(osd_data, 0, sizeof (*osd_data));
91 osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
92}
93
94static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
95 struct page **pages, u64 length, u32 alignment,
96 bool pages_from_pool, bool own_pages)
97{
98 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
99 osd_data->pages = pages;
100 osd_data->length = length;
101 osd_data->alignment = alignment;
102 osd_data->pages_from_pool = pages_from_pool;
103 osd_data->own_pages = own_pages;
104}
105
106static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
107 struct ceph_pagelist *pagelist)
108{
109 osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
110 osd_data->pagelist = pagelist;
111}
112
113#ifdef CONFIG_BLOCK
114static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
115 struct bio *bio, size_t bio_length)
116{
117 osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
118 osd_data->bio = bio;
119 osd_data->bio_length = bio_length;
120}
121#endif /* CONFIG_BLOCK */
122
123#define osd_req_op_data(oreq, whch, typ, fld) \
124({ \
125 struct ceph_osd_request *__oreq = (oreq); \
126 unsigned int __whch = (whch); \
127 BUG_ON(__whch >= __oreq->r_num_ops); \
128 &__oreq->r_ops[__whch].typ.fld; \
129})
130
131static struct ceph_osd_data *
132osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
133{
134 BUG_ON(which >= osd_req->r_num_ops);
135
136 return &osd_req->r_ops[which].raw_data_in;
137}
138
139struct ceph_osd_data *
140osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
141 unsigned int which)
142{
143 return osd_req_op_data(osd_req, which, extent, osd_data);
144}
145EXPORT_SYMBOL(osd_req_op_extent_osd_data);
146
147struct ceph_osd_data *
148osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
149 unsigned int which)
150{
151 return osd_req_op_data(osd_req, which, cls, response_data);
152}
153EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */
154
155void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
156 unsigned int which, struct page **pages,
157 u64 length, u32 alignment,
158 bool pages_from_pool, bool own_pages)
159{
160 struct ceph_osd_data *osd_data;
161
162 osd_data = osd_req_op_raw_data_in(osd_req, which);
163 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
164 pages_from_pool, own_pages);
165}
166EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
167
168void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
169 unsigned int which, struct page **pages,
170 u64 length, u32 alignment,
171 bool pages_from_pool, bool own_pages)
172{
173 struct ceph_osd_data *osd_data;
174
175 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
176 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
177 pages_from_pool, own_pages);
178}
179EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
180
181void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
182 unsigned int which, struct ceph_pagelist *pagelist)
183{
184 struct ceph_osd_data *osd_data;
185
186 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
187 ceph_osd_data_pagelist_init(osd_data, pagelist);
188}
189EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
190
191#ifdef CONFIG_BLOCK
192void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
193 unsigned int which, struct bio *bio, size_t bio_length)
194{
195 struct ceph_osd_data *osd_data;
196
197 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
198 ceph_osd_data_bio_init(osd_data, bio, bio_length);
199}
200EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
201#endif /* CONFIG_BLOCK */
202
203static void osd_req_op_cls_request_info_pagelist(
204 struct ceph_osd_request *osd_req,
205 unsigned int which, struct ceph_pagelist *pagelist)
206{
207 struct ceph_osd_data *osd_data;
208
209 osd_data = osd_req_op_data(osd_req, which, cls, request_info);
210 ceph_osd_data_pagelist_init(osd_data, pagelist);
211}
212
213void osd_req_op_cls_request_data_pagelist(
214 struct ceph_osd_request *osd_req,
215 unsigned int which, struct ceph_pagelist *pagelist)
216{
217 struct ceph_osd_data *osd_data;
218
219 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
220 ceph_osd_data_pagelist_init(osd_data, pagelist);
221}
222EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
223
224void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
225 unsigned int which, struct page **pages, u64 length,
226 u32 alignment, bool pages_from_pool, bool own_pages)
227{
228 struct ceph_osd_data *osd_data;
229
230 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
231 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
232 pages_from_pool, own_pages);
233}
234EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
235
236void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
237 unsigned int which, struct page **pages, u64 length,
238 u32 alignment, bool pages_from_pool, bool own_pages)
239{
240 struct ceph_osd_data *osd_data;
241
242 osd_data = osd_req_op_data(osd_req, which, cls, response_data);
243 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
244 pages_from_pool, own_pages);
245}
246EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
247
248static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
249{
250 switch (osd_data->type) {
251 case CEPH_OSD_DATA_TYPE_NONE:
252 return 0;
253 case CEPH_OSD_DATA_TYPE_PAGES:
254 return osd_data->length;
255 case CEPH_OSD_DATA_TYPE_PAGELIST:
256 return (u64)osd_data->pagelist->length;
257#ifdef CONFIG_BLOCK
258 case CEPH_OSD_DATA_TYPE_BIO:
259 return (u64)osd_data->bio_length;
260#endif /* CONFIG_BLOCK */
261 default:
262 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
263 return 0;
264 }
265}
266
267static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
268{
269 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
270 int num_pages;
271
272 num_pages = calc_pages_for((u64)osd_data->alignment,
273 (u64)osd_data->length);
274 ceph_release_page_vector(osd_data->pages, num_pages);
275 }
276 ceph_osd_data_init(osd_data);
277}
278
279static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
280 unsigned int which)
281{
282 struct ceph_osd_req_op *op;
283
284 BUG_ON(which >= osd_req->r_num_ops);
285 op = &osd_req->r_ops[which];
286
287 switch (op->op) {
288 case CEPH_OSD_OP_READ:
289 case CEPH_OSD_OP_WRITE:
290 case CEPH_OSD_OP_WRITEFULL:
291 ceph_osd_data_release(&op->extent.osd_data);
292 break;
293 case CEPH_OSD_OP_CALL:
294 ceph_osd_data_release(&op->cls.request_info);
295 ceph_osd_data_release(&op->cls.request_data);
296 ceph_osd_data_release(&op->cls.response_data);
297 break;
298 case CEPH_OSD_OP_SETXATTR:
299 case CEPH_OSD_OP_CMPXATTR:
300 ceph_osd_data_release(&op->xattr.osd_data);
301 break;
302 case CEPH_OSD_OP_STAT:
303 ceph_osd_data_release(&op->raw_data_in);
304 break;
305 default:
306 break;
307 }
308}
309
310/*
311 * requests
312 */
313static void ceph_osdc_release_request(struct kref *kref)
314{
315 struct ceph_osd_request *req = container_of(kref,
316 struct ceph_osd_request, r_kref);
317 unsigned int which;
318
319 dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
320 req->r_request, req->r_reply);
321 WARN_ON(!RB_EMPTY_NODE(&req->r_node));
322 WARN_ON(!list_empty(&req->r_req_lru_item));
323 WARN_ON(!list_empty(&req->r_osd_item));
324 WARN_ON(!list_empty(&req->r_linger_item));
325 WARN_ON(!list_empty(&req->r_linger_osd_item));
326 WARN_ON(req->r_osd);
327
328 if (req->r_request)
329 ceph_msg_put(req->r_request);
330 if (req->r_reply) {
331 ceph_msg_revoke_incoming(req->r_reply);
332 ceph_msg_put(req->r_reply);
333 }
334
335 for (which = 0; which < req->r_num_ops; which++)
336 osd_req_op_data_release(req, which);
337
338 ceph_put_snap_context(req->r_snapc);
339 if (req->r_mempool)
340 mempool_free(req, req->r_osdc->req_mempool);
341 else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS)
342 kmem_cache_free(ceph_osd_request_cache, req);
343 else
344 kfree(req);
345}
346
347void ceph_osdc_get_request(struct ceph_osd_request *req)
348{
349 dout("%s %p (was %d)\n", __func__, req,
350 atomic_read(&req->r_kref.refcount));
351 kref_get(&req->r_kref);
352}
353EXPORT_SYMBOL(ceph_osdc_get_request);
354
355void ceph_osdc_put_request(struct ceph_osd_request *req)
356{
357 dout("%s %p (was %d)\n", __func__, req,
358 atomic_read(&req->r_kref.refcount));
359 kref_put(&req->r_kref, ceph_osdc_release_request);
360}
361EXPORT_SYMBOL(ceph_osdc_put_request);
362
363struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
364 struct ceph_snap_context *snapc,
365 unsigned int num_ops,
366 bool use_mempool,
367 gfp_t gfp_flags)
368{
369 struct ceph_osd_request *req;
370 struct ceph_msg *msg;
371 size_t msg_size;
372
373 if (use_mempool) {
374 BUG_ON(num_ops > CEPH_OSD_SLAB_OPS);
375 req = mempool_alloc(osdc->req_mempool, gfp_flags);
376 } else if (num_ops <= CEPH_OSD_SLAB_OPS) {
377 req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags);
378 } else {
379 BUG_ON(num_ops > CEPH_OSD_MAX_OPS);
380 req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]),
381 gfp_flags);
382 }
383 if (unlikely(!req))
384 return NULL;
385
386 /* req only, each op is zeroed in _osd_req_op_init() */
387 memset(req, 0, sizeof(*req));
388
389 req->r_osdc = osdc;
390 req->r_mempool = use_mempool;
391 req->r_num_ops = num_ops;
392
393 kref_init(&req->r_kref);
394 init_completion(&req->r_completion);
395 init_completion(&req->r_safe_completion);
396 RB_CLEAR_NODE(&req->r_node);
397 INIT_LIST_HEAD(&req->r_unsafe_item);
398 INIT_LIST_HEAD(&req->r_linger_item);
399 INIT_LIST_HEAD(&req->r_linger_osd_item);
400 INIT_LIST_HEAD(&req->r_req_lru_item);
401 INIT_LIST_HEAD(&req->r_osd_item);
402
403 req->r_base_oloc.pool = -1;
404 req->r_target_oloc.pool = -1;
405
406 msg_size = OSD_OPREPLY_FRONT_LEN;
407 if (num_ops > CEPH_OSD_SLAB_OPS) {
408 /* ceph_osd_op and rval */
409 msg_size += (num_ops - CEPH_OSD_SLAB_OPS) *
410 (sizeof(struct ceph_osd_op) + 4);
411 }
412
413 /* create reply message */
414 if (use_mempool)
415 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
416 else
417 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size,
418 gfp_flags, true);
419 if (!msg) {
420 ceph_osdc_put_request(req);
421 return NULL;
422 }
423 req->r_reply = msg;
424
425 msg_size = 4 + 4 + 4; /* client_inc, osdmap_epoch, flags */
426 msg_size += 4 + 4 + 4 + 8; /* mtime, reassert_version */
427 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
428 msg_size += 1 + 8 + 4 + 4; /* pgid */
429 msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
430 msg_size += 2 + num_ops * sizeof(struct ceph_osd_op);
431 msg_size += 8; /* snapid */
432 msg_size += 8; /* snap_seq */
433 msg_size += 4 + 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
434 msg_size += 4; /* retry_attempt */
435
436 /* create request message; allow space for oid */
437 if (use_mempool)
438 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
439 else
440 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
441 if (!msg) {
442 ceph_osdc_put_request(req);
443 return NULL;
444 }
445
446 memset(msg->front.iov_base, 0, msg->front.iov_len);
447
448 req->r_request = msg;
449
450 return req;
451}
452EXPORT_SYMBOL(ceph_osdc_alloc_request);
453
454static bool osd_req_opcode_valid(u16 opcode)
455{
456 switch (opcode) {
457#define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true;
458__CEPH_FORALL_OSD_OPS(GENERATE_CASE)
459#undef GENERATE_CASE
460 default:
461 return false;
462 }
463}
464
465/*
466 * This is an osd op init function for opcodes that have no data or
467 * other information associated with them. It also serves as a
468 * common init routine for all the other init functions, below.
469 */
470static struct ceph_osd_req_op *
471_osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
472 u16 opcode, u32 flags)
473{
474 struct ceph_osd_req_op *op;
475
476 BUG_ON(which >= osd_req->r_num_ops);
477 BUG_ON(!osd_req_opcode_valid(opcode));
478
479 op = &osd_req->r_ops[which];
480 memset(op, 0, sizeof (*op));
481 op->op = opcode;
482 op->flags = flags;
483
484 return op;
485}
486
487void osd_req_op_init(struct ceph_osd_request *osd_req,
488 unsigned int which, u16 opcode, u32 flags)
489{
490 (void)_osd_req_op_init(osd_req, which, opcode, flags);
491}
492EXPORT_SYMBOL(osd_req_op_init);
493
494void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
495 unsigned int which, u16 opcode,
496 u64 offset, u64 length,
497 u64 truncate_size, u32 truncate_seq)
498{
499 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
500 opcode, 0);
501 size_t payload_len = 0;
502
503 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
504 opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO &&
505 opcode != CEPH_OSD_OP_TRUNCATE);
506
507 op->extent.offset = offset;
508 op->extent.length = length;
509 op->extent.truncate_size = truncate_size;
510 op->extent.truncate_seq = truncate_seq;
511 if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL)
512 payload_len += length;
513
514 op->indata_len = payload_len;
515}
516EXPORT_SYMBOL(osd_req_op_extent_init);
517
518void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
519 unsigned int which, u64 length)
520{
521 struct ceph_osd_req_op *op;
522 u64 previous;
523
524 BUG_ON(which >= osd_req->r_num_ops);
525 op = &osd_req->r_ops[which];
526 previous = op->extent.length;
527
528 if (length == previous)
529 return; /* Nothing to do */
530 BUG_ON(length > previous);
531
532 op->extent.length = length;
533 op->indata_len -= previous - length;
534}
535EXPORT_SYMBOL(osd_req_op_extent_update);
536
537void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
538 unsigned int which, u64 offset_inc)
539{
540 struct ceph_osd_req_op *op, *prev_op;
541
542 BUG_ON(which + 1 >= osd_req->r_num_ops);
543
544 prev_op = &osd_req->r_ops[which];
545 op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags);
546 /* dup previous one */
547 op->indata_len = prev_op->indata_len;
548 op->outdata_len = prev_op->outdata_len;
549 op->extent = prev_op->extent;
550 /* adjust offset */
551 op->extent.offset += offset_inc;
552 op->extent.length -= offset_inc;
553
554 if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
555 op->indata_len -= offset_inc;
556}
557EXPORT_SYMBOL(osd_req_op_extent_dup_last);
558
559void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
560 u16 opcode, const char *class, const char *method)
561{
562 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
563 opcode, 0);
564 struct ceph_pagelist *pagelist;
565 size_t payload_len = 0;
566 size_t size;
567
568 BUG_ON(opcode != CEPH_OSD_OP_CALL);
569
570 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
571 BUG_ON(!pagelist);
572 ceph_pagelist_init(pagelist);
573
574 op->cls.class_name = class;
575 size = strlen(class);
576 BUG_ON(size > (size_t) U8_MAX);
577 op->cls.class_len = size;
578 ceph_pagelist_append(pagelist, class, size);
579 payload_len += size;
580
581 op->cls.method_name = method;
582 size = strlen(method);
583 BUG_ON(size > (size_t) U8_MAX);
584 op->cls.method_len = size;
585 ceph_pagelist_append(pagelist, method, size);
586 payload_len += size;
587
588 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
589
590 op->cls.argc = 0; /* currently unused */
591
592 op->indata_len = payload_len;
593}
594EXPORT_SYMBOL(osd_req_op_cls_init);
595
596int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
597 u16 opcode, const char *name, const void *value,
598 size_t size, u8 cmp_op, u8 cmp_mode)
599{
600 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
601 opcode, 0);
602 struct ceph_pagelist *pagelist;
603 size_t payload_len;
604
605 BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
606
607 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
608 if (!pagelist)
609 return -ENOMEM;
610
611 ceph_pagelist_init(pagelist);
612
613 payload_len = strlen(name);
614 op->xattr.name_len = payload_len;
615 ceph_pagelist_append(pagelist, name, payload_len);
616
617 op->xattr.value_len = size;
618 ceph_pagelist_append(pagelist, value, size);
619 payload_len += size;
620
621 op->xattr.cmp_op = cmp_op;
622 op->xattr.cmp_mode = cmp_mode;
623
624 ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
625 op->indata_len = payload_len;
626 return 0;
627}
628EXPORT_SYMBOL(osd_req_op_xattr_init);
629
630void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
631 unsigned int which, u16 opcode,
632 u64 cookie, u64 version, int flag)
633{
634 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
635 opcode, 0);
636
637 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
638
639 op->watch.cookie = cookie;
640 op->watch.ver = version;
641 if (opcode == CEPH_OSD_OP_WATCH && flag)
642 op->watch.flag = (u8)1;
643}
644EXPORT_SYMBOL(osd_req_op_watch_init);
645
646void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
647 unsigned int which,
648 u64 expected_object_size,
649 u64 expected_write_size)
650{
651 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
652 CEPH_OSD_OP_SETALLOCHINT,
653 0);
654
655 op->alloc_hint.expected_object_size = expected_object_size;
656 op->alloc_hint.expected_write_size = expected_write_size;
657
658 /*
659 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
660 * not worth a feature bit. Set FAILOK per-op flag to make
661 * sure older osds don't trip over an unsupported opcode.
662 */
663 op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
664}
665EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
666
667static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
668 struct ceph_osd_data *osd_data)
669{
670 u64 length = ceph_osd_data_length(osd_data);
671
672 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
673 BUG_ON(length > (u64) SIZE_MAX);
674 if (length)
675 ceph_msg_data_add_pages(msg, osd_data->pages,
676 length, osd_data->alignment);
677 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
678 BUG_ON(!length);
679 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
680#ifdef CONFIG_BLOCK
681 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
682 ceph_msg_data_add_bio(msg, osd_data->bio, length);
683#endif
684 } else {
685 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
686 }
687}
688
689static u64 osd_req_encode_op(struct ceph_osd_request *req,
690 struct ceph_osd_op *dst, unsigned int which)
691{
692 struct ceph_osd_req_op *src;
693 struct ceph_osd_data *osd_data;
694 u64 request_data_len = 0;
695 u64 data_length;
696
697 BUG_ON(which >= req->r_num_ops);
698 src = &req->r_ops[which];
699 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
700 pr_err("unrecognized osd opcode %d\n", src->op);
701
702 return 0;
703 }
704
705 switch (src->op) {
706 case CEPH_OSD_OP_STAT:
707 osd_data = &src->raw_data_in;
708 ceph_osdc_msg_data_add(req->r_reply, osd_data);
709 break;
710 case CEPH_OSD_OP_READ:
711 case CEPH_OSD_OP_WRITE:
712 case CEPH_OSD_OP_WRITEFULL:
713 case CEPH_OSD_OP_ZERO:
714 case CEPH_OSD_OP_TRUNCATE:
715 if (src->op == CEPH_OSD_OP_WRITE ||
716 src->op == CEPH_OSD_OP_WRITEFULL)
717 request_data_len = src->extent.length;
718 dst->extent.offset = cpu_to_le64(src->extent.offset);
719 dst->extent.length = cpu_to_le64(src->extent.length);
720 dst->extent.truncate_size =
721 cpu_to_le64(src->extent.truncate_size);
722 dst->extent.truncate_seq =
723 cpu_to_le32(src->extent.truncate_seq);
724 osd_data = &src->extent.osd_data;
725 if (src->op == CEPH_OSD_OP_WRITE ||
726 src->op == CEPH_OSD_OP_WRITEFULL)
727 ceph_osdc_msg_data_add(req->r_request, osd_data);
728 else
729 ceph_osdc_msg_data_add(req->r_reply, osd_data);
730 break;
731 case CEPH_OSD_OP_CALL:
732 dst->cls.class_len = src->cls.class_len;
733 dst->cls.method_len = src->cls.method_len;
734 osd_data = &src->cls.request_info;
735 ceph_osdc_msg_data_add(req->r_request, osd_data);
736 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
737 request_data_len = osd_data->pagelist->length;
738
739 osd_data = &src->cls.request_data;
740 data_length = ceph_osd_data_length(osd_data);
741 if (data_length) {
742 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
743 dst->cls.indata_len = cpu_to_le32(data_length);
744 ceph_osdc_msg_data_add(req->r_request, osd_data);
745 src->indata_len += data_length;
746 request_data_len += data_length;
747 }
748 osd_data = &src->cls.response_data;
749 ceph_osdc_msg_data_add(req->r_reply, osd_data);
750 break;
751 case CEPH_OSD_OP_STARTSYNC:
752 break;
753 case CEPH_OSD_OP_NOTIFY_ACK:
754 case CEPH_OSD_OP_WATCH:
755 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
756 dst->watch.ver = cpu_to_le64(src->watch.ver);
757 dst->watch.flag = src->watch.flag;
758 break;
759 case CEPH_OSD_OP_SETALLOCHINT:
760 dst->alloc_hint.expected_object_size =
761 cpu_to_le64(src->alloc_hint.expected_object_size);
762 dst->alloc_hint.expected_write_size =
763 cpu_to_le64(src->alloc_hint.expected_write_size);
764 break;
765 case CEPH_OSD_OP_SETXATTR:
766 case CEPH_OSD_OP_CMPXATTR:
767 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
768 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
769 dst->xattr.cmp_op = src->xattr.cmp_op;
770 dst->xattr.cmp_mode = src->xattr.cmp_mode;
771 osd_data = &src->xattr.osd_data;
772 ceph_osdc_msg_data_add(req->r_request, osd_data);
773 request_data_len = osd_data->pagelist->length;
774 break;
775 case CEPH_OSD_OP_CREATE:
776 case CEPH_OSD_OP_DELETE:
777 break;
778 default:
779 pr_err("unsupported osd opcode %s\n",
780 ceph_osd_op_name(src->op));
781 WARN_ON(1);
782
783 return 0;
784 }
785
786 dst->op = cpu_to_le16(src->op);
787 dst->flags = cpu_to_le32(src->flags);
788 dst->payload_len = cpu_to_le32(src->indata_len);
789
790 return request_data_len;
791}
792
793/*
794 * build new request AND message, calculate layout, and adjust file
795 * extent as needed.
796 *
797 * if the file was recently truncated, we include information about its
798 * old and new size so that the object can be updated appropriately. (we
799 * avoid synchronously deleting truncated objects because it's slow.)
800 *
801 * if @do_sync, include a 'startsync' command so that the osd will flush
802 * data quickly.
803 */
804struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
805 struct ceph_file_layout *layout,
806 struct ceph_vino vino,
807 u64 off, u64 *plen,
808 unsigned int which, int num_ops,
809 int opcode, int flags,
810 struct ceph_snap_context *snapc,
811 u32 truncate_seq,
812 u64 truncate_size,
813 bool use_mempool)
814{
815 struct ceph_osd_request *req;
816 u64 objnum = 0;
817 u64 objoff = 0;
818 u64 objlen = 0;
819 int r;
820
821 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
822 opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
823 opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
824
825 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
826 GFP_NOFS);
827 if (!req)
828 return ERR_PTR(-ENOMEM);
829
830 req->r_flags = flags;
831
832 /* calculate max write size */
833 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
834 if (r < 0) {
835 ceph_osdc_put_request(req);
836 return ERR_PTR(r);
837 }
838
839 if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
840 osd_req_op_init(req, which, opcode, 0);
841 } else {
842 u32 object_size = le32_to_cpu(layout->fl_object_size);
843 u32 object_base = off - objoff;
844 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
845 if (truncate_size <= object_base) {
846 truncate_size = 0;
847 } else {
848 truncate_size -= object_base;
849 if (truncate_size > object_size)
850 truncate_size = object_size;
851 }
852 }
853 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
854 truncate_size, truncate_seq);
855 }
856
857 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
858
859 snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
860 "%llx.%08llx", vino.ino, objnum);
861 req->r_base_oid.name_len = strlen(req->r_base_oid.name);
862
863 return req;
864}
865EXPORT_SYMBOL(ceph_osdc_new_request);
866
867/*
868 * We keep osd requests in an rbtree, sorted by ->r_tid.
869 */
870static void __insert_request(struct ceph_osd_client *osdc,
871 struct ceph_osd_request *new)
872{
873 struct rb_node **p = &osdc->requests.rb_node;
874 struct rb_node *parent = NULL;
875 struct ceph_osd_request *req = NULL;
876
877 while (*p) {
878 parent = *p;
879 req = rb_entry(parent, struct ceph_osd_request, r_node);
880 if (new->r_tid < req->r_tid)
881 p = &(*p)->rb_left;
882 else if (new->r_tid > req->r_tid)
883 p = &(*p)->rb_right;
884 else
885 BUG();
886 }
887
888 rb_link_node(&new->r_node, parent, p);
889 rb_insert_color(&new->r_node, &osdc->requests);
890}
891
892static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
893 u64 tid)
894{
895 struct ceph_osd_request *req;
896 struct rb_node *n = osdc->requests.rb_node;
897
898 while (n) {
899 req = rb_entry(n, struct ceph_osd_request, r_node);
900 if (tid < req->r_tid)
901 n = n->rb_left;
902 else if (tid > req->r_tid)
903 n = n->rb_right;
904 else
905 return req;
906 }
907 return NULL;
908}
909
910static struct ceph_osd_request *
911__lookup_request_ge(struct ceph_osd_client *osdc,
912 u64 tid)
913{
914 struct ceph_osd_request *req;
915 struct rb_node *n = osdc->requests.rb_node;
916
917 while (n) {
918 req = rb_entry(n, struct ceph_osd_request, r_node);
919 if (tid < req->r_tid) {
920 if (!n->rb_left)
921 return req;
922 n = n->rb_left;
923 } else if (tid > req->r_tid) {
924 n = n->rb_right;
925 } else {
926 return req;
927 }
928 }
929 return NULL;
930}
931
932static void __kick_linger_request(struct ceph_osd_request *req)
933{
934 struct ceph_osd_client *osdc = req->r_osdc;
935 struct ceph_osd *osd = req->r_osd;
936
937 /*
938 * Linger requests need to be resent with a new tid to avoid
939 * the dup op detection logic on the OSDs. Achieve this with
940 * a re-register dance instead of open-coding.
941 */
942 ceph_osdc_get_request(req);
943 if (!list_empty(&req->r_linger_item))
944 __unregister_linger_request(osdc, req);
945 else
946 __unregister_request(osdc, req);
947 __register_request(osdc, req);
948 ceph_osdc_put_request(req);
949
950 /*
951 * Unless request has been registered as both normal and
952 * lingering, __unregister{,_linger}_request clears r_osd.
953 * However, here we need to preserve r_osd to make sure we
954 * requeue on the same OSD.
955 */
956 WARN_ON(req->r_osd || !osd);
957 req->r_osd = osd;
958
959 dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
960 __enqueue_request(req);
961}
962
963/*
964 * Resubmit requests pending on the given osd.
965 */
966static void __kick_osd_requests(struct ceph_osd_client *osdc,
967 struct ceph_osd *osd)
968{
969 struct ceph_osd_request *req, *nreq;
970 LIST_HEAD(resend);
971 LIST_HEAD(resend_linger);
972 int err;
973
974 dout("%s osd%d\n", __func__, osd->o_osd);
975 err = __reset_osd(osdc, osd);
976 if (err)
977 return;
978
979 /*
980 * Build up a list of requests to resend by traversing the
981 * osd's list of requests. Requests for a given object are
982 * sent in tid order, and that is also the order they're
983 * kept on this list. Therefore all requests that are in
984 * flight will be found first, followed by all requests that
985 * have not yet been sent. And to resend requests while
986 * preserving this order we will want to put any sent
987 * requests back on the front of the osd client's unsent
988 * list.
989 *
990 * So we build a separate ordered list of already-sent
991 * requests for the affected osd and splice it onto the
992 * front of the osd client's unsent list. Once we've seen a
993 * request that has not yet been sent we're done. Those
994 * requests are already sitting right where they belong.
995 */
996 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
997 if (!req->r_sent)
998 break;
999
1000 if (!req->r_linger) {
1001 dout("%s requeueing %p tid %llu\n", __func__, req,
1002 req->r_tid);
1003 list_move_tail(&req->r_req_lru_item, &resend);
1004 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1005 } else {
1006 list_move_tail(&req->r_req_lru_item, &resend_linger);
1007 }
1008 }
1009 list_splice(&resend, &osdc->req_unsent);
1010
1011 /*
1012 * Both registered and not yet registered linger requests are
1013 * enqueued with a new tid on the same OSD. We add/move them
1014 * to req_unsent/o_requests at the end to keep things in tid
1015 * order.
1016 */
1017 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
1018 r_linger_osd_item) {
1019 WARN_ON(!list_empty(&req->r_req_lru_item));
1020 __kick_linger_request(req);
1021 }
1022
1023 list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
1024 __kick_linger_request(req);
1025}
1026
1027/*
1028 * If the osd connection drops, we need to resubmit all requests.
1029 */
1030static void osd_reset(struct ceph_connection *con)
1031{
1032 struct ceph_osd *osd = con->private;
1033 struct ceph_osd_client *osdc;
1034
1035 if (!osd)
1036 return;
1037 dout("osd_reset osd%d\n", osd->o_osd);
1038 osdc = osd->o_osdc;
1039 down_read(&osdc->map_sem);
1040 mutex_lock(&osdc->request_mutex);
1041 __kick_osd_requests(osdc, osd);
1042 __send_queued(osdc);
1043 mutex_unlock(&osdc->request_mutex);
1044 up_read(&osdc->map_sem);
1045}
1046
1047/*
1048 * Track open sessions with osds.
1049 */
1050static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1051{
1052 struct ceph_osd *osd;
1053
1054 osd = kzalloc(sizeof(*osd), GFP_NOFS);
1055 if (!osd)
1056 return NULL;
1057
1058 atomic_set(&osd->o_ref, 1);
1059 osd->o_osdc = osdc;
1060 osd->o_osd = onum;
1061 RB_CLEAR_NODE(&osd->o_node);
1062 INIT_LIST_HEAD(&osd->o_requests);
1063 INIT_LIST_HEAD(&osd->o_linger_requests);
1064 INIT_LIST_HEAD(&osd->o_osd_lru);
1065 osd->o_incarnation = 1;
1066
1067 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1068
1069 INIT_LIST_HEAD(&osd->o_keepalive_item);
1070 return osd;
1071}
1072
1073static struct ceph_osd *get_osd(struct ceph_osd *osd)
1074{
1075 if (atomic_inc_not_zero(&osd->o_ref)) {
1076 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1077 atomic_read(&osd->o_ref));
1078 return osd;
1079 } else {
1080 dout("get_osd %p FAIL\n", osd);
1081 return NULL;
1082 }
1083}
1084
1085static void put_osd(struct ceph_osd *osd)
1086{
1087 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1088 atomic_read(&osd->o_ref) - 1);
1089 if (atomic_dec_and_test(&osd->o_ref)) {
1090 if (osd->o_auth.authorizer)
1091 ceph_auth_destroy_authorizer(osd->o_auth.authorizer);
1092 kfree(osd);
1093 }
1094}
1095
1096/*
1097 * remove an osd from our map
1098 */
1099static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1100{
1101 dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1102 WARN_ON(!list_empty(&osd->o_requests));
1103 WARN_ON(!list_empty(&osd->o_linger_requests));
1104
1105 list_del_init(&osd->o_osd_lru);
1106 rb_erase(&osd->o_node, &osdc->osds);
1107 RB_CLEAR_NODE(&osd->o_node);
1108}
1109
1110static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1111{
1112 dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1113
1114 if (!RB_EMPTY_NODE(&osd->o_node)) {
1115 ceph_con_close(&osd->o_con);
1116 __remove_osd(osdc, osd);
1117 put_osd(osd);
1118 }
1119}
1120
1121static void remove_all_osds(struct ceph_osd_client *osdc)
1122{
1123 dout("%s %p\n", __func__, osdc);
1124 mutex_lock(&osdc->request_mutex);
1125 while (!RB_EMPTY_ROOT(&osdc->osds)) {
1126 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1127 struct ceph_osd, o_node);
1128 remove_osd(osdc, osd);
1129 }
1130 mutex_unlock(&osdc->request_mutex);
1131}
1132
1133static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1134 struct ceph_osd *osd)
1135{
1136 dout("%s %p\n", __func__, osd);
1137 BUG_ON(!list_empty(&osd->o_osd_lru));
1138
1139 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1140 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1141}
1142
1143static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1144 struct ceph_osd *osd)
1145{
1146 dout("%s %p\n", __func__, osd);
1147
1148 if (list_empty(&osd->o_requests) &&
1149 list_empty(&osd->o_linger_requests))
1150 __move_osd_to_lru(osdc, osd);
1151}
1152
1153static void __remove_osd_from_lru(struct ceph_osd *osd)
1154{
1155 dout("__remove_osd_from_lru %p\n", osd);
1156 if (!list_empty(&osd->o_osd_lru))
1157 list_del_init(&osd->o_osd_lru);
1158}
1159
1160static void remove_old_osds(struct ceph_osd_client *osdc)
1161{
1162 struct ceph_osd *osd, *nosd;
1163
1164 dout("__remove_old_osds %p\n", osdc);
1165 mutex_lock(&osdc->request_mutex);
1166 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1167 if (time_before(jiffies, osd->lru_ttl))
1168 break;
1169 remove_osd(osdc, osd);
1170 }
1171 mutex_unlock(&osdc->request_mutex);
1172}
1173
1174/*
1175 * reset osd connect
1176 */
1177static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1178{
1179 struct ceph_entity_addr *peer_addr;
1180
1181 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1182 if (list_empty(&osd->o_requests) &&
1183 list_empty(&osd->o_linger_requests)) {
1184 remove_osd(osdc, osd);
1185 return -ENODEV;
1186 }
1187
1188 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1189 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1190 !ceph_con_opened(&osd->o_con)) {
1191 struct ceph_osd_request *req;
1192
1193 dout("osd addr hasn't changed and connection never opened, "
1194 "letting msgr retry\n");
1195 /* touch each r_stamp for handle_timeout()'s benfit */
1196 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1197 req->r_stamp = jiffies;
1198
1199 return -EAGAIN;
1200 }
1201
1202 ceph_con_close(&osd->o_con);
1203 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1204 osd->o_incarnation++;
1205
1206 return 0;
1207}
1208
1209static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1210{
1211 struct rb_node **p = &osdc->osds.rb_node;
1212 struct rb_node *parent = NULL;
1213 struct ceph_osd *osd = NULL;
1214
1215 dout("__insert_osd %p osd%d\n", new, new->o_osd);
1216 while (*p) {
1217 parent = *p;
1218 osd = rb_entry(parent, struct ceph_osd, o_node);
1219 if (new->o_osd < osd->o_osd)
1220 p = &(*p)->rb_left;
1221 else if (new->o_osd > osd->o_osd)
1222 p = &(*p)->rb_right;
1223 else
1224 BUG();
1225 }
1226
1227 rb_link_node(&new->o_node, parent, p);
1228 rb_insert_color(&new->o_node, &osdc->osds);
1229}
1230
1231static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1232{
1233 struct ceph_osd *osd;
1234 struct rb_node *n = osdc->osds.rb_node;
1235
1236 while (n) {
1237 osd = rb_entry(n, struct ceph_osd, o_node);
1238 if (o < osd->o_osd)
1239 n = n->rb_left;
1240 else if (o > osd->o_osd)
1241 n = n->rb_right;
1242 else
1243 return osd;
1244 }
1245 return NULL;
1246}
1247
1248static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1249{
1250 schedule_delayed_work(&osdc->timeout_work,
1251 osdc->client->options->osd_keepalive_timeout);
1252}
1253
1254static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1255{
1256 cancel_delayed_work(&osdc->timeout_work);
1257}
1258
1259/*
1260 * Register request, assign tid. If this is the first request, set up
1261 * the timeout event.
1262 */
1263static void __register_request(struct ceph_osd_client *osdc,
1264 struct ceph_osd_request *req)
1265{
1266 req->r_tid = ++osdc->last_tid;
1267 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1268 dout("__register_request %p tid %lld\n", req, req->r_tid);
1269 __insert_request(osdc, req);
1270 ceph_osdc_get_request(req);
1271 osdc->num_requests++;
1272 if (osdc->num_requests == 1) {
1273 dout(" first request, scheduling timeout\n");
1274 __schedule_osd_timeout(osdc);
1275 }
1276}
1277
1278/*
1279 * called under osdc->request_mutex
1280 */
1281static void __unregister_request(struct ceph_osd_client *osdc,
1282 struct ceph_osd_request *req)
1283{
1284 if (RB_EMPTY_NODE(&req->r_node)) {
1285 dout("__unregister_request %p tid %lld not registered\n",
1286 req, req->r_tid);
1287 return;
1288 }
1289
1290 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1291 rb_erase(&req->r_node, &osdc->requests);
1292 RB_CLEAR_NODE(&req->r_node);
1293 osdc->num_requests--;
1294
1295 if (req->r_osd) {
1296 /* make sure the original request isn't in flight. */
1297 ceph_msg_revoke(req->r_request);
1298
1299 list_del_init(&req->r_osd_item);
1300 maybe_move_osd_to_lru(osdc, req->r_osd);
1301 if (list_empty(&req->r_linger_osd_item))
1302 req->r_osd = NULL;
1303 }
1304
1305 list_del_init(&req->r_req_lru_item);
1306 ceph_osdc_put_request(req);
1307
1308 if (osdc->num_requests == 0) {
1309 dout(" no requests, canceling timeout\n");
1310 __cancel_osd_timeout(osdc);
1311 }
1312}
1313
1314/*
1315 * Cancel a previously queued request message
1316 */
1317static void __cancel_request(struct ceph_osd_request *req)
1318{
1319 if (req->r_sent && req->r_osd) {
1320 ceph_msg_revoke(req->r_request);
1321 req->r_sent = 0;
1322 }
1323}
1324
1325static void __register_linger_request(struct ceph_osd_client *osdc,
1326 struct ceph_osd_request *req)
1327{
1328 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1329 WARN_ON(!req->r_linger);
1330
1331 ceph_osdc_get_request(req);
1332 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1333 if (req->r_osd)
1334 list_add_tail(&req->r_linger_osd_item,
1335 &req->r_osd->o_linger_requests);
1336}
1337
1338static void __unregister_linger_request(struct ceph_osd_client *osdc,
1339 struct ceph_osd_request *req)
1340{
1341 WARN_ON(!req->r_linger);
1342
1343 if (list_empty(&req->r_linger_item)) {
1344 dout("%s %p tid %llu not registered\n", __func__, req,
1345 req->r_tid);
1346 return;
1347 }
1348
1349 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1350 list_del_init(&req->r_linger_item);
1351
1352 if (req->r_osd) {
1353 list_del_init(&req->r_linger_osd_item);
1354 maybe_move_osd_to_lru(osdc, req->r_osd);
1355 if (list_empty(&req->r_osd_item))
1356 req->r_osd = NULL;
1357 }
1358 ceph_osdc_put_request(req);
1359}
1360
1361void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1362 struct ceph_osd_request *req)
1363{
1364 if (!req->r_linger) {
1365 dout("set_request_linger %p\n", req);
1366 req->r_linger = 1;
1367 }
1368}
1369EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1370
1371/*
1372 * Returns whether a request should be blocked from being sent
1373 * based on the current osdmap and osd_client settings.
1374 *
1375 * Caller should hold map_sem for read.
1376 */
1377static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1378 struct ceph_osd_request *req)
1379{
1380 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1381 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1382 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1383 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1384 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1385}
1386
1387/*
1388 * Calculate mapping of a request to a PG. Takes tiering into account.
1389 */
1390static int __calc_request_pg(struct ceph_osdmap *osdmap,
1391 struct ceph_osd_request *req,
1392 struct ceph_pg *pg_out)
1393{
1394 bool need_check_tiering;
1395
1396 need_check_tiering = false;
1397 if (req->r_target_oloc.pool == -1) {
1398 req->r_target_oloc = req->r_base_oloc; /* struct */
1399 need_check_tiering = true;
1400 }
1401 if (req->r_target_oid.name_len == 0) {
1402 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1403 need_check_tiering = true;
1404 }
1405
1406 if (need_check_tiering &&
1407 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1408 struct ceph_pg_pool_info *pi;
1409
1410 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1411 if (pi) {
1412 if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1413 pi->read_tier >= 0)
1414 req->r_target_oloc.pool = pi->read_tier;
1415 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1416 pi->write_tier >= 0)
1417 req->r_target_oloc.pool = pi->write_tier;
1418 }
1419 /* !pi is caught in ceph_oloc_oid_to_pg() */
1420 }
1421
1422 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1423 &req->r_target_oid, pg_out);
1424}
1425
1426static void __enqueue_request(struct ceph_osd_request *req)
1427{
1428 struct ceph_osd_client *osdc = req->r_osdc;
1429
1430 dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1431 req->r_osd ? req->r_osd->o_osd : -1);
1432
1433 if (req->r_osd) {
1434 __remove_osd_from_lru(req->r_osd);
1435 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1436 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1437 } else {
1438 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1439 }
1440}
1441
1442/*
1443 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1444 * (as needed), and set the request r_osd appropriately. If there is
1445 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1446 * (unsent, homeless) or leave on in-flight lru.
1447 *
1448 * Return 0 if unchanged, 1 if changed, or negative on error.
1449 *
1450 * Caller should hold map_sem for read and request_mutex.
1451 */
1452static int __map_request(struct ceph_osd_client *osdc,
1453 struct ceph_osd_request *req, int force_resend)
1454{
1455 struct ceph_pg pgid;
1456 int acting[CEPH_PG_MAX_SIZE];
1457 int num, o;
1458 int err;
1459 bool was_paused;
1460
1461 dout("map_request %p tid %lld\n", req, req->r_tid);
1462
1463 err = __calc_request_pg(osdc->osdmap, req, &pgid);
1464 if (err) {
1465 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1466 return err;
1467 }
1468 req->r_pgid = pgid;
1469
1470 num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1471 if (num < 0)
1472 num = 0;
1473
1474 was_paused = req->r_paused;
1475 req->r_paused = __req_should_be_paused(osdc, req);
1476 if (was_paused && !req->r_paused)
1477 force_resend = 1;
1478
1479 if ((!force_resend &&
1480 req->r_osd && req->r_osd->o_osd == o &&
1481 req->r_sent >= req->r_osd->o_incarnation &&
1482 req->r_num_pg_osds == num &&
1483 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1484 (req->r_osd == NULL && o == -1) ||
1485 req->r_paused)
1486 return 0; /* no change */
1487
1488 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1489 req->r_tid, pgid.pool, pgid.seed, o,
1490 req->r_osd ? req->r_osd->o_osd : -1);
1491
1492 /* record full pg acting set */
1493 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1494 req->r_num_pg_osds = num;
1495
1496 if (req->r_osd) {
1497 __cancel_request(req);
1498 list_del_init(&req->r_osd_item);
1499 list_del_init(&req->r_linger_osd_item);
1500 req->r_osd = NULL;
1501 }
1502
1503 req->r_osd = __lookup_osd(osdc, o);
1504 if (!req->r_osd && o >= 0) {
1505 err = -ENOMEM;
1506 req->r_osd = create_osd(osdc, o);
1507 if (!req->r_osd) {
1508 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1509 goto out;
1510 }
1511
1512 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1513 __insert_osd(osdc, req->r_osd);
1514
1515 ceph_con_open(&req->r_osd->o_con,
1516 CEPH_ENTITY_TYPE_OSD, o,
1517 &osdc->osdmap->osd_addr[o]);
1518 }
1519
1520 __enqueue_request(req);
1521 err = 1; /* osd or pg changed */
1522
1523out:
1524 return err;
1525}
1526
1527/*
1528 * caller should hold map_sem (for read) and request_mutex
1529 */
1530static void __send_request(struct ceph_osd_client *osdc,
1531 struct ceph_osd_request *req)
1532{
1533 void *p;
1534
1535 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1536 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1537 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1538
1539 /* fill in message content that changes each time we send it */
1540 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1541 put_unaligned_le32(req->r_flags, req->r_request_flags);
1542 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1543 p = req->r_request_pgid;
1544 ceph_encode_64(&p, req->r_pgid.pool);
1545 ceph_encode_32(&p, req->r_pgid.seed);
1546 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1547 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1548 sizeof(req->r_reassert_version));
1549
1550 req->r_stamp = jiffies;
1551 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1552
1553 ceph_msg_get(req->r_request); /* send consumes a ref */
1554
1555 req->r_sent = req->r_osd->o_incarnation;
1556
1557 ceph_con_send(&req->r_osd->o_con, req->r_request);
1558}
1559
1560/*
1561 * Send any requests in the queue (req_unsent).
1562 */
1563static void __send_queued(struct ceph_osd_client *osdc)
1564{
1565 struct ceph_osd_request *req, *tmp;
1566
1567 dout("__send_queued\n");
1568 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1569 __send_request(osdc, req);
1570}
1571
1572/*
1573 * Caller should hold map_sem for read and request_mutex.
1574 */
1575static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1576 struct ceph_osd_request *req,
1577 bool nofail)
1578{
1579 int rc;
1580
1581 __register_request(osdc, req);
1582 req->r_sent = 0;
1583 req->r_got_reply = 0;
1584 rc = __map_request(osdc, req, 0);
1585 if (rc < 0) {
1586 if (nofail) {
1587 dout("osdc_start_request failed map, "
1588 " will retry %lld\n", req->r_tid);
1589 rc = 0;
1590 } else {
1591 __unregister_request(osdc, req);
1592 }
1593 return rc;
1594 }
1595
1596 if (req->r_osd == NULL) {
1597 dout("send_request %p no up osds in pg\n", req);
1598 ceph_monc_request_next_osdmap(&osdc->client->monc);
1599 } else {
1600 __send_queued(osdc);
1601 }
1602
1603 return 0;
1604}
1605
1606/*
1607 * Timeout callback, called every N seconds when 1 or more osd
1608 * requests has been active for more than N seconds. When this
1609 * happens, we ping all OSDs with requests who have timed out to
1610 * ensure any communications channel reset is detected. Reset the
1611 * request timeouts another N seconds in the future as we go.
1612 * Reschedule the timeout event another N seconds in future (unless
1613 * there are no open requests).
1614 */
1615static void handle_timeout(struct work_struct *work)
1616{
1617 struct ceph_osd_client *osdc =
1618 container_of(work, struct ceph_osd_client, timeout_work.work);
1619 struct ceph_options *opts = osdc->client->options;
1620 struct ceph_osd_request *req;
1621 struct ceph_osd *osd;
1622 struct list_head slow_osds;
1623 dout("timeout\n");
1624 down_read(&osdc->map_sem);
1625
1626 ceph_monc_request_next_osdmap(&osdc->client->monc);
1627
1628 mutex_lock(&osdc->request_mutex);
1629
1630 /*
1631 * ping osds that are a bit slow. this ensures that if there
1632 * is a break in the TCP connection we will notice, and reopen
1633 * a connection with that osd (from the fault callback).
1634 */
1635 INIT_LIST_HEAD(&slow_osds);
1636 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1637 if (time_before(jiffies,
1638 req->r_stamp + opts->osd_keepalive_timeout))
1639 break;
1640
1641 osd = req->r_osd;
1642 BUG_ON(!osd);
1643 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1644 req->r_tid, osd->o_osd);
1645 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1646 }
1647 while (!list_empty(&slow_osds)) {
1648 osd = list_entry(slow_osds.next, struct ceph_osd,
1649 o_keepalive_item);
1650 list_del_init(&osd->o_keepalive_item);
1651 ceph_con_keepalive(&osd->o_con);
1652 }
1653
1654 __schedule_osd_timeout(osdc);
1655 __send_queued(osdc);
1656 mutex_unlock(&osdc->request_mutex);
1657 up_read(&osdc->map_sem);
1658}
1659
1660static void handle_osds_timeout(struct work_struct *work)
1661{
1662 struct ceph_osd_client *osdc =
1663 container_of(work, struct ceph_osd_client,
1664 osds_timeout_work.work);
1665 unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
1666
1667 dout("osds timeout\n");
1668 down_read(&osdc->map_sem);
1669 remove_old_osds(osdc);
1670 up_read(&osdc->map_sem);
1671
1672 schedule_delayed_work(&osdc->osds_timeout_work,
1673 round_jiffies_relative(delay));
1674}
1675
1676static int ceph_oloc_decode(void **p, void *end,
1677 struct ceph_object_locator *oloc)
1678{
1679 u8 struct_v, struct_cv;
1680 u32 len;
1681 void *struct_end;
1682 int ret = 0;
1683
1684 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1685 struct_v = ceph_decode_8(p);
1686 struct_cv = ceph_decode_8(p);
1687 if (struct_v < 3) {
1688 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1689 struct_v, struct_cv);
1690 goto e_inval;
1691 }
1692 if (struct_cv > 6) {
1693 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1694 struct_v, struct_cv);
1695 goto e_inval;
1696 }
1697 len = ceph_decode_32(p);
1698 ceph_decode_need(p, end, len, e_inval);
1699 struct_end = *p + len;
1700
1701 oloc->pool = ceph_decode_64(p);
1702 *p += 4; /* skip preferred */
1703
1704 len = ceph_decode_32(p);
1705 if (len > 0) {
1706 pr_warn("ceph_object_locator::key is set\n");
1707 goto e_inval;
1708 }
1709
1710 if (struct_v >= 5) {
1711 len = ceph_decode_32(p);
1712 if (len > 0) {
1713 pr_warn("ceph_object_locator::nspace is set\n");
1714 goto e_inval;
1715 }
1716 }
1717
1718 if (struct_v >= 6) {
1719 s64 hash = ceph_decode_64(p);
1720 if (hash != -1) {
1721 pr_warn("ceph_object_locator::hash is set\n");
1722 goto e_inval;
1723 }
1724 }
1725
1726 /* skip the rest */
1727 *p = struct_end;
1728out:
1729 return ret;
1730
1731e_inval:
1732 ret = -EINVAL;
1733 goto out;
1734}
1735
1736static int ceph_redirect_decode(void **p, void *end,
1737 struct ceph_request_redirect *redir)
1738{
1739 u8 struct_v, struct_cv;
1740 u32 len;
1741 void *struct_end;
1742 int ret;
1743
1744 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1745 struct_v = ceph_decode_8(p);
1746 struct_cv = ceph_decode_8(p);
1747 if (struct_cv > 1) {
1748 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1749 struct_v, struct_cv);
1750 goto e_inval;
1751 }
1752 len = ceph_decode_32(p);
1753 ceph_decode_need(p, end, len, e_inval);
1754 struct_end = *p + len;
1755
1756 ret = ceph_oloc_decode(p, end, &redir->oloc);
1757 if (ret)
1758 goto out;
1759
1760 len = ceph_decode_32(p);
1761 if (len > 0) {
1762 pr_warn("ceph_request_redirect::object_name is set\n");
1763 goto e_inval;
1764 }
1765
1766 len = ceph_decode_32(p);
1767 *p += len; /* skip osd_instructions */
1768
1769 /* skip the rest */
1770 *p = struct_end;
1771out:
1772 return ret;
1773
1774e_inval:
1775 ret = -EINVAL;
1776 goto out;
1777}
1778
1779static void complete_request(struct ceph_osd_request *req)
1780{
1781 complete_all(&req->r_safe_completion); /* fsync waiter */
1782}
1783
1784/*
1785 * handle osd op reply. either call the callback if it is specified,
1786 * or do the completion to wake up the waiting thread.
1787 */
1788static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1789{
1790 void *p, *end;
1791 struct ceph_osd_request *req;
1792 struct ceph_request_redirect redir;
1793 u64 tid;
1794 int object_len;
1795 unsigned int numops;
1796 int payload_len, flags;
1797 s32 result;
1798 s32 retry_attempt;
1799 struct ceph_pg pg;
1800 int err;
1801 u32 reassert_epoch;
1802 u64 reassert_version;
1803 u32 osdmap_epoch;
1804 int already_completed;
1805 u32 bytes;
1806 u8 decode_redir;
1807 unsigned int i;
1808
1809 tid = le64_to_cpu(msg->hdr.tid);
1810 dout("handle_reply %p tid %llu\n", msg, tid);
1811
1812 p = msg->front.iov_base;
1813 end = p + msg->front.iov_len;
1814
1815 ceph_decode_need(&p, end, 4, bad);
1816 object_len = ceph_decode_32(&p);
1817 ceph_decode_need(&p, end, object_len, bad);
1818 p += object_len;
1819
1820 err = ceph_decode_pgid(&p, end, &pg);
1821 if (err)
1822 goto bad;
1823
1824 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1825 flags = ceph_decode_64(&p);
1826 result = ceph_decode_32(&p);
1827 reassert_epoch = ceph_decode_32(&p);
1828 reassert_version = ceph_decode_64(&p);
1829 osdmap_epoch = ceph_decode_32(&p);
1830
1831 /* lookup */
1832 down_read(&osdc->map_sem);
1833 mutex_lock(&osdc->request_mutex);
1834 req = __lookup_request(osdc, tid);
1835 if (req == NULL) {
1836 dout("handle_reply tid %llu dne\n", tid);
1837 goto bad_mutex;
1838 }
1839 ceph_osdc_get_request(req);
1840
1841 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1842 req, result);
1843
1844 ceph_decode_need(&p, end, 4, bad_put);
1845 numops = ceph_decode_32(&p);
1846 if (numops > CEPH_OSD_MAX_OPS)
1847 goto bad_put;
1848 if (numops != req->r_num_ops)
1849 goto bad_put;
1850 payload_len = 0;
1851 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1852 for (i = 0; i < numops; i++) {
1853 struct ceph_osd_op *op = p;
1854 int len;
1855
1856 len = le32_to_cpu(op->payload_len);
1857 req->r_ops[i].outdata_len = len;
1858 dout(" op %d has %d bytes\n", i, len);
1859 payload_len += len;
1860 p += sizeof(*op);
1861 }
1862 bytes = le32_to_cpu(msg->hdr.data_len);
1863 if (payload_len != bytes) {
1864 pr_warn("sum of op payload lens %d != data_len %d\n",
1865 payload_len, bytes);
1866 goto bad_put;
1867 }
1868
1869 ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1870 retry_attempt = ceph_decode_32(&p);
1871 for (i = 0; i < numops; i++)
1872 req->r_ops[i].rval = ceph_decode_32(&p);
1873
1874 if (le16_to_cpu(msg->hdr.version) >= 6) {
1875 p += 8 + 4; /* skip replay_version */
1876 p += 8; /* skip user_version */
1877
1878 if (le16_to_cpu(msg->hdr.version) >= 7)
1879 ceph_decode_8_safe(&p, end, decode_redir, bad_put);
1880 else
1881 decode_redir = 1;
1882 } else {
1883 decode_redir = 0;
1884 }
1885
1886 if (decode_redir) {
1887 err = ceph_redirect_decode(&p, end, &redir);
1888 if (err)
1889 goto bad_put;
1890 } else {
1891 redir.oloc.pool = -1;
1892 }
1893
1894 if (redir.oloc.pool != -1) {
1895 dout("redirect pool %lld\n", redir.oloc.pool);
1896
1897 __unregister_request(osdc, req);
1898
1899 req->r_target_oloc = redir.oloc; /* struct */
1900
1901 /*
1902 * Start redirect requests with nofail=true. If
1903 * mapping fails, request will end up on the notarget
1904 * list, waiting for the new osdmap (which can take
1905 * a while), even though the original request mapped
1906 * successfully. In the future we might want to follow
1907 * original request's nofail setting here.
1908 */
1909 err = __ceph_osdc_start_request(osdc, req, true);
1910 BUG_ON(err);
1911
1912 goto out_unlock;
1913 }
1914
1915 already_completed = req->r_got_reply;
1916 if (!req->r_got_reply) {
1917 req->r_result = result;
1918 dout("handle_reply result %d bytes %d\n", req->r_result,
1919 bytes);
1920 if (req->r_result == 0)
1921 req->r_result = bytes;
1922
1923 /* in case this is a write and we need to replay, */
1924 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1925 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1926
1927 req->r_got_reply = 1;
1928 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1929 dout("handle_reply tid %llu dup ack\n", tid);
1930 goto out_unlock;
1931 }
1932
1933 dout("handle_reply tid %llu flags %d\n", tid, flags);
1934
1935 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1936 __register_linger_request(osdc, req);
1937
1938 /* either this is a read, or we got the safe response */
1939 if (result < 0 ||
1940 (flags & CEPH_OSD_FLAG_ONDISK) ||
1941 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1942 __unregister_request(osdc, req);
1943
1944 mutex_unlock(&osdc->request_mutex);
1945 up_read(&osdc->map_sem);
1946
1947 if (!already_completed) {
1948 if (req->r_unsafe_callback &&
1949 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1950 req->r_unsafe_callback(req, true);
1951 if (req->r_callback)
1952 req->r_callback(req, msg);
1953 else
1954 complete_all(&req->r_completion);
1955 }
1956
1957 if (flags & CEPH_OSD_FLAG_ONDISK) {
1958 if (req->r_unsafe_callback && already_completed)
1959 req->r_unsafe_callback(req, false);
1960 complete_request(req);
1961 }
1962
1963out:
1964 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1965 ceph_osdc_put_request(req);
1966 return;
1967out_unlock:
1968 mutex_unlock(&osdc->request_mutex);
1969 up_read(&osdc->map_sem);
1970 goto out;
1971
1972bad_put:
1973 req->r_result = -EIO;
1974 __unregister_request(osdc, req);
1975 if (req->r_callback)
1976 req->r_callback(req, msg);
1977 else
1978 complete_all(&req->r_completion);
1979 complete_request(req);
1980 ceph_osdc_put_request(req);
1981bad_mutex:
1982 mutex_unlock(&osdc->request_mutex);
1983 up_read(&osdc->map_sem);
1984bad:
1985 pr_err("corrupt osd_op_reply got %d %d\n",
1986 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1987 ceph_msg_dump(msg);
1988}
1989
1990static void reset_changed_osds(struct ceph_osd_client *osdc)
1991{
1992 struct rb_node *p, *n;
1993
1994 dout("%s %p\n", __func__, osdc);
1995 for (p = rb_first(&osdc->osds); p; p = n) {
1996 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1997
1998 n = rb_next(p);
1999 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
2000 memcmp(&osd->o_con.peer_addr,
2001 ceph_osd_addr(osdc->osdmap,
2002 osd->o_osd),
2003 sizeof(struct ceph_entity_addr)) != 0)
2004 __reset_osd(osdc, osd);
2005 }
2006}
2007
2008/*
2009 * Requeue requests whose mapping to an OSD has changed. If requests map to
2010 * no osd, request a new map.
2011 *
2012 * Caller should hold map_sem for read.
2013 */
2014static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
2015 bool force_resend_writes)
2016{
2017 struct ceph_osd_request *req, *nreq;
2018 struct rb_node *p;
2019 int needmap = 0;
2020 int err;
2021 bool force_resend_req;
2022
2023 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
2024 force_resend_writes ? " (force resend writes)" : "");
2025 mutex_lock(&osdc->request_mutex);
2026 for (p = rb_first(&osdc->requests); p; ) {
2027 req = rb_entry(p, struct ceph_osd_request, r_node);
2028 p = rb_next(p);
2029
2030 /*
2031 * For linger requests that have not yet been
2032 * registered, move them to the linger list; they'll
2033 * be sent to the osd in the loop below. Unregister
2034 * the request before re-registering it as a linger
2035 * request to ensure the __map_request() below
2036 * will decide it needs to be sent.
2037 */
2038 if (req->r_linger && list_empty(&req->r_linger_item)) {
2039 dout("%p tid %llu restart on osd%d\n",
2040 req, req->r_tid,
2041 req->r_osd ? req->r_osd->o_osd : -1);
2042 ceph_osdc_get_request(req);
2043 __unregister_request(osdc, req);
2044 __register_linger_request(osdc, req);
2045 ceph_osdc_put_request(req);
2046 continue;
2047 }
2048
2049 force_resend_req = force_resend ||
2050 (force_resend_writes &&
2051 req->r_flags & CEPH_OSD_FLAG_WRITE);
2052 err = __map_request(osdc, req, force_resend_req);
2053 if (err < 0)
2054 continue; /* error */
2055 if (req->r_osd == NULL) {
2056 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2057 needmap++; /* request a newer map */
2058 } else if (err > 0) {
2059 if (!req->r_linger) {
2060 dout("%p tid %llu requeued on osd%d\n", req,
2061 req->r_tid,
2062 req->r_osd ? req->r_osd->o_osd : -1);
2063 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2064 }
2065 }
2066 }
2067
2068 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2069 r_linger_item) {
2070 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2071
2072 err = __map_request(osdc, req,
2073 force_resend || force_resend_writes);
2074 dout("__map_request returned %d\n", err);
2075 if (err < 0)
2076 continue; /* hrm! */
2077 if (req->r_osd == NULL || err > 0) {
2078 if (req->r_osd == NULL) {
2079 dout("lingering %p tid %llu maps to no osd\n",
2080 req, req->r_tid);
2081 /*
2082 * A homeless lingering request makes
2083 * no sense, as it's job is to keep
2084 * a particular OSD connection open.
2085 * Request a newer map and kick the
2086 * request, knowing that it won't be
2087 * resent until we actually get a map
2088 * that can tell us where to send it.
2089 */
2090 needmap++;
2091 }
2092
2093 dout("kicking lingering %p tid %llu osd%d\n", req,
2094 req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2095 __register_request(osdc, req);
2096 __unregister_linger_request(osdc, req);
2097 }
2098 }
2099 reset_changed_osds(osdc);
2100 mutex_unlock(&osdc->request_mutex);
2101
2102 if (needmap) {
2103 dout("%d requests for down osds, need new map\n", needmap);
2104 ceph_monc_request_next_osdmap(&osdc->client->monc);
2105 }
2106}
2107
2108
2109/*
2110 * Process updated osd map.
2111 *
2112 * The message contains any number of incremental and full maps, normally
2113 * indicating some sort of topology change in the cluster. Kick requests
2114 * off to different OSDs as needed.
2115 */
2116void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2117{
2118 void *p, *end, *next;
2119 u32 nr_maps, maplen;
2120 u32 epoch;
2121 struct ceph_osdmap *newmap = NULL, *oldmap;
2122 int err;
2123 struct ceph_fsid fsid;
2124 bool was_full;
2125
2126 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2127 p = msg->front.iov_base;
2128 end = p + msg->front.iov_len;
2129
2130 /* verify fsid */
2131 ceph_decode_need(&p, end, sizeof(fsid), bad);
2132 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2133 if (ceph_check_fsid(osdc->client, &fsid) < 0)
2134 return;
2135
2136 down_write(&osdc->map_sem);
2137
2138 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2139
2140 /* incremental maps */
2141 ceph_decode_32_safe(&p, end, nr_maps, bad);
2142 dout(" %d inc maps\n", nr_maps);
2143 while (nr_maps > 0) {
2144 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2145 epoch = ceph_decode_32(&p);
2146 maplen = ceph_decode_32(&p);
2147 ceph_decode_need(&p, end, maplen, bad);
2148 next = p + maplen;
2149 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2150 dout("applying incremental map %u len %d\n",
2151 epoch, maplen);
2152 newmap = osdmap_apply_incremental(&p, next,
2153 osdc->osdmap,
2154 &osdc->client->msgr);
2155 if (IS_ERR(newmap)) {
2156 err = PTR_ERR(newmap);
2157 goto bad;
2158 }
2159 BUG_ON(!newmap);
2160 if (newmap != osdc->osdmap) {
2161 ceph_osdmap_destroy(osdc->osdmap);
2162 osdc->osdmap = newmap;
2163 }
2164 was_full = was_full ||
2165 ceph_osdmap_flag(osdc->osdmap,
2166 CEPH_OSDMAP_FULL);
2167 kick_requests(osdc, 0, was_full);
2168 } else {
2169 dout("ignoring incremental map %u len %d\n",
2170 epoch, maplen);
2171 }
2172 p = next;
2173 nr_maps--;
2174 }
2175 if (newmap)
2176 goto done;
2177
2178 /* full maps */
2179 ceph_decode_32_safe(&p, end, nr_maps, bad);
2180 dout(" %d full maps\n", nr_maps);
2181 while (nr_maps) {
2182 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2183 epoch = ceph_decode_32(&p);
2184 maplen = ceph_decode_32(&p);
2185 ceph_decode_need(&p, end, maplen, bad);
2186 if (nr_maps > 1) {
2187 dout("skipping non-latest full map %u len %d\n",
2188 epoch, maplen);
2189 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2190 dout("skipping full map %u len %d, "
2191 "older than our %u\n", epoch, maplen,
2192 osdc->osdmap->epoch);
2193 } else {
2194 int skipped_map = 0;
2195
2196 dout("taking full map %u len %d\n", epoch, maplen);
2197 newmap = ceph_osdmap_decode(&p, p+maplen);
2198 if (IS_ERR(newmap)) {
2199 err = PTR_ERR(newmap);
2200 goto bad;
2201 }
2202 BUG_ON(!newmap);
2203 oldmap = osdc->osdmap;
2204 osdc->osdmap = newmap;
2205 if (oldmap) {
2206 if (oldmap->epoch + 1 < newmap->epoch)
2207 skipped_map = 1;
2208 ceph_osdmap_destroy(oldmap);
2209 }
2210 was_full = was_full ||
2211 ceph_osdmap_flag(osdc->osdmap,
2212 CEPH_OSDMAP_FULL);
2213 kick_requests(osdc, skipped_map, was_full);
2214 }
2215 p += maplen;
2216 nr_maps--;
2217 }
2218
2219 if (!osdc->osdmap)
2220 goto bad;
2221done:
2222 downgrade_write(&osdc->map_sem);
2223 ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
2224 osdc->osdmap->epoch);
2225
2226 /*
2227 * subscribe to subsequent osdmap updates if full to ensure
2228 * we find out when we are no longer full and stop returning
2229 * ENOSPC.
2230 */
2231 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2232 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2233 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2234 ceph_monc_request_next_osdmap(&osdc->client->monc);
2235
2236 mutex_lock(&osdc->request_mutex);
2237 __send_queued(osdc);
2238 mutex_unlock(&osdc->request_mutex);
2239 up_read(&osdc->map_sem);
2240 wake_up_all(&osdc->client->auth_wq);
2241 return;
2242
2243bad:
2244 pr_err("osdc handle_map corrupt msg\n");
2245 ceph_msg_dump(msg);
2246 up_write(&osdc->map_sem);
2247}
2248
2249/*
2250 * watch/notify callback event infrastructure
2251 *
2252 * These callbacks are used both for watch and notify operations.
2253 */
2254static void __release_event(struct kref *kref)
2255{
2256 struct ceph_osd_event *event =
2257 container_of(kref, struct ceph_osd_event, kref);
2258
2259 dout("__release_event %p\n", event);
2260 kfree(event);
2261}
2262
2263static void get_event(struct ceph_osd_event *event)
2264{
2265 kref_get(&event->kref);
2266}
2267
2268void ceph_osdc_put_event(struct ceph_osd_event *event)
2269{
2270 kref_put(&event->kref, __release_event);
2271}
2272EXPORT_SYMBOL(ceph_osdc_put_event);
2273
2274static void __insert_event(struct ceph_osd_client *osdc,
2275 struct ceph_osd_event *new)
2276{
2277 struct rb_node **p = &osdc->event_tree.rb_node;
2278 struct rb_node *parent = NULL;
2279 struct ceph_osd_event *event = NULL;
2280
2281 while (*p) {
2282 parent = *p;
2283 event = rb_entry(parent, struct ceph_osd_event, node);
2284 if (new->cookie < event->cookie)
2285 p = &(*p)->rb_left;
2286 else if (new->cookie > event->cookie)
2287 p = &(*p)->rb_right;
2288 else
2289 BUG();
2290 }
2291
2292 rb_link_node(&new->node, parent, p);
2293 rb_insert_color(&new->node, &osdc->event_tree);
2294}
2295
2296static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2297 u64 cookie)
2298{
2299 struct rb_node **p = &osdc->event_tree.rb_node;
2300 struct rb_node *parent = NULL;
2301 struct ceph_osd_event *event = NULL;
2302
2303 while (*p) {
2304 parent = *p;
2305 event = rb_entry(parent, struct ceph_osd_event, node);
2306 if (cookie < event->cookie)
2307 p = &(*p)->rb_left;
2308 else if (cookie > event->cookie)
2309 p = &(*p)->rb_right;
2310 else
2311 return event;
2312 }
2313 return NULL;
2314}
2315
2316static void __remove_event(struct ceph_osd_event *event)
2317{
2318 struct ceph_osd_client *osdc = event->osdc;
2319
2320 if (!RB_EMPTY_NODE(&event->node)) {
2321 dout("__remove_event removed %p\n", event);
2322 rb_erase(&event->node, &osdc->event_tree);
2323 ceph_osdc_put_event(event);
2324 } else {
2325 dout("__remove_event didn't remove %p\n", event);
2326 }
2327}
2328
2329int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2330 void (*event_cb)(u64, u64, u8, void *),
2331 void *data, struct ceph_osd_event **pevent)
2332{
2333 struct ceph_osd_event *event;
2334
2335 event = kmalloc(sizeof(*event), GFP_NOIO);
2336 if (!event)
2337 return -ENOMEM;
2338
2339 dout("create_event %p\n", event);
2340 event->cb = event_cb;
2341 event->one_shot = 0;
2342 event->data = data;
2343 event->osdc = osdc;
2344 INIT_LIST_HEAD(&event->osd_node);
2345 RB_CLEAR_NODE(&event->node);
2346 kref_init(&event->kref); /* one ref for us */
2347 kref_get(&event->kref); /* one ref for the caller */
2348
2349 spin_lock(&osdc->event_lock);
2350 event->cookie = ++osdc->event_count;
2351 __insert_event(osdc, event);
2352 spin_unlock(&osdc->event_lock);
2353
2354 *pevent = event;
2355 return 0;
2356}
2357EXPORT_SYMBOL(ceph_osdc_create_event);
2358
2359void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2360{
2361 struct ceph_osd_client *osdc = event->osdc;
2362
2363 dout("cancel_event %p\n", event);
2364 spin_lock(&osdc->event_lock);
2365 __remove_event(event);
2366 spin_unlock(&osdc->event_lock);
2367 ceph_osdc_put_event(event); /* caller's */
2368}
2369EXPORT_SYMBOL(ceph_osdc_cancel_event);
2370
2371
2372static void do_event_work(struct work_struct *work)
2373{
2374 struct ceph_osd_event_work *event_work =
2375 container_of(work, struct ceph_osd_event_work, work);
2376 struct ceph_osd_event *event = event_work->event;
2377 u64 ver = event_work->ver;
2378 u64 notify_id = event_work->notify_id;
2379 u8 opcode = event_work->opcode;
2380
2381 dout("do_event_work completing %p\n", event);
2382 event->cb(ver, notify_id, opcode, event->data);
2383 dout("do_event_work completed %p\n", event);
2384 ceph_osdc_put_event(event);
2385 kfree(event_work);
2386}
2387
2388
2389/*
2390 * Process osd watch notifications
2391 */
2392static void handle_watch_notify(struct ceph_osd_client *osdc,
2393 struct ceph_msg *msg)
2394{
2395 void *p, *end;
2396 u8 proto_ver;
2397 u64 cookie, ver, notify_id;
2398 u8 opcode;
2399 struct ceph_osd_event *event;
2400 struct ceph_osd_event_work *event_work;
2401
2402 p = msg->front.iov_base;
2403 end = p + msg->front.iov_len;
2404
2405 ceph_decode_8_safe(&p, end, proto_ver, bad);
2406 ceph_decode_8_safe(&p, end, opcode, bad);
2407 ceph_decode_64_safe(&p, end, cookie, bad);
2408 ceph_decode_64_safe(&p, end, ver, bad);
2409 ceph_decode_64_safe(&p, end, notify_id, bad);
2410
2411 spin_lock(&osdc->event_lock);
2412 event = __find_event(osdc, cookie);
2413 if (event) {
2414 BUG_ON(event->one_shot);
2415 get_event(event);
2416 }
2417 spin_unlock(&osdc->event_lock);
2418 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2419 cookie, ver, event);
2420 if (event) {
2421 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2422 if (!event_work) {
2423 pr_err("couldn't allocate event_work\n");
2424 ceph_osdc_put_event(event);
2425 return;
2426 }
2427 INIT_WORK(&event_work->work, do_event_work);
2428 event_work->event = event;
2429 event_work->ver = ver;
2430 event_work->notify_id = notify_id;
2431 event_work->opcode = opcode;
2432
2433 queue_work(osdc->notify_wq, &event_work->work);
2434 }
2435
2436 return;
2437
2438bad:
2439 pr_err("osdc handle_watch_notify corrupt msg\n");
2440}
2441
2442/*
2443 * build new request AND message
2444 *
2445 */
2446void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2447 struct ceph_snap_context *snapc, u64 snap_id,
2448 struct timespec *mtime)
2449{
2450 struct ceph_msg *msg = req->r_request;
2451 void *p;
2452 size_t msg_size;
2453 int flags = req->r_flags;
2454 u64 data_len;
2455 unsigned int i;
2456
2457 req->r_snapid = snap_id;
2458 req->r_snapc = ceph_get_snap_context(snapc);
2459
2460 /* encode request */
2461 msg->hdr.version = cpu_to_le16(4);
2462
2463 p = msg->front.iov_base;
2464 ceph_encode_32(&p, 1); /* client_inc is always 1 */
2465 req->r_request_osdmap_epoch = p;
2466 p += 4;
2467 req->r_request_flags = p;
2468 p += 4;
2469 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2470 ceph_encode_timespec(p, mtime);
2471 p += sizeof(struct ceph_timespec);
2472 req->r_request_reassert_version = p;
2473 p += sizeof(struct ceph_eversion); /* will get filled in */
2474
2475 /* oloc */
2476 ceph_encode_8(&p, 4);
2477 ceph_encode_8(&p, 4);
2478 ceph_encode_32(&p, 8 + 4 + 4);
2479 req->r_request_pool = p;
2480 p += 8;
2481 ceph_encode_32(&p, -1); /* preferred */
2482 ceph_encode_32(&p, 0); /* key len */
2483
2484 ceph_encode_8(&p, 1);
2485 req->r_request_pgid = p;
2486 p += 8 + 4;
2487 ceph_encode_32(&p, -1); /* preferred */
2488
2489 /* oid */
2490 ceph_encode_32(&p, req->r_base_oid.name_len);
2491 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2492 dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2493 req->r_base_oid.name, req->r_base_oid.name_len);
2494 p += req->r_base_oid.name_len;
2495
2496 /* ops--can imply data */
2497 ceph_encode_16(&p, (u16)req->r_num_ops);
2498 data_len = 0;
2499 for (i = 0; i < req->r_num_ops; i++) {
2500 data_len += osd_req_encode_op(req, p, i);
2501 p += sizeof(struct ceph_osd_op);
2502 }
2503
2504 /* snaps */
2505 ceph_encode_64(&p, req->r_snapid);
2506 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2507 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2508 if (req->r_snapc) {
2509 for (i = 0; i < snapc->num_snaps; i++) {
2510 ceph_encode_64(&p, req->r_snapc->snaps[i]);
2511 }
2512 }
2513
2514 req->r_request_attempts = p;
2515 p += 4;
2516
2517 /* data */
2518 if (flags & CEPH_OSD_FLAG_WRITE) {
2519 u16 data_off;
2520
2521 /*
2522 * The header "data_off" is a hint to the receiver
2523 * allowing it to align received data into its
2524 * buffers such that there's no need to re-copy
2525 * it before writing it to disk (direct I/O).
2526 */
2527 data_off = (u16) (off & 0xffff);
2528 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2529 }
2530 req->r_request->hdr.data_len = cpu_to_le32(data_len);
2531
2532 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2533 msg_size = p - msg->front.iov_base;
2534 msg->front.iov_len = msg_size;
2535 msg->hdr.front_len = cpu_to_le32(msg_size);
2536
2537 dout("build_request msg_size was %d\n", (int)msg_size);
2538}
2539EXPORT_SYMBOL(ceph_osdc_build_request);
2540
2541/*
2542 * Register request, send initial attempt.
2543 */
2544int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2545 struct ceph_osd_request *req,
2546 bool nofail)
2547{
2548 int rc;
2549
2550 down_read(&osdc->map_sem);
2551 mutex_lock(&osdc->request_mutex);
2552
2553 rc = __ceph_osdc_start_request(osdc, req, nofail);
2554
2555 mutex_unlock(&osdc->request_mutex);
2556 up_read(&osdc->map_sem);
2557
2558 return rc;
2559}
2560EXPORT_SYMBOL(ceph_osdc_start_request);
2561
2562/*
2563 * Unregister a registered request. The request is not completed (i.e.
2564 * no callbacks or wakeups) - higher layers are supposed to know what
2565 * they are canceling.
2566 */
2567void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2568{
2569 struct ceph_osd_client *osdc = req->r_osdc;
2570
2571 mutex_lock(&osdc->request_mutex);
2572 if (req->r_linger)
2573 __unregister_linger_request(osdc, req);
2574 __unregister_request(osdc, req);
2575 mutex_unlock(&osdc->request_mutex);
2576
2577 dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2578}
2579EXPORT_SYMBOL(ceph_osdc_cancel_request);
2580
2581/*
2582 * wait for a request to complete
2583 */
2584int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2585 struct ceph_osd_request *req)
2586{
2587 int rc;
2588
2589 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2590
2591 rc = wait_for_completion_interruptible(&req->r_completion);
2592 if (rc < 0) {
2593 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2594 ceph_osdc_cancel_request(req);
2595 complete_request(req);
2596 return rc;
2597 }
2598
2599 dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2600 req->r_result);
2601 return req->r_result;
2602}
2603EXPORT_SYMBOL(ceph_osdc_wait_request);
2604
2605/*
2606 * sync - wait for all in-flight requests to flush. avoid starvation.
2607 */
2608void ceph_osdc_sync(struct ceph_osd_client *osdc)
2609{
2610 struct ceph_osd_request *req;
2611 u64 last_tid, next_tid = 0;
2612
2613 mutex_lock(&osdc->request_mutex);
2614 last_tid = osdc->last_tid;
2615 while (1) {
2616 req = __lookup_request_ge(osdc, next_tid);
2617 if (!req)
2618 break;
2619 if (req->r_tid > last_tid)
2620 break;
2621
2622 next_tid = req->r_tid + 1;
2623 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2624 continue;
2625
2626 ceph_osdc_get_request(req);
2627 mutex_unlock(&osdc->request_mutex);
2628 dout("sync waiting on tid %llu (last is %llu)\n",
2629 req->r_tid, last_tid);
2630 wait_for_completion(&req->r_safe_completion);
2631 mutex_lock(&osdc->request_mutex);
2632 ceph_osdc_put_request(req);
2633 }
2634 mutex_unlock(&osdc->request_mutex);
2635 dout("sync done (thru tid %llu)\n", last_tid);
2636}
2637EXPORT_SYMBOL(ceph_osdc_sync);
2638
2639/*
2640 * Call all pending notify callbacks - for use after a watch is
2641 * unregistered, to make sure no more callbacks for it will be invoked
2642 */
2643void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2644{
2645 flush_workqueue(osdc->notify_wq);
2646}
2647EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2648
2649
2650/*
2651 * init, shutdown
2652 */
2653int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2654{
2655 int err;
2656
2657 dout("init\n");
2658 osdc->client = client;
2659 osdc->osdmap = NULL;
2660 init_rwsem(&osdc->map_sem);
2661 init_completion(&osdc->map_waiters);
2662 osdc->last_requested_map = 0;
2663 mutex_init(&osdc->request_mutex);
2664 osdc->last_tid = 0;
2665 osdc->osds = RB_ROOT;
2666 INIT_LIST_HEAD(&osdc->osd_lru);
2667 osdc->requests = RB_ROOT;
2668 INIT_LIST_HEAD(&osdc->req_lru);
2669 INIT_LIST_HEAD(&osdc->req_unsent);
2670 INIT_LIST_HEAD(&osdc->req_notarget);
2671 INIT_LIST_HEAD(&osdc->req_linger);
2672 osdc->num_requests = 0;
2673 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2674 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2675 spin_lock_init(&osdc->event_lock);
2676 osdc->event_tree = RB_ROOT;
2677 osdc->event_count = 0;
2678
2679 schedule_delayed_work(&osdc->osds_timeout_work,
2680 round_jiffies_relative(osdc->client->options->osd_idle_ttl));
2681
2682 err = -ENOMEM;
2683 osdc->req_mempool = mempool_create_slab_pool(10,
2684 ceph_osd_request_cache);
2685 if (!osdc->req_mempool)
2686 goto out;
2687
2688 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2689 OSD_OP_FRONT_LEN, 10, true,
2690 "osd_op");
2691 if (err < 0)
2692 goto out_mempool;
2693 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2694 OSD_OPREPLY_FRONT_LEN, 10, true,
2695 "osd_op_reply");
2696 if (err < 0)
2697 goto out_msgpool;
2698
2699 err = -ENOMEM;
2700 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2701 if (!osdc->notify_wq)
2702 goto out_msgpool_reply;
2703
2704 return 0;
2705
2706out_msgpool_reply:
2707 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2708out_msgpool:
2709 ceph_msgpool_destroy(&osdc->msgpool_op);
2710out_mempool:
2711 mempool_destroy(osdc->req_mempool);
2712out:
2713 return err;
2714}
2715
2716void ceph_osdc_stop(struct ceph_osd_client *osdc)
2717{
2718 flush_workqueue(osdc->notify_wq);
2719 destroy_workqueue(osdc->notify_wq);
2720 cancel_delayed_work_sync(&osdc->timeout_work);
2721 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2722 if (osdc->osdmap) {
2723 ceph_osdmap_destroy(osdc->osdmap);
2724 osdc->osdmap = NULL;
2725 }
2726 remove_all_osds(osdc);
2727 mempool_destroy(osdc->req_mempool);
2728 ceph_msgpool_destroy(&osdc->msgpool_op);
2729 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2730}
2731
2732/*
2733 * Read some contiguous pages. If we cross a stripe boundary, shorten
2734 * *plen. Return number of bytes read, or error.
2735 */
2736int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2737 struct ceph_vino vino, struct ceph_file_layout *layout,
2738 u64 off, u64 *plen,
2739 u32 truncate_seq, u64 truncate_size,
2740 struct page **pages, int num_pages, int page_align)
2741{
2742 struct ceph_osd_request *req;
2743 int rc = 0;
2744
2745 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2746 vino.snap, off, *plen);
2747 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2748 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2749 NULL, truncate_seq, truncate_size,
2750 false);
2751 if (IS_ERR(req))
2752 return PTR_ERR(req);
2753
2754 /* it may be a short read due to an object boundary */
2755
2756 osd_req_op_extent_osd_data_pages(req, 0,
2757 pages, *plen, page_align, false, false);
2758
2759 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2760 off, *plen, *plen, page_align);
2761
2762 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2763
2764 rc = ceph_osdc_start_request(osdc, req, false);
2765 if (!rc)
2766 rc = ceph_osdc_wait_request(osdc, req);
2767
2768 ceph_osdc_put_request(req);
2769 dout("readpages result %d\n", rc);
2770 return rc;
2771}
2772EXPORT_SYMBOL(ceph_osdc_readpages);
2773
2774/*
2775 * do a synchronous write on N pages
2776 */
2777int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2778 struct ceph_file_layout *layout,
2779 struct ceph_snap_context *snapc,
2780 u64 off, u64 len,
2781 u32 truncate_seq, u64 truncate_size,
2782 struct timespec *mtime,
2783 struct page **pages, int num_pages)
2784{
2785 struct ceph_osd_request *req;
2786 int rc = 0;
2787 int page_align = off & ~PAGE_MASK;
2788
2789 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */
2790 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2791 CEPH_OSD_OP_WRITE,
2792 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2793 snapc, truncate_seq, truncate_size,
2794 true);
2795 if (IS_ERR(req))
2796 return PTR_ERR(req);
2797
2798 /* it may be a short write due to an object boundary */
2799 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2800 false, false);
2801 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2802
2803 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2804
2805 rc = ceph_osdc_start_request(osdc, req, true);
2806 if (!rc)
2807 rc = ceph_osdc_wait_request(osdc, req);
2808
2809 ceph_osdc_put_request(req);
2810 if (rc == 0)
2811 rc = len;
2812 dout("writepages result %d\n", rc);
2813 return rc;
2814}
2815EXPORT_SYMBOL(ceph_osdc_writepages);
2816
2817int ceph_osdc_setup(void)
2818{
2819 size_t size = sizeof(struct ceph_osd_request) +
2820 CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op);
2821
2822 BUG_ON(ceph_osd_request_cache);
2823 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size,
2824 0, 0, NULL);
2825
2826 return ceph_osd_request_cache ? 0 : -ENOMEM;
2827}
2828EXPORT_SYMBOL(ceph_osdc_setup);
2829
2830void ceph_osdc_cleanup(void)
2831{
2832 BUG_ON(!ceph_osd_request_cache);
2833 kmem_cache_destroy(ceph_osd_request_cache);
2834 ceph_osd_request_cache = NULL;
2835}
2836EXPORT_SYMBOL(ceph_osdc_cleanup);
2837
2838/*
2839 * handle incoming message
2840 */
2841static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2842{
2843 struct ceph_osd *osd = con->private;
2844 struct ceph_osd_client *osdc;
2845 int type = le16_to_cpu(msg->hdr.type);
2846
2847 if (!osd)
2848 goto out;
2849 osdc = osd->o_osdc;
2850
2851 switch (type) {
2852 case CEPH_MSG_OSD_MAP:
2853 ceph_osdc_handle_map(osdc, msg);
2854 break;
2855 case CEPH_MSG_OSD_OPREPLY:
2856 handle_reply(osdc, msg);
2857 break;
2858 case CEPH_MSG_WATCH_NOTIFY:
2859 handle_watch_notify(osdc, msg);
2860 break;
2861
2862 default:
2863 pr_err("received unknown message type %d %s\n", type,
2864 ceph_msg_type_name(type));
2865 }
2866out:
2867 ceph_msg_put(msg);
2868}
2869
2870/*
2871 * Lookup and return message for incoming reply. Don't try to do
2872 * anything about a larger than preallocated data portion of the
2873 * message at the moment - for now, just skip the message.
2874 */
2875static struct ceph_msg *get_reply(struct ceph_connection *con,
2876 struct ceph_msg_header *hdr,
2877 int *skip)
2878{
2879 struct ceph_osd *osd = con->private;
2880 struct ceph_osd_client *osdc = osd->o_osdc;
2881 struct ceph_msg *m;
2882 struct ceph_osd_request *req;
2883 int front_len = le32_to_cpu(hdr->front_len);
2884 int data_len = le32_to_cpu(hdr->data_len);
2885 u64 tid;
2886
2887 tid = le64_to_cpu(hdr->tid);
2888 mutex_lock(&osdc->request_mutex);
2889 req = __lookup_request(osdc, tid);
2890 if (!req) {
2891 dout("%s osd%d tid %llu unknown, skipping\n", __func__,
2892 osd->o_osd, tid);
2893 m = NULL;
2894 *skip = 1;
2895 goto out;
2896 }
2897
2898 ceph_msg_revoke_incoming(req->r_reply);
2899
2900 if (front_len > req->r_reply->front_alloc_len) {
2901 pr_warn("%s osd%d tid %llu front %d > preallocated %d\n",
2902 __func__, osd->o_osd, req->r_tid, front_len,
2903 req->r_reply->front_alloc_len);
2904 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2905 false);
2906 if (!m)
2907 goto out;
2908 ceph_msg_put(req->r_reply);
2909 req->r_reply = m;
2910 }
2911
2912 if (data_len > req->r_reply->data_length) {
2913 pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n",
2914 __func__, osd->o_osd, req->r_tid, data_len,
2915 req->r_reply->data_length);
2916 m = NULL;
2917 *skip = 1;
2918 goto out;
2919 }
2920
2921 m = ceph_msg_get(req->r_reply);
2922 dout("get_reply tid %lld %p\n", tid, m);
2923
2924out:
2925 mutex_unlock(&osdc->request_mutex);
2926 return m;
2927}
2928
2929static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2930 struct ceph_msg_header *hdr,
2931 int *skip)
2932{
2933 struct ceph_osd *osd = con->private;
2934 int type = le16_to_cpu(hdr->type);
2935 int front = le32_to_cpu(hdr->front_len);
2936
2937 *skip = 0;
2938 switch (type) {
2939 case CEPH_MSG_OSD_MAP:
2940 case CEPH_MSG_WATCH_NOTIFY:
2941 return ceph_msg_new(type, front, GFP_NOFS, false);
2942 case CEPH_MSG_OSD_OPREPLY:
2943 return get_reply(con, hdr, skip);
2944 default:
2945 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2946 osd->o_osd);
2947 *skip = 1;
2948 return NULL;
2949 }
2950}
2951
2952/*
2953 * Wrappers to refcount containing ceph_osd struct
2954 */
2955static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2956{
2957 struct ceph_osd *osd = con->private;
2958 if (get_osd(osd))
2959 return con;
2960 return NULL;
2961}
2962
2963static void put_osd_con(struct ceph_connection *con)
2964{
2965 struct ceph_osd *osd = con->private;
2966 put_osd(osd);
2967}
2968
2969/*
2970 * authentication
2971 */
2972/*
2973 * Note: returned pointer is the address of a structure that's
2974 * managed separately. Caller must *not* attempt to free it.
2975 */
2976static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2977 int *proto, int force_new)
2978{
2979 struct ceph_osd *o = con->private;
2980 struct ceph_osd_client *osdc = o->o_osdc;
2981 struct ceph_auth_client *ac = osdc->client->monc.auth;
2982 struct ceph_auth_handshake *auth = &o->o_auth;
2983
2984 if (force_new && auth->authorizer) {
2985 ceph_auth_destroy_authorizer(auth->authorizer);
2986 auth->authorizer = NULL;
2987 }
2988 if (!auth->authorizer) {
2989 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2990 auth);
2991 if (ret)
2992 return ERR_PTR(ret);
2993 } else {
2994 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2995 auth);
2996 if (ret)
2997 return ERR_PTR(ret);
2998 }
2999 *proto = ac->protocol;
3000
3001 return auth;
3002}
3003
3004
3005static int verify_authorizer_reply(struct ceph_connection *con, int len)
3006{
3007 struct ceph_osd *o = con->private;
3008 struct ceph_osd_client *osdc = o->o_osdc;
3009 struct ceph_auth_client *ac = osdc->client->monc.auth;
3010
3011 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
3012}
3013
3014static int invalidate_authorizer(struct ceph_connection *con)
3015{
3016 struct ceph_osd *o = con->private;
3017 struct ceph_osd_client *osdc = o->o_osdc;
3018 struct ceph_auth_client *ac = osdc->client->monc.auth;
3019
3020 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
3021 return ceph_monc_validate_auth(&osdc->client->monc);
3022}
3023
3024static int osd_sign_message(struct ceph_msg *msg)
3025{
3026 struct ceph_osd *o = msg->con->private;
3027 struct ceph_auth_handshake *auth = &o->o_auth;
3028
3029 return ceph_auth_sign_message(auth, msg);
3030}
3031
3032static int osd_check_message_signature(struct ceph_msg *msg)
3033{
3034 struct ceph_osd *o = msg->con->private;
3035 struct ceph_auth_handshake *auth = &o->o_auth;
3036
3037 return ceph_auth_check_message_signature(auth, msg);
3038}
3039
3040static const struct ceph_connection_operations osd_con_ops = {
3041 .get = get_osd_con,
3042 .put = put_osd_con,
3043 .dispatch = dispatch,
3044 .get_authorizer = get_authorizer,
3045 .verify_authorizer_reply = verify_authorizer_reply,
3046 .invalidate_authorizer = invalidate_authorizer,
3047 .alloc_msg = alloc_msg,
3048 .sign_message = osd_sign_message,
3049 .check_message_signature = osd_check_message_signature,
3050 .fault = osd_reset,
3051};
Linux 6.x block layer and io_uring tree(s)