]> Pileus Git - ~andy/linux/blob - net/ceph/osd_client.c
libceph: use osd_req_op_extent_init()
[~andy/linux] / net / ceph / osd_client.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/module.h>
4 #include <linux/err.h>
5 #include <linux/highmem.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/slab.h>
9 #include <linux/uaccess.h>
10 #ifdef CONFIG_BLOCK
11 #include <linux/bio.h>
12 #endif
13
14 #include <linux/ceph/libceph.h>
15 #include <linux/ceph/osd_client.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/auth.h>
19 #include <linux/ceph/pagelist.h>
20
21 #define OSD_OP_FRONT_LEN        4096
22 #define OSD_OPREPLY_FRONT_LEN   512
23
24 static const struct ceph_connection_operations osd_con_ops;
25
26 static void __send_queued(struct ceph_osd_client *osdc);
27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
28 static void __register_request(struct ceph_osd_client *osdc,
29                                struct ceph_osd_request *req);
30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
31                                         struct ceph_osd_request *req);
32 static void __send_request(struct ceph_osd_client *osdc,
33                            struct ceph_osd_request *req);
34
35 /*
36  * Implement client access to distributed object storage cluster.
37  *
38  * All data objects are stored within a cluster/cloud of OSDs, or
39  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
40  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
41  * remote daemons serving up and coordinating consistent and safe
42  * access to storage.
43  *
44  * Cluster membership and the mapping of data objects onto storage devices
45  * are described by the osd map.
46  *
47  * We keep track of pending OSD requests (read, write), resubmit
48  * requests to different OSDs when the cluster topology/data layout
49  * change, or retry the affected requests when the communications
50  * channel with an OSD is reset.
51  */
52
53 /*
54  * calculate the mapping of a file extent onto an object, and fill out the
55  * request accordingly.  shorten extent as necessary if it crosses an
56  * object boundary.
57  *
58  * fill osd op in request message.
59  */
60 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
61                         u64 *objnum, u64 *objoff, u64 *objlen)
62 {
63         u64 orig_len = *plen;
64         int r;
65
66         /* object extent? */
67         r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
68                                           objoff, objlen);
69         if (r < 0)
70                 return r;
71         if (*objlen < orig_len) {
72                 *plen = *objlen;
73                 dout(" skipping last %llu, final file extent %llu~%llu\n",
74                      orig_len - *plen, off, *plen);
75         }
76
77         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
78
79         return 0;
80 }
81
82 /*
83  * requests
84  */
85 void ceph_osdc_release_request(struct kref *kref)
86 {
87         int num_pages;
88         struct ceph_osd_request *req = container_of(kref,
89                                                     struct ceph_osd_request,
90                                                     r_kref);
91
92         if (req->r_request)
93                 ceph_msg_put(req->r_request);
94         if (req->r_con_filling_msg) {
95                 dout("%s revoking msg %p from con %p\n", __func__,
96                      req->r_reply, req->r_con_filling_msg);
97                 ceph_msg_revoke_incoming(req->r_reply);
98                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
99                 req->r_con_filling_msg = NULL;
100         }
101         if (req->r_reply)
102                 ceph_msg_put(req->r_reply);
103
104         if (req->r_data_in.type == CEPH_OSD_DATA_TYPE_PAGES &&
105                         req->r_data_in.own_pages) {
106                 num_pages = calc_pages_for((u64)req->r_data_in.alignment,
107                                                 (u64)req->r_data_in.length);
108                 ceph_release_page_vector(req->r_data_in.pages, num_pages);
109         }
110         if (req->r_data_out.type == CEPH_OSD_DATA_TYPE_PAGES &&
111                         req->r_data_out.own_pages) {
112                 num_pages = calc_pages_for((u64)req->r_data_out.alignment,
113                                                 (u64)req->r_data_out.length);
114                 ceph_release_page_vector(req->r_data_out.pages, num_pages);
115         }
116
117         ceph_put_snap_context(req->r_snapc);
118         if (req->r_mempool)
119                 mempool_free(req, req->r_osdc->req_mempool);
120         else
121                 kfree(req);
122 }
123 EXPORT_SYMBOL(ceph_osdc_release_request);
124
125 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
126                                                struct ceph_snap_context *snapc,
127                                                unsigned int num_ops,
128                                                bool use_mempool,
129                                                gfp_t gfp_flags)
130 {
131         struct ceph_osd_request *req;
132         struct ceph_msg *msg;
133         size_t msg_size;
134
135         msg_size = 4 + 4 + 8 + 8 + 4+8;
136         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
137         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
138         msg_size += 4 + MAX_OBJ_NAME_SIZE;
139         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
140         msg_size += 8;  /* snapid */
141         msg_size += 8;  /* snap_seq */
142         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
143         msg_size += 4;
144
145         if (use_mempool) {
146                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
147                 memset(req, 0, sizeof(*req));
148         } else {
149                 req = kzalloc(sizeof(*req), gfp_flags);
150         }
151         if (req == NULL)
152                 return NULL;
153
154         req->r_osdc = osdc;
155         req->r_mempool = use_mempool;
156
157         kref_init(&req->r_kref);
158         init_completion(&req->r_completion);
159         init_completion(&req->r_safe_completion);
160         RB_CLEAR_NODE(&req->r_node);
161         INIT_LIST_HEAD(&req->r_unsafe_item);
162         INIT_LIST_HEAD(&req->r_linger_item);
163         INIT_LIST_HEAD(&req->r_linger_osd);
164         INIT_LIST_HEAD(&req->r_req_lru_item);
165         INIT_LIST_HEAD(&req->r_osd_item);
166
167         /* create reply message */
168         if (use_mempool)
169                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
170         else
171                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
172                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
173         if (!msg) {
174                 ceph_osdc_put_request(req);
175                 return NULL;
176         }
177         req->r_reply = msg;
178
179         req->r_data_in.type = CEPH_OSD_DATA_TYPE_NONE;
180         req->r_data_out.type = CEPH_OSD_DATA_TYPE_NONE;
181
182         /* create request message; allow space for oid */
183         if (use_mempool)
184                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
185         else
186                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
187         if (!msg) {
188                 ceph_osdc_put_request(req);
189                 return NULL;
190         }
191
192         memset(msg->front.iov_base, 0, msg->front.iov_len);
193
194         req->r_request = msg;
195
196         return req;
197 }
198 EXPORT_SYMBOL(ceph_osdc_alloc_request);
199
200 static bool osd_req_opcode_valid(u16 opcode)
201 {
202         switch (opcode) {
203         case CEPH_OSD_OP_READ:
204         case CEPH_OSD_OP_STAT:
205         case CEPH_OSD_OP_MAPEXT:
206         case CEPH_OSD_OP_MASKTRUNC:
207         case CEPH_OSD_OP_SPARSE_READ:
208         case CEPH_OSD_OP_NOTIFY:
209         case CEPH_OSD_OP_NOTIFY_ACK:
210         case CEPH_OSD_OP_ASSERT_VER:
211         case CEPH_OSD_OP_WRITE:
212         case CEPH_OSD_OP_WRITEFULL:
213         case CEPH_OSD_OP_TRUNCATE:
214         case CEPH_OSD_OP_ZERO:
215         case CEPH_OSD_OP_DELETE:
216         case CEPH_OSD_OP_APPEND:
217         case CEPH_OSD_OP_STARTSYNC:
218         case CEPH_OSD_OP_SETTRUNC:
219         case CEPH_OSD_OP_TRIMTRUNC:
220         case CEPH_OSD_OP_TMAPUP:
221         case CEPH_OSD_OP_TMAPPUT:
222         case CEPH_OSD_OP_TMAPGET:
223         case CEPH_OSD_OP_CREATE:
224         case CEPH_OSD_OP_ROLLBACK:
225         case CEPH_OSD_OP_WATCH:
226         case CEPH_OSD_OP_OMAPGETKEYS:
227         case CEPH_OSD_OP_OMAPGETVALS:
228         case CEPH_OSD_OP_OMAPGETHEADER:
229         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
230         case CEPH_OSD_OP_OMAPSETVALS:
231         case CEPH_OSD_OP_OMAPSETHEADER:
232         case CEPH_OSD_OP_OMAPCLEAR:
233         case CEPH_OSD_OP_OMAPRMKEYS:
234         case CEPH_OSD_OP_OMAP_CMP:
235         case CEPH_OSD_OP_CLONERANGE:
236         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
237         case CEPH_OSD_OP_SRC_CMPXATTR:
238         case CEPH_OSD_OP_GETXATTR:
239         case CEPH_OSD_OP_GETXATTRS:
240         case CEPH_OSD_OP_CMPXATTR:
241         case CEPH_OSD_OP_SETXATTR:
242         case CEPH_OSD_OP_SETXATTRS:
243         case CEPH_OSD_OP_RESETXATTRS:
244         case CEPH_OSD_OP_RMXATTR:
245         case CEPH_OSD_OP_PULL:
246         case CEPH_OSD_OP_PUSH:
247         case CEPH_OSD_OP_BALANCEREADS:
248         case CEPH_OSD_OP_UNBALANCEREADS:
249         case CEPH_OSD_OP_SCRUB:
250         case CEPH_OSD_OP_SCRUB_RESERVE:
251         case CEPH_OSD_OP_SCRUB_UNRESERVE:
252         case CEPH_OSD_OP_SCRUB_STOP:
253         case CEPH_OSD_OP_SCRUB_MAP:
254         case CEPH_OSD_OP_WRLOCK:
255         case CEPH_OSD_OP_WRUNLOCK:
256         case CEPH_OSD_OP_RDLOCK:
257         case CEPH_OSD_OP_RDUNLOCK:
258         case CEPH_OSD_OP_UPLOCK:
259         case CEPH_OSD_OP_DNLOCK:
260         case CEPH_OSD_OP_CALL:
261         case CEPH_OSD_OP_PGLS:
262         case CEPH_OSD_OP_PGLS_FILTER:
263                 return true;
264         default:
265                 return false;
266         }
267 }
268
269 /*
270  * This is an osd op init function for opcodes that have no data or
271  * other information associated with them.  It also serves as a
272  * common init routine for all the other init functions, below.
273  */
274 void osd_req_op_init(struct ceph_osd_req_op *op, u16 opcode)
275 {
276         BUG_ON(!osd_req_opcode_valid(opcode));
277
278         memset(op, 0, sizeof (*op));
279
280         op->op = opcode;
281 }
282
283 void osd_req_op_extent_init(struct ceph_osd_req_op *op, u16 opcode,
284                                 u64 offset, u64 length,
285                                 u64 truncate_size, u32 truncate_seq)
286 {
287         size_t payload_len = 0;
288
289         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE);
290
291         osd_req_op_init(op, opcode);
292
293         op->extent.offset = offset;
294         op->extent.length = length;
295         op->extent.truncate_size = truncate_size;
296         op->extent.truncate_seq = truncate_seq;
297         if (opcode == CEPH_OSD_OP_WRITE)
298                 payload_len += length;
299
300         op->payload_len = payload_len;
301 }
302 EXPORT_SYMBOL(osd_req_op_extent_init);
303
304 void osd_req_op_cls_init(struct ceph_osd_req_op *op, u16 opcode,
305                         const char *class, const char *method,
306                         const void *request_data, size_t request_data_size)
307 {
308         size_t payload_len = 0;
309         size_t size;
310
311         BUG_ON(opcode != CEPH_OSD_OP_CALL);
312
313         osd_req_op_init(op, opcode);
314
315         op->cls.class_name = class;
316         size = strlen(class);
317         BUG_ON(size > (size_t) U8_MAX);
318         op->cls.class_len = size;
319         payload_len += size;
320
321         op->cls.method_name = method;
322         size = strlen(method);
323         BUG_ON(size > (size_t) U8_MAX);
324         op->cls.method_len = size;
325         payload_len += size;
326
327         op->cls.indata = request_data;
328         BUG_ON(request_data_size > (size_t) U32_MAX);
329         op->cls.indata_len = (u32) request_data_size;
330         payload_len += request_data_size;
331
332         op->cls.argc = 0;       /* currently unused */
333
334         op->payload_len = payload_len;
335 }
336 EXPORT_SYMBOL(osd_req_op_cls_init);
337
338 void osd_req_op_watch_init(struct ceph_osd_req_op *op, u16 opcode,
339                                 u64 cookie, u64 version, int flag)
340 {
341         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
342
343         osd_req_op_init(op, opcode);
344
345         op->watch.cookie = cookie;
346         /* op->watch.ver = version; */  /* XXX 3847 */
347         op->watch.ver = cpu_to_le64(version);
348         if (opcode == CEPH_OSD_OP_WATCH && flag)
349                 op->watch.flag = (u8) 1;
350 }
351 EXPORT_SYMBOL(osd_req_op_watch_init);
352
353 static u64 osd_req_encode_op(struct ceph_osd_request *req,
354                               struct ceph_osd_op *dst,
355                               struct ceph_osd_req_op *src)
356 {
357         u64 out_data_len = 0;
358         struct ceph_pagelist *pagelist;
359
360         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
361                 pr_err("unrecognized osd opcode %d\n", src->op);
362
363                 return 0;
364         }
365
366         switch (src->op) {
367         case CEPH_OSD_OP_STAT:
368                 break;
369         case CEPH_OSD_OP_READ:
370         case CEPH_OSD_OP_WRITE:
371                 if (src->op == CEPH_OSD_OP_WRITE)
372                         out_data_len = src->extent.length;
373                 dst->extent.offset = cpu_to_le64(src->extent.offset);
374                 dst->extent.length = cpu_to_le64(src->extent.length);
375                 dst->extent.truncate_size =
376                         cpu_to_le64(src->extent.truncate_size);
377                 dst->extent.truncate_seq =
378                         cpu_to_le32(src->extent.truncate_seq);
379                 break;
380         case CEPH_OSD_OP_CALL:
381                 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
382                 BUG_ON(!pagelist);
383                 ceph_pagelist_init(pagelist);
384
385                 dst->cls.class_len = src->cls.class_len;
386                 dst->cls.method_len = src->cls.method_len;
387                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
388                 ceph_pagelist_append(pagelist, src->cls.class_name,
389                                      src->cls.class_len);
390                 ceph_pagelist_append(pagelist, src->cls.method_name,
391                                      src->cls.method_len);
392                 ceph_pagelist_append(pagelist, src->cls.indata,
393                                      src->cls.indata_len);
394
395                 req->r_data_out.type = CEPH_OSD_DATA_TYPE_PAGELIST;
396                 req->r_data_out.pagelist = pagelist;
397                 out_data_len = pagelist->length;
398                 break;
399         case CEPH_OSD_OP_STARTSYNC:
400                 break;
401         case CEPH_OSD_OP_NOTIFY_ACK:
402         case CEPH_OSD_OP_WATCH:
403                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
404                 dst->watch.ver = cpu_to_le64(src->watch.ver);
405                 dst->watch.flag = src->watch.flag;
406                 break;
407         default:
408                 pr_err("unsupported osd opcode %s\n",
409                         ceph_osd_op_name(src->op));
410                 WARN_ON(1);
411
412                 return 0;
413         }
414         dst->op = cpu_to_le16(src->op);
415         dst->payload_len = cpu_to_le32(src->payload_len);
416
417         return out_data_len;
418 }
419
420 /*
421  * build new request AND message
422  *
423  */
424 void ceph_osdc_build_request(struct ceph_osd_request *req,
425                              u64 off, unsigned int num_ops,
426                              struct ceph_osd_req_op *src_ops,
427                              struct ceph_snap_context *snapc, u64 snap_id,
428                              struct timespec *mtime)
429 {
430         struct ceph_msg *msg = req->r_request;
431         struct ceph_osd_req_op *src_op;
432         void *p;
433         size_t msg_size;
434         int flags = req->r_flags;
435         u64 data_len;
436         int i;
437
438         req->r_num_ops = num_ops;
439         req->r_snapid = snap_id;
440         req->r_snapc = ceph_get_snap_context(snapc);
441
442         /* encode request */
443         msg->hdr.version = cpu_to_le16(4);
444
445         p = msg->front.iov_base;
446         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
447         req->r_request_osdmap_epoch = p;
448         p += 4;
449         req->r_request_flags = p;
450         p += 4;
451         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
452                 ceph_encode_timespec(p, mtime);
453         p += sizeof(struct ceph_timespec);
454         req->r_request_reassert_version = p;
455         p += sizeof(struct ceph_eversion); /* will get filled in */
456
457         /* oloc */
458         ceph_encode_8(&p, 4);
459         ceph_encode_8(&p, 4);
460         ceph_encode_32(&p, 8 + 4 + 4);
461         req->r_request_pool = p;
462         p += 8;
463         ceph_encode_32(&p, -1);  /* preferred */
464         ceph_encode_32(&p, 0);   /* key len */
465
466         ceph_encode_8(&p, 1);
467         req->r_request_pgid = p;
468         p += 8 + 4;
469         ceph_encode_32(&p, -1);  /* preferred */
470
471         /* oid */
472         ceph_encode_32(&p, req->r_oid_len);
473         memcpy(p, req->r_oid, req->r_oid_len);
474         dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len);
475         p += req->r_oid_len;
476
477         /* ops--can imply data */
478         ceph_encode_16(&p, num_ops);
479         src_op = src_ops;
480         req->r_request_ops = p;
481         data_len = 0;
482         for (i = 0; i < num_ops; i++, src_op++) {
483                 data_len += osd_req_encode_op(req, p, src_op);
484                 p += sizeof(struct ceph_osd_op);
485         }
486
487         /* snaps */
488         ceph_encode_64(&p, req->r_snapid);
489         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
490         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
491         if (req->r_snapc) {
492                 for (i = 0; i < snapc->num_snaps; i++) {
493                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
494                 }
495         }
496
497         req->r_request_attempts = p;
498         p += 4;
499
500         /* data */
501         if (flags & CEPH_OSD_FLAG_WRITE) {
502                 u16 data_off;
503
504                 /*
505                  * The header "data_off" is a hint to the receiver
506                  * allowing it to align received data into its
507                  * buffers such that there's no need to re-copy
508                  * it before writing it to disk (direct I/O).
509                  */
510                 data_off = (u16) (off & 0xffff);
511                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
512         }
513         req->r_request->hdr.data_len = cpu_to_le32(data_len);
514
515         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
516         msg_size = p - msg->front.iov_base;
517         msg->front.iov_len = msg_size;
518         msg->hdr.front_len = cpu_to_le32(msg_size);
519
520         dout("build_request msg_size was %d num_ops %d\n", (int)msg_size,
521              num_ops);
522         return;
523 }
524 EXPORT_SYMBOL(ceph_osdc_build_request);
525
526 /*
527  * build new request AND message, calculate layout, and adjust file
528  * extent as needed.
529  *
530  * if the file was recently truncated, we include information about its
531  * old and new size so that the object can be updated appropriately.  (we
532  * avoid synchronously deleting truncated objects because it's slow.)
533  *
534  * if @do_sync, include a 'startsync' command so that the osd will flush
535  * data quickly.
536  */
537 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
538                                                struct ceph_file_layout *layout,
539                                                struct ceph_vino vino,
540                                                u64 off, u64 *plen,
541                                                int opcode, int flags,
542                                                struct ceph_snap_context *snapc,
543                                                int do_sync,
544                                                u32 truncate_seq,
545                                                u64 truncate_size,
546                                                struct timespec *mtime,
547                                                bool use_mempool)
548 {
549         struct ceph_osd_req_op ops[2];
550         struct ceph_osd_request *req;
551         unsigned int num_op = do_sync ? 2 : 1;
552         u64 objnum = 0;
553         u64 objoff = 0;
554         u64 objlen = 0;
555         u32 object_size;
556         u64 object_base;
557         int r;
558
559         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE);
560
561         req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
562                                         GFP_NOFS);
563         if (!req)
564                 return ERR_PTR(-ENOMEM);
565         req->r_flags = flags;
566
567         /* calculate max write size */
568         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
569         if (r < 0) {
570                 ceph_osdc_put_request(req);
571                 return ERR_PTR(r);
572         }
573
574         object_size = le32_to_cpu(layout->fl_object_size);
575         object_base = off - objoff;
576         if (truncate_size <= object_base) {
577                 truncate_size = 0;
578         } else {
579                 truncate_size -= object_base;
580                 if (truncate_size > object_size)
581                         truncate_size = object_size;
582         }
583
584         osd_req_op_extent_init(&ops[0], opcode, objoff, objlen,
585                                 truncate_size, truncate_seq);
586         if (do_sync)
587                 osd_req_op_init(&ops[1], CEPH_OSD_OP_STARTSYNC);
588
589         req->r_file_layout = *layout;  /* keep a copy */
590
591         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx",
592                 vino.ino, objnum);
593         req->r_oid_len = strlen(req->r_oid);
594
595         ceph_osdc_build_request(req, off, num_op, ops,
596                                 snapc, vino.snap, mtime);
597
598         return req;
599 }
600 EXPORT_SYMBOL(ceph_osdc_new_request);
601
602 /*
603  * We keep osd requests in an rbtree, sorted by ->r_tid.
604  */
605 static void __insert_request(struct ceph_osd_client *osdc,
606                              struct ceph_osd_request *new)
607 {
608         struct rb_node **p = &osdc->requests.rb_node;
609         struct rb_node *parent = NULL;
610         struct ceph_osd_request *req = NULL;
611
612         while (*p) {
613                 parent = *p;
614                 req = rb_entry(parent, struct ceph_osd_request, r_node);
615                 if (new->r_tid < req->r_tid)
616                         p = &(*p)->rb_left;
617                 else if (new->r_tid > req->r_tid)
618                         p = &(*p)->rb_right;
619                 else
620                         BUG();
621         }
622
623         rb_link_node(&new->r_node, parent, p);
624         rb_insert_color(&new->r_node, &osdc->requests);
625 }
626
627 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
628                                                  u64 tid)
629 {
630         struct ceph_osd_request *req;
631         struct rb_node *n = osdc->requests.rb_node;
632
633         while (n) {
634                 req = rb_entry(n, struct ceph_osd_request, r_node);
635                 if (tid < req->r_tid)
636                         n = n->rb_left;
637                 else if (tid > req->r_tid)
638                         n = n->rb_right;
639                 else
640                         return req;
641         }
642         return NULL;
643 }
644
645 static struct ceph_osd_request *
646 __lookup_request_ge(struct ceph_osd_client *osdc,
647                     u64 tid)
648 {
649         struct ceph_osd_request *req;
650         struct rb_node *n = osdc->requests.rb_node;
651
652         while (n) {
653                 req = rb_entry(n, struct ceph_osd_request, r_node);
654                 if (tid < req->r_tid) {
655                         if (!n->rb_left)
656                                 return req;
657                         n = n->rb_left;
658                 } else if (tid > req->r_tid) {
659                         n = n->rb_right;
660                 } else {
661                         return req;
662                 }
663         }
664         return NULL;
665 }
666
667 /*
668  * Resubmit requests pending on the given osd.
669  */
670 static void __kick_osd_requests(struct ceph_osd_client *osdc,
671                                 struct ceph_osd *osd)
672 {
673         struct ceph_osd_request *req, *nreq;
674         LIST_HEAD(resend);
675         int err;
676
677         dout("__kick_osd_requests osd%d\n", osd->o_osd);
678         err = __reset_osd(osdc, osd);
679         if (err)
680                 return;
681         /*
682          * Build up a list of requests to resend by traversing the
683          * osd's list of requests.  Requests for a given object are
684          * sent in tid order, and that is also the order they're
685          * kept on this list.  Therefore all requests that are in
686          * flight will be found first, followed by all requests that
687          * have not yet been sent.  And to resend requests while
688          * preserving this order we will want to put any sent
689          * requests back on the front of the osd client's unsent
690          * list.
691          *
692          * So we build a separate ordered list of already-sent
693          * requests for the affected osd and splice it onto the
694          * front of the osd client's unsent list.  Once we've seen a
695          * request that has not yet been sent we're done.  Those
696          * requests are already sitting right where they belong.
697          */
698         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
699                 if (!req->r_sent)
700                         break;
701                 list_move_tail(&req->r_req_lru_item, &resend);
702                 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
703                      osd->o_osd);
704                 if (!req->r_linger)
705                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
706         }
707         list_splice(&resend, &osdc->req_unsent);
708
709         /*
710          * Linger requests are re-registered before sending, which
711          * sets up a new tid for each.  We add them to the unsent
712          * list at the end to keep things in tid order.
713          */
714         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
715                                  r_linger_osd) {
716                 /*
717                  * reregister request prior to unregistering linger so
718                  * that r_osd is preserved.
719                  */
720                 BUG_ON(!list_empty(&req->r_req_lru_item));
721                 __register_request(osdc, req);
722                 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
723                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
724                 __unregister_linger_request(osdc, req);
725                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
726                      osd->o_osd);
727         }
728 }
729
730 /*
731  * If the osd connection drops, we need to resubmit all requests.
732  */
733 static void osd_reset(struct ceph_connection *con)
734 {
735         struct ceph_osd *osd = con->private;
736         struct ceph_osd_client *osdc;
737
738         if (!osd)
739                 return;
740         dout("osd_reset osd%d\n", osd->o_osd);
741         osdc = osd->o_osdc;
742         down_read(&osdc->map_sem);
743         mutex_lock(&osdc->request_mutex);
744         __kick_osd_requests(osdc, osd);
745         __send_queued(osdc);
746         mutex_unlock(&osdc->request_mutex);
747         up_read(&osdc->map_sem);
748 }
749
750 /*
751  * Track open sessions with osds.
752  */
753 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
754 {
755         struct ceph_osd *osd;
756
757         osd = kzalloc(sizeof(*osd), GFP_NOFS);
758         if (!osd)
759                 return NULL;
760
761         atomic_set(&osd->o_ref, 1);
762         osd->o_osdc = osdc;
763         osd->o_osd = onum;
764         RB_CLEAR_NODE(&osd->o_node);
765         INIT_LIST_HEAD(&osd->o_requests);
766         INIT_LIST_HEAD(&osd->o_linger_requests);
767         INIT_LIST_HEAD(&osd->o_osd_lru);
768         osd->o_incarnation = 1;
769
770         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
771
772         INIT_LIST_HEAD(&osd->o_keepalive_item);
773         return osd;
774 }
775
776 static struct ceph_osd *get_osd(struct ceph_osd *osd)
777 {
778         if (atomic_inc_not_zero(&osd->o_ref)) {
779                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
780                      atomic_read(&osd->o_ref));
781                 return osd;
782         } else {
783                 dout("get_osd %p FAIL\n", osd);
784                 return NULL;
785         }
786 }
787
788 static void put_osd(struct ceph_osd *osd)
789 {
790         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
791              atomic_read(&osd->o_ref) - 1);
792         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
793                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
794
795                 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
796                 kfree(osd);
797         }
798 }
799
800 /*
801  * remove an osd from our map
802  */
803 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
804 {
805         dout("__remove_osd %p\n", osd);
806         BUG_ON(!list_empty(&osd->o_requests));
807         rb_erase(&osd->o_node, &osdc->osds);
808         list_del_init(&osd->o_osd_lru);
809         ceph_con_close(&osd->o_con);
810         put_osd(osd);
811 }
812
813 static void remove_all_osds(struct ceph_osd_client *osdc)
814 {
815         dout("%s %p\n", __func__, osdc);
816         mutex_lock(&osdc->request_mutex);
817         while (!RB_EMPTY_ROOT(&osdc->osds)) {
818                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
819                                                 struct ceph_osd, o_node);
820                 __remove_osd(osdc, osd);
821         }
822         mutex_unlock(&osdc->request_mutex);
823 }
824
825 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
826                               struct ceph_osd *osd)
827 {
828         dout("__move_osd_to_lru %p\n", osd);
829         BUG_ON(!list_empty(&osd->o_osd_lru));
830         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
831         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
832 }
833
834 static void __remove_osd_from_lru(struct ceph_osd *osd)
835 {
836         dout("__remove_osd_from_lru %p\n", osd);
837         if (!list_empty(&osd->o_osd_lru))
838                 list_del_init(&osd->o_osd_lru);
839 }
840
841 static void remove_old_osds(struct ceph_osd_client *osdc)
842 {
843         struct ceph_osd *osd, *nosd;
844
845         dout("__remove_old_osds %p\n", osdc);
846         mutex_lock(&osdc->request_mutex);
847         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
848                 if (time_before(jiffies, osd->lru_ttl))
849                         break;
850                 __remove_osd(osdc, osd);
851         }
852         mutex_unlock(&osdc->request_mutex);
853 }
854
855 /*
856  * reset osd connect
857  */
858 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
859 {
860         struct ceph_entity_addr *peer_addr;
861
862         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
863         if (list_empty(&osd->o_requests) &&
864             list_empty(&osd->o_linger_requests)) {
865                 __remove_osd(osdc, osd);
866
867                 return -ENODEV;
868         }
869
870         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
871         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
872                         !ceph_con_opened(&osd->o_con)) {
873                 struct ceph_osd_request *req;
874
875                 dout(" osd addr hasn't changed and connection never opened,"
876                      " letting msgr retry");
877                 /* touch each r_stamp for handle_timeout()'s benfit */
878                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
879                         req->r_stamp = jiffies;
880
881                 return -EAGAIN;
882         }
883
884         ceph_con_close(&osd->o_con);
885         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
886         osd->o_incarnation++;
887
888         return 0;
889 }
890
891 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
892 {
893         struct rb_node **p = &osdc->osds.rb_node;
894         struct rb_node *parent = NULL;
895         struct ceph_osd *osd = NULL;
896
897         dout("__insert_osd %p osd%d\n", new, new->o_osd);
898         while (*p) {
899                 parent = *p;
900                 osd = rb_entry(parent, struct ceph_osd, o_node);
901                 if (new->o_osd < osd->o_osd)
902                         p = &(*p)->rb_left;
903                 else if (new->o_osd > osd->o_osd)
904                         p = &(*p)->rb_right;
905                 else
906                         BUG();
907         }
908
909         rb_link_node(&new->o_node, parent, p);
910         rb_insert_color(&new->o_node, &osdc->osds);
911 }
912
913 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
914 {
915         struct ceph_osd *osd;
916         struct rb_node *n = osdc->osds.rb_node;
917
918         while (n) {
919                 osd = rb_entry(n, struct ceph_osd, o_node);
920                 if (o < osd->o_osd)
921                         n = n->rb_left;
922                 else if (o > osd->o_osd)
923                         n = n->rb_right;
924                 else
925                         return osd;
926         }
927         return NULL;
928 }
929
930 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
931 {
932         schedule_delayed_work(&osdc->timeout_work,
933                         osdc->client->options->osd_keepalive_timeout * HZ);
934 }
935
936 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
937 {
938         cancel_delayed_work(&osdc->timeout_work);
939 }
940
941 /*
942  * Register request, assign tid.  If this is the first request, set up
943  * the timeout event.
944  */
945 static void __register_request(struct ceph_osd_client *osdc,
946                                struct ceph_osd_request *req)
947 {
948         req->r_tid = ++osdc->last_tid;
949         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
950         dout("__register_request %p tid %lld\n", req, req->r_tid);
951         __insert_request(osdc, req);
952         ceph_osdc_get_request(req);
953         osdc->num_requests++;
954         if (osdc->num_requests == 1) {
955                 dout(" first request, scheduling timeout\n");
956                 __schedule_osd_timeout(osdc);
957         }
958 }
959
960 /*
961  * called under osdc->request_mutex
962  */
963 static void __unregister_request(struct ceph_osd_client *osdc,
964                                  struct ceph_osd_request *req)
965 {
966         if (RB_EMPTY_NODE(&req->r_node)) {
967                 dout("__unregister_request %p tid %lld not registered\n",
968                         req, req->r_tid);
969                 return;
970         }
971
972         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
973         rb_erase(&req->r_node, &osdc->requests);
974         osdc->num_requests--;
975
976         if (req->r_osd) {
977                 /* make sure the original request isn't in flight. */
978                 ceph_msg_revoke(req->r_request);
979
980                 list_del_init(&req->r_osd_item);
981                 if (list_empty(&req->r_osd->o_requests) &&
982                     list_empty(&req->r_osd->o_linger_requests)) {
983                         dout("moving osd to %p lru\n", req->r_osd);
984                         __move_osd_to_lru(osdc, req->r_osd);
985                 }
986                 if (list_empty(&req->r_linger_item))
987                         req->r_osd = NULL;
988         }
989
990         list_del_init(&req->r_req_lru_item);
991         ceph_osdc_put_request(req);
992
993         if (osdc->num_requests == 0) {
994                 dout(" no requests, canceling timeout\n");
995                 __cancel_osd_timeout(osdc);
996         }
997 }
998
999 /*
1000  * Cancel a previously queued request message
1001  */
1002 static void __cancel_request(struct ceph_osd_request *req)
1003 {
1004         if (req->r_sent && req->r_osd) {
1005                 ceph_msg_revoke(req->r_request);
1006                 req->r_sent = 0;
1007         }
1008 }
1009
1010 static void __register_linger_request(struct ceph_osd_client *osdc,
1011                                     struct ceph_osd_request *req)
1012 {
1013         dout("__register_linger_request %p\n", req);
1014         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1015         if (req->r_osd)
1016                 list_add_tail(&req->r_linger_osd,
1017                               &req->r_osd->o_linger_requests);
1018 }
1019
1020 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1021                                         struct ceph_osd_request *req)
1022 {
1023         dout("__unregister_linger_request %p\n", req);
1024         list_del_init(&req->r_linger_item);
1025         if (req->r_osd) {
1026                 list_del_init(&req->r_linger_osd);
1027
1028                 if (list_empty(&req->r_osd->o_requests) &&
1029                     list_empty(&req->r_osd->o_linger_requests)) {
1030                         dout("moving osd to %p lru\n", req->r_osd);
1031                         __move_osd_to_lru(osdc, req->r_osd);
1032                 }
1033                 if (list_empty(&req->r_osd_item))
1034                         req->r_osd = NULL;
1035         }
1036 }
1037
1038 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1039                                          struct ceph_osd_request *req)
1040 {
1041         mutex_lock(&osdc->request_mutex);
1042         if (req->r_linger) {
1043                 __unregister_linger_request(osdc, req);
1044                 ceph_osdc_put_request(req);
1045         }
1046         mutex_unlock(&osdc->request_mutex);
1047 }
1048 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1049
1050 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1051                                   struct ceph_osd_request *req)
1052 {
1053         if (!req->r_linger) {
1054                 dout("set_request_linger %p\n", req);
1055                 req->r_linger = 1;
1056                 /*
1057                  * caller is now responsible for calling
1058                  * unregister_linger_request
1059                  */
1060                 ceph_osdc_get_request(req);
1061         }
1062 }
1063 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1064
1065 /*
1066  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1067  * (as needed), and set the request r_osd appropriately.  If there is
1068  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1069  * (unsent, homeless) or leave on in-flight lru.
1070  *
1071  * Return 0 if unchanged, 1 if changed, or negative on error.
1072  *
1073  * Caller should hold map_sem for read and request_mutex.
1074  */
1075 static int __map_request(struct ceph_osd_client *osdc,
1076                          struct ceph_osd_request *req, int force_resend)
1077 {
1078         struct ceph_pg pgid;
1079         int acting[CEPH_PG_MAX_SIZE];
1080         int o = -1, num = 0;
1081         int err;
1082
1083         dout("map_request %p tid %lld\n", req, req->r_tid);
1084         err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap,
1085                                 ceph_file_layout_pg_pool(req->r_file_layout));
1086         if (err) {
1087                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1088                 return err;
1089         }
1090         req->r_pgid = pgid;
1091
1092         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
1093         if (err > 0) {
1094                 o = acting[0];
1095                 num = err;
1096         }
1097
1098         if ((!force_resend &&
1099              req->r_osd && req->r_osd->o_osd == o &&
1100              req->r_sent >= req->r_osd->o_incarnation &&
1101              req->r_num_pg_osds == num &&
1102              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1103             (req->r_osd == NULL && o == -1))
1104                 return 0;  /* no change */
1105
1106         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1107              req->r_tid, pgid.pool, pgid.seed, o,
1108              req->r_osd ? req->r_osd->o_osd : -1);
1109
1110         /* record full pg acting set */
1111         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1112         req->r_num_pg_osds = num;
1113
1114         if (req->r_osd) {
1115                 __cancel_request(req);
1116                 list_del_init(&req->r_osd_item);
1117                 req->r_osd = NULL;
1118         }
1119
1120         req->r_osd = __lookup_osd(osdc, o);
1121         if (!req->r_osd && o >= 0) {
1122                 err = -ENOMEM;
1123                 req->r_osd = create_osd(osdc, o);
1124                 if (!req->r_osd) {
1125                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1126                         goto out;
1127                 }
1128
1129                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1130                 __insert_osd(osdc, req->r_osd);
1131
1132                 ceph_con_open(&req->r_osd->o_con,
1133                               CEPH_ENTITY_TYPE_OSD, o,
1134                               &osdc->osdmap->osd_addr[o]);
1135         }
1136
1137         if (req->r_osd) {
1138                 __remove_osd_from_lru(req->r_osd);
1139                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1140                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1141         } else {
1142                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1143         }
1144         err = 1;   /* osd or pg changed */
1145
1146 out:
1147         return err;
1148 }
1149
1150 /*
1151  * caller should hold map_sem (for read) and request_mutex
1152  */
1153 static void __send_request(struct ceph_osd_client *osdc,
1154                            struct ceph_osd_request *req)
1155 {
1156         void *p;
1157
1158         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1159              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1160              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1161
1162         /* fill in message content that changes each time we send it */
1163         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1164         put_unaligned_le32(req->r_flags, req->r_request_flags);
1165         put_unaligned_le64(req->r_pgid.pool, req->r_request_pool);
1166         p = req->r_request_pgid;
1167         ceph_encode_64(&p, req->r_pgid.pool);
1168         ceph_encode_32(&p, req->r_pgid.seed);
1169         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1170         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1171                sizeof(req->r_reassert_version));
1172
1173         req->r_stamp = jiffies;
1174         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1175
1176         ceph_msg_get(req->r_request); /* send consumes a ref */
1177         ceph_con_send(&req->r_osd->o_con, req->r_request);
1178         req->r_sent = req->r_osd->o_incarnation;
1179 }
1180
1181 /*
1182  * Send any requests in the queue (req_unsent).
1183  */
1184 static void __send_queued(struct ceph_osd_client *osdc)
1185 {
1186         struct ceph_osd_request *req, *tmp;
1187
1188         dout("__send_queued\n");
1189         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1190                 __send_request(osdc, req);
1191 }
1192
1193 /*
1194  * Timeout callback, called every N seconds when 1 or more osd
1195  * requests has been active for more than N seconds.  When this
1196  * happens, we ping all OSDs with requests who have timed out to
1197  * ensure any communications channel reset is detected.  Reset the
1198  * request timeouts another N seconds in the future as we go.
1199  * Reschedule the timeout event another N seconds in future (unless
1200  * there are no open requests).
1201  */
1202 static void handle_timeout(struct work_struct *work)
1203 {
1204         struct ceph_osd_client *osdc =
1205                 container_of(work, struct ceph_osd_client, timeout_work.work);
1206         struct ceph_osd_request *req;
1207         struct ceph_osd *osd;
1208         unsigned long keepalive =
1209                 osdc->client->options->osd_keepalive_timeout * HZ;
1210         struct list_head slow_osds;
1211         dout("timeout\n");
1212         down_read(&osdc->map_sem);
1213
1214         ceph_monc_request_next_osdmap(&osdc->client->monc);
1215
1216         mutex_lock(&osdc->request_mutex);
1217
1218         /*
1219          * ping osds that are a bit slow.  this ensures that if there
1220          * is a break in the TCP connection we will notice, and reopen
1221          * a connection with that osd (from the fault callback).
1222          */
1223         INIT_LIST_HEAD(&slow_osds);
1224         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1225                 if (time_before(jiffies, req->r_stamp + keepalive))
1226                         break;
1227
1228                 osd = req->r_osd;
1229                 BUG_ON(!osd);
1230                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1231                      req->r_tid, osd->o_osd);
1232                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1233         }
1234         while (!list_empty(&slow_osds)) {
1235                 osd = list_entry(slow_osds.next, struct ceph_osd,
1236                                  o_keepalive_item);
1237                 list_del_init(&osd->o_keepalive_item);
1238                 ceph_con_keepalive(&osd->o_con);
1239         }
1240
1241         __schedule_osd_timeout(osdc);
1242         __send_queued(osdc);
1243         mutex_unlock(&osdc->request_mutex);
1244         up_read(&osdc->map_sem);
1245 }
1246
1247 static void handle_osds_timeout(struct work_struct *work)
1248 {
1249         struct ceph_osd_client *osdc =
1250                 container_of(work, struct ceph_osd_client,
1251                              osds_timeout_work.work);
1252         unsigned long delay =
1253                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1254
1255         dout("osds timeout\n");
1256         down_read(&osdc->map_sem);
1257         remove_old_osds(osdc);
1258         up_read(&osdc->map_sem);
1259
1260         schedule_delayed_work(&osdc->osds_timeout_work,
1261                               round_jiffies_relative(delay));
1262 }
1263
1264 static void complete_request(struct ceph_osd_request *req)
1265 {
1266         if (req->r_safe_callback)
1267                 req->r_safe_callback(req, NULL);
1268         complete_all(&req->r_safe_completion);  /* fsync waiter */
1269 }
1270
1271 static int __decode_pgid(void **p, void *end, struct ceph_pg *pgid)
1272 {
1273         __u8 v;
1274
1275         ceph_decode_need(p, end, 1 + 8 + 4 + 4, bad);
1276         v = ceph_decode_8(p);
1277         if (v > 1) {
1278                 pr_warning("do not understand pg encoding %d > 1", v);
1279                 return -EINVAL;
1280         }
1281         pgid->pool = ceph_decode_64(p);
1282         pgid->seed = ceph_decode_32(p);
1283         *p += 4;
1284         return 0;
1285
1286 bad:
1287         pr_warning("incomplete pg encoding");
1288         return -EINVAL;
1289 }
1290
1291 /*
1292  * handle osd op reply.  either call the callback if it is specified,
1293  * or do the completion to wake up the waiting thread.
1294  */
1295 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1296                          struct ceph_connection *con)
1297 {
1298         void *p, *end;
1299         struct ceph_osd_request *req;
1300         u64 tid;
1301         int object_len;
1302         int numops, payload_len, flags;
1303         s32 result;
1304         s32 retry_attempt;
1305         struct ceph_pg pg;
1306         int err;
1307         u32 reassert_epoch;
1308         u64 reassert_version;
1309         u32 osdmap_epoch;
1310         int already_completed;
1311         int i;
1312
1313         tid = le64_to_cpu(msg->hdr.tid);
1314         dout("handle_reply %p tid %llu\n", msg, tid);
1315
1316         p = msg->front.iov_base;
1317         end = p + msg->front.iov_len;
1318
1319         ceph_decode_need(&p, end, 4, bad);
1320         object_len = ceph_decode_32(&p);
1321         ceph_decode_need(&p, end, object_len, bad);
1322         p += object_len;
1323
1324         err = __decode_pgid(&p, end, &pg);
1325         if (err)
1326                 goto bad;
1327
1328         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1329         flags = ceph_decode_64(&p);
1330         result = ceph_decode_32(&p);
1331         reassert_epoch = ceph_decode_32(&p);
1332         reassert_version = ceph_decode_64(&p);
1333         osdmap_epoch = ceph_decode_32(&p);
1334
1335         /* lookup */
1336         mutex_lock(&osdc->request_mutex);
1337         req = __lookup_request(osdc, tid);
1338         if (req == NULL) {
1339                 dout("handle_reply tid %llu dne\n", tid);
1340                 mutex_unlock(&osdc->request_mutex);
1341                 return;
1342         }
1343         ceph_osdc_get_request(req);
1344
1345         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1346              req, result);
1347
1348         ceph_decode_need(&p, end, 4, bad);
1349         numops = ceph_decode_32(&p);
1350         if (numops > CEPH_OSD_MAX_OP)
1351                 goto bad_put;
1352         if (numops != req->r_num_ops)
1353                 goto bad_put;
1354         payload_len = 0;
1355         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad);
1356         for (i = 0; i < numops; i++) {
1357                 struct ceph_osd_op *op = p;
1358                 int len;
1359
1360                 len = le32_to_cpu(op->payload_len);
1361                 req->r_reply_op_len[i] = len;
1362                 dout(" op %d has %d bytes\n", i, len);
1363                 payload_len += len;
1364                 p += sizeof(*op);
1365         }
1366         if (payload_len != le32_to_cpu(msg->hdr.data_len)) {
1367                 pr_warning("sum of op payload lens %d != data_len %d",
1368                            payload_len, le32_to_cpu(msg->hdr.data_len));
1369                 goto bad_put;
1370         }
1371
1372         ceph_decode_need(&p, end, 4 + numops * 4, bad);
1373         retry_attempt = ceph_decode_32(&p);
1374         for (i = 0; i < numops; i++)
1375                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1376
1377         /*
1378          * if this connection filled our message, drop our reference now, to
1379          * avoid a (safe but slower) revoke later.
1380          */
1381         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1382                 dout(" dropping con_filling_msg ref %p\n", con);
1383                 req->r_con_filling_msg = NULL;
1384                 con->ops->put(con);
1385         }
1386
1387         if (!req->r_got_reply) {
1388                 unsigned int bytes;
1389
1390                 req->r_result = result;
1391                 bytes = le32_to_cpu(msg->hdr.data_len);
1392                 dout("handle_reply result %d bytes %d\n", req->r_result,
1393                      bytes);
1394                 if (req->r_result == 0)
1395                         req->r_result = bytes;
1396
1397                 /* in case this is a write and we need to replay, */
1398                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1399                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1400
1401                 req->r_got_reply = 1;
1402         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1403                 dout("handle_reply tid %llu dup ack\n", tid);
1404                 mutex_unlock(&osdc->request_mutex);
1405                 goto done;
1406         }
1407
1408         dout("handle_reply tid %llu flags %d\n", tid, flags);
1409
1410         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1411                 __register_linger_request(osdc, req);
1412
1413         /* either this is a read, or we got the safe response */
1414         if (result < 0 ||
1415             (flags & CEPH_OSD_FLAG_ONDISK) ||
1416             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1417                 __unregister_request(osdc, req);
1418
1419         already_completed = req->r_completed;
1420         req->r_completed = 1;
1421         mutex_unlock(&osdc->request_mutex);
1422         if (already_completed)
1423                 goto done;
1424
1425         if (req->r_callback)
1426                 req->r_callback(req, msg);
1427         else
1428                 complete_all(&req->r_completion);
1429
1430         if (flags & CEPH_OSD_FLAG_ONDISK)
1431                 complete_request(req);
1432
1433 done:
1434         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1435         ceph_osdc_put_request(req);
1436         return;
1437
1438 bad_put:
1439         ceph_osdc_put_request(req);
1440 bad:
1441         pr_err("corrupt osd_op_reply got %d %d\n",
1442                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1443         ceph_msg_dump(msg);
1444 }
1445
1446 static void reset_changed_osds(struct ceph_osd_client *osdc)
1447 {
1448         struct rb_node *p, *n;
1449
1450         for (p = rb_first(&osdc->osds); p; p = n) {
1451                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1452
1453                 n = rb_next(p);
1454                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1455                     memcmp(&osd->o_con.peer_addr,
1456                            ceph_osd_addr(osdc->osdmap,
1457                                          osd->o_osd),
1458                            sizeof(struct ceph_entity_addr)) != 0)
1459                         __reset_osd(osdc, osd);
1460         }
1461 }
1462
1463 /*
1464  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1465  * no osd, request a new map.
1466  *
1467  * Caller should hold map_sem for read.
1468  */
1469 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1470 {
1471         struct ceph_osd_request *req, *nreq;
1472         struct rb_node *p;
1473         int needmap = 0;
1474         int err;
1475
1476         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1477         mutex_lock(&osdc->request_mutex);
1478         for (p = rb_first(&osdc->requests); p; ) {
1479                 req = rb_entry(p, struct ceph_osd_request, r_node);
1480                 p = rb_next(p);
1481
1482                 /*
1483                  * For linger requests that have not yet been
1484                  * registered, move them to the linger list; they'll
1485                  * be sent to the osd in the loop below.  Unregister
1486                  * the request before re-registering it as a linger
1487                  * request to ensure the __map_request() below
1488                  * will decide it needs to be sent.
1489                  */
1490                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1491                         dout("%p tid %llu restart on osd%d\n",
1492                              req, req->r_tid,
1493                              req->r_osd ? req->r_osd->o_osd : -1);
1494                         __unregister_request(osdc, req);
1495                         __register_linger_request(osdc, req);
1496                         continue;
1497                 }
1498
1499                 err = __map_request(osdc, req, force_resend);
1500                 if (err < 0)
1501                         continue;  /* error */
1502                 if (req->r_osd == NULL) {
1503                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1504                         needmap++;  /* request a newer map */
1505                 } else if (err > 0) {
1506                         if (!req->r_linger) {
1507                                 dout("%p tid %llu requeued on osd%d\n", req,
1508                                      req->r_tid,
1509                                      req->r_osd ? req->r_osd->o_osd : -1);
1510                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1511                         }
1512                 }
1513         }
1514
1515         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1516                                  r_linger_item) {
1517                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1518
1519                 err = __map_request(osdc, req, force_resend);
1520                 dout("__map_request returned %d\n", err);
1521                 if (err == 0)
1522                         continue;  /* no change and no osd was specified */
1523                 if (err < 0)
1524                         continue;  /* hrm! */
1525                 if (req->r_osd == NULL) {
1526                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1527                         needmap++;  /* request a newer map */
1528                         continue;
1529                 }
1530
1531                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1532                      req->r_osd ? req->r_osd->o_osd : -1);
1533                 __register_request(osdc, req);
1534                 __unregister_linger_request(osdc, req);
1535         }
1536         mutex_unlock(&osdc->request_mutex);
1537
1538         if (needmap) {
1539                 dout("%d requests for down osds, need new map\n", needmap);
1540                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1541         }
1542         reset_changed_osds(osdc);
1543 }
1544
1545
1546 /*
1547  * Process updated osd map.
1548  *
1549  * The message contains any number of incremental and full maps, normally
1550  * indicating some sort of topology change in the cluster.  Kick requests
1551  * off to different OSDs as needed.
1552  */
1553 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1554 {
1555         void *p, *end, *next;
1556         u32 nr_maps, maplen;
1557         u32 epoch;
1558         struct ceph_osdmap *newmap = NULL, *oldmap;
1559         int err;
1560         struct ceph_fsid fsid;
1561
1562         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1563         p = msg->front.iov_base;
1564         end = p + msg->front.iov_len;
1565
1566         /* verify fsid */
1567         ceph_decode_need(&p, end, sizeof(fsid), bad);
1568         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1569         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1570                 return;
1571
1572         down_write(&osdc->map_sem);
1573
1574         /* incremental maps */
1575         ceph_decode_32_safe(&p, end, nr_maps, bad);
1576         dout(" %d inc maps\n", nr_maps);
1577         while (nr_maps > 0) {
1578                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1579                 epoch = ceph_decode_32(&p);
1580                 maplen = ceph_decode_32(&p);
1581                 ceph_decode_need(&p, end, maplen, bad);
1582                 next = p + maplen;
1583                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1584                         dout("applying incremental map %u len %d\n",
1585                              epoch, maplen);
1586                         newmap = osdmap_apply_incremental(&p, next,
1587                                                           osdc->osdmap,
1588                                                           &osdc->client->msgr);
1589                         if (IS_ERR(newmap)) {
1590                                 err = PTR_ERR(newmap);
1591                                 goto bad;
1592                         }
1593                         BUG_ON(!newmap);
1594                         if (newmap != osdc->osdmap) {
1595                                 ceph_osdmap_destroy(osdc->osdmap);
1596                                 osdc->osdmap = newmap;
1597                         }
1598                         kick_requests(osdc, 0);
1599                 } else {
1600                         dout("ignoring incremental map %u len %d\n",
1601                              epoch, maplen);
1602                 }
1603                 p = next;
1604                 nr_maps--;
1605         }
1606         if (newmap)
1607                 goto done;
1608
1609         /* full maps */
1610         ceph_decode_32_safe(&p, end, nr_maps, bad);
1611         dout(" %d full maps\n", nr_maps);
1612         while (nr_maps) {
1613                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1614                 epoch = ceph_decode_32(&p);
1615                 maplen = ceph_decode_32(&p);
1616                 ceph_decode_need(&p, end, maplen, bad);
1617                 if (nr_maps > 1) {
1618                         dout("skipping non-latest full map %u len %d\n",
1619                              epoch, maplen);
1620                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1621                         dout("skipping full map %u len %d, "
1622                              "older than our %u\n", epoch, maplen,
1623                              osdc->osdmap->epoch);
1624                 } else {
1625                         int skipped_map = 0;
1626
1627                         dout("taking full map %u len %d\n", epoch, maplen);
1628                         newmap = osdmap_decode(&p, p+maplen);
1629                         if (IS_ERR(newmap)) {
1630                                 err = PTR_ERR(newmap);
1631                                 goto bad;
1632                         }
1633                         BUG_ON(!newmap);
1634                         oldmap = osdc->osdmap;
1635                         osdc->osdmap = newmap;
1636                         if (oldmap) {
1637                                 if (oldmap->epoch + 1 < newmap->epoch)
1638                                         skipped_map = 1;
1639                                 ceph_osdmap_destroy(oldmap);
1640                         }
1641                         kick_requests(osdc, skipped_map);
1642                 }
1643                 p += maplen;
1644                 nr_maps--;
1645         }
1646
1647 done:
1648         downgrade_write(&osdc->map_sem);
1649         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1650
1651         /*
1652          * subscribe to subsequent osdmap updates if full to ensure
1653          * we find out when we are no longer full and stop returning
1654          * ENOSPC.
1655          */
1656         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1657                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1658
1659         mutex_lock(&osdc->request_mutex);
1660         __send_queued(osdc);
1661         mutex_unlock(&osdc->request_mutex);
1662         up_read(&osdc->map_sem);
1663         wake_up_all(&osdc->client->auth_wq);
1664         return;
1665
1666 bad:
1667         pr_err("osdc handle_map corrupt msg\n");
1668         ceph_msg_dump(msg);
1669         up_write(&osdc->map_sem);
1670         return;
1671 }
1672
1673 /*
1674  * watch/notify callback event infrastructure
1675  *
1676  * These callbacks are used both for watch and notify operations.
1677  */
1678 static void __release_event(struct kref *kref)
1679 {
1680         struct ceph_osd_event *event =
1681                 container_of(kref, struct ceph_osd_event, kref);
1682
1683         dout("__release_event %p\n", event);
1684         kfree(event);
1685 }
1686
1687 static void get_event(struct ceph_osd_event *event)
1688 {
1689         kref_get(&event->kref);
1690 }
1691
1692 void ceph_osdc_put_event(struct ceph_osd_event *event)
1693 {
1694         kref_put(&event->kref, __release_event);
1695 }
1696 EXPORT_SYMBOL(ceph_osdc_put_event);
1697
1698 static void __insert_event(struct ceph_osd_client *osdc,
1699                              struct ceph_osd_event *new)
1700 {
1701         struct rb_node **p = &osdc->event_tree.rb_node;
1702         struct rb_node *parent = NULL;
1703         struct ceph_osd_event *event = NULL;
1704
1705         while (*p) {
1706                 parent = *p;
1707                 event = rb_entry(parent, struct ceph_osd_event, node);
1708                 if (new->cookie < event->cookie)
1709                         p = &(*p)->rb_left;
1710                 else if (new->cookie > event->cookie)
1711                         p = &(*p)->rb_right;
1712                 else
1713                         BUG();
1714         }
1715
1716         rb_link_node(&new->node, parent, p);
1717         rb_insert_color(&new->node, &osdc->event_tree);
1718 }
1719
1720 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1721                                                 u64 cookie)
1722 {
1723         struct rb_node **p = &osdc->event_tree.rb_node;
1724         struct rb_node *parent = NULL;
1725         struct ceph_osd_event *event = NULL;
1726
1727         while (*p) {
1728                 parent = *p;
1729                 event = rb_entry(parent, struct ceph_osd_event, node);
1730                 if (cookie < event->cookie)
1731                         p = &(*p)->rb_left;
1732                 else if (cookie > event->cookie)
1733                         p = &(*p)->rb_right;
1734                 else
1735                         return event;
1736         }
1737         return NULL;
1738 }
1739
1740 static void __remove_event(struct ceph_osd_event *event)
1741 {
1742         struct ceph_osd_client *osdc = event->osdc;
1743
1744         if (!RB_EMPTY_NODE(&event->node)) {
1745                 dout("__remove_event removed %p\n", event);
1746                 rb_erase(&event->node, &osdc->event_tree);
1747                 ceph_osdc_put_event(event);
1748         } else {
1749                 dout("__remove_event didn't remove %p\n", event);
1750         }
1751 }
1752
1753 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1754                            void (*event_cb)(u64, u64, u8, void *),
1755                            void *data, struct ceph_osd_event **pevent)
1756 {
1757         struct ceph_osd_event *event;
1758
1759         event = kmalloc(sizeof(*event), GFP_NOIO);
1760         if (!event)
1761                 return -ENOMEM;
1762
1763         dout("create_event %p\n", event);
1764         event->cb = event_cb;
1765         event->one_shot = 0;
1766         event->data = data;
1767         event->osdc = osdc;
1768         INIT_LIST_HEAD(&event->osd_node);
1769         RB_CLEAR_NODE(&event->node);
1770         kref_init(&event->kref);   /* one ref for us */
1771         kref_get(&event->kref);    /* one ref for the caller */
1772
1773         spin_lock(&osdc->event_lock);
1774         event->cookie = ++osdc->event_count;
1775         __insert_event(osdc, event);
1776         spin_unlock(&osdc->event_lock);
1777
1778         *pevent = event;
1779         return 0;
1780 }
1781 EXPORT_SYMBOL(ceph_osdc_create_event);
1782
1783 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1784 {
1785         struct ceph_osd_client *osdc = event->osdc;
1786
1787         dout("cancel_event %p\n", event);
1788         spin_lock(&osdc->event_lock);
1789         __remove_event(event);
1790         spin_unlock(&osdc->event_lock);
1791         ceph_osdc_put_event(event); /* caller's */
1792 }
1793 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1794
1795
1796 static void do_event_work(struct work_struct *work)
1797 {
1798         struct ceph_osd_event_work *event_work =
1799                 container_of(work, struct ceph_osd_event_work, work);
1800         struct ceph_osd_event *event = event_work->event;
1801         u64 ver = event_work->ver;
1802         u64 notify_id = event_work->notify_id;
1803         u8 opcode = event_work->opcode;
1804
1805         dout("do_event_work completing %p\n", event);
1806         event->cb(ver, notify_id, opcode, event->data);
1807         dout("do_event_work completed %p\n", event);
1808         ceph_osdc_put_event(event);
1809         kfree(event_work);
1810 }
1811
1812
1813 /*
1814  * Process osd watch notifications
1815  */
1816 static void handle_watch_notify(struct ceph_osd_client *osdc,
1817                                 struct ceph_msg *msg)
1818 {
1819         void *p, *end;
1820         u8 proto_ver;
1821         u64 cookie, ver, notify_id;
1822         u8 opcode;
1823         struct ceph_osd_event *event;
1824         struct ceph_osd_event_work *event_work;
1825
1826         p = msg->front.iov_base;
1827         end = p + msg->front.iov_len;
1828
1829         ceph_decode_8_safe(&p, end, proto_ver, bad);
1830         ceph_decode_8_safe(&p, end, opcode, bad);
1831         ceph_decode_64_safe(&p, end, cookie, bad);
1832         ceph_decode_64_safe(&p, end, ver, bad);
1833         ceph_decode_64_safe(&p, end, notify_id, bad);
1834
1835         spin_lock(&osdc->event_lock);
1836         event = __find_event(osdc, cookie);
1837         if (event) {
1838                 BUG_ON(event->one_shot);
1839                 get_event(event);
1840         }
1841         spin_unlock(&osdc->event_lock);
1842         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1843              cookie, ver, event);
1844         if (event) {
1845                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1846                 if (!event_work) {
1847                         dout("ERROR: could not allocate event_work\n");
1848                         goto done_err;
1849                 }
1850                 INIT_WORK(&event_work->work, do_event_work);
1851                 event_work->event = event;
1852                 event_work->ver = ver;
1853                 event_work->notify_id = notify_id;
1854                 event_work->opcode = opcode;
1855                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1856                         dout("WARNING: failed to queue notify event work\n");
1857                         goto done_err;
1858                 }
1859         }
1860
1861         return;
1862
1863 done_err:
1864         ceph_osdc_put_event(event);
1865         return;
1866
1867 bad:
1868         pr_err("osdc handle_watch_notify corrupt msg\n");
1869         return;
1870 }
1871
1872 static void ceph_osdc_msg_data_set(struct ceph_msg *msg,
1873                                 struct ceph_osd_data *osd_data)
1874 {
1875         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
1876                 BUG_ON(osd_data->length > (u64) SIZE_MAX);
1877                 if (osd_data->length)
1878                         ceph_msg_data_set_pages(msg, osd_data->pages,
1879                                 osd_data->length, osd_data->alignment);
1880         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
1881                 BUG_ON(!osd_data->pagelist->length);
1882                 ceph_msg_data_set_pagelist(msg, osd_data->pagelist);
1883 #ifdef CONFIG_BLOCK
1884         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
1885                 ceph_msg_data_set_bio(msg, osd_data->bio);
1886 #endif
1887         } else {
1888                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
1889         }
1890 }
1891
1892 /*
1893  * Register request, send initial attempt.
1894  */
1895 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1896                             struct ceph_osd_request *req,
1897                             bool nofail)
1898 {
1899         int rc = 0;
1900
1901         /* Set up response incoming data and request outgoing data fields */
1902
1903         ceph_osdc_msg_data_set(req->r_reply, &req->r_data_in);
1904         ceph_osdc_msg_data_set(req->r_request, &req->r_data_out);
1905
1906         down_read(&osdc->map_sem);
1907         mutex_lock(&osdc->request_mutex);
1908         __register_request(osdc, req);
1909         WARN_ON(req->r_sent);
1910         rc = __map_request(osdc, req, 0);
1911         if (rc < 0) {
1912                 if (nofail) {
1913                         dout("osdc_start_request failed map, "
1914                                 " will retry %lld\n", req->r_tid);
1915                         rc = 0;
1916                 }
1917                 goto out_unlock;
1918         }
1919         if (req->r_osd == NULL) {
1920                 dout("send_request %p no up osds in pg\n", req);
1921                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1922         } else {
1923                 __send_queued(osdc);
1924         }
1925         rc = 0;
1926 out_unlock:
1927         mutex_unlock(&osdc->request_mutex);
1928         up_read(&osdc->map_sem);
1929         return rc;
1930 }
1931 EXPORT_SYMBOL(ceph_osdc_start_request);
1932
1933 /*
1934  * wait for a request to complete
1935  */
1936 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1937                            struct ceph_osd_request *req)
1938 {
1939         int rc;
1940
1941         rc = wait_for_completion_interruptible(&req->r_completion);
1942         if (rc < 0) {
1943                 mutex_lock(&osdc->request_mutex);
1944                 __cancel_request(req);
1945                 __unregister_request(osdc, req);
1946                 mutex_unlock(&osdc->request_mutex);
1947                 complete_request(req);
1948                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1949                 return rc;
1950         }
1951
1952         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1953         return req->r_result;
1954 }
1955 EXPORT_SYMBOL(ceph_osdc_wait_request);
1956
1957 /*
1958  * sync - wait for all in-flight requests to flush.  avoid starvation.
1959  */
1960 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1961 {
1962         struct ceph_osd_request *req;
1963         u64 last_tid, next_tid = 0;
1964
1965         mutex_lock(&osdc->request_mutex);
1966         last_tid = osdc->last_tid;
1967         while (1) {
1968                 req = __lookup_request_ge(osdc, next_tid);
1969                 if (!req)
1970                         break;
1971                 if (req->r_tid > last_tid)
1972                         break;
1973
1974                 next_tid = req->r_tid + 1;
1975                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1976                         continue;
1977
1978                 ceph_osdc_get_request(req);
1979                 mutex_unlock(&osdc->request_mutex);
1980                 dout("sync waiting on tid %llu (last is %llu)\n",
1981                      req->r_tid, last_tid);
1982                 wait_for_completion(&req->r_safe_completion);
1983                 mutex_lock(&osdc->request_mutex);
1984                 ceph_osdc_put_request(req);
1985         }
1986         mutex_unlock(&osdc->request_mutex);
1987         dout("sync done (thru tid %llu)\n", last_tid);
1988 }
1989 EXPORT_SYMBOL(ceph_osdc_sync);
1990
1991 /*
1992  * init, shutdown
1993  */
1994 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1995 {
1996         int err;
1997
1998         dout("init\n");
1999         osdc->client = client;
2000         osdc->osdmap = NULL;
2001         init_rwsem(&osdc->map_sem);
2002         init_completion(&osdc->map_waiters);
2003         osdc->last_requested_map = 0;
2004         mutex_init(&osdc->request_mutex);
2005         osdc->last_tid = 0;
2006         osdc->osds = RB_ROOT;
2007         INIT_LIST_HEAD(&osdc->osd_lru);
2008         osdc->requests = RB_ROOT;
2009         INIT_LIST_HEAD(&osdc->req_lru);
2010         INIT_LIST_HEAD(&osdc->req_unsent);
2011         INIT_LIST_HEAD(&osdc->req_notarget);
2012         INIT_LIST_HEAD(&osdc->req_linger);
2013         osdc->num_requests = 0;
2014         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2015         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2016         spin_lock_init(&osdc->event_lock);
2017         osdc->event_tree = RB_ROOT;
2018         osdc->event_count = 0;
2019
2020         schedule_delayed_work(&osdc->osds_timeout_work,
2021            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2022
2023         err = -ENOMEM;
2024         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2025                                         sizeof(struct ceph_osd_request));
2026         if (!osdc->req_mempool)
2027                 goto out;
2028
2029         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2030                                 OSD_OP_FRONT_LEN, 10, true,
2031                                 "osd_op");
2032         if (err < 0)
2033                 goto out_mempool;
2034         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2035                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2036                                 "osd_op_reply");
2037         if (err < 0)
2038                 goto out_msgpool;
2039
2040         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2041         if (IS_ERR(osdc->notify_wq)) {
2042                 err = PTR_ERR(osdc->notify_wq);
2043                 osdc->notify_wq = NULL;
2044                 goto out_msgpool;
2045         }
2046         return 0;
2047
2048 out_msgpool:
2049         ceph_msgpool_destroy(&osdc->msgpool_op);
2050 out_mempool:
2051         mempool_destroy(osdc->req_mempool);
2052 out:
2053         return err;
2054 }
2055
2056 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2057 {
2058         flush_workqueue(osdc->notify_wq);
2059         destroy_workqueue(osdc->notify_wq);
2060         cancel_delayed_work_sync(&osdc->timeout_work);
2061         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2062         if (osdc->osdmap) {
2063                 ceph_osdmap_destroy(osdc->osdmap);
2064                 osdc->osdmap = NULL;
2065         }
2066         remove_all_osds(osdc);
2067         mempool_destroy(osdc->req_mempool);
2068         ceph_msgpool_destroy(&osdc->msgpool_op);
2069         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2070 }
2071
2072 /*
2073  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2074  * *plen.  Return number of bytes read, or error.
2075  */
2076 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2077                         struct ceph_vino vino, struct ceph_file_layout *layout,
2078                         u64 off, u64 *plen,
2079                         u32 truncate_seq, u64 truncate_size,
2080                         struct page **pages, int num_pages, int page_align)
2081 {
2082         struct ceph_osd_request *req;
2083         struct ceph_osd_data *osd_data;
2084         int rc = 0;
2085
2086         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2087              vino.snap, off, *plen);
2088         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
2089                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2090                                     NULL, 0, truncate_seq, truncate_size, NULL,
2091                                     false);
2092         if (IS_ERR(req))
2093                 return PTR_ERR(req);
2094
2095         /* it may be a short read due to an object boundary */
2096
2097         osd_data = &req->r_data_in;
2098         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
2099         osd_data->pages = pages;
2100         osd_data->length = *plen;
2101         osd_data->alignment = page_align;
2102
2103         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2104              off, *plen, osd_data->length, page_align);
2105
2106         rc = ceph_osdc_start_request(osdc, req, false);
2107         if (!rc)
2108                 rc = ceph_osdc_wait_request(osdc, req);
2109
2110         ceph_osdc_put_request(req);
2111         dout("readpages result %d\n", rc);
2112         return rc;
2113 }
2114 EXPORT_SYMBOL(ceph_osdc_readpages);
2115
2116 /*
2117  * do a synchronous write on N pages
2118  */
2119 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2120                          struct ceph_file_layout *layout,
2121                          struct ceph_snap_context *snapc,
2122                          u64 off, u64 len,
2123                          u32 truncate_seq, u64 truncate_size,
2124                          struct timespec *mtime,
2125                          struct page **pages, int num_pages)
2126 {
2127         struct ceph_osd_request *req;
2128         struct ceph_osd_data *osd_data;
2129         int rc = 0;
2130         int page_align = off & ~PAGE_MASK;
2131
2132         BUG_ON(vino.snap != CEPH_NOSNAP);
2133         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
2134                                     CEPH_OSD_OP_WRITE,
2135                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2136                                     snapc, 0,
2137                                     truncate_seq, truncate_size, mtime,
2138                                     true);
2139         if (IS_ERR(req))
2140                 return PTR_ERR(req);
2141
2142         /* it may be a short write due to an object boundary */
2143         osd_data = &req->r_data_out;
2144         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
2145         osd_data->pages = pages;
2146         osd_data->length = len;
2147         osd_data->alignment = page_align;
2148         dout("writepages %llu~%llu (%llu bytes)\n", off, len, osd_data->length);
2149
2150         rc = ceph_osdc_start_request(osdc, req, true);
2151         if (!rc)
2152                 rc = ceph_osdc_wait_request(osdc, req);
2153
2154         ceph_osdc_put_request(req);
2155         if (rc == 0)
2156                 rc = len;
2157         dout("writepages result %d\n", rc);
2158         return rc;
2159 }
2160 EXPORT_SYMBOL(ceph_osdc_writepages);
2161
2162 /*
2163  * handle incoming message
2164  */
2165 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2166 {
2167         struct ceph_osd *osd = con->private;
2168         struct ceph_osd_client *osdc;
2169         int type = le16_to_cpu(msg->hdr.type);
2170
2171         if (!osd)
2172                 goto out;
2173         osdc = osd->o_osdc;
2174
2175         switch (type) {
2176         case CEPH_MSG_OSD_MAP:
2177                 ceph_osdc_handle_map(osdc, msg);
2178                 break;
2179         case CEPH_MSG_OSD_OPREPLY:
2180                 handle_reply(osdc, msg, con);
2181                 break;
2182         case CEPH_MSG_WATCH_NOTIFY:
2183                 handle_watch_notify(osdc, msg);
2184                 break;
2185
2186         default:
2187                 pr_err("received unknown message type %d %s\n", type,
2188                        ceph_msg_type_name(type));
2189         }
2190 out:
2191         ceph_msg_put(msg);
2192 }
2193
2194 /*
2195  * lookup and return message for incoming reply.  set up reply message
2196  * pages.
2197  */
2198 static struct ceph_msg *get_reply(struct ceph_connection *con,
2199                                   struct ceph_msg_header *hdr,
2200                                   int *skip)
2201 {
2202         struct ceph_osd *osd = con->private;
2203         struct ceph_osd_client *osdc = osd->o_osdc;
2204         struct ceph_msg *m;
2205         struct ceph_osd_request *req;
2206         int front = le32_to_cpu(hdr->front_len);
2207         int data_len = le32_to_cpu(hdr->data_len);
2208         u64 tid;
2209
2210         tid = le64_to_cpu(hdr->tid);
2211         mutex_lock(&osdc->request_mutex);
2212         req = __lookup_request(osdc, tid);
2213         if (!req) {
2214                 *skip = 1;
2215                 m = NULL;
2216                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2217                      osd->o_osd);
2218                 goto out;
2219         }
2220
2221         if (req->r_con_filling_msg) {
2222                 dout("%s revoking msg %p from old con %p\n", __func__,
2223                      req->r_reply, req->r_con_filling_msg);
2224                 ceph_msg_revoke_incoming(req->r_reply);
2225                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2226                 req->r_con_filling_msg = NULL;
2227         }
2228
2229         if (front > req->r_reply->front.iov_len) {
2230                 pr_warning("get_reply front %d > preallocated %d\n",
2231                            front, (int)req->r_reply->front.iov_len);
2232                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2233                 if (!m)
2234                         goto out;
2235                 ceph_msg_put(req->r_reply);
2236                 req->r_reply = m;
2237         }
2238         m = ceph_msg_get(req->r_reply);
2239
2240         if (data_len > 0) {
2241                 struct ceph_osd_data *osd_data = &req->r_data_in;
2242
2243                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2244                         if (osd_data->pages &&
2245                                 unlikely(osd_data->length < data_len)) {
2246
2247                                 pr_warning("tid %lld reply has %d bytes "
2248                                         "we had only %llu bytes ready\n",
2249                                         tid, data_len, osd_data->length);
2250                                 *skip = 1;
2251                                 ceph_msg_put(m);
2252                                 m = NULL;
2253                                 goto out;
2254                         }
2255                 }
2256         }
2257         *skip = 0;
2258         req->r_con_filling_msg = con->ops->get(con);
2259         dout("get_reply tid %lld %p\n", tid, m);
2260
2261 out:
2262         mutex_unlock(&osdc->request_mutex);
2263         return m;
2264
2265 }
2266
2267 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2268                                   struct ceph_msg_header *hdr,
2269                                   int *skip)
2270 {
2271         struct ceph_osd *osd = con->private;
2272         int type = le16_to_cpu(hdr->type);
2273         int front = le32_to_cpu(hdr->front_len);
2274
2275         *skip = 0;
2276         switch (type) {
2277         case CEPH_MSG_OSD_MAP:
2278         case CEPH_MSG_WATCH_NOTIFY:
2279                 return ceph_msg_new(type, front, GFP_NOFS, false);
2280         case CEPH_MSG_OSD_OPREPLY:
2281                 return get_reply(con, hdr, skip);
2282         default:
2283                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2284                         osd->o_osd);
2285                 *skip = 1;
2286                 return NULL;
2287         }
2288 }
2289
2290 /*
2291  * Wrappers to refcount containing ceph_osd struct
2292  */
2293 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2294 {
2295         struct ceph_osd *osd = con->private;
2296         if (get_osd(osd))
2297                 return con;
2298         return NULL;
2299 }
2300
2301 static void put_osd_con(struct ceph_connection *con)
2302 {
2303         struct ceph_osd *osd = con->private;
2304         put_osd(osd);
2305 }
2306
2307 /*
2308  * authentication
2309  */
2310 /*
2311  * Note: returned pointer is the address of a structure that's
2312  * managed separately.  Caller must *not* attempt to free it.
2313  */
2314 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2315                                         int *proto, int force_new)
2316 {
2317         struct ceph_osd *o = con->private;
2318         struct ceph_osd_client *osdc = o->o_osdc;
2319         struct ceph_auth_client *ac = osdc->client->monc.auth;
2320         struct ceph_auth_handshake *auth = &o->o_auth;
2321
2322         if (force_new && auth->authorizer) {
2323                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2324                 auth->authorizer = NULL;
2325         }
2326         if (!auth->authorizer) {
2327                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2328                                                       auth);
2329                 if (ret)
2330                         return ERR_PTR(ret);
2331         } else {
2332                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2333                                                      auth);
2334                 if (ret)
2335                         return ERR_PTR(ret);
2336         }
2337         *proto = ac->protocol;
2338
2339         return auth;
2340 }
2341
2342
2343 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2344 {
2345         struct ceph_osd *o = con->private;
2346         struct ceph_osd_client *osdc = o->o_osdc;
2347         struct ceph_auth_client *ac = osdc->client->monc.auth;
2348
2349         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2350 }
2351
2352 static int invalidate_authorizer(struct ceph_connection *con)
2353 {
2354         struct ceph_osd *o = con->private;
2355         struct ceph_osd_client *osdc = o->o_osdc;
2356         struct ceph_auth_client *ac = osdc->client->monc.auth;
2357
2358         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2359         return ceph_monc_validate_auth(&osdc->client->monc);
2360 }
2361
2362 static const struct ceph_connection_operations osd_con_ops = {
2363         .get = get_osd_con,
2364         .put = put_osd_con,
2365         .dispatch = dispatch,
2366         .get_authorizer = get_authorizer,
2367         .verify_authorizer_reply = verify_authorizer_reply,
2368         .invalidate_authorizer = invalidate_authorizer,
2369         .alloc_msg = alloc_msg,
2370         .fault = osd_reset,
2371 };