xref: /linux/net/ceph/osd_client.c (revision 092e0e7e520a1fca03e13c9f2d157432a8657ff2)
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 static int __kick_requests(struct ceph_osd_client *osdc,
26 			  struct ceph_osd *kickosd);
27 
28 static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
29 
30 static int op_needs_trail(int op)
31 {
32 	switch (op) {
33 	case CEPH_OSD_OP_GETXATTR:
34 	case CEPH_OSD_OP_SETXATTR:
35 	case CEPH_OSD_OP_CMPXATTR:
36 	case CEPH_OSD_OP_CALL:
37 		return 1;
38 	default:
39 		return 0;
40 	}
41 }
42 
43 static int op_has_extent(int op)
44 {
45 	return (op == CEPH_OSD_OP_READ ||
46 		op == CEPH_OSD_OP_WRITE);
47 }
48 
49 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
50 			struct ceph_file_layout *layout,
51 			u64 snapid,
52 			u64 off, u64 *plen, u64 *bno,
53 			struct ceph_osd_request *req,
54 			struct ceph_osd_req_op *op)
55 {
56 	struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
57 	u64 orig_len = *plen;
58 	u64 objoff, objlen;    /* extent in object */
59 
60 	reqhead->snapid = cpu_to_le64(snapid);
61 
62 	/* object extent? */
63 	ceph_calc_file_object_mapping(layout, off, plen, bno,
64 				      &objoff, &objlen);
65 	if (*plen < orig_len)
66 		dout(" skipping last %llu, final file extent %llu~%llu\n",
67 		     orig_len - *plen, off, *plen);
68 
69 	if (op_has_extent(op->op)) {
70 		op->extent.offset = objoff;
71 		op->extent.length = objlen;
72 	}
73 	req->r_num_pages = calc_pages_for(off, *plen);
74 	if (op->op == CEPH_OSD_OP_WRITE)
75 		op->payload_len = *plen;
76 
77 	dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
78 	     *bno, objoff, objlen, req->r_num_pages);
79 
80 }
81 EXPORT_SYMBOL(ceph_calc_raw_layout);
82 
83 /*
84  * Implement client access to distributed object storage cluster.
85  *
86  * All data objects are stored within a cluster/cloud of OSDs, or
87  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
88  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
89  * remote daemons serving up and coordinating consistent and safe
90  * access to storage.
91  *
92  * Cluster membership and the mapping of data objects onto storage devices
93  * are described by the osd map.
94  *
95  * We keep track of pending OSD requests (read, write), resubmit
96  * requests to different OSDs when the cluster topology/data layout
97  * change, or retry the affected requests when the communications
98  * channel with an OSD is reset.
99  */
100 
101 /*
102  * calculate the mapping of a file extent onto an object, and fill out the
103  * request accordingly.  shorten extent as necessary if it crosses an
104  * object boundary.
105  *
106  * fill osd op in request message.
107  */
108 static void calc_layout(struct ceph_osd_client *osdc,
109 			struct ceph_vino vino,
110 			struct ceph_file_layout *layout,
111 			u64 off, u64 *plen,
112 			struct ceph_osd_request *req,
113 			struct ceph_osd_req_op *op)
114 {
115 	u64 bno;
116 
117 	ceph_calc_raw_layout(osdc, layout, vino.snap, off,
118 			     plen, &bno, req, op);
119 
120 	sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
121 	req->r_oid_len = strlen(req->r_oid);
122 }
123 
124 /*
125  * requests
126  */
127 void ceph_osdc_release_request(struct kref *kref)
128 {
129 	struct ceph_osd_request *req = container_of(kref,
130 						    struct ceph_osd_request,
131 						    r_kref);
132 
133 	if (req->r_request)
134 		ceph_msg_put(req->r_request);
135 	if (req->r_reply)
136 		ceph_msg_put(req->r_reply);
137 	if (req->r_con_filling_msg) {
138 		dout("release_request revoking pages %p from con %p\n",
139 		     req->r_pages, req->r_con_filling_msg);
140 		ceph_con_revoke_message(req->r_con_filling_msg,
141 				      req->r_reply);
142 		ceph_con_put(req->r_con_filling_msg);
143 	}
144 	if (req->r_own_pages)
145 		ceph_release_page_vector(req->r_pages,
146 					 req->r_num_pages);
147 #ifdef CONFIG_BLOCK
148 	if (req->r_bio)
149 		bio_put(req->r_bio);
150 #endif
151 	ceph_put_snap_context(req->r_snapc);
152 	if (req->r_trail) {
153 		ceph_pagelist_release(req->r_trail);
154 		kfree(req->r_trail);
155 	}
156 	if (req->r_mempool)
157 		mempool_free(req, req->r_osdc->req_mempool);
158 	else
159 		kfree(req);
160 }
161 EXPORT_SYMBOL(ceph_osdc_release_request);
162 
163 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
164 {
165 	int i = 0;
166 
167 	if (needs_trail)
168 		*needs_trail = 0;
169 	while (ops[i].op) {
170 		if (needs_trail && op_needs_trail(ops[i].op))
171 			*needs_trail = 1;
172 		i++;
173 	}
174 
175 	return i;
176 }
177 
178 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
179 					       int flags,
180 					       struct ceph_snap_context *snapc,
181 					       struct ceph_osd_req_op *ops,
182 					       bool use_mempool,
183 					       gfp_t gfp_flags,
184 					       struct page **pages,
185 					       struct bio *bio)
186 {
187 	struct ceph_osd_request *req;
188 	struct ceph_msg *msg;
189 	int needs_trail;
190 	int num_op = get_num_ops(ops, &needs_trail);
191 	size_t msg_size = sizeof(struct ceph_osd_request_head);
192 
193 	msg_size += num_op*sizeof(struct ceph_osd_op);
194 
195 	if (use_mempool) {
196 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
197 		memset(req, 0, sizeof(*req));
198 	} else {
199 		req = kzalloc(sizeof(*req), gfp_flags);
200 	}
201 	if (req == NULL)
202 		return NULL;
203 
204 	req->r_osdc = osdc;
205 	req->r_mempool = use_mempool;
206 
207 	kref_init(&req->r_kref);
208 	init_completion(&req->r_completion);
209 	init_completion(&req->r_safe_completion);
210 	INIT_LIST_HEAD(&req->r_unsafe_item);
211 	req->r_flags = flags;
212 
213 	WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
214 
215 	/* create reply message */
216 	if (use_mempool)
217 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
218 	else
219 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
220 				   OSD_OPREPLY_FRONT_LEN, gfp_flags);
221 	if (!msg) {
222 		ceph_osdc_put_request(req);
223 		return NULL;
224 	}
225 	req->r_reply = msg;
226 
227 	/* allocate space for the trailing data */
228 	if (needs_trail) {
229 		req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
230 		if (!req->r_trail) {
231 			ceph_osdc_put_request(req);
232 			return NULL;
233 		}
234 		ceph_pagelist_init(req->r_trail);
235 	}
236 	/* create request message; allow space for oid */
237 	msg_size += 40;
238 	if (snapc)
239 		msg_size += sizeof(u64) * snapc->num_snaps;
240 	if (use_mempool)
241 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
242 	else
243 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags);
244 	if (!msg) {
245 		ceph_osdc_put_request(req);
246 		return NULL;
247 	}
248 
249 	msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
250 	memset(msg->front.iov_base, 0, msg->front.iov_len);
251 
252 	req->r_request = msg;
253 	req->r_pages = pages;
254 #ifdef CONFIG_BLOCK
255 	if (bio) {
256 		req->r_bio = bio;
257 		bio_get(req->r_bio);
258 	}
259 #endif
260 
261 	return req;
262 }
263 EXPORT_SYMBOL(ceph_osdc_alloc_request);
264 
265 static void osd_req_encode_op(struct ceph_osd_request *req,
266 			      struct ceph_osd_op *dst,
267 			      struct ceph_osd_req_op *src)
268 {
269 	dst->op = cpu_to_le16(src->op);
270 
271 	switch (dst->op) {
272 	case CEPH_OSD_OP_READ:
273 	case CEPH_OSD_OP_WRITE:
274 		dst->extent.offset =
275 			cpu_to_le64(src->extent.offset);
276 		dst->extent.length =
277 			cpu_to_le64(src->extent.length);
278 		dst->extent.truncate_size =
279 			cpu_to_le64(src->extent.truncate_size);
280 		dst->extent.truncate_seq =
281 			cpu_to_le32(src->extent.truncate_seq);
282 		break;
283 
284 	case CEPH_OSD_OP_GETXATTR:
285 	case CEPH_OSD_OP_SETXATTR:
286 	case CEPH_OSD_OP_CMPXATTR:
287 		BUG_ON(!req->r_trail);
288 
289 		dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
290 		dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
291 		dst->xattr.cmp_op = src->xattr.cmp_op;
292 		dst->xattr.cmp_mode = src->xattr.cmp_mode;
293 		ceph_pagelist_append(req->r_trail, src->xattr.name,
294 				     src->xattr.name_len);
295 		ceph_pagelist_append(req->r_trail, src->xattr.val,
296 				     src->xattr.value_len);
297 		break;
298 	case CEPH_OSD_OP_CALL:
299 		BUG_ON(!req->r_trail);
300 
301 		dst->cls.class_len = src->cls.class_len;
302 		dst->cls.method_len = src->cls.method_len;
303 		dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
304 
305 		ceph_pagelist_append(req->r_trail, src->cls.class_name,
306 				     src->cls.class_len);
307 		ceph_pagelist_append(req->r_trail, src->cls.method_name,
308 				     src->cls.method_len);
309 		ceph_pagelist_append(req->r_trail, src->cls.indata,
310 				     src->cls.indata_len);
311 		break;
312 	case CEPH_OSD_OP_ROLLBACK:
313 		dst->snap.snapid = cpu_to_le64(src->snap.snapid);
314 		break;
315 	case CEPH_OSD_OP_STARTSYNC:
316 		break;
317 	default:
318 		pr_err("unrecognized osd opcode %d\n", dst->op);
319 		WARN_ON(1);
320 		break;
321 	}
322 	dst->payload_len = cpu_to_le32(src->payload_len);
323 }
324 
325 /*
326  * build new request AND message
327  *
328  */
329 void ceph_osdc_build_request(struct ceph_osd_request *req,
330 			     u64 off, u64 *plen,
331 			     struct ceph_osd_req_op *src_ops,
332 			     struct ceph_snap_context *snapc,
333 			     struct timespec *mtime,
334 			     const char *oid,
335 			     int oid_len)
336 {
337 	struct ceph_msg *msg = req->r_request;
338 	struct ceph_osd_request_head *head;
339 	struct ceph_osd_req_op *src_op;
340 	struct ceph_osd_op *op;
341 	void *p;
342 	int num_op = get_num_ops(src_ops, NULL);
343 	size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
344 	int flags = req->r_flags;
345 	u64 data_len = 0;
346 	int i;
347 
348 	head = msg->front.iov_base;
349 	op = (void *)(head + 1);
350 	p = (void *)(op + num_op);
351 
352 	req->r_snapc = ceph_get_snap_context(snapc);
353 
354 	head->client_inc = cpu_to_le32(1); /* always, for now. */
355 	head->flags = cpu_to_le32(flags);
356 	if (flags & CEPH_OSD_FLAG_WRITE)
357 		ceph_encode_timespec(&head->mtime, mtime);
358 	head->num_ops = cpu_to_le16(num_op);
359 
360 
361 	/* fill in oid */
362 	head->object_len = cpu_to_le32(oid_len);
363 	memcpy(p, oid, oid_len);
364 	p += oid_len;
365 
366 	src_op = src_ops;
367 	while (src_op->op) {
368 		osd_req_encode_op(req, op, src_op);
369 		src_op++;
370 		op++;
371 	}
372 
373 	if (req->r_trail)
374 		data_len += req->r_trail->length;
375 
376 	if (snapc) {
377 		head->snap_seq = cpu_to_le64(snapc->seq);
378 		head->num_snaps = cpu_to_le32(snapc->num_snaps);
379 		for (i = 0; i < snapc->num_snaps; i++) {
380 			put_unaligned_le64(snapc->snaps[i], p);
381 			p += sizeof(u64);
382 		}
383 	}
384 
385 	if (flags & CEPH_OSD_FLAG_WRITE) {
386 		req->r_request->hdr.data_off = cpu_to_le16(off);
387 		req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
388 	} else if (data_len) {
389 		req->r_request->hdr.data_off = 0;
390 		req->r_request->hdr.data_len = cpu_to_le32(data_len);
391 	}
392 
393 	BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
394 	msg_size = p - msg->front.iov_base;
395 	msg->front.iov_len = msg_size;
396 	msg->hdr.front_len = cpu_to_le32(msg_size);
397 	return;
398 }
399 EXPORT_SYMBOL(ceph_osdc_build_request);
400 
401 /*
402  * build new request AND message, calculate layout, and adjust file
403  * extent as needed.
404  *
405  * if the file was recently truncated, we include information about its
406  * old and new size so that the object can be updated appropriately.  (we
407  * avoid synchronously deleting truncated objects because it's slow.)
408  *
409  * if @do_sync, include a 'startsync' command so that the osd will flush
410  * data quickly.
411  */
412 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
413 					       struct ceph_file_layout *layout,
414 					       struct ceph_vino vino,
415 					       u64 off, u64 *plen,
416 					       int opcode, int flags,
417 					       struct ceph_snap_context *snapc,
418 					       int do_sync,
419 					       u32 truncate_seq,
420 					       u64 truncate_size,
421 					       struct timespec *mtime,
422 					       bool use_mempool, int num_reply)
423 {
424 	struct ceph_osd_req_op ops[3];
425 	struct ceph_osd_request *req;
426 
427 	ops[0].op = opcode;
428 	ops[0].extent.truncate_seq = truncate_seq;
429 	ops[0].extent.truncate_size = truncate_size;
430 	ops[0].payload_len = 0;
431 
432 	if (do_sync) {
433 		ops[1].op = CEPH_OSD_OP_STARTSYNC;
434 		ops[1].payload_len = 0;
435 		ops[2].op = 0;
436 	} else
437 		ops[1].op = 0;
438 
439 	req = ceph_osdc_alloc_request(osdc, flags,
440 					 snapc, ops,
441 					 use_mempool,
442 					 GFP_NOFS, NULL, NULL);
443 	if (IS_ERR(req))
444 		return req;
445 
446 	/* calculate max write size */
447 	calc_layout(osdc, vino, layout, off, plen, req, ops);
448 	req->r_file_layout = *layout;  /* keep a copy */
449 
450 	ceph_osdc_build_request(req, off, plen, ops,
451 				snapc,
452 				mtime,
453 				req->r_oid, req->r_oid_len);
454 
455 	return req;
456 }
457 EXPORT_SYMBOL(ceph_osdc_new_request);
458 
459 /*
460  * We keep osd requests in an rbtree, sorted by ->r_tid.
461  */
462 static void __insert_request(struct ceph_osd_client *osdc,
463 			     struct ceph_osd_request *new)
464 {
465 	struct rb_node **p = &osdc->requests.rb_node;
466 	struct rb_node *parent = NULL;
467 	struct ceph_osd_request *req = NULL;
468 
469 	while (*p) {
470 		parent = *p;
471 		req = rb_entry(parent, struct ceph_osd_request, r_node);
472 		if (new->r_tid < req->r_tid)
473 			p = &(*p)->rb_left;
474 		else if (new->r_tid > req->r_tid)
475 			p = &(*p)->rb_right;
476 		else
477 			BUG();
478 	}
479 
480 	rb_link_node(&new->r_node, parent, p);
481 	rb_insert_color(&new->r_node, &osdc->requests);
482 }
483 
484 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
485 						 u64 tid)
486 {
487 	struct ceph_osd_request *req;
488 	struct rb_node *n = osdc->requests.rb_node;
489 
490 	while (n) {
491 		req = rb_entry(n, struct ceph_osd_request, r_node);
492 		if (tid < req->r_tid)
493 			n = n->rb_left;
494 		else if (tid > req->r_tid)
495 			n = n->rb_right;
496 		else
497 			return req;
498 	}
499 	return NULL;
500 }
501 
502 static struct ceph_osd_request *
503 __lookup_request_ge(struct ceph_osd_client *osdc,
504 		    u64 tid)
505 {
506 	struct ceph_osd_request *req;
507 	struct rb_node *n = osdc->requests.rb_node;
508 
509 	while (n) {
510 		req = rb_entry(n, struct ceph_osd_request, r_node);
511 		if (tid < req->r_tid) {
512 			if (!n->rb_left)
513 				return req;
514 			n = n->rb_left;
515 		} else if (tid > req->r_tid) {
516 			n = n->rb_right;
517 		} else {
518 			return req;
519 		}
520 	}
521 	return NULL;
522 }
523 
524 
525 /*
526  * If the osd connection drops, we need to resubmit all requests.
527  */
528 static void osd_reset(struct ceph_connection *con)
529 {
530 	struct ceph_osd *osd = con->private;
531 	struct ceph_osd_client *osdc;
532 
533 	if (!osd)
534 		return;
535 	dout("osd_reset osd%d\n", osd->o_osd);
536 	osdc = osd->o_osdc;
537 	down_read(&osdc->map_sem);
538 	kick_requests(osdc, osd);
539 	up_read(&osdc->map_sem);
540 }
541 
542 /*
543  * Track open sessions with osds.
544  */
545 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
546 {
547 	struct ceph_osd *osd;
548 
549 	osd = kzalloc(sizeof(*osd), GFP_NOFS);
550 	if (!osd)
551 		return NULL;
552 
553 	atomic_set(&osd->o_ref, 1);
554 	osd->o_osdc = osdc;
555 	INIT_LIST_HEAD(&osd->o_requests);
556 	INIT_LIST_HEAD(&osd->o_osd_lru);
557 	osd->o_incarnation = 1;
558 
559 	ceph_con_init(osdc->client->msgr, &osd->o_con);
560 	osd->o_con.private = osd;
561 	osd->o_con.ops = &osd_con_ops;
562 	osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
563 
564 	INIT_LIST_HEAD(&osd->o_keepalive_item);
565 	return osd;
566 }
567 
568 static struct ceph_osd *get_osd(struct ceph_osd *osd)
569 {
570 	if (atomic_inc_not_zero(&osd->o_ref)) {
571 		dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
572 		     atomic_read(&osd->o_ref));
573 		return osd;
574 	} else {
575 		dout("get_osd %p FAIL\n", osd);
576 		return NULL;
577 	}
578 }
579 
580 static void put_osd(struct ceph_osd *osd)
581 {
582 	dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
583 	     atomic_read(&osd->o_ref) - 1);
584 	if (atomic_dec_and_test(&osd->o_ref)) {
585 		struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
586 
587 		if (osd->o_authorizer)
588 			ac->ops->destroy_authorizer(ac, osd->o_authorizer);
589 		kfree(osd);
590 	}
591 }
592 
593 /*
594  * remove an osd from our map
595  */
596 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
597 {
598 	dout("__remove_osd %p\n", osd);
599 	BUG_ON(!list_empty(&osd->o_requests));
600 	rb_erase(&osd->o_node, &osdc->osds);
601 	list_del_init(&osd->o_osd_lru);
602 	ceph_con_close(&osd->o_con);
603 	put_osd(osd);
604 }
605 
606 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
607 			      struct ceph_osd *osd)
608 {
609 	dout("__move_osd_to_lru %p\n", osd);
610 	BUG_ON(!list_empty(&osd->o_osd_lru));
611 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
612 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
613 }
614 
615 static void __remove_osd_from_lru(struct ceph_osd *osd)
616 {
617 	dout("__remove_osd_from_lru %p\n", osd);
618 	if (!list_empty(&osd->o_osd_lru))
619 		list_del_init(&osd->o_osd_lru);
620 }
621 
622 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
623 {
624 	struct ceph_osd *osd, *nosd;
625 
626 	dout("__remove_old_osds %p\n", osdc);
627 	mutex_lock(&osdc->request_mutex);
628 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
629 		if (!remove_all && time_before(jiffies, osd->lru_ttl))
630 			break;
631 		__remove_osd(osdc, osd);
632 	}
633 	mutex_unlock(&osdc->request_mutex);
634 }
635 
636 /*
637  * reset osd connect
638  */
639 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
640 {
641 	struct ceph_osd_request *req;
642 	int ret = 0;
643 
644 	dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
645 	if (list_empty(&osd->o_requests)) {
646 		__remove_osd(osdc, osd);
647 	} else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
648 			  &osd->o_con.peer_addr,
649 			  sizeof(osd->o_con.peer_addr)) == 0 &&
650 		   !ceph_con_opened(&osd->o_con)) {
651 		dout(" osd addr hasn't changed and connection never opened,"
652 		     " letting msgr retry");
653 		/* touch each r_stamp for handle_timeout()'s benfit */
654 		list_for_each_entry(req, &osd->o_requests, r_osd_item)
655 			req->r_stamp = jiffies;
656 		ret = -EAGAIN;
657 	} else {
658 		ceph_con_close(&osd->o_con);
659 		ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
660 		osd->o_incarnation++;
661 	}
662 	return ret;
663 }
664 
665 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
666 {
667 	struct rb_node **p = &osdc->osds.rb_node;
668 	struct rb_node *parent = NULL;
669 	struct ceph_osd *osd = NULL;
670 
671 	while (*p) {
672 		parent = *p;
673 		osd = rb_entry(parent, struct ceph_osd, o_node);
674 		if (new->o_osd < osd->o_osd)
675 			p = &(*p)->rb_left;
676 		else if (new->o_osd > osd->o_osd)
677 			p = &(*p)->rb_right;
678 		else
679 			BUG();
680 	}
681 
682 	rb_link_node(&new->o_node, parent, p);
683 	rb_insert_color(&new->o_node, &osdc->osds);
684 }
685 
686 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
687 {
688 	struct ceph_osd *osd;
689 	struct rb_node *n = osdc->osds.rb_node;
690 
691 	while (n) {
692 		osd = rb_entry(n, struct ceph_osd, o_node);
693 		if (o < osd->o_osd)
694 			n = n->rb_left;
695 		else if (o > osd->o_osd)
696 			n = n->rb_right;
697 		else
698 			return osd;
699 	}
700 	return NULL;
701 }
702 
703 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
704 {
705 	schedule_delayed_work(&osdc->timeout_work,
706 			osdc->client->options->osd_keepalive_timeout * HZ);
707 }
708 
709 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
710 {
711 	cancel_delayed_work(&osdc->timeout_work);
712 }
713 
714 /*
715  * Register request, assign tid.  If this is the first request, set up
716  * the timeout event.
717  */
718 static void register_request(struct ceph_osd_client *osdc,
719 			     struct ceph_osd_request *req)
720 {
721 	mutex_lock(&osdc->request_mutex);
722 	req->r_tid = ++osdc->last_tid;
723 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
724 	INIT_LIST_HEAD(&req->r_req_lru_item);
725 
726 	dout("register_request %p tid %lld\n", req, req->r_tid);
727 	__insert_request(osdc, req);
728 	ceph_osdc_get_request(req);
729 	osdc->num_requests++;
730 
731 	if (osdc->num_requests == 1) {
732 		dout(" first request, scheduling timeout\n");
733 		__schedule_osd_timeout(osdc);
734 	}
735 	mutex_unlock(&osdc->request_mutex);
736 }
737 
738 /*
739  * called under osdc->request_mutex
740  */
741 static void __unregister_request(struct ceph_osd_client *osdc,
742 				 struct ceph_osd_request *req)
743 {
744 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
745 	rb_erase(&req->r_node, &osdc->requests);
746 	osdc->num_requests--;
747 
748 	if (req->r_osd) {
749 		/* make sure the original request isn't in flight. */
750 		ceph_con_revoke(&req->r_osd->o_con, req->r_request);
751 
752 		list_del_init(&req->r_osd_item);
753 		if (list_empty(&req->r_osd->o_requests))
754 			__move_osd_to_lru(osdc, req->r_osd);
755 		req->r_osd = NULL;
756 	}
757 
758 	ceph_osdc_put_request(req);
759 
760 	list_del_init(&req->r_req_lru_item);
761 	if (osdc->num_requests == 0) {
762 		dout(" no requests, canceling timeout\n");
763 		__cancel_osd_timeout(osdc);
764 	}
765 }
766 
767 /*
768  * Cancel a previously queued request message
769  */
770 static void __cancel_request(struct ceph_osd_request *req)
771 {
772 	if (req->r_sent && req->r_osd) {
773 		ceph_con_revoke(&req->r_osd->o_con, req->r_request);
774 		req->r_sent = 0;
775 	}
776 	list_del_init(&req->r_req_lru_item);
777 }
778 
779 /*
780  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
781  * (as needed), and set the request r_osd appropriately.  If there is
782  * no up osd, set r_osd to NULL.
783  *
784  * Return 0 if unchanged, 1 if changed, or negative on error.
785  *
786  * Caller should hold map_sem for read and request_mutex.
787  */
788 static int __map_osds(struct ceph_osd_client *osdc,
789 		      struct ceph_osd_request *req)
790 {
791 	struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
792 	struct ceph_pg pgid;
793 	int acting[CEPH_PG_MAX_SIZE];
794 	int o = -1, num = 0;
795 	int err;
796 
797 	dout("map_osds %p tid %lld\n", req, req->r_tid);
798 	err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
799 				      &req->r_file_layout, osdc->osdmap);
800 	if (err)
801 		return err;
802 	pgid = reqhead->layout.ol_pgid;
803 	req->r_pgid = pgid;
804 
805 	err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
806 	if (err > 0) {
807 		o = acting[0];
808 		num = err;
809 	}
810 
811 	if ((req->r_osd && req->r_osd->o_osd == o &&
812 	     req->r_sent >= req->r_osd->o_incarnation &&
813 	     req->r_num_pg_osds == num &&
814 	     memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
815 	    (req->r_osd == NULL && o == -1))
816 		return 0;  /* no change */
817 
818 	dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
819 	     req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
820 	     req->r_osd ? req->r_osd->o_osd : -1);
821 
822 	/* record full pg acting set */
823 	memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
824 	req->r_num_pg_osds = num;
825 
826 	if (req->r_osd) {
827 		__cancel_request(req);
828 		list_del_init(&req->r_osd_item);
829 		req->r_osd = NULL;
830 	}
831 
832 	req->r_osd = __lookup_osd(osdc, o);
833 	if (!req->r_osd && o >= 0) {
834 		err = -ENOMEM;
835 		req->r_osd = create_osd(osdc);
836 		if (!req->r_osd)
837 			goto out;
838 
839 		dout("map_osds osd %p is osd%d\n", req->r_osd, o);
840 		req->r_osd->o_osd = o;
841 		req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
842 		__insert_osd(osdc, req->r_osd);
843 
844 		ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
845 	}
846 
847 	if (req->r_osd) {
848 		__remove_osd_from_lru(req->r_osd);
849 		list_add(&req->r_osd_item, &req->r_osd->o_requests);
850 	}
851 	err = 1;   /* osd or pg changed */
852 
853 out:
854 	return err;
855 }
856 
857 /*
858  * caller should hold map_sem (for read) and request_mutex
859  */
860 static int __send_request(struct ceph_osd_client *osdc,
861 			  struct ceph_osd_request *req)
862 {
863 	struct ceph_osd_request_head *reqhead;
864 	int err;
865 
866 	err = __map_osds(osdc, req);
867 	if (err < 0)
868 		return err;
869 	if (req->r_osd == NULL) {
870 		dout("send_request %p no up osds in pg\n", req);
871 		ceph_monc_request_next_osdmap(&osdc->client->monc);
872 		return 0;
873 	}
874 
875 	dout("send_request %p tid %llu to osd%d flags %d\n",
876 	     req, req->r_tid, req->r_osd->o_osd, req->r_flags);
877 
878 	reqhead = req->r_request->front.iov_base;
879 	reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
880 	reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
881 	reqhead->reassert_version = req->r_reassert_version;
882 
883 	req->r_stamp = jiffies;
884 	list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
885 
886 	ceph_msg_get(req->r_request); /* send consumes a ref */
887 	ceph_con_send(&req->r_osd->o_con, req->r_request);
888 	req->r_sent = req->r_osd->o_incarnation;
889 	return 0;
890 }
891 
892 /*
893  * Timeout callback, called every N seconds when 1 or more osd
894  * requests has been active for more than N seconds.  When this
895  * happens, we ping all OSDs with requests who have timed out to
896  * ensure any communications channel reset is detected.  Reset the
897  * request timeouts another N seconds in the future as we go.
898  * Reschedule the timeout event another N seconds in future (unless
899  * there are no open requests).
900  */
901 static void handle_timeout(struct work_struct *work)
902 {
903 	struct ceph_osd_client *osdc =
904 		container_of(work, struct ceph_osd_client, timeout_work.work);
905 	struct ceph_osd_request *req, *last_req = NULL;
906 	struct ceph_osd *osd;
907 	unsigned long timeout = osdc->client->options->osd_timeout * HZ;
908 	unsigned long keepalive =
909 		osdc->client->options->osd_keepalive_timeout * HZ;
910 	unsigned long last_stamp = 0;
911 	struct rb_node *p;
912 	struct list_head slow_osds;
913 
914 	dout("timeout\n");
915 	down_read(&osdc->map_sem);
916 
917 	ceph_monc_request_next_osdmap(&osdc->client->monc);
918 
919 	mutex_lock(&osdc->request_mutex);
920 	for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
921 		req = rb_entry(p, struct ceph_osd_request, r_node);
922 
923 		if (req->r_resend) {
924 			int err;
925 
926 			dout("osdc resending prev failed %lld\n", req->r_tid);
927 			err = __send_request(osdc, req);
928 			if (err)
929 				dout("osdc failed again on %lld\n", req->r_tid);
930 			else
931 				req->r_resend = false;
932 			continue;
933 		}
934 	}
935 
936 	/*
937 	 * reset osds that appear to be _really_ unresponsive.  this
938 	 * is a failsafe measure.. we really shouldn't be getting to
939 	 * this point if the system is working properly.  the monitors
940 	 * should mark the osd as failed and we should find out about
941 	 * it from an updated osd map.
942 	 */
943 	while (timeout && !list_empty(&osdc->req_lru)) {
944 		req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
945 				 r_req_lru_item);
946 
947 		if (time_before(jiffies, req->r_stamp + timeout))
948 			break;
949 
950 		BUG_ON(req == last_req && req->r_stamp == last_stamp);
951 		last_req = req;
952 		last_stamp = req->r_stamp;
953 
954 		osd = req->r_osd;
955 		BUG_ON(!osd);
956 		pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
957 			   req->r_tid, osd->o_osd);
958 		__kick_requests(osdc, osd);
959 	}
960 
961 	/*
962 	 * ping osds that are a bit slow.  this ensures that if there
963 	 * is a break in the TCP connection we will notice, and reopen
964 	 * a connection with that osd (from the fault callback).
965 	 */
966 	INIT_LIST_HEAD(&slow_osds);
967 	list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
968 		if (time_before(jiffies, req->r_stamp + keepalive))
969 			break;
970 
971 		osd = req->r_osd;
972 		BUG_ON(!osd);
973 		dout(" tid %llu is slow, will send keepalive on osd%d\n",
974 		     req->r_tid, osd->o_osd);
975 		list_move_tail(&osd->o_keepalive_item, &slow_osds);
976 	}
977 	while (!list_empty(&slow_osds)) {
978 		osd = list_entry(slow_osds.next, struct ceph_osd,
979 				 o_keepalive_item);
980 		list_del_init(&osd->o_keepalive_item);
981 		ceph_con_keepalive(&osd->o_con);
982 	}
983 
984 	__schedule_osd_timeout(osdc);
985 	mutex_unlock(&osdc->request_mutex);
986 
987 	up_read(&osdc->map_sem);
988 }
989 
990 static void handle_osds_timeout(struct work_struct *work)
991 {
992 	struct ceph_osd_client *osdc =
993 		container_of(work, struct ceph_osd_client,
994 			     osds_timeout_work.work);
995 	unsigned long delay =
996 		osdc->client->options->osd_idle_ttl * HZ >> 2;
997 
998 	dout("osds timeout\n");
999 	down_read(&osdc->map_sem);
1000 	remove_old_osds(osdc, 0);
1001 	up_read(&osdc->map_sem);
1002 
1003 	schedule_delayed_work(&osdc->osds_timeout_work,
1004 			      round_jiffies_relative(delay));
1005 }
1006 
1007 /*
1008  * handle osd op reply.  either call the callback if it is specified,
1009  * or do the completion to wake up the waiting thread.
1010  */
1011 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1012 			 struct ceph_connection *con)
1013 {
1014 	struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1015 	struct ceph_osd_request *req;
1016 	u64 tid;
1017 	int numops, object_len, flags;
1018 	s32 result;
1019 
1020 	tid = le64_to_cpu(msg->hdr.tid);
1021 	if (msg->front.iov_len < sizeof(*rhead))
1022 		goto bad;
1023 	numops = le32_to_cpu(rhead->num_ops);
1024 	object_len = le32_to_cpu(rhead->object_len);
1025 	result = le32_to_cpu(rhead->result);
1026 	if (msg->front.iov_len != sizeof(*rhead) + object_len +
1027 	    numops * sizeof(struct ceph_osd_op))
1028 		goto bad;
1029 	dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1030 
1031 	/* lookup */
1032 	mutex_lock(&osdc->request_mutex);
1033 	req = __lookup_request(osdc, tid);
1034 	if (req == NULL) {
1035 		dout("handle_reply tid %llu dne\n", tid);
1036 		mutex_unlock(&osdc->request_mutex);
1037 		return;
1038 	}
1039 	ceph_osdc_get_request(req);
1040 	flags = le32_to_cpu(rhead->flags);
1041 
1042 	/*
1043 	 * if this connection filled our message, drop our reference now, to
1044 	 * avoid a (safe but slower) revoke later.
1045 	 */
1046 	if (req->r_con_filling_msg == con && req->r_reply == msg) {
1047 		dout(" dropping con_filling_msg ref %p\n", con);
1048 		req->r_con_filling_msg = NULL;
1049 		ceph_con_put(con);
1050 	}
1051 
1052 	if (!req->r_got_reply) {
1053 		unsigned bytes;
1054 
1055 		req->r_result = le32_to_cpu(rhead->result);
1056 		bytes = le32_to_cpu(msg->hdr.data_len);
1057 		dout("handle_reply result %d bytes %d\n", req->r_result,
1058 		     bytes);
1059 		if (req->r_result == 0)
1060 			req->r_result = bytes;
1061 
1062 		/* in case this is a write and we need to replay, */
1063 		req->r_reassert_version = rhead->reassert_version;
1064 
1065 		req->r_got_reply = 1;
1066 	} else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1067 		dout("handle_reply tid %llu dup ack\n", tid);
1068 		mutex_unlock(&osdc->request_mutex);
1069 		goto done;
1070 	}
1071 
1072 	dout("handle_reply tid %llu flags %d\n", tid, flags);
1073 
1074 	/* either this is a read, or we got the safe response */
1075 	if (result < 0 ||
1076 	    (flags & CEPH_OSD_FLAG_ONDISK) ||
1077 	    ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1078 		__unregister_request(osdc, req);
1079 
1080 	mutex_unlock(&osdc->request_mutex);
1081 
1082 	if (req->r_callback)
1083 		req->r_callback(req, msg);
1084 	else
1085 		complete_all(&req->r_completion);
1086 
1087 	if (flags & CEPH_OSD_FLAG_ONDISK) {
1088 		if (req->r_safe_callback)
1089 			req->r_safe_callback(req, msg);
1090 		complete_all(&req->r_safe_completion);  /* fsync waiter */
1091 	}
1092 
1093 done:
1094 	ceph_osdc_put_request(req);
1095 	return;
1096 
1097 bad:
1098 	pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1099 	       (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1100 	       (int)sizeof(*rhead));
1101 	ceph_msg_dump(msg);
1102 }
1103 
1104 
1105 static int __kick_requests(struct ceph_osd_client *osdc,
1106 			  struct ceph_osd *kickosd)
1107 {
1108 	struct ceph_osd_request *req;
1109 	struct rb_node *p, *n;
1110 	int needmap = 0;
1111 	int err;
1112 
1113 	dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
1114 	if (kickosd) {
1115 		err = __reset_osd(osdc, kickosd);
1116 		if (err == -EAGAIN)
1117 			return 1;
1118 	} else {
1119 		for (p = rb_first(&osdc->osds); p; p = n) {
1120 			struct ceph_osd *osd =
1121 				rb_entry(p, struct ceph_osd, o_node);
1122 
1123 			n = rb_next(p);
1124 			if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1125 			    memcmp(&osd->o_con.peer_addr,
1126 				   ceph_osd_addr(osdc->osdmap,
1127 						 osd->o_osd),
1128 				   sizeof(struct ceph_entity_addr)) != 0)
1129 				__reset_osd(osdc, osd);
1130 		}
1131 	}
1132 
1133 	for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1134 		req = rb_entry(p, struct ceph_osd_request, r_node);
1135 
1136 		if (req->r_resend) {
1137 			dout(" r_resend set on tid %llu\n", req->r_tid);
1138 			__cancel_request(req);
1139 			goto kick;
1140 		}
1141 		if (req->r_osd && kickosd == req->r_osd) {
1142 			__cancel_request(req);
1143 			goto kick;
1144 		}
1145 
1146 		err = __map_osds(osdc, req);
1147 		if (err == 0)
1148 			continue;  /* no change */
1149 		if (err < 0) {
1150 			/*
1151 			 * FIXME: really, we should set the request
1152 			 * error and fail if this isn't a 'nofail'
1153 			 * request, but that's a fair bit more
1154 			 * complicated to do.  So retry!
1155 			 */
1156 			dout(" setting r_resend on %llu\n", req->r_tid);
1157 			req->r_resend = true;
1158 			continue;
1159 		}
1160 		if (req->r_osd == NULL) {
1161 			dout("tid %llu maps to no valid osd\n", req->r_tid);
1162 			needmap++;  /* request a newer map */
1163 			continue;
1164 		}
1165 
1166 kick:
1167 		dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
1168 		     req->r_osd ? req->r_osd->o_osd : -1);
1169 		req->r_flags |= CEPH_OSD_FLAG_RETRY;
1170 		err = __send_request(osdc, req);
1171 		if (err) {
1172 			dout(" setting r_resend on %llu\n", req->r_tid);
1173 			req->r_resend = true;
1174 		}
1175 	}
1176 
1177 	return needmap;
1178 }
1179 
1180 /*
1181  * Resubmit osd requests whose osd or osd address has changed.  Request
1182  * a new osd map if osds are down, or we are otherwise unable to determine
1183  * how to direct a request.
1184  *
1185  * Close connections to down osds.
1186  *
1187  * If @who is specified, resubmit requests for that specific osd.
1188  *
1189  * Caller should hold map_sem for read and request_mutex.
1190  */
1191 static void kick_requests(struct ceph_osd_client *osdc,
1192 			  struct ceph_osd *kickosd)
1193 {
1194 	int needmap;
1195 
1196 	mutex_lock(&osdc->request_mutex);
1197 	needmap = __kick_requests(osdc, kickosd);
1198 	mutex_unlock(&osdc->request_mutex);
1199 
1200 	if (needmap) {
1201 		dout("%d requests for down osds, need new map\n", needmap);
1202 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1203 	}
1204 
1205 }
1206 /*
1207  * Process updated osd map.
1208  *
1209  * The message contains any number of incremental and full maps, normally
1210  * indicating some sort of topology change in the cluster.  Kick requests
1211  * off to different OSDs as needed.
1212  */
1213 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1214 {
1215 	void *p, *end, *next;
1216 	u32 nr_maps, maplen;
1217 	u32 epoch;
1218 	struct ceph_osdmap *newmap = NULL, *oldmap;
1219 	int err;
1220 	struct ceph_fsid fsid;
1221 
1222 	dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1223 	p = msg->front.iov_base;
1224 	end = p + msg->front.iov_len;
1225 
1226 	/* verify fsid */
1227 	ceph_decode_need(&p, end, sizeof(fsid), bad);
1228 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
1229 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
1230 		return;
1231 
1232 	down_write(&osdc->map_sem);
1233 
1234 	/* incremental maps */
1235 	ceph_decode_32_safe(&p, end, nr_maps, bad);
1236 	dout(" %d inc maps\n", nr_maps);
1237 	while (nr_maps > 0) {
1238 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1239 		epoch = ceph_decode_32(&p);
1240 		maplen = ceph_decode_32(&p);
1241 		ceph_decode_need(&p, end, maplen, bad);
1242 		next = p + maplen;
1243 		if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1244 			dout("applying incremental map %u len %d\n",
1245 			     epoch, maplen);
1246 			newmap = osdmap_apply_incremental(&p, next,
1247 							  osdc->osdmap,
1248 							  osdc->client->msgr);
1249 			if (IS_ERR(newmap)) {
1250 				err = PTR_ERR(newmap);
1251 				goto bad;
1252 			}
1253 			BUG_ON(!newmap);
1254 			if (newmap != osdc->osdmap) {
1255 				ceph_osdmap_destroy(osdc->osdmap);
1256 				osdc->osdmap = newmap;
1257 			}
1258 		} else {
1259 			dout("ignoring incremental map %u len %d\n",
1260 			     epoch, maplen);
1261 		}
1262 		p = next;
1263 		nr_maps--;
1264 	}
1265 	if (newmap)
1266 		goto done;
1267 
1268 	/* full maps */
1269 	ceph_decode_32_safe(&p, end, nr_maps, bad);
1270 	dout(" %d full maps\n", nr_maps);
1271 	while (nr_maps) {
1272 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1273 		epoch = ceph_decode_32(&p);
1274 		maplen = ceph_decode_32(&p);
1275 		ceph_decode_need(&p, end, maplen, bad);
1276 		if (nr_maps > 1) {
1277 			dout("skipping non-latest full map %u len %d\n",
1278 			     epoch, maplen);
1279 		} else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1280 			dout("skipping full map %u len %d, "
1281 			     "older than our %u\n", epoch, maplen,
1282 			     osdc->osdmap->epoch);
1283 		} else {
1284 			dout("taking full map %u len %d\n", epoch, maplen);
1285 			newmap = osdmap_decode(&p, p+maplen);
1286 			if (IS_ERR(newmap)) {
1287 				err = PTR_ERR(newmap);
1288 				goto bad;
1289 			}
1290 			BUG_ON(!newmap);
1291 			oldmap = osdc->osdmap;
1292 			osdc->osdmap = newmap;
1293 			if (oldmap)
1294 				ceph_osdmap_destroy(oldmap);
1295 		}
1296 		p += maplen;
1297 		nr_maps--;
1298 	}
1299 
1300 done:
1301 	downgrade_write(&osdc->map_sem);
1302 	ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1303 	if (newmap)
1304 		kick_requests(osdc, NULL);
1305 	up_read(&osdc->map_sem);
1306 	wake_up_all(&osdc->client->auth_wq);
1307 	return;
1308 
1309 bad:
1310 	pr_err("osdc handle_map corrupt msg\n");
1311 	ceph_msg_dump(msg);
1312 	up_write(&osdc->map_sem);
1313 	return;
1314 }
1315 
1316 /*
1317  * Register request, send initial attempt.
1318  */
1319 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1320 			    struct ceph_osd_request *req,
1321 			    bool nofail)
1322 {
1323 	int rc = 0;
1324 
1325 	req->r_request->pages = req->r_pages;
1326 	req->r_request->nr_pages = req->r_num_pages;
1327 #ifdef CONFIG_BLOCK
1328 	req->r_request->bio = req->r_bio;
1329 #endif
1330 	req->r_request->trail = req->r_trail;
1331 
1332 	register_request(osdc, req);
1333 
1334 	down_read(&osdc->map_sem);
1335 	mutex_lock(&osdc->request_mutex);
1336 	/*
1337 	 * a racing kick_requests() may have sent the message for us
1338 	 * while we dropped request_mutex above, so only send now if
1339 	 * the request still han't been touched yet.
1340 	 */
1341 	if (req->r_sent == 0) {
1342 		rc = __send_request(osdc, req);
1343 		if (rc) {
1344 			if (nofail) {
1345 				dout("osdc_start_request failed send, "
1346 				     " marking %lld\n", req->r_tid);
1347 				req->r_resend = true;
1348 				rc = 0;
1349 			} else {
1350 				__unregister_request(osdc, req);
1351 			}
1352 		}
1353 	}
1354 	mutex_unlock(&osdc->request_mutex);
1355 	up_read(&osdc->map_sem);
1356 	return rc;
1357 }
1358 EXPORT_SYMBOL(ceph_osdc_start_request);
1359 
1360 /*
1361  * wait for a request to complete
1362  */
1363 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1364 			   struct ceph_osd_request *req)
1365 {
1366 	int rc;
1367 
1368 	rc = wait_for_completion_interruptible(&req->r_completion);
1369 	if (rc < 0) {
1370 		mutex_lock(&osdc->request_mutex);
1371 		__cancel_request(req);
1372 		__unregister_request(osdc, req);
1373 		mutex_unlock(&osdc->request_mutex);
1374 		dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1375 		return rc;
1376 	}
1377 
1378 	dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1379 	return req->r_result;
1380 }
1381 EXPORT_SYMBOL(ceph_osdc_wait_request);
1382 
1383 /*
1384  * sync - wait for all in-flight requests to flush.  avoid starvation.
1385  */
1386 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1387 {
1388 	struct ceph_osd_request *req;
1389 	u64 last_tid, next_tid = 0;
1390 
1391 	mutex_lock(&osdc->request_mutex);
1392 	last_tid = osdc->last_tid;
1393 	while (1) {
1394 		req = __lookup_request_ge(osdc, next_tid);
1395 		if (!req)
1396 			break;
1397 		if (req->r_tid > last_tid)
1398 			break;
1399 
1400 		next_tid = req->r_tid + 1;
1401 		if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1402 			continue;
1403 
1404 		ceph_osdc_get_request(req);
1405 		mutex_unlock(&osdc->request_mutex);
1406 		dout("sync waiting on tid %llu (last is %llu)\n",
1407 		     req->r_tid, last_tid);
1408 		wait_for_completion(&req->r_safe_completion);
1409 		mutex_lock(&osdc->request_mutex);
1410 		ceph_osdc_put_request(req);
1411 	}
1412 	mutex_unlock(&osdc->request_mutex);
1413 	dout("sync done (thru tid %llu)\n", last_tid);
1414 }
1415 EXPORT_SYMBOL(ceph_osdc_sync);
1416 
1417 /*
1418  * init, shutdown
1419  */
1420 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1421 {
1422 	int err;
1423 
1424 	dout("init\n");
1425 	osdc->client = client;
1426 	osdc->osdmap = NULL;
1427 	init_rwsem(&osdc->map_sem);
1428 	init_completion(&osdc->map_waiters);
1429 	osdc->last_requested_map = 0;
1430 	mutex_init(&osdc->request_mutex);
1431 	osdc->last_tid = 0;
1432 	osdc->osds = RB_ROOT;
1433 	INIT_LIST_HEAD(&osdc->osd_lru);
1434 	osdc->requests = RB_ROOT;
1435 	INIT_LIST_HEAD(&osdc->req_lru);
1436 	osdc->num_requests = 0;
1437 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1438 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1439 
1440 	schedule_delayed_work(&osdc->osds_timeout_work,
1441 	   round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1442 
1443 	err = -ENOMEM;
1444 	osdc->req_mempool = mempool_create_kmalloc_pool(10,
1445 					sizeof(struct ceph_osd_request));
1446 	if (!osdc->req_mempool)
1447 		goto out;
1448 
1449 	err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1450 				"osd_op");
1451 	if (err < 0)
1452 		goto out_mempool;
1453 	err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1454 				OSD_OPREPLY_FRONT_LEN, 10, true,
1455 				"osd_op_reply");
1456 	if (err < 0)
1457 		goto out_msgpool;
1458 	return 0;
1459 
1460 out_msgpool:
1461 	ceph_msgpool_destroy(&osdc->msgpool_op);
1462 out_mempool:
1463 	mempool_destroy(osdc->req_mempool);
1464 out:
1465 	return err;
1466 }
1467 EXPORT_SYMBOL(ceph_osdc_init);
1468 
1469 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1470 {
1471 	cancel_delayed_work_sync(&osdc->timeout_work);
1472 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
1473 	if (osdc->osdmap) {
1474 		ceph_osdmap_destroy(osdc->osdmap);
1475 		osdc->osdmap = NULL;
1476 	}
1477 	remove_old_osds(osdc, 1);
1478 	mempool_destroy(osdc->req_mempool);
1479 	ceph_msgpool_destroy(&osdc->msgpool_op);
1480 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1481 }
1482 EXPORT_SYMBOL(ceph_osdc_stop);
1483 
1484 /*
1485  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1486  * *plen.  Return number of bytes read, or error.
1487  */
1488 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1489 			struct ceph_vino vino, struct ceph_file_layout *layout,
1490 			u64 off, u64 *plen,
1491 			u32 truncate_seq, u64 truncate_size,
1492 			struct page **pages, int num_pages)
1493 {
1494 	struct ceph_osd_request *req;
1495 	int rc = 0;
1496 
1497 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1498 	     vino.snap, off, *plen);
1499 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1500 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1501 				    NULL, 0, truncate_seq, truncate_size, NULL,
1502 				    false, 1);
1503 	if (!req)
1504 		return -ENOMEM;
1505 
1506 	/* it may be a short read due to an object boundary */
1507 	req->r_pages = pages;
1508 
1509 	dout("readpages  final extent is %llu~%llu (%d pages)\n",
1510 	     off, *plen, req->r_num_pages);
1511 
1512 	rc = ceph_osdc_start_request(osdc, req, false);
1513 	if (!rc)
1514 		rc = ceph_osdc_wait_request(osdc, req);
1515 
1516 	ceph_osdc_put_request(req);
1517 	dout("readpages result %d\n", rc);
1518 	return rc;
1519 }
1520 EXPORT_SYMBOL(ceph_osdc_readpages);
1521 
1522 /*
1523  * do a synchronous write on N pages
1524  */
1525 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1526 			 struct ceph_file_layout *layout,
1527 			 struct ceph_snap_context *snapc,
1528 			 u64 off, u64 len,
1529 			 u32 truncate_seq, u64 truncate_size,
1530 			 struct timespec *mtime,
1531 			 struct page **pages, int num_pages,
1532 			 int flags, int do_sync, bool nofail)
1533 {
1534 	struct ceph_osd_request *req;
1535 	int rc = 0;
1536 
1537 	BUG_ON(vino.snap != CEPH_NOSNAP);
1538 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1539 				    CEPH_OSD_OP_WRITE,
1540 				    flags | CEPH_OSD_FLAG_ONDISK |
1541 					    CEPH_OSD_FLAG_WRITE,
1542 				    snapc, do_sync,
1543 				    truncate_seq, truncate_size, mtime,
1544 				    nofail, 1);
1545 	if (!req)
1546 		return -ENOMEM;
1547 
1548 	/* it may be a short write due to an object boundary */
1549 	req->r_pages = pages;
1550 	dout("writepages %llu~%llu (%d pages)\n", off, len,
1551 	     req->r_num_pages);
1552 
1553 	rc = ceph_osdc_start_request(osdc, req, nofail);
1554 	if (!rc)
1555 		rc = ceph_osdc_wait_request(osdc, req);
1556 
1557 	ceph_osdc_put_request(req);
1558 	if (rc == 0)
1559 		rc = len;
1560 	dout("writepages result %d\n", rc);
1561 	return rc;
1562 }
1563 EXPORT_SYMBOL(ceph_osdc_writepages);
1564 
1565 /*
1566  * handle incoming message
1567  */
1568 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1569 {
1570 	struct ceph_osd *osd = con->private;
1571 	struct ceph_osd_client *osdc;
1572 	int type = le16_to_cpu(msg->hdr.type);
1573 
1574 	if (!osd)
1575 		goto out;
1576 	osdc = osd->o_osdc;
1577 
1578 	switch (type) {
1579 	case CEPH_MSG_OSD_MAP:
1580 		ceph_osdc_handle_map(osdc, msg);
1581 		break;
1582 	case CEPH_MSG_OSD_OPREPLY:
1583 		handle_reply(osdc, msg, con);
1584 		break;
1585 
1586 	default:
1587 		pr_err("received unknown message type %d %s\n", type,
1588 		       ceph_msg_type_name(type));
1589 	}
1590 out:
1591 	ceph_msg_put(msg);
1592 }
1593 
1594 /*
1595  * lookup and return message for incoming reply.  set up reply message
1596  * pages.
1597  */
1598 static struct ceph_msg *get_reply(struct ceph_connection *con,
1599 				  struct ceph_msg_header *hdr,
1600 				  int *skip)
1601 {
1602 	struct ceph_osd *osd = con->private;
1603 	struct ceph_osd_client *osdc = osd->o_osdc;
1604 	struct ceph_msg *m;
1605 	struct ceph_osd_request *req;
1606 	int front = le32_to_cpu(hdr->front_len);
1607 	int data_len = le32_to_cpu(hdr->data_len);
1608 	u64 tid;
1609 
1610 	tid = le64_to_cpu(hdr->tid);
1611 	mutex_lock(&osdc->request_mutex);
1612 	req = __lookup_request(osdc, tid);
1613 	if (!req) {
1614 		*skip = 1;
1615 		m = NULL;
1616 		pr_info("get_reply unknown tid %llu from osd%d\n", tid,
1617 			osd->o_osd);
1618 		goto out;
1619 	}
1620 
1621 	if (req->r_con_filling_msg) {
1622 		dout("get_reply revoking msg %p from old con %p\n",
1623 		     req->r_reply, req->r_con_filling_msg);
1624 		ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
1625 		ceph_con_put(req->r_con_filling_msg);
1626 		req->r_con_filling_msg = NULL;
1627 	}
1628 
1629 	if (front > req->r_reply->front.iov_len) {
1630 		pr_warning("get_reply front %d > preallocated %d\n",
1631 			   front, (int)req->r_reply->front.iov_len);
1632 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS);
1633 		if (!m)
1634 			goto out;
1635 		ceph_msg_put(req->r_reply);
1636 		req->r_reply = m;
1637 	}
1638 	m = ceph_msg_get(req->r_reply);
1639 
1640 	if (data_len > 0) {
1641 		unsigned data_off = le16_to_cpu(hdr->data_off);
1642 		int want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1643 
1644 		if (unlikely(req->r_num_pages < want)) {
1645 			pr_warning("tid %lld reply %d > expected %d pages\n",
1646 				   tid, want, m->nr_pages);
1647 			*skip = 1;
1648 			ceph_msg_put(m);
1649 			m = NULL;
1650 			goto out;
1651 		}
1652 		m->pages = req->r_pages;
1653 		m->nr_pages = req->r_num_pages;
1654 #ifdef CONFIG_BLOCK
1655 		m->bio = req->r_bio;
1656 #endif
1657 	}
1658 	*skip = 0;
1659 	req->r_con_filling_msg = ceph_con_get(con);
1660 	dout("get_reply tid %lld %p\n", tid, m);
1661 
1662 out:
1663 	mutex_unlock(&osdc->request_mutex);
1664 	return m;
1665 
1666 }
1667 
1668 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1669 				  struct ceph_msg_header *hdr,
1670 				  int *skip)
1671 {
1672 	struct ceph_osd *osd = con->private;
1673 	int type = le16_to_cpu(hdr->type);
1674 	int front = le32_to_cpu(hdr->front_len);
1675 
1676 	switch (type) {
1677 	case CEPH_MSG_OSD_MAP:
1678 		return ceph_msg_new(type, front, GFP_NOFS);
1679 	case CEPH_MSG_OSD_OPREPLY:
1680 		return get_reply(con, hdr, skip);
1681 	default:
1682 		pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
1683 			osd->o_osd);
1684 		*skip = 1;
1685 		return NULL;
1686 	}
1687 }
1688 
1689 /*
1690  * Wrappers to refcount containing ceph_osd struct
1691  */
1692 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1693 {
1694 	struct ceph_osd *osd = con->private;
1695 	if (get_osd(osd))
1696 		return con;
1697 	return NULL;
1698 }
1699 
1700 static void put_osd_con(struct ceph_connection *con)
1701 {
1702 	struct ceph_osd *osd = con->private;
1703 	put_osd(osd);
1704 }
1705 
1706 /*
1707  * authentication
1708  */
1709 static int get_authorizer(struct ceph_connection *con,
1710 			  void **buf, int *len, int *proto,
1711 			  void **reply_buf, int *reply_len, int force_new)
1712 {
1713 	struct ceph_osd *o = con->private;
1714 	struct ceph_osd_client *osdc = o->o_osdc;
1715 	struct ceph_auth_client *ac = osdc->client->monc.auth;
1716 	int ret = 0;
1717 
1718 	if (force_new && o->o_authorizer) {
1719 		ac->ops->destroy_authorizer(ac, o->o_authorizer);
1720 		o->o_authorizer = NULL;
1721 	}
1722 	if (o->o_authorizer == NULL) {
1723 		ret = ac->ops->create_authorizer(
1724 			ac, CEPH_ENTITY_TYPE_OSD,
1725 			&o->o_authorizer,
1726 			&o->o_authorizer_buf,
1727 			&o->o_authorizer_buf_len,
1728 			&o->o_authorizer_reply_buf,
1729 			&o->o_authorizer_reply_buf_len);
1730 		if (ret)
1731 			return ret;
1732 	}
1733 
1734 	*proto = ac->protocol;
1735 	*buf = o->o_authorizer_buf;
1736 	*len = o->o_authorizer_buf_len;
1737 	*reply_buf = o->o_authorizer_reply_buf;
1738 	*reply_len = o->o_authorizer_reply_buf_len;
1739 	return 0;
1740 }
1741 
1742 
1743 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1744 {
1745 	struct ceph_osd *o = con->private;
1746 	struct ceph_osd_client *osdc = o->o_osdc;
1747 	struct ceph_auth_client *ac = osdc->client->monc.auth;
1748 
1749 	return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1750 }
1751 
1752 static int invalidate_authorizer(struct ceph_connection *con)
1753 {
1754 	struct ceph_osd *o = con->private;
1755 	struct ceph_osd_client *osdc = o->o_osdc;
1756 	struct ceph_auth_client *ac = osdc->client->monc.auth;
1757 
1758 	if (ac->ops->invalidate_authorizer)
1759 		ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
1760 
1761 	return ceph_monc_validate_auth(&osdc->client->monc);
1762 }
1763 
1764 static const struct ceph_connection_operations osd_con_ops = {
1765 	.get = get_osd_con,
1766 	.put = put_osd_con,
1767 	.dispatch = dispatch,
1768 	.get_authorizer = get_authorizer,
1769 	.verify_authorizer_reply = verify_authorizer_reply,
1770 	.invalidate_authorizer = invalidate_authorizer,
1771 	.alloc_msg = alloc_msg,
1772 	.fault = osd_reset,
1773 };
1774