xref: /openbmc/linux/net/ceph/osd_client.c (revision 6774def6)
1 
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/module.h>
5 #include <linux/err.h>
6 #include <linux/highmem.h>
7 #include <linux/mm.h>
8 #include <linux/pagemap.h>
9 #include <linux/slab.h>
10 #include <linux/uaccess.h>
11 #ifdef CONFIG_BLOCK
12 #include <linux/bio.h>
13 #endif
14 
15 #include <linux/ceph/libceph.h>
16 #include <linux/ceph/osd_client.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/pagelist.h>
21 
22 #define OSD_OP_FRONT_LEN	4096
23 #define OSD_OPREPLY_FRONT_LEN	512
24 
25 static struct kmem_cache	*ceph_osd_request_cache;
26 
27 static const struct ceph_connection_operations osd_con_ops;
28 
29 static void __send_queued(struct ceph_osd_client *osdc);
30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
31 static void __register_request(struct ceph_osd_client *osdc,
32 			       struct ceph_osd_request *req);
33 static void __unregister_request(struct ceph_osd_client *osdc,
34 				 struct ceph_osd_request *req);
35 static void __unregister_linger_request(struct ceph_osd_client *osdc,
36 					struct ceph_osd_request *req);
37 static void __enqueue_request(struct ceph_osd_request *req);
38 static void __send_request(struct ceph_osd_client *osdc,
39 			   struct ceph_osd_request *req);
40 
41 /*
42  * Implement client access to distributed object storage cluster.
43  *
44  * All data objects are stored within a cluster/cloud of OSDs, or
45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
47  * remote daemons serving up and coordinating consistent and safe
48  * access to storage.
49  *
50  * Cluster membership and the mapping of data objects onto storage devices
51  * are described by the osd map.
52  *
53  * We keep track of pending OSD requests (read, write), resubmit
54  * requests to different OSDs when the cluster topology/data layout
55  * change, or retry the affected requests when the communications
56  * channel with an OSD is reset.
57  */
58 
59 /*
60  * calculate the mapping of a file extent onto an object, and fill out the
61  * request accordingly.  shorten extent as necessary if it crosses an
62  * object boundary.
63  *
64  * fill osd op in request message.
65  */
66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
67 			u64 *objnum, u64 *objoff, u64 *objlen)
68 {
69 	u64 orig_len = *plen;
70 	int r;
71 
72 	/* object extent? */
73 	r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
74 					  objoff, objlen);
75 	if (r < 0)
76 		return r;
77 	if (*objlen < orig_len) {
78 		*plen = *objlen;
79 		dout(" skipping last %llu, final file extent %llu~%llu\n",
80 		     orig_len - *plen, off, *plen);
81 	}
82 
83 	dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
84 
85 	return 0;
86 }
87 
88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
89 {
90 	memset(osd_data, 0, sizeof (*osd_data));
91 	osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
92 }
93 
94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
95 			struct page **pages, u64 length, u32 alignment,
96 			bool pages_from_pool, bool own_pages)
97 {
98 	osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
99 	osd_data->pages = pages;
100 	osd_data->length = length;
101 	osd_data->alignment = alignment;
102 	osd_data->pages_from_pool = pages_from_pool;
103 	osd_data->own_pages = own_pages;
104 }
105 
106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
107 			struct ceph_pagelist *pagelist)
108 {
109 	osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
110 	osd_data->pagelist = pagelist;
111 }
112 
113 #ifdef CONFIG_BLOCK
114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
115 			struct bio *bio, size_t bio_length)
116 {
117 	osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
118 	osd_data->bio = bio;
119 	osd_data->bio_length = bio_length;
120 }
121 #endif /* CONFIG_BLOCK */
122 
123 #define osd_req_op_data(oreq, whch, typ, fld)	\
124 	({						\
125 		BUG_ON(whch >= (oreq)->r_num_ops);	\
126 		&(oreq)->r_ops[whch].typ.fld;		\
127 	})
128 
129 static struct ceph_osd_data *
130 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
131 {
132 	BUG_ON(which >= osd_req->r_num_ops);
133 
134 	return &osd_req->r_ops[which].raw_data_in;
135 }
136 
137 struct ceph_osd_data *
138 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
139 			unsigned int which)
140 {
141 	return osd_req_op_data(osd_req, which, extent, osd_data);
142 }
143 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
144 
145 struct ceph_osd_data *
146 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
147 			unsigned int which)
148 {
149 	return osd_req_op_data(osd_req, which, cls, response_data);
150 }
151 EXPORT_SYMBOL(osd_req_op_cls_response_data);	/* ??? */
152 
153 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
154 			unsigned int which, struct page **pages,
155 			u64 length, u32 alignment,
156 			bool pages_from_pool, bool own_pages)
157 {
158 	struct ceph_osd_data *osd_data;
159 
160 	osd_data = osd_req_op_raw_data_in(osd_req, which);
161 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
162 				pages_from_pool, own_pages);
163 }
164 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
165 
166 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
167 			unsigned int which, struct page **pages,
168 			u64 length, u32 alignment,
169 			bool pages_from_pool, bool own_pages)
170 {
171 	struct ceph_osd_data *osd_data;
172 
173 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
174 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
175 				pages_from_pool, own_pages);
176 }
177 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
178 
179 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
180 			unsigned int which, struct ceph_pagelist *pagelist)
181 {
182 	struct ceph_osd_data *osd_data;
183 
184 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
185 	ceph_osd_data_pagelist_init(osd_data, pagelist);
186 }
187 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
188 
189 #ifdef CONFIG_BLOCK
190 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
191 			unsigned int which, struct bio *bio, size_t bio_length)
192 {
193 	struct ceph_osd_data *osd_data;
194 
195 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
196 	ceph_osd_data_bio_init(osd_data, bio, bio_length);
197 }
198 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
199 #endif /* CONFIG_BLOCK */
200 
201 static void osd_req_op_cls_request_info_pagelist(
202 			struct ceph_osd_request *osd_req,
203 			unsigned int which, struct ceph_pagelist *pagelist)
204 {
205 	struct ceph_osd_data *osd_data;
206 
207 	osd_data = osd_req_op_data(osd_req, which, cls, request_info);
208 	ceph_osd_data_pagelist_init(osd_data, pagelist);
209 }
210 
211 void osd_req_op_cls_request_data_pagelist(
212 			struct ceph_osd_request *osd_req,
213 			unsigned int which, struct ceph_pagelist *pagelist)
214 {
215 	struct ceph_osd_data *osd_data;
216 
217 	osd_data = osd_req_op_data(osd_req, which, cls, request_data);
218 	ceph_osd_data_pagelist_init(osd_data, pagelist);
219 }
220 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
221 
222 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
223 			unsigned int which, struct page **pages, u64 length,
224 			u32 alignment, bool pages_from_pool, bool own_pages)
225 {
226 	struct ceph_osd_data *osd_data;
227 
228 	osd_data = osd_req_op_data(osd_req, which, cls, request_data);
229 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
230 				pages_from_pool, own_pages);
231 }
232 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
233 
234 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
235 			unsigned int which, struct page **pages, u64 length,
236 			u32 alignment, bool pages_from_pool, bool own_pages)
237 {
238 	struct ceph_osd_data *osd_data;
239 
240 	osd_data = osd_req_op_data(osd_req, which, cls, response_data);
241 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
242 				pages_from_pool, own_pages);
243 }
244 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
245 
246 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
247 {
248 	switch (osd_data->type) {
249 	case CEPH_OSD_DATA_TYPE_NONE:
250 		return 0;
251 	case CEPH_OSD_DATA_TYPE_PAGES:
252 		return osd_data->length;
253 	case CEPH_OSD_DATA_TYPE_PAGELIST:
254 		return (u64)osd_data->pagelist->length;
255 #ifdef CONFIG_BLOCK
256 	case CEPH_OSD_DATA_TYPE_BIO:
257 		return (u64)osd_data->bio_length;
258 #endif /* CONFIG_BLOCK */
259 	default:
260 		WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
261 		return 0;
262 	}
263 }
264 
265 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
266 {
267 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
268 		int num_pages;
269 
270 		num_pages = calc_pages_for((u64)osd_data->alignment,
271 						(u64)osd_data->length);
272 		ceph_release_page_vector(osd_data->pages, num_pages);
273 	}
274 	ceph_osd_data_init(osd_data);
275 }
276 
277 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
278 			unsigned int which)
279 {
280 	struct ceph_osd_req_op *op;
281 
282 	BUG_ON(which >= osd_req->r_num_ops);
283 	op = &osd_req->r_ops[which];
284 
285 	switch (op->op) {
286 	case CEPH_OSD_OP_READ:
287 	case CEPH_OSD_OP_WRITE:
288 		ceph_osd_data_release(&op->extent.osd_data);
289 		break;
290 	case CEPH_OSD_OP_CALL:
291 		ceph_osd_data_release(&op->cls.request_info);
292 		ceph_osd_data_release(&op->cls.request_data);
293 		ceph_osd_data_release(&op->cls.response_data);
294 		break;
295 	default:
296 		break;
297 	}
298 }
299 
300 /*
301  * requests
302  */
303 static void ceph_osdc_release_request(struct kref *kref)
304 {
305 	struct ceph_osd_request *req = container_of(kref,
306 					    struct ceph_osd_request, r_kref);
307 	unsigned int which;
308 
309 	dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
310 	     req->r_request, req->r_reply);
311 	WARN_ON(!RB_EMPTY_NODE(&req->r_node));
312 	WARN_ON(!list_empty(&req->r_req_lru_item));
313 	WARN_ON(!list_empty(&req->r_osd_item));
314 	WARN_ON(!list_empty(&req->r_linger_item));
315 	WARN_ON(!list_empty(&req->r_linger_osd_item));
316 	WARN_ON(req->r_osd);
317 
318 	if (req->r_request)
319 		ceph_msg_put(req->r_request);
320 	if (req->r_reply) {
321 		ceph_msg_revoke_incoming(req->r_reply);
322 		ceph_msg_put(req->r_reply);
323 	}
324 
325 	for (which = 0; which < req->r_num_ops; which++)
326 		osd_req_op_data_release(req, which);
327 
328 	ceph_put_snap_context(req->r_snapc);
329 	if (req->r_mempool)
330 		mempool_free(req, req->r_osdc->req_mempool);
331 	else
332 		kmem_cache_free(ceph_osd_request_cache, req);
333 
334 }
335 
336 void ceph_osdc_get_request(struct ceph_osd_request *req)
337 {
338 	dout("%s %p (was %d)\n", __func__, req,
339 	     atomic_read(&req->r_kref.refcount));
340 	kref_get(&req->r_kref);
341 }
342 EXPORT_SYMBOL(ceph_osdc_get_request);
343 
344 void ceph_osdc_put_request(struct ceph_osd_request *req)
345 {
346 	dout("%s %p (was %d)\n", __func__, req,
347 	     atomic_read(&req->r_kref.refcount));
348 	kref_put(&req->r_kref, ceph_osdc_release_request);
349 }
350 EXPORT_SYMBOL(ceph_osdc_put_request);
351 
352 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
353 					       struct ceph_snap_context *snapc,
354 					       unsigned int num_ops,
355 					       bool use_mempool,
356 					       gfp_t gfp_flags)
357 {
358 	struct ceph_osd_request *req;
359 	struct ceph_msg *msg;
360 	size_t msg_size;
361 
362 	BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
363 	BUG_ON(num_ops > CEPH_OSD_MAX_OP);
364 
365 	msg_size = 4 + 4 + 8 + 8 + 4+8;
366 	msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
367 	msg_size += 1 + 8 + 4 + 4;     /* pg_t */
368 	msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
369 	msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
370 	msg_size += 8;  /* snapid */
371 	msg_size += 8;  /* snap_seq */
372 	msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
373 	msg_size += 4;
374 
375 	if (use_mempool) {
376 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
377 		memset(req, 0, sizeof(*req));
378 	} else {
379 		req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
380 	}
381 	if (req == NULL)
382 		return NULL;
383 
384 	req->r_osdc = osdc;
385 	req->r_mempool = use_mempool;
386 	req->r_num_ops = num_ops;
387 
388 	kref_init(&req->r_kref);
389 	init_completion(&req->r_completion);
390 	init_completion(&req->r_safe_completion);
391 	RB_CLEAR_NODE(&req->r_node);
392 	INIT_LIST_HEAD(&req->r_unsafe_item);
393 	INIT_LIST_HEAD(&req->r_linger_item);
394 	INIT_LIST_HEAD(&req->r_linger_osd_item);
395 	INIT_LIST_HEAD(&req->r_req_lru_item);
396 	INIT_LIST_HEAD(&req->r_osd_item);
397 
398 	req->r_base_oloc.pool = -1;
399 	req->r_target_oloc.pool = -1;
400 
401 	/* create reply message */
402 	if (use_mempool)
403 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
404 	else
405 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
406 				   OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
407 	if (!msg) {
408 		ceph_osdc_put_request(req);
409 		return NULL;
410 	}
411 	req->r_reply = msg;
412 
413 	/* create request message; allow space for oid */
414 	if (use_mempool)
415 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
416 	else
417 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
418 	if (!msg) {
419 		ceph_osdc_put_request(req);
420 		return NULL;
421 	}
422 
423 	memset(msg->front.iov_base, 0, msg->front.iov_len);
424 
425 	req->r_request = msg;
426 
427 	return req;
428 }
429 EXPORT_SYMBOL(ceph_osdc_alloc_request);
430 
431 static bool osd_req_opcode_valid(u16 opcode)
432 {
433 	switch (opcode) {
434 #define GENERATE_CASE(op, opcode, str)	case CEPH_OSD_OP_##op: return true;
435 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
436 #undef GENERATE_CASE
437 	default:
438 		return false;
439 	}
440 }
441 
442 /*
443  * This is an osd op init function for opcodes that have no data or
444  * other information associated with them.  It also serves as a
445  * common init routine for all the other init functions, below.
446  */
447 static struct ceph_osd_req_op *
448 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
449 				u16 opcode)
450 {
451 	struct ceph_osd_req_op *op;
452 
453 	BUG_ON(which >= osd_req->r_num_ops);
454 	BUG_ON(!osd_req_opcode_valid(opcode));
455 
456 	op = &osd_req->r_ops[which];
457 	memset(op, 0, sizeof (*op));
458 	op->op = opcode;
459 
460 	return op;
461 }
462 
463 void osd_req_op_init(struct ceph_osd_request *osd_req,
464 				unsigned int which, u16 opcode)
465 {
466 	(void)_osd_req_op_init(osd_req, which, opcode);
467 }
468 EXPORT_SYMBOL(osd_req_op_init);
469 
470 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
471 				unsigned int which, u16 opcode,
472 				u64 offset, u64 length,
473 				u64 truncate_size, u32 truncate_seq)
474 {
475 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
476 	size_t payload_len = 0;
477 
478 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
479 	       opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
480 	       opcode != CEPH_OSD_OP_TRUNCATE);
481 
482 	op->extent.offset = offset;
483 	op->extent.length = length;
484 	op->extent.truncate_size = truncate_size;
485 	op->extent.truncate_seq = truncate_seq;
486 	if (opcode == CEPH_OSD_OP_WRITE)
487 		payload_len += length;
488 
489 	op->payload_len = payload_len;
490 }
491 EXPORT_SYMBOL(osd_req_op_extent_init);
492 
493 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
494 				unsigned int which, u64 length)
495 {
496 	struct ceph_osd_req_op *op;
497 	u64 previous;
498 
499 	BUG_ON(which >= osd_req->r_num_ops);
500 	op = &osd_req->r_ops[which];
501 	previous = op->extent.length;
502 
503 	if (length == previous)
504 		return;		/* Nothing to do */
505 	BUG_ON(length > previous);
506 
507 	op->extent.length = length;
508 	op->payload_len -= previous - length;
509 }
510 EXPORT_SYMBOL(osd_req_op_extent_update);
511 
512 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
513 			u16 opcode, const char *class, const char *method)
514 {
515 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
516 	struct ceph_pagelist *pagelist;
517 	size_t payload_len = 0;
518 	size_t size;
519 
520 	BUG_ON(opcode != CEPH_OSD_OP_CALL);
521 
522 	pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
523 	BUG_ON(!pagelist);
524 	ceph_pagelist_init(pagelist);
525 
526 	op->cls.class_name = class;
527 	size = strlen(class);
528 	BUG_ON(size > (size_t) U8_MAX);
529 	op->cls.class_len = size;
530 	ceph_pagelist_append(pagelist, class, size);
531 	payload_len += size;
532 
533 	op->cls.method_name = method;
534 	size = strlen(method);
535 	BUG_ON(size > (size_t) U8_MAX);
536 	op->cls.method_len = size;
537 	ceph_pagelist_append(pagelist, method, size);
538 	payload_len += size;
539 
540 	osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
541 
542 	op->cls.argc = 0;	/* currently unused */
543 
544 	op->payload_len = payload_len;
545 }
546 EXPORT_SYMBOL(osd_req_op_cls_init);
547 
548 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
549 				unsigned int which, u16 opcode,
550 				u64 cookie, u64 version, int flag)
551 {
552 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
553 
554 	BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
555 
556 	op->watch.cookie = cookie;
557 	op->watch.ver = version;
558 	if (opcode == CEPH_OSD_OP_WATCH && flag)
559 		op->watch.flag = (u8)1;
560 }
561 EXPORT_SYMBOL(osd_req_op_watch_init);
562 
563 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
564 				unsigned int which,
565 				u64 expected_object_size,
566 				u64 expected_write_size)
567 {
568 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
569 						      CEPH_OSD_OP_SETALLOCHINT);
570 
571 	op->alloc_hint.expected_object_size = expected_object_size;
572 	op->alloc_hint.expected_write_size = expected_write_size;
573 
574 	/*
575 	 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
576 	 * not worth a feature bit.  Set FAILOK per-op flag to make
577 	 * sure older osds don't trip over an unsupported opcode.
578 	 */
579 	op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
580 }
581 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
582 
583 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
584 				struct ceph_osd_data *osd_data)
585 {
586 	u64 length = ceph_osd_data_length(osd_data);
587 
588 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
589 		BUG_ON(length > (u64) SIZE_MAX);
590 		if (length)
591 			ceph_msg_data_add_pages(msg, osd_data->pages,
592 					length, osd_data->alignment);
593 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
594 		BUG_ON(!length);
595 		ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
596 #ifdef CONFIG_BLOCK
597 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
598 		ceph_msg_data_add_bio(msg, osd_data->bio, length);
599 #endif
600 	} else {
601 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
602 	}
603 }
604 
605 static u64 osd_req_encode_op(struct ceph_osd_request *req,
606 			      struct ceph_osd_op *dst, unsigned int which)
607 {
608 	struct ceph_osd_req_op *src;
609 	struct ceph_osd_data *osd_data;
610 	u64 request_data_len = 0;
611 	u64 data_length;
612 
613 	BUG_ON(which >= req->r_num_ops);
614 	src = &req->r_ops[which];
615 	if (WARN_ON(!osd_req_opcode_valid(src->op))) {
616 		pr_err("unrecognized osd opcode %d\n", src->op);
617 
618 		return 0;
619 	}
620 
621 	switch (src->op) {
622 	case CEPH_OSD_OP_STAT:
623 		osd_data = &src->raw_data_in;
624 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
625 		break;
626 	case CEPH_OSD_OP_READ:
627 	case CEPH_OSD_OP_WRITE:
628 	case CEPH_OSD_OP_ZERO:
629 	case CEPH_OSD_OP_DELETE:
630 	case CEPH_OSD_OP_TRUNCATE:
631 		if (src->op == CEPH_OSD_OP_WRITE)
632 			request_data_len = src->extent.length;
633 		dst->extent.offset = cpu_to_le64(src->extent.offset);
634 		dst->extent.length = cpu_to_le64(src->extent.length);
635 		dst->extent.truncate_size =
636 			cpu_to_le64(src->extent.truncate_size);
637 		dst->extent.truncate_seq =
638 			cpu_to_le32(src->extent.truncate_seq);
639 		osd_data = &src->extent.osd_data;
640 		if (src->op == CEPH_OSD_OP_WRITE)
641 			ceph_osdc_msg_data_add(req->r_request, osd_data);
642 		else
643 			ceph_osdc_msg_data_add(req->r_reply, osd_data);
644 		break;
645 	case CEPH_OSD_OP_CALL:
646 		dst->cls.class_len = src->cls.class_len;
647 		dst->cls.method_len = src->cls.method_len;
648 		osd_data = &src->cls.request_info;
649 		ceph_osdc_msg_data_add(req->r_request, osd_data);
650 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
651 		request_data_len = osd_data->pagelist->length;
652 
653 		osd_data = &src->cls.request_data;
654 		data_length = ceph_osd_data_length(osd_data);
655 		if (data_length) {
656 			BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
657 			dst->cls.indata_len = cpu_to_le32(data_length);
658 			ceph_osdc_msg_data_add(req->r_request, osd_data);
659 			src->payload_len += data_length;
660 			request_data_len += data_length;
661 		}
662 		osd_data = &src->cls.response_data;
663 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
664 		break;
665 	case CEPH_OSD_OP_STARTSYNC:
666 		break;
667 	case CEPH_OSD_OP_NOTIFY_ACK:
668 	case CEPH_OSD_OP_WATCH:
669 		dst->watch.cookie = cpu_to_le64(src->watch.cookie);
670 		dst->watch.ver = cpu_to_le64(src->watch.ver);
671 		dst->watch.flag = src->watch.flag;
672 		break;
673 	case CEPH_OSD_OP_SETALLOCHINT:
674 		dst->alloc_hint.expected_object_size =
675 		    cpu_to_le64(src->alloc_hint.expected_object_size);
676 		dst->alloc_hint.expected_write_size =
677 		    cpu_to_le64(src->alloc_hint.expected_write_size);
678 		break;
679 	default:
680 		pr_err("unsupported osd opcode %s\n",
681 			ceph_osd_op_name(src->op));
682 		WARN_ON(1);
683 
684 		return 0;
685 	}
686 
687 	dst->op = cpu_to_le16(src->op);
688 	dst->flags = cpu_to_le32(src->flags);
689 	dst->payload_len = cpu_to_le32(src->payload_len);
690 
691 	return request_data_len;
692 }
693 
694 /*
695  * build new request AND message, calculate layout, and adjust file
696  * extent as needed.
697  *
698  * if the file was recently truncated, we include information about its
699  * old and new size so that the object can be updated appropriately.  (we
700  * avoid synchronously deleting truncated objects because it's slow.)
701  *
702  * if @do_sync, include a 'startsync' command so that the osd will flush
703  * data quickly.
704  */
705 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
706 					       struct ceph_file_layout *layout,
707 					       struct ceph_vino vino,
708 					       u64 off, u64 *plen, int num_ops,
709 					       int opcode, int flags,
710 					       struct ceph_snap_context *snapc,
711 					       u32 truncate_seq,
712 					       u64 truncate_size,
713 					       bool use_mempool)
714 {
715 	struct ceph_osd_request *req;
716 	u64 objnum = 0;
717 	u64 objoff = 0;
718 	u64 objlen = 0;
719 	u32 object_size;
720 	u64 object_base;
721 	int r;
722 
723 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
724 	       opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
725 	       opcode != CEPH_OSD_OP_TRUNCATE);
726 
727 	req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
728 					GFP_NOFS);
729 	if (!req)
730 		return ERR_PTR(-ENOMEM);
731 
732 	req->r_flags = flags;
733 
734 	/* calculate max write size */
735 	r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
736 	if (r < 0) {
737 		ceph_osdc_put_request(req);
738 		return ERR_PTR(r);
739 	}
740 
741 	object_size = le32_to_cpu(layout->fl_object_size);
742 	object_base = off - objoff;
743 	if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
744 		if (truncate_size <= object_base) {
745 			truncate_size = 0;
746 		} else {
747 			truncate_size -= object_base;
748 			if (truncate_size > object_size)
749 				truncate_size = object_size;
750 		}
751 	}
752 
753 	osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
754 				truncate_size, truncate_seq);
755 
756 	/*
757 	 * A second op in the ops array means the caller wants to
758 	 * also issue a include a 'startsync' command so that the
759 	 * osd will flush data quickly.
760 	 */
761 	if (num_ops > 1)
762 		osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
763 
764 	req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
765 
766 	snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
767 		 "%llx.%08llx", vino.ino, objnum);
768 	req->r_base_oid.name_len = strlen(req->r_base_oid.name);
769 
770 	return req;
771 }
772 EXPORT_SYMBOL(ceph_osdc_new_request);
773 
774 /*
775  * We keep osd requests in an rbtree, sorted by ->r_tid.
776  */
777 static void __insert_request(struct ceph_osd_client *osdc,
778 			     struct ceph_osd_request *new)
779 {
780 	struct rb_node **p = &osdc->requests.rb_node;
781 	struct rb_node *parent = NULL;
782 	struct ceph_osd_request *req = NULL;
783 
784 	while (*p) {
785 		parent = *p;
786 		req = rb_entry(parent, struct ceph_osd_request, r_node);
787 		if (new->r_tid < req->r_tid)
788 			p = &(*p)->rb_left;
789 		else if (new->r_tid > req->r_tid)
790 			p = &(*p)->rb_right;
791 		else
792 			BUG();
793 	}
794 
795 	rb_link_node(&new->r_node, parent, p);
796 	rb_insert_color(&new->r_node, &osdc->requests);
797 }
798 
799 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
800 						 u64 tid)
801 {
802 	struct ceph_osd_request *req;
803 	struct rb_node *n = osdc->requests.rb_node;
804 
805 	while (n) {
806 		req = rb_entry(n, struct ceph_osd_request, r_node);
807 		if (tid < req->r_tid)
808 			n = n->rb_left;
809 		else if (tid > req->r_tid)
810 			n = n->rb_right;
811 		else
812 			return req;
813 	}
814 	return NULL;
815 }
816 
817 static struct ceph_osd_request *
818 __lookup_request_ge(struct ceph_osd_client *osdc,
819 		    u64 tid)
820 {
821 	struct ceph_osd_request *req;
822 	struct rb_node *n = osdc->requests.rb_node;
823 
824 	while (n) {
825 		req = rb_entry(n, struct ceph_osd_request, r_node);
826 		if (tid < req->r_tid) {
827 			if (!n->rb_left)
828 				return req;
829 			n = n->rb_left;
830 		} else if (tid > req->r_tid) {
831 			n = n->rb_right;
832 		} else {
833 			return req;
834 		}
835 	}
836 	return NULL;
837 }
838 
839 static void __kick_linger_request(struct ceph_osd_request *req)
840 {
841 	struct ceph_osd_client *osdc = req->r_osdc;
842 	struct ceph_osd *osd = req->r_osd;
843 
844 	/*
845 	 * Linger requests need to be resent with a new tid to avoid
846 	 * the dup op detection logic on the OSDs.  Achieve this with
847 	 * a re-register dance instead of open-coding.
848 	 */
849 	ceph_osdc_get_request(req);
850 	if (!list_empty(&req->r_linger_item))
851 		__unregister_linger_request(osdc, req);
852 	else
853 		__unregister_request(osdc, req);
854 	__register_request(osdc, req);
855 	ceph_osdc_put_request(req);
856 
857 	/*
858 	 * Unless request has been registered as both normal and
859 	 * lingering, __unregister{,_linger}_request clears r_osd.
860 	 * However, here we need to preserve r_osd to make sure we
861 	 * requeue on the same OSD.
862 	 */
863 	WARN_ON(req->r_osd || !osd);
864 	req->r_osd = osd;
865 
866 	dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
867 	__enqueue_request(req);
868 }
869 
870 /*
871  * Resubmit requests pending on the given osd.
872  */
873 static void __kick_osd_requests(struct ceph_osd_client *osdc,
874 				struct ceph_osd *osd)
875 {
876 	struct ceph_osd_request *req, *nreq;
877 	LIST_HEAD(resend);
878 	LIST_HEAD(resend_linger);
879 	int err;
880 
881 	dout("%s osd%d\n", __func__, osd->o_osd);
882 	err = __reset_osd(osdc, osd);
883 	if (err)
884 		return;
885 
886 	/*
887 	 * Build up a list of requests to resend by traversing the
888 	 * osd's list of requests.  Requests for a given object are
889 	 * sent in tid order, and that is also the order they're
890 	 * kept on this list.  Therefore all requests that are in
891 	 * flight will be found first, followed by all requests that
892 	 * have not yet been sent.  And to resend requests while
893 	 * preserving this order we will want to put any sent
894 	 * requests back on the front of the osd client's unsent
895 	 * list.
896 	 *
897 	 * So we build a separate ordered list of already-sent
898 	 * requests for the affected osd and splice it onto the
899 	 * front of the osd client's unsent list.  Once we've seen a
900 	 * request that has not yet been sent we're done.  Those
901 	 * requests are already sitting right where they belong.
902 	 */
903 	list_for_each_entry(req, &osd->o_requests, r_osd_item) {
904 		if (!req->r_sent)
905 			break;
906 
907 		if (!req->r_linger) {
908 			dout("%s requeueing %p tid %llu\n", __func__, req,
909 			     req->r_tid);
910 			list_move_tail(&req->r_req_lru_item, &resend);
911 			req->r_flags |= CEPH_OSD_FLAG_RETRY;
912 		} else {
913 			list_move_tail(&req->r_req_lru_item, &resend_linger);
914 		}
915 	}
916 	list_splice(&resend, &osdc->req_unsent);
917 
918 	/*
919 	 * Both registered and not yet registered linger requests are
920 	 * enqueued with a new tid on the same OSD.  We add/move them
921 	 * to req_unsent/o_requests at the end to keep things in tid
922 	 * order.
923 	 */
924 	list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
925 				 r_linger_osd_item) {
926 		WARN_ON(!list_empty(&req->r_req_lru_item));
927 		__kick_linger_request(req);
928 	}
929 
930 	list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
931 		__kick_linger_request(req);
932 }
933 
934 /*
935  * If the osd connection drops, we need to resubmit all requests.
936  */
937 static void osd_reset(struct ceph_connection *con)
938 {
939 	struct ceph_osd *osd = con->private;
940 	struct ceph_osd_client *osdc;
941 
942 	if (!osd)
943 		return;
944 	dout("osd_reset osd%d\n", osd->o_osd);
945 	osdc = osd->o_osdc;
946 	down_read(&osdc->map_sem);
947 	mutex_lock(&osdc->request_mutex);
948 	__kick_osd_requests(osdc, osd);
949 	__send_queued(osdc);
950 	mutex_unlock(&osdc->request_mutex);
951 	up_read(&osdc->map_sem);
952 }
953 
954 /*
955  * Track open sessions with osds.
956  */
957 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
958 {
959 	struct ceph_osd *osd;
960 
961 	osd = kzalloc(sizeof(*osd), GFP_NOFS);
962 	if (!osd)
963 		return NULL;
964 
965 	atomic_set(&osd->o_ref, 1);
966 	osd->o_osdc = osdc;
967 	osd->o_osd = onum;
968 	RB_CLEAR_NODE(&osd->o_node);
969 	INIT_LIST_HEAD(&osd->o_requests);
970 	INIT_LIST_HEAD(&osd->o_linger_requests);
971 	INIT_LIST_HEAD(&osd->o_osd_lru);
972 	osd->o_incarnation = 1;
973 
974 	ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
975 
976 	INIT_LIST_HEAD(&osd->o_keepalive_item);
977 	return osd;
978 }
979 
980 static struct ceph_osd *get_osd(struct ceph_osd *osd)
981 {
982 	if (atomic_inc_not_zero(&osd->o_ref)) {
983 		dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
984 		     atomic_read(&osd->o_ref));
985 		return osd;
986 	} else {
987 		dout("get_osd %p FAIL\n", osd);
988 		return NULL;
989 	}
990 }
991 
992 static void put_osd(struct ceph_osd *osd)
993 {
994 	dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
995 	     atomic_read(&osd->o_ref) - 1);
996 	if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
997 		struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
998 
999 		ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1000 		kfree(osd);
1001 	}
1002 }
1003 
1004 /*
1005  * remove an osd from our map
1006  */
1007 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1008 {
1009 	dout("__remove_osd %p\n", osd);
1010 	WARN_ON(!list_empty(&osd->o_requests));
1011 	WARN_ON(!list_empty(&osd->o_linger_requests));
1012 
1013 	rb_erase(&osd->o_node, &osdc->osds);
1014 	list_del_init(&osd->o_osd_lru);
1015 	ceph_con_close(&osd->o_con);
1016 	put_osd(osd);
1017 }
1018 
1019 static void remove_all_osds(struct ceph_osd_client *osdc)
1020 {
1021 	dout("%s %p\n", __func__, osdc);
1022 	mutex_lock(&osdc->request_mutex);
1023 	while (!RB_EMPTY_ROOT(&osdc->osds)) {
1024 		struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1025 						struct ceph_osd, o_node);
1026 		__remove_osd(osdc, osd);
1027 	}
1028 	mutex_unlock(&osdc->request_mutex);
1029 }
1030 
1031 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1032 			      struct ceph_osd *osd)
1033 {
1034 	dout("%s %p\n", __func__, osd);
1035 	BUG_ON(!list_empty(&osd->o_osd_lru));
1036 
1037 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1038 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1039 }
1040 
1041 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1042 				  struct ceph_osd *osd)
1043 {
1044 	dout("%s %p\n", __func__, osd);
1045 
1046 	if (list_empty(&osd->o_requests) &&
1047 	    list_empty(&osd->o_linger_requests))
1048 		__move_osd_to_lru(osdc, osd);
1049 }
1050 
1051 static void __remove_osd_from_lru(struct ceph_osd *osd)
1052 {
1053 	dout("__remove_osd_from_lru %p\n", osd);
1054 	if (!list_empty(&osd->o_osd_lru))
1055 		list_del_init(&osd->o_osd_lru);
1056 }
1057 
1058 static void remove_old_osds(struct ceph_osd_client *osdc)
1059 {
1060 	struct ceph_osd *osd, *nosd;
1061 
1062 	dout("__remove_old_osds %p\n", osdc);
1063 	mutex_lock(&osdc->request_mutex);
1064 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1065 		if (time_before(jiffies, osd->lru_ttl))
1066 			break;
1067 		__remove_osd(osdc, osd);
1068 	}
1069 	mutex_unlock(&osdc->request_mutex);
1070 }
1071 
1072 /*
1073  * reset osd connect
1074  */
1075 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1076 {
1077 	struct ceph_entity_addr *peer_addr;
1078 
1079 	dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1080 	if (list_empty(&osd->o_requests) &&
1081 	    list_empty(&osd->o_linger_requests)) {
1082 		__remove_osd(osdc, osd);
1083 
1084 		return -ENODEV;
1085 	}
1086 
1087 	peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1088 	if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1089 			!ceph_con_opened(&osd->o_con)) {
1090 		struct ceph_osd_request *req;
1091 
1092 		dout("osd addr hasn't changed and connection never opened, "
1093 		     "letting msgr retry\n");
1094 		/* touch each r_stamp for handle_timeout()'s benfit */
1095 		list_for_each_entry(req, &osd->o_requests, r_osd_item)
1096 			req->r_stamp = jiffies;
1097 
1098 		return -EAGAIN;
1099 	}
1100 
1101 	ceph_con_close(&osd->o_con);
1102 	ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1103 	osd->o_incarnation++;
1104 
1105 	return 0;
1106 }
1107 
1108 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1109 {
1110 	struct rb_node **p = &osdc->osds.rb_node;
1111 	struct rb_node *parent = NULL;
1112 	struct ceph_osd *osd = NULL;
1113 
1114 	dout("__insert_osd %p osd%d\n", new, new->o_osd);
1115 	while (*p) {
1116 		parent = *p;
1117 		osd = rb_entry(parent, struct ceph_osd, o_node);
1118 		if (new->o_osd < osd->o_osd)
1119 			p = &(*p)->rb_left;
1120 		else if (new->o_osd > osd->o_osd)
1121 			p = &(*p)->rb_right;
1122 		else
1123 			BUG();
1124 	}
1125 
1126 	rb_link_node(&new->o_node, parent, p);
1127 	rb_insert_color(&new->o_node, &osdc->osds);
1128 }
1129 
1130 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1131 {
1132 	struct ceph_osd *osd;
1133 	struct rb_node *n = osdc->osds.rb_node;
1134 
1135 	while (n) {
1136 		osd = rb_entry(n, struct ceph_osd, o_node);
1137 		if (o < osd->o_osd)
1138 			n = n->rb_left;
1139 		else if (o > osd->o_osd)
1140 			n = n->rb_right;
1141 		else
1142 			return osd;
1143 	}
1144 	return NULL;
1145 }
1146 
1147 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1148 {
1149 	schedule_delayed_work(&osdc->timeout_work,
1150 			osdc->client->options->osd_keepalive_timeout * HZ);
1151 }
1152 
1153 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1154 {
1155 	cancel_delayed_work(&osdc->timeout_work);
1156 }
1157 
1158 /*
1159  * Register request, assign tid.  If this is the first request, set up
1160  * the timeout event.
1161  */
1162 static void __register_request(struct ceph_osd_client *osdc,
1163 			       struct ceph_osd_request *req)
1164 {
1165 	req->r_tid = ++osdc->last_tid;
1166 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1167 	dout("__register_request %p tid %lld\n", req, req->r_tid);
1168 	__insert_request(osdc, req);
1169 	ceph_osdc_get_request(req);
1170 	osdc->num_requests++;
1171 	if (osdc->num_requests == 1) {
1172 		dout(" first request, scheduling timeout\n");
1173 		__schedule_osd_timeout(osdc);
1174 	}
1175 }
1176 
1177 /*
1178  * called under osdc->request_mutex
1179  */
1180 static void __unregister_request(struct ceph_osd_client *osdc,
1181 				 struct ceph_osd_request *req)
1182 {
1183 	if (RB_EMPTY_NODE(&req->r_node)) {
1184 		dout("__unregister_request %p tid %lld not registered\n",
1185 			req, req->r_tid);
1186 		return;
1187 	}
1188 
1189 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1190 	rb_erase(&req->r_node, &osdc->requests);
1191 	RB_CLEAR_NODE(&req->r_node);
1192 	osdc->num_requests--;
1193 
1194 	if (req->r_osd) {
1195 		/* make sure the original request isn't in flight. */
1196 		ceph_msg_revoke(req->r_request);
1197 
1198 		list_del_init(&req->r_osd_item);
1199 		maybe_move_osd_to_lru(osdc, req->r_osd);
1200 		if (list_empty(&req->r_linger_osd_item))
1201 			req->r_osd = NULL;
1202 	}
1203 
1204 	list_del_init(&req->r_req_lru_item);
1205 	ceph_osdc_put_request(req);
1206 
1207 	if (osdc->num_requests == 0) {
1208 		dout(" no requests, canceling timeout\n");
1209 		__cancel_osd_timeout(osdc);
1210 	}
1211 }
1212 
1213 /*
1214  * Cancel a previously queued request message
1215  */
1216 static void __cancel_request(struct ceph_osd_request *req)
1217 {
1218 	if (req->r_sent && req->r_osd) {
1219 		ceph_msg_revoke(req->r_request);
1220 		req->r_sent = 0;
1221 	}
1222 }
1223 
1224 static void __register_linger_request(struct ceph_osd_client *osdc,
1225 				    struct ceph_osd_request *req)
1226 {
1227 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1228 	WARN_ON(!req->r_linger);
1229 
1230 	ceph_osdc_get_request(req);
1231 	list_add_tail(&req->r_linger_item, &osdc->req_linger);
1232 	if (req->r_osd)
1233 		list_add_tail(&req->r_linger_osd_item,
1234 			      &req->r_osd->o_linger_requests);
1235 }
1236 
1237 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1238 					struct ceph_osd_request *req)
1239 {
1240 	WARN_ON(!req->r_linger);
1241 
1242 	if (list_empty(&req->r_linger_item)) {
1243 		dout("%s %p tid %llu not registered\n", __func__, req,
1244 		     req->r_tid);
1245 		return;
1246 	}
1247 
1248 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1249 	list_del_init(&req->r_linger_item);
1250 
1251 	if (req->r_osd) {
1252 		list_del_init(&req->r_linger_osd_item);
1253 		maybe_move_osd_to_lru(osdc, req->r_osd);
1254 		if (list_empty(&req->r_osd_item))
1255 			req->r_osd = NULL;
1256 	}
1257 
1258 	list_del_init(&req->r_req_lru_item); /* can be on notarget */
1259 	ceph_osdc_put_request(req);
1260 }
1261 
1262 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1263 				  struct ceph_osd_request *req)
1264 {
1265 	if (!req->r_linger) {
1266 		dout("set_request_linger %p\n", req);
1267 		req->r_linger = 1;
1268 	}
1269 }
1270 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1271 
1272 /*
1273  * Returns whether a request should be blocked from being sent
1274  * based on the current osdmap and osd_client settings.
1275  *
1276  * Caller should hold map_sem for read.
1277  */
1278 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1279 				   struct ceph_osd_request *req)
1280 {
1281 	bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1282 	bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1283 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1284 	return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1285 		(req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1286 }
1287 
1288 /*
1289  * Calculate mapping of a request to a PG.  Takes tiering into account.
1290  */
1291 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1292 			     struct ceph_osd_request *req,
1293 			     struct ceph_pg *pg_out)
1294 {
1295 	bool need_check_tiering;
1296 
1297 	need_check_tiering = false;
1298 	if (req->r_target_oloc.pool == -1) {
1299 		req->r_target_oloc = req->r_base_oloc; /* struct */
1300 		need_check_tiering = true;
1301 	}
1302 	if (req->r_target_oid.name_len == 0) {
1303 		ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1304 		need_check_tiering = true;
1305 	}
1306 
1307 	if (need_check_tiering &&
1308 	    (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1309 		struct ceph_pg_pool_info *pi;
1310 
1311 		pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1312 		if (pi) {
1313 			if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1314 			    pi->read_tier >= 0)
1315 				req->r_target_oloc.pool = pi->read_tier;
1316 			if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1317 			    pi->write_tier >= 0)
1318 				req->r_target_oloc.pool = pi->write_tier;
1319 		}
1320 		/* !pi is caught in ceph_oloc_oid_to_pg() */
1321 	}
1322 
1323 	return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1324 				   &req->r_target_oid, pg_out);
1325 }
1326 
1327 static void __enqueue_request(struct ceph_osd_request *req)
1328 {
1329 	struct ceph_osd_client *osdc = req->r_osdc;
1330 
1331 	dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1332 	     req->r_osd ? req->r_osd->o_osd : -1);
1333 
1334 	if (req->r_osd) {
1335 		__remove_osd_from_lru(req->r_osd);
1336 		list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1337 		list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1338 	} else {
1339 		list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1340 	}
1341 }
1342 
1343 /*
1344  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1345  * (as needed), and set the request r_osd appropriately.  If there is
1346  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1347  * (unsent, homeless) or leave on in-flight lru.
1348  *
1349  * Return 0 if unchanged, 1 if changed, or negative on error.
1350  *
1351  * Caller should hold map_sem for read and request_mutex.
1352  */
1353 static int __map_request(struct ceph_osd_client *osdc,
1354 			 struct ceph_osd_request *req, int force_resend)
1355 {
1356 	struct ceph_pg pgid;
1357 	int acting[CEPH_PG_MAX_SIZE];
1358 	int num, o;
1359 	int err;
1360 	bool was_paused;
1361 
1362 	dout("map_request %p tid %lld\n", req, req->r_tid);
1363 
1364 	err = __calc_request_pg(osdc->osdmap, req, &pgid);
1365 	if (err) {
1366 		list_move(&req->r_req_lru_item, &osdc->req_notarget);
1367 		return err;
1368 	}
1369 	req->r_pgid = pgid;
1370 
1371 	num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1372 	if (num < 0)
1373 		num = 0;
1374 
1375 	was_paused = req->r_paused;
1376 	req->r_paused = __req_should_be_paused(osdc, req);
1377 	if (was_paused && !req->r_paused)
1378 		force_resend = 1;
1379 
1380 	if ((!force_resend &&
1381 	     req->r_osd && req->r_osd->o_osd == o &&
1382 	     req->r_sent >= req->r_osd->o_incarnation &&
1383 	     req->r_num_pg_osds == num &&
1384 	     memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1385 	    (req->r_osd == NULL && o == -1) ||
1386 	    req->r_paused)
1387 		return 0;  /* no change */
1388 
1389 	dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1390 	     req->r_tid, pgid.pool, pgid.seed, o,
1391 	     req->r_osd ? req->r_osd->o_osd : -1);
1392 
1393 	/* record full pg acting set */
1394 	memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1395 	req->r_num_pg_osds = num;
1396 
1397 	if (req->r_osd) {
1398 		__cancel_request(req);
1399 		list_del_init(&req->r_osd_item);
1400 		list_del_init(&req->r_linger_osd_item);
1401 		req->r_osd = NULL;
1402 	}
1403 
1404 	req->r_osd = __lookup_osd(osdc, o);
1405 	if (!req->r_osd && o >= 0) {
1406 		err = -ENOMEM;
1407 		req->r_osd = create_osd(osdc, o);
1408 		if (!req->r_osd) {
1409 			list_move(&req->r_req_lru_item, &osdc->req_notarget);
1410 			goto out;
1411 		}
1412 
1413 		dout("map_request osd %p is osd%d\n", req->r_osd, o);
1414 		__insert_osd(osdc, req->r_osd);
1415 
1416 		ceph_con_open(&req->r_osd->o_con,
1417 			      CEPH_ENTITY_TYPE_OSD, o,
1418 			      &osdc->osdmap->osd_addr[o]);
1419 	}
1420 
1421 	__enqueue_request(req);
1422 	err = 1;   /* osd or pg changed */
1423 
1424 out:
1425 	return err;
1426 }
1427 
1428 /*
1429  * caller should hold map_sem (for read) and request_mutex
1430  */
1431 static void __send_request(struct ceph_osd_client *osdc,
1432 			   struct ceph_osd_request *req)
1433 {
1434 	void *p;
1435 
1436 	dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1437 	     req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1438 	     (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1439 
1440 	/* fill in message content that changes each time we send it */
1441 	put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1442 	put_unaligned_le32(req->r_flags, req->r_request_flags);
1443 	put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1444 	p = req->r_request_pgid;
1445 	ceph_encode_64(&p, req->r_pgid.pool);
1446 	ceph_encode_32(&p, req->r_pgid.seed);
1447 	put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1448 	memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1449 	       sizeof(req->r_reassert_version));
1450 
1451 	req->r_stamp = jiffies;
1452 	list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1453 
1454 	ceph_msg_get(req->r_request); /* send consumes a ref */
1455 
1456 	req->r_sent = req->r_osd->o_incarnation;
1457 
1458 	ceph_con_send(&req->r_osd->o_con, req->r_request);
1459 }
1460 
1461 /*
1462  * Send any requests in the queue (req_unsent).
1463  */
1464 static void __send_queued(struct ceph_osd_client *osdc)
1465 {
1466 	struct ceph_osd_request *req, *tmp;
1467 
1468 	dout("__send_queued\n");
1469 	list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1470 		__send_request(osdc, req);
1471 }
1472 
1473 /*
1474  * Caller should hold map_sem for read and request_mutex.
1475  */
1476 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1477 				     struct ceph_osd_request *req,
1478 				     bool nofail)
1479 {
1480 	int rc;
1481 
1482 	__register_request(osdc, req);
1483 	req->r_sent = 0;
1484 	req->r_got_reply = 0;
1485 	rc = __map_request(osdc, req, 0);
1486 	if (rc < 0) {
1487 		if (nofail) {
1488 			dout("osdc_start_request failed map, "
1489 				" will retry %lld\n", req->r_tid);
1490 			rc = 0;
1491 		} else {
1492 			__unregister_request(osdc, req);
1493 		}
1494 		return rc;
1495 	}
1496 
1497 	if (req->r_osd == NULL) {
1498 		dout("send_request %p no up osds in pg\n", req);
1499 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1500 	} else {
1501 		__send_queued(osdc);
1502 	}
1503 
1504 	return 0;
1505 }
1506 
1507 /*
1508  * Timeout callback, called every N seconds when 1 or more osd
1509  * requests has been active for more than N seconds.  When this
1510  * happens, we ping all OSDs with requests who have timed out to
1511  * ensure any communications channel reset is detected.  Reset the
1512  * request timeouts another N seconds in the future as we go.
1513  * Reschedule the timeout event another N seconds in future (unless
1514  * there are no open requests).
1515  */
1516 static void handle_timeout(struct work_struct *work)
1517 {
1518 	struct ceph_osd_client *osdc =
1519 		container_of(work, struct ceph_osd_client, timeout_work.work);
1520 	struct ceph_osd_request *req;
1521 	struct ceph_osd *osd;
1522 	unsigned long keepalive =
1523 		osdc->client->options->osd_keepalive_timeout * HZ;
1524 	struct list_head slow_osds;
1525 	dout("timeout\n");
1526 	down_read(&osdc->map_sem);
1527 
1528 	ceph_monc_request_next_osdmap(&osdc->client->monc);
1529 
1530 	mutex_lock(&osdc->request_mutex);
1531 
1532 	/*
1533 	 * ping osds that are a bit slow.  this ensures that if there
1534 	 * is a break in the TCP connection we will notice, and reopen
1535 	 * a connection with that osd (from the fault callback).
1536 	 */
1537 	INIT_LIST_HEAD(&slow_osds);
1538 	list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1539 		if (time_before(jiffies, req->r_stamp + keepalive))
1540 			break;
1541 
1542 		osd = req->r_osd;
1543 		BUG_ON(!osd);
1544 		dout(" tid %llu is slow, will send keepalive on osd%d\n",
1545 		     req->r_tid, osd->o_osd);
1546 		list_move_tail(&osd->o_keepalive_item, &slow_osds);
1547 	}
1548 	while (!list_empty(&slow_osds)) {
1549 		osd = list_entry(slow_osds.next, struct ceph_osd,
1550 				 o_keepalive_item);
1551 		list_del_init(&osd->o_keepalive_item);
1552 		ceph_con_keepalive(&osd->o_con);
1553 	}
1554 
1555 	__schedule_osd_timeout(osdc);
1556 	__send_queued(osdc);
1557 	mutex_unlock(&osdc->request_mutex);
1558 	up_read(&osdc->map_sem);
1559 }
1560 
1561 static void handle_osds_timeout(struct work_struct *work)
1562 {
1563 	struct ceph_osd_client *osdc =
1564 		container_of(work, struct ceph_osd_client,
1565 			     osds_timeout_work.work);
1566 	unsigned long delay =
1567 		osdc->client->options->osd_idle_ttl * HZ >> 2;
1568 
1569 	dout("osds timeout\n");
1570 	down_read(&osdc->map_sem);
1571 	remove_old_osds(osdc);
1572 	up_read(&osdc->map_sem);
1573 
1574 	schedule_delayed_work(&osdc->osds_timeout_work,
1575 			      round_jiffies_relative(delay));
1576 }
1577 
1578 static int ceph_oloc_decode(void **p, void *end,
1579 			    struct ceph_object_locator *oloc)
1580 {
1581 	u8 struct_v, struct_cv;
1582 	u32 len;
1583 	void *struct_end;
1584 	int ret = 0;
1585 
1586 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1587 	struct_v = ceph_decode_8(p);
1588 	struct_cv = ceph_decode_8(p);
1589 	if (struct_v < 3) {
1590 		pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1591 			struct_v, struct_cv);
1592 		goto e_inval;
1593 	}
1594 	if (struct_cv > 6) {
1595 		pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1596 			struct_v, struct_cv);
1597 		goto e_inval;
1598 	}
1599 	len = ceph_decode_32(p);
1600 	ceph_decode_need(p, end, len, e_inval);
1601 	struct_end = *p + len;
1602 
1603 	oloc->pool = ceph_decode_64(p);
1604 	*p += 4; /* skip preferred */
1605 
1606 	len = ceph_decode_32(p);
1607 	if (len > 0) {
1608 		pr_warn("ceph_object_locator::key is set\n");
1609 		goto e_inval;
1610 	}
1611 
1612 	if (struct_v >= 5) {
1613 		len = ceph_decode_32(p);
1614 		if (len > 0) {
1615 			pr_warn("ceph_object_locator::nspace is set\n");
1616 			goto e_inval;
1617 		}
1618 	}
1619 
1620 	if (struct_v >= 6) {
1621 		s64 hash = ceph_decode_64(p);
1622 		if (hash != -1) {
1623 			pr_warn("ceph_object_locator::hash is set\n");
1624 			goto e_inval;
1625 		}
1626 	}
1627 
1628 	/* skip the rest */
1629 	*p = struct_end;
1630 out:
1631 	return ret;
1632 
1633 e_inval:
1634 	ret = -EINVAL;
1635 	goto out;
1636 }
1637 
1638 static int ceph_redirect_decode(void **p, void *end,
1639 				struct ceph_request_redirect *redir)
1640 {
1641 	u8 struct_v, struct_cv;
1642 	u32 len;
1643 	void *struct_end;
1644 	int ret;
1645 
1646 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1647 	struct_v = ceph_decode_8(p);
1648 	struct_cv = ceph_decode_8(p);
1649 	if (struct_cv > 1) {
1650 		pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1651 			struct_v, struct_cv);
1652 		goto e_inval;
1653 	}
1654 	len = ceph_decode_32(p);
1655 	ceph_decode_need(p, end, len, e_inval);
1656 	struct_end = *p + len;
1657 
1658 	ret = ceph_oloc_decode(p, end, &redir->oloc);
1659 	if (ret)
1660 		goto out;
1661 
1662 	len = ceph_decode_32(p);
1663 	if (len > 0) {
1664 		pr_warn("ceph_request_redirect::object_name is set\n");
1665 		goto e_inval;
1666 	}
1667 
1668 	len = ceph_decode_32(p);
1669 	*p += len; /* skip osd_instructions */
1670 
1671 	/* skip the rest */
1672 	*p = struct_end;
1673 out:
1674 	return ret;
1675 
1676 e_inval:
1677 	ret = -EINVAL;
1678 	goto out;
1679 }
1680 
1681 static void complete_request(struct ceph_osd_request *req)
1682 {
1683 	complete_all(&req->r_safe_completion);  /* fsync waiter */
1684 }
1685 
1686 /*
1687  * handle osd op reply.  either call the callback if it is specified,
1688  * or do the completion to wake up the waiting thread.
1689  */
1690 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1691 			 struct ceph_connection *con)
1692 {
1693 	void *p, *end;
1694 	struct ceph_osd_request *req;
1695 	struct ceph_request_redirect redir;
1696 	u64 tid;
1697 	int object_len;
1698 	unsigned int numops;
1699 	int payload_len, flags;
1700 	s32 result;
1701 	s32 retry_attempt;
1702 	struct ceph_pg pg;
1703 	int err;
1704 	u32 reassert_epoch;
1705 	u64 reassert_version;
1706 	u32 osdmap_epoch;
1707 	int already_completed;
1708 	u32 bytes;
1709 	unsigned int i;
1710 
1711 	tid = le64_to_cpu(msg->hdr.tid);
1712 	dout("handle_reply %p tid %llu\n", msg, tid);
1713 
1714 	p = msg->front.iov_base;
1715 	end = p + msg->front.iov_len;
1716 
1717 	ceph_decode_need(&p, end, 4, bad);
1718 	object_len = ceph_decode_32(&p);
1719 	ceph_decode_need(&p, end, object_len, bad);
1720 	p += object_len;
1721 
1722 	err = ceph_decode_pgid(&p, end, &pg);
1723 	if (err)
1724 		goto bad;
1725 
1726 	ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1727 	flags = ceph_decode_64(&p);
1728 	result = ceph_decode_32(&p);
1729 	reassert_epoch = ceph_decode_32(&p);
1730 	reassert_version = ceph_decode_64(&p);
1731 	osdmap_epoch = ceph_decode_32(&p);
1732 
1733 	/* lookup */
1734 	down_read(&osdc->map_sem);
1735 	mutex_lock(&osdc->request_mutex);
1736 	req = __lookup_request(osdc, tid);
1737 	if (req == NULL) {
1738 		dout("handle_reply tid %llu dne\n", tid);
1739 		goto bad_mutex;
1740 	}
1741 	ceph_osdc_get_request(req);
1742 
1743 	dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1744 	     req, result);
1745 
1746 	ceph_decode_need(&p, end, 4, bad_put);
1747 	numops = ceph_decode_32(&p);
1748 	if (numops > CEPH_OSD_MAX_OP)
1749 		goto bad_put;
1750 	if (numops != req->r_num_ops)
1751 		goto bad_put;
1752 	payload_len = 0;
1753 	ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1754 	for (i = 0; i < numops; i++) {
1755 		struct ceph_osd_op *op = p;
1756 		int len;
1757 
1758 		len = le32_to_cpu(op->payload_len);
1759 		req->r_reply_op_len[i] = len;
1760 		dout(" op %d has %d bytes\n", i, len);
1761 		payload_len += len;
1762 		p += sizeof(*op);
1763 	}
1764 	bytes = le32_to_cpu(msg->hdr.data_len);
1765 	if (payload_len != bytes) {
1766 		pr_warn("sum of op payload lens %d != data_len %d\n",
1767 			payload_len, bytes);
1768 		goto bad_put;
1769 	}
1770 
1771 	ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1772 	retry_attempt = ceph_decode_32(&p);
1773 	for (i = 0; i < numops; i++)
1774 		req->r_reply_op_result[i] = ceph_decode_32(&p);
1775 
1776 	if (le16_to_cpu(msg->hdr.version) >= 6) {
1777 		p += 8 + 4; /* skip replay_version */
1778 		p += 8; /* skip user_version */
1779 
1780 		err = ceph_redirect_decode(&p, end, &redir);
1781 		if (err)
1782 			goto bad_put;
1783 	} else {
1784 		redir.oloc.pool = -1;
1785 	}
1786 
1787 	if (redir.oloc.pool != -1) {
1788 		dout("redirect pool %lld\n", redir.oloc.pool);
1789 
1790 		__unregister_request(osdc, req);
1791 
1792 		req->r_target_oloc = redir.oloc; /* struct */
1793 
1794 		/*
1795 		 * Start redirect requests with nofail=true.  If
1796 		 * mapping fails, request will end up on the notarget
1797 		 * list, waiting for the new osdmap (which can take
1798 		 * a while), even though the original request mapped
1799 		 * successfully.  In the future we might want to follow
1800 		 * original request's nofail setting here.
1801 		 */
1802 		err = __ceph_osdc_start_request(osdc, req, true);
1803 		BUG_ON(err);
1804 
1805 		goto out_unlock;
1806 	}
1807 
1808 	already_completed = req->r_got_reply;
1809 	if (!req->r_got_reply) {
1810 		req->r_result = result;
1811 		dout("handle_reply result %d bytes %d\n", req->r_result,
1812 		     bytes);
1813 		if (req->r_result == 0)
1814 			req->r_result = bytes;
1815 
1816 		/* in case this is a write and we need to replay, */
1817 		req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1818 		req->r_reassert_version.version = cpu_to_le64(reassert_version);
1819 
1820 		req->r_got_reply = 1;
1821 	} else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1822 		dout("handle_reply tid %llu dup ack\n", tid);
1823 		goto out_unlock;
1824 	}
1825 
1826 	dout("handle_reply tid %llu flags %d\n", tid, flags);
1827 
1828 	if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1829 		__register_linger_request(osdc, req);
1830 
1831 	/* either this is a read, or we got the safe response */
1832 	if (result < 0 ||
1833 	    (flags & CEPH_OSD_FLAG_ONDISK) ||
1834 	    ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1835 		__unregister_request(osdc, req);
1836 
1837 	mutex_unlock(&osdc->request_mutex);
1838 	up_read(&osdc->map_sem);
1839 
1840 	if (!already_completed) {
1841 		if (req->r_unsafe_callback &&
1842 		    result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1843 			req->r_unsafe_callback(req, true);
1844 		if (req->r_callback)
1845 			req->r_callback(req, msg);
1846 		else
1847 			complete_all(&req->r_completion);
1848 	}
1849 
1850 	if (flags & CEPH_OSD_FLAG_ONDISK) {
1851 		if (req->r_unsafe_callback && already_completed)
1852 			req->r_unsafe_callback(req, false);
1853 		complete_request(req);
1854 	}
1855 
1856 out:
1857 	dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1858 	ceph_osdc_put_request(req);
1859 	return;
1860 out_unlock:
1861 	mutex_unlock(&osdc->request_mutex);
1862 	up_read(&osdc->map_sem);
1863 	goto out;
1864 
1865 bad_put:
1866 	req->r_result = -EIO;
1867 	__unregister_request(osdc, req);
1868 	if (req->r_callback)
1869 		req->r_callback(req, msg);
1870 	else
1871 		complete_all(&req->r_completion);
1872 	complete_request(req);
1873 	ceph_osdc_put_request(req);
1874 bad_mutex:
1875 	mutex_unlock(&osdc->request_mutex);
1876 	up_read(&osdc->map_sem);
1877 bad:
1878 	pr_err("corrupt osd_op_reply got %d %d\n",
1879 	       (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1880 	ceph_msg_dump(msg);
1881 }
1882 
1883 static void reset_changed_osds(struct ceph_osd_client *osdc)
1884 {
1885 	struct rb_node *p, *n;
1886 
1887 	for (p = rb_first(&osdc->osds); p; p = n) {
1888 		struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1889 
1890 		n = rb_next(p);
1891 		if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1892 		    memcmp(&osd->o_con.peer_addr,
1893 			   ceph_osd_addr(osdc->osdmap,
1894 					 osd->o_osd),
1895 			   sizeof(struct ceph_entity_addr)) != 0)
1896 			__reset_osd(osdc, osd);
1897 	}
1898 }
1899 
1900 /*
1901  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1902  * no osd, request a new map.
1903  *
1904  * Caller should hold map_sem for read.
1905  */
1906 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1907 			  bool force_resend_writes)
1908 {
1909 	struct ceph_osd_request *req, *nreq;
1910 	struct rb_node *p;
1911 	int needmap = 0;
1912 	int err;
1913 	bool force_resend_req;
1914 
1915 	dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1916 		force_resend_writes ? " (force resend writes)" : "");
1917 	mutex_lock(&osdc->request_mutex);
1918 	for (p = rb_first(&osdc->requests); p; ) {
1919 		req = rb_entry(p, struct ceph_osd_request, r_node);
1920 		p = rb_next(p);
1921 
1922 		/*
1923 		 * For linger requests that have not yet been
1924 		 * registered, move them to the linger list; they'll
1925 		 * be sent to the osd in the loop below.  Unregister
1926 		 * the request before re-registering it as a linger
1927 		 * request to ensure the __map_request() below
1928 		 * will decide it needs to be sent.
1929 		 */
1930 		if (req->r_linger && list_empty(&req->r_linger_item)) {
1931 			dout("%p tid %llu restart on osd%d\n",
1932 			     req, req->r_tid,
1933 			     req->r_osd ? req->r_osd->o_osd : -1);
1934 			ceph_osdc_get_request(req);
1935 			__unregister_request(osdc, req);
1936 			__register_linger_request(osdc, req);
1937 			ceph_osdc_put_request(req);
1938 			continue;
1939 		}
1940 
1941 		force_resend_req = force_resend ||
1942 			(force_resend_writes &&
1943 				req->r_flags & CEPH_OSD_FLAG_WRITE);
1944 		err = __map_request(osdc, req, force_resend_req);
1945 		if (err < 0)
1946 			continue;  /* error */
1947 		if (req->r_osd == NULL) {
1948 			dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1949 			needmap++;  /* request a newer map */
1950 		} else if (err > 0) {
1951 			if (!req->r_linger) {
1952 				dout("%p tid %llu requeued on osd%d\n", req,
1953 				     req->r_tid,
1954 				     req->r_osd ? req->r_osd->o_osd : -1);
1955 				req->r_flags |= CEPH_OSD_FLAG_RETRY;
1956 			}
1957 		}
1958 	}
1959 
1960 	list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1961 				 r_linger_item) {
1962 		dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1963 
1964 		err = __map_request(osdc, req,
1965 				    force_resend || force_resend_writes);
1966 		dout("__map_request returned %d\n", err);
1967 		if (err == 0)
1968 			continue;  /* no change and no osd was specified */
1969 		if (err < 0)
1970 			continue;  /* hrm! */
1971 		if (req->r_osd == NULL) {
1972 			dout("tid %llu maps to no valid osd\n", req->r_tid);
1973 			needmap++;  /* request a newer map */
1974 			continue;
1975 		}
1976 
1977 		dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1978 		     req->r_osd ? req->r_osd->o_osd : -1);
1979 		__register_request(osdc, req);
1980 		__unregister_linger_request(osdc, req);
1981 	}
1982 	reset_changed_osds(osdc);
1983 	mutex_unlock(&osdc->request_mutex);
1984 
1985 	if (needmap) {
1986 		dout("%d requests for down osds, need new map\n", needmap);
1987 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1988 	}
1989 }
1990 
1991 
1992 /*
1993  * Process updated osd map.
1994  *
1995  * The message contains any number of incremental and full maps, normally
1996  * indicating some sort of topology change in the cluster.  Kick requests
1997  * off to different OSDs as needed.
1998  */
1999 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2000 {
2001 	void *p, *end, *next;
2002 	u32 nr_maps, maplen;
2003 	u32 epoch;
2004 	struct ceph_osdmap *newmap = NULL, *oldmap;
2005 	int err;
2006 	struct ceph_fsid fsid;
2007 	bool was_full;
2008 
2009 	dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2010 	p = msg->front.iov_base;
2011 	end = p + msg->front.iov_len;
2012 
2013 	/* verify fsid */
2014 	ceph_decode_need(&p, end, sizeof(fsid), bad);
2015 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
2016 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
2017 		return;
2018 
2019 	down_write(&osdc->map_sem);
2020 
2021 	was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2022 
2023 	/* incremental maps */
2024 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2025 	dout(" %d inc maps\n", nr_maps);
2026 	while (nr_maps > 0) {
2027 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2028 		epoch = ceph_decode_32(&p);
2029 		maplen = ceph_decode_32(&p);
2030 		ceph_decode_need(&p, end, maplen, bad);
2031 		next = p + maplen;
2032 		if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2033 			dout("applying incremental map %u len %d\n",
2034 			     epoch, maplen);
2035 			newmap = osdmap_apply_incremental(&p, next,
2036 							  osdc->osdmap,
2037 							  &osdc->client->msgr);
2038 			if (IS_ERR(newmap)) {
2039 				err = PTR_ERR(newmap);
2040 				goto bad;
2041 			}
2042 			BUG_ON(!newmap);
2043 			if (newmap != osdc->osdmap) {
2044 				ceph_osdmap_destroy(osdc->osdmap);
2045 				osdc->osdmap = newmap;
2046 			}
2047 			was_full = was_full ||
2048 				ceph_osdmap_flag(osdc->osdmap,
2049 						 CEPH_OSDMAP_FULL);
2050 			kick_requests(osdc, 0, was_full);
2051 		} else {
2052 			dout("ignoring incremental map %u len %d\n",
2053 			     epoch, maplen);
2054 		}
2055 		p = next;
2056 		nr_maps--;
2057 	}
2058 	if (newmap)
2059 		goto done;
2060 
2061 	/* full maps */
2062 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2063 	dout(" %d full maps\n", nr_maps);
2064 	while (nr_maps) {
2065 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2066 		epoch = ceph_decode_32(&p);
2067 		maplen = ceph_decode_32(&p);
2068 		ceph_decode_need(&p, end, maplen, bad);
2069 		if (nr_maps > 1) {
2070 			dout("skipping non-latest full map %u len %d\n",
2071 			     epoch, maplen);
2072 		} else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2073 			dout("skipping full map %u len %d, "
2074 			     "older than our %u\n", epoch, maplen,
2075 			     osdc->osdmap->epoch);
2076 		} else {
2077 			int skipped_map = 0;
2078 
2079 			dout("taking full map %u len %d\n", epoch, maplen);
2080 			newmap = ceph_osdmap_decode(&p, p+maplen);
2081 			if (IS_ERR(newmap)) {
2082 				err = PTR_ERR(newmap);
2083 				goto bad;
2084 			}
2085 			BUG_ON(!newmap);
2086 			oldmap = osdc->osdmap;
2087 			osdc->osdmap = newmap;
2088 			if (oldmap) {
2089 				if (oldmap->epoch + 1 < newmap->epoch)
2090 					skipped_map = 1;
2091 				ceph_osdmap_destroy(oldmap);
2092 			}
2093 			was_full = was_full ||
2094 				ceph_osdmap_flag(osdc->osdmap,
2095 						 CEPH_OSDMAP_FULL);
2096 			kick_requests(osdc, skipped_map, was_full);
2097 		}
2098 		p += maplen;
2099 		nr_maps--;
2100 	}
2101 
2102 	if (!osdc->osdmap)
2103 		goto bad;
2104 done:
2105 	downgrade_write(&osdc->map_sem);
2106 	ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2107 
2108 	/*
2109 	 * subscribe to subsequent osdmap updates if full to ensure
2110 	 * we find out when we are no longer full and stop returning
2111 	 * ENOSPC.
2112 	 */
2113 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2114 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2115 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2116 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2117 
2118 	mutex_lock(&osdc->request_mutex);
2119 	__send_queued(osdc);
2120 	mutex_unlock(&osdc->request_mutex);
2121 	up_read(&osdc->map_sem);
2122 	wake_up_all(&osdc->client->auth_wq);
2123 	return;
2124 
2125 bad:
2126 	pr_err("osdc handle_map corrupt msg\n");
2127 	ceph_msg_dump(msg);
2128 	up_write(&osdc->map_sem);
2129 }
2130 
2131 /*
2132  * watch/notify callback event infrastructure
2133  *
2134  * These callbacks are used both for watch and notify operations.
2135  */
2136 static void __release_event(struct kref *kref)
2137 {
2138 	struct ceph_osd_event *event =
2139 		container_of(kref, struct ceph_osd_event, kref);
2140 
2141 	dout("__release_event %p\n", event);
2142 	kfree(event);
2143 }
2144 
2145 static void get_event(struct ceph_osd_event *event)
2146 {
2147 	kref_get(&event->kref);
2148 }
2149 
2150 void ceph_osdc_put_event(struct ceph_osd_event *event)
2151 {
2152 	kref_put(&event->kref, __release_event);
2153 }
2154 EXPORT_SYMBOL(ceph_osdc_put_event);
2155 
2156 static void __insert_event(struct ceph_osd_client *osdc,
2157 			     struct ceph_osd_event *new)
2158 {
2159 	struct rb_node **p = &osdc->event_tree.rb_node;
2160 	struct rb_node *parent = NULL;
2161 	struct ceph_osd_event *event = NULL;
2162 
2163 	while (*p) {
2164 		parent = *p;
2165 		event = rb_entry(parent, struct ceph_osd_event, node);
2166 		if (new->cookie < event->cookie)
2167 			p = &(*p)->rb_left;
2168 		else if (new->cookie > event->cookie)
2169 			p = &(*p)->rb_right;
2170 		else
2171 			BUG();
2172 	}
2173 
2174 	rb_link_node(&new->node, parent, p);
2175 	rb_insert_color(&new->node, &osdc->event_tree);
2176 }
2177 
2178 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2179 					        u64 cookie)
2180 {
2181 	struct rb_node **p = &osdc->event_tree.rb_node;
2182 	struct rb_node *parent = NULL;
2183 	struct ceph_osd_event *event = NULL;
2184 
2185 	while (*p) {
2186 		parent = *p;
2187 		event = rb_entry(parent, struct ceph_osd_event, node);
2188 		if (cookie < event->cookie)
2189 			p = &(*p)->rb_left;
2190 		else if (cookie > event->cookie)
2191 			p = &(*p)->rb_right;
2192 		else
2193 			return event;
2194 	}
2195 	return NULL;
2196 }
2197 
2198 static void __remove_event(struct ceph_osd_event *event)
2199 {
2200 	struct ceph_osd_client *osdc = event->osdc;
2201 
2202 	if (!RB_EMPTY_NODE(&event->node)) {
2203 		dout("__remove_event removed %p\n", event);
2204 		rb_erase(&event->node, &osdc->event_tree);
2205 		ceph_osdc_put_event(event);
2206 	} else {
2207 		dout("__remove_event didn't remove %p\n", event);
2208 	}
2209 }
2210 
2211 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2212 			   void (*event_cb)(u64, u64, u8, void *),
2213 			   void *data, struct ceph_osd_event **pevent)
2214 {
2215 	struct ceph_osd_event *event;
2216 
2217 	event = kmalloc(sizeof(*event), GFP_NOIO);
2218 	if (!event)
2219 		return -ENOMEM;
2220 
2221 	dout("create_event %p\n", event);
2222 	event->cb = event_cb;
2223 	event->one_shot = 0;
2224 	event->data = data;
2225 	event->osdc = osdc;
2226 	INIT_LIST_HEAD(&event->osd_node);
2227 	RB_CLEAR_NODE(&event->node);
2228 	kref_init(&event->kref);   /* one ref for us */
2229 	kref_get(&event->kref);    /* one ref for the caller */
2230 
2231 	spin_lock(&osdc->event_lock);
2232 	event->cookie = ++osdc->event_count;
2233 	__insert_event(osdc, event);
2234 	spin_unlock(&osdc->event_lock);
2235 
2236 	*pevent = event;
2237 	return 0;
2238 }
2239 EXPORT_SYMBOL(ceph_osdc_create_event);
2240 
2241 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2242 {
2243 	struct ceph_osd_client *osdc = event->osdc;
2244 
2245 	dout("cancel_event %p\n", event);
2246 	spin_lock(&osdc->event_lock);
2247 	__remove_event(event);
2248 	spin_unlock(&osdc->event_lock);
2249 	ceph_osdc_put_event(event); /* caller's */
2250 }
2251 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2252 
2253 
2254 static void do_event_work(struct work_struct *work)
2255 {
2256 	struct ceph_osd_event_work *event_work =
2257 		container_of(work, struct ceph_osd_event_work, work);
2258 	struct ceph_osd_event *event = event_work->event;
2259 	u64 ver = event_work->ver;
2260 	u64 notify_id = event_work->notify_id;
2261 	u8 opcode = event_work->opcode;
2262 
2263 	dout("do_event_work completing %p\n", event);
2264 	event->cb(ver, notify_id, opcode, event->data);
2265 	dout("do_event_work completed %p\n", event);
2266 	ceph_osdc_put_event(event);
2267 	kfree(event_work);
2268 }
2269 
2270 
2271 /*
2272  * Process osd watch notifications
2273  */
2274 static void handle_watch_notify(struct ceph_osd_client *osdc,
2275 				struct ceph_msg *msg)
2276 {
2277 	void *p, *end;
2278 	u8 proto_ver;
2279 	u64 cookie, ver, notify_id;
2280 	u8 opcode;
2281 	struct ceph_osd_event *event;
2282 	struct ceph_osd_event_work *event_work;
2283 
2284 	p = msg->front.iov_base;
2285 	end = p + msg->front.iov_len;
2286 
2287 	ceph_decode_8_safe(&p, end, proto_ver, bad);
2288 	ceph_decode_8_safe(&p, end, opcode, bad);
2289 	ceph_decode_64_safe(&p, end, cookie, bad);
2290 	ceph_decode_64_safe(&p, end, ver, bad);
2291 	ceph_decode_64_safe(&p, end, notify_id, bad);
2292 
2293 	spin_lock(&osdc->event_lock);
2294 	event = __find_event(osdc, cookie);
2295 	if (event) {
2296 		BUG_ON(event->one_shot);
2297 		get_event(event);
2298 	}
2299 	spin_unlock(&osdc->event_lock);
2300 	dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2301 	     cookie, ver, event);
2302 	if (event) {
2303 		event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2304 		if (!event_work) {
2305 			pr_err("couldn't allocate event_work\n");
2306 			ceph_osdc_put_event(event);
2307 			return;
2308 		}
2309 		INIT_WORK(&event_work->work, do_event_work);
2310 		event_work->event = event;
2311 		event_work->ver = ver;
2312 		event_work->notify_id = notify_id;
2313 		event_work->opcode = opcode;
2314 
2315 		queue_work(osdc->notify_wq, &event_work->work);
2316 	}
2317 
2318 	return;
2319 
2320 bad:
2321 	pr_err("osdc handle_watch_notify corrupt msg\n");
2322 }
2323 
2324 /*
2325  * build new request AND message
2326  *
2327  */
2328 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2329 				struct ceph_snap_context *snapc, u64 snap_id,
2330 				struct timespec *mtime)
2331 {
2332 	struct ceph_msg *msg = req->r_request;
2333 	void *p;
2334 	size_t msg_size;
2335 	int flags = req->r_flags;
2336 	u64 data_len;
2337 	unsigned int i;
2338 
2339 	req->r_snapid = snap_id;
2340 	req->r_snapc = ceph_get_snap_context(snapc);
2341 
2342 	/* encode request */
2343 	msg->hdr.version = cpu_to_le16(4);
2344 
2345 	p = msg->front.iov_base;
2346 	ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2347 	req->r_request_osdmap_epoch = p;
2348 	p += 4;
2349 	req->r_request_flags = p;
2350 	p += 4;
2351 	if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2352 		ceph_encode_timespec(p, mtime);
2353 	p += sizeof(struct ceph_timespec);
2354 	req->r_request_reassert_version = p;
2355 	p += sizeof(struct ceph_eversion); /* will get filled in */
2356 
2357 	/* oloc */
2358 	ceph_encode_8(&p, 4);
2359 	ceph_encode_8(&p, 4);
2360 	ceph_encode_32(&p, 8 + 4 + 4);
2361 	req->r_request_pool = p;
2362 	p += 8;
2363 	ceph_encode_32(&p, -1);  /* preferred */
2364 	ceph_encode_32(&p, 0);   /* key len */
2365 
2366 	ceph_encode_8(&p, 1);
2367 	req->r_request_pgid = p;
2368 	p += 8 + 4;
2369 	ceph_encode_32(&p, -1);  /* preferred */
2370 
2371 	/* oid */
2372 	ceph_encode_32(&p, req->r_base_oid.name_len);
2373 	memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2374 	dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2375 	     req->r_base_oid.name, req->r_base_oid.name_len);
2376 	p += req->r_base_oid.name_len;
2377 
2378 	/* ops--can imply data */
2379 	ceph_encode_16(&p, (u16)req->r_num_ops);
2380 	data_len = 0;
2381 	for (i = 0; i < req->r_num_ops; i++) {
2382 		data_len += osd_req_encode_op(req, p, i);
2383 		p += sizeof(struct ceph_osd_op);
2384 	}
2385 
2386 	/* snaps */
2387 	ceph_encode_64(&p, req->r_snapid);
2388 	ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2389 	ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2390 	if (req->r_snapc) {
2391 		for (i = 0; i < snapc->num_snaps; i++) {
2392 			ceph_encode_64(&p, req->r_snapc->snaps[i]);
2393 		}
2394 	}
2395 
2396 	req->r_request_attempts = p;
2397 	p += 4;
2398 
2399 	/* data */
2400 	if (flags & CEPH_OSD_FLAG_WRITE) {
2401 		u16 data_off;
2402 
2403 		/*
2404 		 * The header "data_off" is a hint to the receiver
2405 		 * allowing it to align received data into its
2406 		 * buffers such that there's no need to re-copy
2407 		 * it before writing it to disk (direct I/O).
2408 		 */
2409 		data_off = (u16) (off & 0xffff);
2410 		req->r_request->hdr.data_off = cpu_to_le16(data_off);
2411 	}
2412 	req->r_request->hdr.data_len = cpu_to_le32(data_len);
2413 
2414 	BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2415 	msg_size = p - msg->front.iov_base;
2416 	msg->front.iov_len = msg_size;
2417 	msg->hdr.front_len = cpu_to_le32(msg_size);
2418 
2419 	dout("build_request msg_size was %d\n", (int)msg_size);
2420 }
2421 EXPORT_SYMBOL(ceph_osdc_build_request);
2422 
2423 /*
2424  * Register request, send initial attempt.
2425  */
2426 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2427 			    struct ceph_osd_request *req,
2428 			    bool nofail)
2429 {
2430 	int rc;
2431 
2432 	down_read(&osdc->map_sem);
2433 	mutex_lock(&osdc->request_mutex);
2434 
2435 	rc = __ceph_osdc_start_request(osdc, req, nofail);
2436 
2437 	mutex_unlock(&osdc->request_mutex);
2438 	up_read(&osdc->map_sem);
2439 
2440 	return rc;
2441 }
2442 EXPORT_SYMBOL(ceph_osdc_start_request);
2443 
2444 /*
2445  * Unregister a registered request.  The request is not completed (i.e.
2446  * no callbacks or wakeups) - higher layers are supposed to know what
2447  * they are canceling.
2448  */
2449 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2450 {
2451 	struct ceph_osd_client *osdc = req->r_osdc;
2452 
2453 	mutex_lock(&osdc->request_mutex);
2454 	if (req->r_linger)
2455 		__unregister_linger_request(osdc, req);
2456 	__unregister_request(osdc, req);
2457 	mutex_unlock(&osdc->request_mutex);
2458 
2459 	dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2460 }
2461 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2462 
2463 /*
2464  * wait for a request to complete
2465  */
2466 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2467 			   struct ceph_osd_request *req)
2468 {
2469 	int rc;
2470 
2471 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2472 
2473 	rc = wait_for_completion_interruptible(&req->r_completion);
2474 	if (rc < 0) {
2475 		dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2476 		ceph_osdc_cancel_request(req);
2477 		complete_request(req);
2478 		return rc;
2479 	}
2480 
2481 	dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2482 	     req->r_result);
2483 	return req->r_result;
2484 }
2485 EXPORT_SYMBOL(ceph_osdc_wait_request);
2486 
2487 /*
2488  * sync - wait for all in-flight requests to flush.  avoid starvation.
2489  */
2490 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2491 {
2492 	struct ceph_osd_request *req;
2493 	u64 last_tid, next_tid = 0;
2494 
2495 	mutex_lock(&osdc->request_mutex);
2496 	last_tid = osdc->last_tid;
2497 	while (1) {
2498 		req = __lookup_request_ge(osdc, next_tid);
2499 		if (!req)
2500 			break;
2501 		if (req->r_tid > last_tid)
2502 			break;
2503 
2504 		next_tid = req->r_tid + 1;
2505 		if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2506 			continue;
2507 
2508 		ceph_osdc_get_request(req);
2509 		mutex_unlock(&osdc->request_mutex);
2510 		dout("sync waiting on tid %llu (last is %llu)\n",
2511 		     req->r_tid, last_tid);
2512 		wait_for_completion(&req->r_safe_completion);
2513 		mutex_lock(&osdc->request_mutex);
2514 		ceph_osdc_put_request(req);
2515 	}
2516 	mutex_unlock(&osdc->request_mutex);
2517 	dout("sync done (thru tid %llu)\n", last_tid);
2518 }
2519 EXPORT_SYMBOL(ceph_osdc_sync);
2520 
2521 /*
2522  * Call all pending notify callbacks - for use after a watch is
2523  * unregistered, to make sure no more callbacks for it will be invoked
2524  */
2525 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2526 {
2527 	flush_workqueue(osdc->notify_wq);
2528 }
2529 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2530 
2531 
2532 /*
2533  * init, shutdown
2534  */
2535 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2536 {
2537 	int err;
2538 
2539 	dout("init\n");
2540 	osdc->client = client;
2541 	osdc->osdmap = NULL;
2542 	init_rwsem(&osdc->map_sem);
2543 	init_completion(&osdc->map_waiters);
2544 	osdc->last_requested_map = 0;
2545 	mutex_init(&osdc->request_mutex);
2546 	osdc->last_tid = 0;
2547 	osdc->osds = RB_ROOT;
2548 	INIT_LIST_HEAD(&osdc->osd_lru);
2549 	osdc->requests = RB_ROOT;
2550 	INIT_LIST_HEAD(&osdc->req_lru);
2551 	INIT_LIST_HEAD(&osdc->req_unsent);
2552 	INIT_LIST_HEAD(&osdc->req_notarget);
2553 	INIT_LIST_HEAD(&osdc->req_linger);
2554 	osdc->num_requests = 0;
2555 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2556 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2557 	spin_lock_init(&osdc->event_lock);
2558 	osdc->event_tree = RB_ROOT;
2559 	osdc->event_count = 0;
2560 
2561 	schedule_delayed_work(&osdc->osds_timeout_work,
2562 	   round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2563 
2564 	err = -ENOMEM;
2565 	osdc->req_mempool = mempool_create_kmalloc_pool(10,
2566 					sizeof(struct ceph_osd_request));
2567 	if (!osdc->req_mempool)
2568 		goto out;
2569 
2570 	err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2571 				OSD_OP_FRONT_LEN, 10, true,
2572 				"osd_op");
2573 	if (err < 0)
2574 		goto out_mempool;
2575 	err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2576 				OSD_OPREPLY_FRONT_LEN, 10, true,
2577 				"osd_op_reply");
2578 	if (err < 0)
2579 		goto out_msgpool;
2580 
2581 	err = -ENOMEM;
2582 	osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2583 	if (!osdc->notify_wq)
2584 		goto out_msgpool_reply;
2585 
2586 	return 0;
2587 
2588 out_msgpool_reply:
2589 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2590 out_msgpool:
2591 	ceph_msgpool_destroy(&osdc->msgpool_op);
2592 out_mempool:
2593 	mempool_destroy(osdc->req_mempool);
2594 out:
2595 	return err;
2596 }
2597 
2598 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2599 {
2600 	flush_workqueue(osdc->notify_wq);
2601 	destroy_workqueue(osdc->notify_wq);
2602 	cancel_delayed_work_sync(&osdc->timeout_work);
2603 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
2604 	if (osdc->osdmap) {
2605 		ceph_osdmap_destroy(osdc->osdmap);
2606 		osdc->osdmap = NULL;
2607 	}
2608 	remove_all_osds(osdc);
2609 	mempool_destroy(osdc->req_mempool);
2610 	ceph_msgpool_destroy(&osdc->msgpool_op);
2611 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2612 }
2613 
2614 /*
2615  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2616  * *plen.  Return number of bytes read, or error.
2617  */
2618 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2619 			struct ceph_vino vino, struct ceph_file_layout *layout,
2620 			u64 off, u64 *plen,
2621 			u32 truncate_seq, u64 truncate_size,
2622 			struct page **pages, int num_pages, int page_align)
2623 {
2624 	struct ceph_osd_request *req;
2625 	int rc = 0;
2626 
2627 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2628 	     vino.snap, off, *plen);
2629 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2630 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2631 				    NULL, truncate_seq, truncate_size,
2632 				    false);
2633 	if (IS_ERR(req))
2634 		return PTR_ERR(req);
2635 
2636 	/* it may be a short read due to an object boundary */
2637 
2638 	osd_req_op_extent_osd_data_pages(req, 0,
2639 				pages, *plen, page_align, false, false);
2640 
2641 	dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2642 	     off, *plen, *plen, page_align);
2643 
2644 	ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2645 
2646 	rc = ceph_osdc_start_request(osdc, req, false);
2647 	if (!rc)
2648 		rc = ceph_osdc_wait_request(osdc, req);
2649 
2650 	ceph_osdc_put_request(req);
2651 	dout("readpages result %d\n", rc);
2652 	return rc;
2653 }
2654 EXPORT_SYMBOL(ceph_osdc_readpages);
2655 
2656 /*
2657  * do a synchronous write on N pages
2658  */
2659 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2660 			 struct ceph_file_layout *layout,
2661 			 struct ceph_snap_context *snapc,
2662 			 u64 off, u64 len,
2663 			 u32 truncate_seq, u64 truncate_size,
2664 			 struct timespec *mtime,
2665 			 struct page **pages, int num_pages)
2666 {
2667 	struct ceph_osd_request *req;
2668 	int rc = 0;
2669 	int page_align = off & ~PAGE_MASK;
2670 
2671 	BUG_ON(vino.snap != CEPH_NOSNAP);	/* snapshots aren't writeable */
2672 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2673 				    CEPH_OSD_OP_WRITE,
2674 				    CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2675 				    snapc, truncate_seq, truncate_size,
2676 				    true);
2677 	if (IS_ERR(req))
2678 		return PTR_ERR(req);
2679 
2680 	/* it may be a short write due to an object boundary */
2681 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2682 				false, false);
2683 	dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2684 
2685 	ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2686 
2687 	rc = ceph_osdc_start_request(osdc, req, true);
2688 	if (!rc)
2689 		rc = ceph_osdc_wait_request(osdc, req);
2690 
2691 	ceph_osdc_put_request(req);
2692 	if (rc == 0)
2693 		rc = len;
2694 	dout("writepages result %d\n", rc);
2695 	return rc;
2696 }
2697 EXPORT_SYMBOL(ceph_osdc_writepages);
2698 
2699 int ceph_osdc_setup(void)
2700 {
2701 	BUG_ON(ceph_osd_request_cache);
2702 	ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2703 					sizeof (struct ceph_osd_request),
2704 					__alignof__(struct ceph_osd_request),
2705 					0, NULL);
2706 
2707 	return ceph_osd_request_cache ? 0 : -ENOMEM;
2708 }
2709 EXPORT_SYMBOL(ceph_osdc_setup);
2710 
2711 void ceph_osdc_cleanup(void)
2712 {
2713 	BUG_ON(!ceph_osd_request_cache);
2714 	kmem_cache_destroy(ceph_osd_request_cache);
2715 	ceph_osd_request_cache = NULL;
2716 }
2717 EXPORT_SYMBOL(ceph_osdc_cleanup);
2718 
2719 /*
2720  * handle incoming message
2721  */
2722 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2723 {
2724 	struct ceph_osd *osd = con->private;
2725 	struct ceph_osd_client *osdc;
2726 	int type = le16_to_cpu(msg->hdr.type);
2727 
2728 	if (!osd)
2729 		goto out;
2730 	osdc = osd->o_osdc;
2731 
2732 	switch (type) {
2733 	case CEPH_MSG_OSD_MAP:
2734 		ceph_osdc_handle_map(osdc, msg);
2735 		break;
2736 	case CEPH_MSG_OSD_OPREPLY:
2737 		handle_reply(osdc, msg, con);
2738 		break;
2739 	case CEPH_MSG_WATCH_NOTIFY:
2740 		handle_watch_notify(osdc, msg);
2741 		break;
2742 
2743 	default:
2744 		pr_err("received unknown message type %d %s\n", type,
2745 		       ceph_msg_type_name(type));
2746 	}
2747 out:
2748 	ceph_msg_put(msg);
2749 }
2750 
2751 /*
2752  * lookup and return message for incoming reply.  set up reply message
2753  * pages.
2754  */
2755 static struct ceph_msg *get_reply(struct ceph_connection *con,
2756 				  struct ceph_msg_header *hdr,
2757 				  int *skip)
2758 {
2759 	struct ceph_osd *osd = con->private;
2760 	struct ceph_osd_client *osdc = osd->o_osdc;
2761 	struct ceph_msg *m;
2762 	struct ceph_osd_request *req;
2763 	int front_len = le32_to_cpu(hdr->front_len);
2764 	int data_len = le32_to_cpu(hdr->data_len);
2765 	u64 tid;
2766 
2767 	tid = le64_to_cpu(hdr->tid);
2768 	mutex_lock(&osdc->request_mutex);
2769 	req = __lookup_request(osdc, tid);
2770 	if (!req) {
2771 		*skip = 1;
2772 		m = NULL;
2773 		dout("get_reply unknown tid %llu from osd%d\n", tid,
2774 		     osd->o_osd);
2775 		goto out;
2776 	}
2777 
2778 	if (req->r_reply->con)
2779 		dout("%s revoking msg %p from old con %p\n", __func__,
2780 		     req->r_reply, req->r_reply->con);
2781 	ceph_msg_revoke_incoming(req->r_reply);
2782 
2783 	if (front_len > req->r_reply->front_alloc_len) {
2784 		pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2785 			front_len, req->r_reply->front_alloc_len,
2786 			(unsigned int)con->peer_name.type,
2787 			le64_to_cpu(con->peer_name.num));
2788 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2789 				 false);
2790 		if (!m)
2791 			goto out;
2792 		ceph_msg_put(req->r_reply);
2793 		req->r_reply = m;
2794 	}
2795 	m = ceph_msg_get(req->r_reply);
2796 
2797 	if (data_len > 0) {
2798 		struct ceph_osd_data *osd_data;
2799 
2800 		/*
2801 		 * XXX This is assuming there is only one op containing
2802 		 * XXX page data.  Probably OK for reads, but this
2803 		 * XXX ought to be done more generally.
2804 		 */
2805 		osd_data = osd_req_op_extent_osd_data(req, 0);
2806 		if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2807 			if (osd_data->pages &&
2808 				unlikely(osd_data->length < data_len)) {
2809 
2810 				pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2811 					tid, data_len, osd_data->length);
2812 				*skip = 1;
2813 				ceph_msg_put(m);
2814 				m = NULL;
2815 				goto out;
2816 			}
2817 		}
2818 	}
2819 	*skip = 0;
2820 	dout("get_reply tid %lld %p\n", tid, m);
2821 
2822 out:
2823 	mutex_unlock(&osdc->request_mutex);
2824 	return m;
2825 
2826 }
2827 
2828 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2829 				  struct ceph_msg_header *hdr,
2830 				  int *skip)
2831 {
2832 	struct ceph_osd *osd = con->private;
2833 	int type = le16_to_cpu(hdr->type);
2834 	int front = le32_to_cpu(hdr->front_len);
2835 
2836 	*skip = 0;
2837 	switch (type) {
2838 	case CEPH_MSG_OSD_MAP:
2839 	case CEPH_MSG_WATCH_NOTIFY:
2840 		return ceph_msg_new(type, front, GFP_NOFS, false);
2841 	case CEPH_MSG_OSD_OPREPLY:
2842 		return get_reply(con, hdr, skip);
2843 	default:
2844 		pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2845 			osd->o_osd);
2846 		*skip = 1;
2847 		return NULL;
2848 	}
2849 }
2850 
2851 /*
2852  * Wrappers to refcount containing ceph_osd struct
2853  */
2854 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2855 {
2856 	struct ceph_osd *osd = con->private;
2857 	if (get_osd(osd))
2858 		return con;
2859 	return NULL;
2860 }
2861 
2862 static void put_osd_con(struct ceph_connection *con)
2863 {
2864 	struct ceph_osd *osd = con->private;
2865 	put_osd(osd);
2866 }
2867 
2868 /*
2869  * authentication
2870  */
2871 /*
2872  * Note: returned pointer is the address of a structure that's
2873  * managed separately.  Caller must *not* attempt to free it.
2874  */
2875 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2876 					int *proto, int force_new)
2877 {
2878 	struct ceph_osd *o = con->private;
2879 	struct ceph_osd_client *osdc = o->o_osdc;
2880 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2881 	struct ceph_auth_handshake *auth = &o->o_auth;
2882 
2883 	if (force_new && auth->authorizer) {
2884 		ceph_auth_destroy_authorizer(ac, auth->authorizer);
2885 		auth->authorizer = NULL;
2886 	}
2887 	if (!auth->authorizer) {
2888 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2889 						      auth);
2890 		if (ret)
2891 			return ERR_PTR(ret);
2892 	} else {
2893 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2894 						     auth);
2895 		if (ret)
2896 			return ERR_PTR(ret);
2897 	}
2898 	*proto = ac->protocol;
2899 
2900 	return auth;
2901 }
2902 
2903 
2904 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2905 {
2906 	struct ceph_osd *o = con->private;
2907 	struct ceph_osd_client *osdc = o->o_osdc;
2908 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2909 
2910 	return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2911 }
2912 
2913 static int invalidate_authorizer(struct ceph_connection *con)
2914 {
2915 	struct ceph_osd *o = con->private;
2916 	struct ceph_osd_client *osdc = o->o_osdc;
2917 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2918 
2919 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2920 	return ceph_monc_validate_auth(&osdc->client->monc);
2921 }
2922 
2923 static const struct ceph_connection_operations osd_con_ops = {
2924 	.get = get_osd_con,
2925 	.put = put_osd_con,
2926 	.dispatch = dispatch,
2927 	.get_authorizer = get_authorizer,
2928 	.verify_authorizer_reply = verify_authorizer_reply,
2929 	.invalidate_authorizer = invalidate_authorizer,
2930 	.alloc_msg = alloc_msg,
2931 	.fault = osd_reset,
2932 };
2933