xref: /openbmc/linux/net/ceph/osd_client.c (revision 65ee8aeb)
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 	case CEPH_OSD_OP_SETXATTR:
296 	case CEPH_OSD_OP_CMPXATTR:
297 		ceph_osd_data_release(&op->xattr.osd_data);
298 		break;
299 	default:
300 		break;
301 	}
302 }
303 
304 /*
305  * requests
306  */
307 static void ceph_osdc_release_request(struct kref *kref)
308 {
309 	struct ceph_osd_request *req = container_of(kref,
310 					    struct ceph_osd_request, r_kref);
311 	unsigned int which;
312 
313 	dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
314 	     req->r_request, req->r_reply);
315 	WARN_ON(!RB_EMPTY_NODE(&req->r_node));
316 	WARN_ON(!list_empty(&req->r_req_lru_item));
317 	WARN_ON(!list_empty(&req->r_osd_item));
318 	WARN_ON(!list_empty(&req->r_linger_item));
319 	WARN_ON(!list_empty(&req->r_linger_osd_item));
320 	WARN_ON(req->r_osd);
321 
322 	if (req->r_request)
323 		ceph_msg_put(req->r_request);
324 	if (req->r_reply) {
325 		ceph_msg_revoke_incoming(req->r_reply);
326 		ceph_msg_put(req->r_reply);
327 	}
328 
329 	for (which = 0; which < req->r_num_ops; which++)
330 		osd_req_op_data_release(req, which);
331 
332 	ceph_put_snap_context(req->r_snapc);
333 	if (req->r_mempool)
334 		mempool_free(req, req->r_osdc->req_mempool);
335 	else
336 		kmem_cache_free(ceph_osd_request_cache, req);
337 
338 }
339 
340 void ceph_osdc_get_request(struct ceph_osd_request *req)
341 {
342 	dout("%s %p (was %d)\n", __func__, req,
343 	     atomic_read(&req->r_kref.refcount));
344 	kref_get(&req->r_kref);
345 }
346 EXPORT_SYMBOL(ceph_osdc_get_request);
347 
348 void ceph_osdc_put_request(struct ceph_osd_request *req)
349 {
350 	dout("%s %p (was %d)\n", __func__, req,
351 	     atomic_read(&req->r_kref.refcount));
352 	kref_put(&req->r_kref, ceph_osdc_release_request);
353 }
354 EXPORT_SYMBOL(ceph_osdc_put_request);
355 
356 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
357 					       struct ceph_snap_context *snapc,
358 					       unsigned int num_ops,
359 					       bool use_mempool,
360 					       gfp_t gfp_flags)
361 {
362 	struct ceph_osd_request *req;
363 	struct ceph_msg *msg;
364 	size_t msg_size;
365 
366 	BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
367 	BUG_ON(num_ops > CEPH_OSD_MAX_OP);
368 
369 	msg_size = 4 + 4 + 8 + 8 + 4+8;
370 	msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
371 	msg_size += 1 + 8 + 4 + 4;     /* pg_t */
372 	msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
373 	msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
374 	msg_size += 8;  /* snapid */
375 	msg_size += 8;  /* snap_seq */
376 	msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
377 	msg_size += 4;
378 
379 	if (use_mempool) {
380 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
381 		memset(req, 0, sizeof(*req));
382 	} else {
383 		req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
384 	}
385 	if (req == NULL)
386 		return NULL;
387 
388 	req->r_osdc = osdc;
389 	req->r_mempool = use_mempool;
390 	req->r_num_ops = num_ops;
391 
392 	kref_init(&req->r_kref);
393 	init_completion(&req->r_completion);
394 	init_completion(&req->r_safe_completion);
395 	RB_CLEAR_NODE(&req->r_node);
396 	INIT_LIST_HEAD(&req->r_unsafe_item);
397 	INIT_LIST_HEAD(&req->r_linger_item);
398 	INIT_LIST_HEAD(&req->r_linger_osd_item);
399 	INIT_LIST_HEAD(&req->r_req_lru_item);
400 	INIT_LIST_HEAD(&req->r_osd_item);
401 
402 	req->r_base_oloc.pool = -1;
403 	req->r_target_oloc.pool = -1;
404 
405 	/* create reply message */
406 	if (use_mempool)
407 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
408 	else
409 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
410 				   OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
411 	if (!msg) {
412 		ceph_osdc_put_request(req);
413 		return NULL;
414 	}
415 	req->r_reply = msg;
416 
417 	/* create request message; allow space for oid */
418 	if (use_mempool)
419 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
420 	else
421 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
422 	if (!msg) {
423 		ceph_osdc_put_request(req);
424 		return NULL;
425 	}
426 
427 	memset(msg->front.iov_base, 0, msg->front.iov_len);
428 
429 	req->r_request = msg;
430 
431 	return req;
432 }
433 EXPORT_SYMBOL(ceph_osdc_alloc_request);
434 
435 static bool osd_req_opcode_valid(u16 opcode)
436 {
437 	switch (opcode) {
438 #define GENERATE_CASE(op, opcode, str)	case CEPH_OSD_OP_##op: return true;
439 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
440 #undef GENERATE_CASE
441 	default:
442 		return false;
443 	}
444 }
445 
446 /*
447  * This is an osd op init function for opcodes that have no data or
448  * other information associated with them.  It also serves as a
449  * common init routine for all the other init functions, below.
450  */
451 static struct ceph_osd_req_op *
452 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
453 				u16 opcode)
454 {
455 	struct ceph_osd_req_op *op;
456 
457 	BUG_ON(which >= osd_req->r_num_ops);
458 	BUG_ON(!osd_req_opcode_valid(opcode));
459 
460 	op = &osd_req->r_ops[which];
461 	memset(op, 0, sizeof (*op));
462 	op->op = opcode;
463 
464 	return op;
465 }
466 
467 void osd_req_op_init(struct ceph_osd_request *osd_req,
468 				unsigned int which, u16 opcode)
469 {
470 	(void)_osd_req_op_init(osd_req, which, opcode);
471 }
472 EXPORT_SYMBOL(osd_req_op_init);
473 
474 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
475 				unsigned int which, u16 opcode,
476 				u64 offset, u64 length,
477 				u64 truncate_size, u32 truncate_seq)
478 {
479 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
480 	size_t payload_len = 0;
481 
482 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
483 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE);
484 
485 	op->extent.offset = offset;
486 	op->extent.length = length;
487 	op->extent.truncate_size = truncate_size;
488 	op->extent.truncate_seq = truncate_seq;
489 	if (opcode == CEPH_OSD_OP_WRITE)
490 		payload_len += length;
491 
492 	op->payload_len = payload_len;
493 }
494 EXPORT_SYMBOL(osd_req_op_extent_init);
495 
496 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
497 				unsigned int which, u64 length)
498 {
499 	struct ceph_osd_req_op *op;
500 	u64 previous;
501 
502 	BUG_ON(which >= osd_req->r_num_ops);
503 	op = &osd_req->r_ops[which];
504 	previous = op->extent.length;
505 
506 	if (length == previous)
507 		return;		/* Nothing to do */
508 	BUG_ON(length > previous);
509 
510 	op->extent.length = length;
511 	op->payload_len -= previous - length;
512 }
513 EXPORT_SYMBOL(osd_req_op_extent_update);
514 
515 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
516 			u16 opcode, const char *class, const char *method)
517 {
518 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
519 	struct ceph_pagelist *pagelist;
520 	size_t payload_len = 0;
521 	size_t size;
522 
523 	BUG_ON(opcode != CEPH_OSD_OP_CALL);
524 
525 	pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
526 	BUG_ON(!pagelist);
527 	ceph_pagelist_init(pagelist);
528 
529 	op->cls.class_name = class;
530 	size = strlen(class);
531 	BUG_ON(size > (size_t) U8_MAX);
532 	op->cls.class_len = size;
533 	ceph_pagelist_append(pagelist, class, size);
534 	payload_len += size;
535 
536 	op->cls.method_name = method;
537 	size = strlen(method);
538 	BUG_ON(size > (size_t) U8_MAX);
539 	op->cls.method_len = size;
540 	ceph_pagelist_append(pagelist, method, size);
541 	payload_len += size;
542 
543 	osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
544 
545 	op->cls.argc = 0;	/* currently unused */
546 
547 	op->payload_len = payload_len;
548 }
549 EXPORT_SYMBOL(osd_req_op_cls_init);
550 
551 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
552 			  u16 opcode, const char *name, const void *value,
553 			  size_t size, u8 cmp_op, u8 cmp_mode)
554 {
555 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
556 	struct ceph_pagelist *pagelist;
557 	size_t payload_len;
558 
559 	BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
560 
561 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
562 	if (!pagelist)
563 		return -ENOMEM;
564 
565 	ceph_pagelist_init(pagelist);
566 
567 	payload_len = strlen(name);
568 	op->xattr.name_len = payload_len;
569 	ceph_pagelist_append(pagelist, name, payload_len);
570 
571 	op->xattr.value_len = size;
572 	ceph_pagelist_append(pagelist, value, size);
573 	payload_len += size;
574 
575 	op->xattr.cmp_op = cmp_op;
576 	op->xattr.cmp_mode = cmp_mode;
577 
578 	ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
579 	op->payload_len = payload_len;
580 	return 0;
581 }
582 EXPORT_SYMBOL(osd_req_op_xattr_init);
583 
584 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
585 				unsigned int which, u16 opcode,
586 				u64 cookie, u64 version, int flag)
587 {
588 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
589 
590 	BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
591 
592 	op->watch.cookie = cookie;
593 	op->watch.ver = version;
594 	if (opcode == CEPH_OSD_OP_WATCH && flag)
595 		op->watch.flag = (u8)1;
596 }
597 EXPORT_SYMBOL(osd_req_op_watch_init);
598 
599 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
600 				unsigned int which,
601 				u64 expected_object_size,
602 				u64 expected_write_size)
603 {
604 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
605 						      CEPH_OSD_OP_SETALLOCHINT);
606 
607 	op->alloc_hint.expected_object_size = expected_object_size;
608 	op->alloc_hint.expected_write_size = expected_write_size;
609 
610 	/*
611 	 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
612 	 * not worth a feature bit.  Set FAILOK per-op flag to make
613 	 * sure older osds don't trip over an unsupported opcode.
614 	 */
615 	op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
616 }
617 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
618 
619 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
620 				struct ceph_osd_data *osd_data)
621 {
622 	u64 length = ceph_osd_data_length(osd_data);
623 
624 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
625 		BUG_ON(length > (u64) SIZE_MAX);
626 		if (length)
627 			ceph_msg_data_add_pages(msg, osd_data->pages,
628 					length, osd_data->alignment);
629 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
630 		BUG_ON(!length);
631 		ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
632 #ifdef CONFIG_BLOCK
633 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
634 		ceph_msg_data_add_bio(msg, osd_data->bio, length);
635 #endif
636 	} else {
637 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
638 	}
639 }
640 
641 static u64 osd_req_encode_op(struct ceph_osd_request *req,
642 			      struct ceph_osd_op *dst, unsigned int which)
643 {
644 	struct ceph_osd_req_op *src;
645 	struct ceph_osd_data *osd_data;
646 	u64 request_data_len = 0;
647 	u64 data_length;
648 
649 	BUG_ON(which >= req->r_num_ops);
650 	src = &req->r_ops[which];
651 	if (WARN_ON(!osd_req_opcode_valid(src->op))) {
652 		pr_err("unrecognized osd opcode %d\n", src->op);
653 
654 		return 0;
655 	}
656 
657 	switch (src->op) {
658 	case CEPH_OSD_OP_STAT:
659 		osd_data = &src->raw_data_in;
660 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
661 		break;
662 	case CEPH_OSD_OP_READ:
663 	case CEPH_OSD_OP_WRITE:
664 	case CEPH_OSD_OP_ZERO:
665 	case CEPH_OSD_OP_TRUNCATE:
666 		if (src->op == CEPH_OSD_OP_WRITE)
667 			request_data_len = src->extent.length;
668 		dst->extent.offset = cpu_to_le64(src->extent.offset);
669 		dst->extent.length = cpu_to_le64(src->extent.length);
670 		dst->extent.truncate_size =
671 			cpu_to_le64(src->extent.truncate_size);
672 		dst->extent.truncate_seq =
673 			cpu_to_le32(src->extent.truncate_seq);
674 		osd_data = &src->extent.osd_data;
675 		if (src->op == CEPH_OSD_OP_WRITE)
676 			ceph_osdc_msg_data_add(req->r_request, osd_data);
677 		else
678 			ceph_osdc_msg_data_add(req->r_reply, osd_data);
679 		break;
680 	case CEPH_OSD_OP_CALL:
681 		dst->cls.class_len = src->cls.class_len;
682 		dst->cls.method_len = src->cls.method_len;
683 		osd_data = &src->cls.request_info;
684 		ceph_osdc_msg_data_add(req->r_request, osd_data);
685 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
686 		request_data_len = osd_data->pagelist->length;
687 
688 		osd_data = &src->cls.request_data;
689 		data_length = ceph_osd_data_length(osd_data);
690 		if (data_length) {
691 			BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
692 			dst->cls.indata_len = cpu_to_le32(data_length);
693 			ceph_osdc_msg_data_add(req->r_request, osd_data);
694 			src->payload_len += data_length;
695 			request_data_len += data_length;
696 		}
697 		osd_data = &src->cls.response_data;
698 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
699 		break;
700 	case CEPH_OSD_OP_STARTSYNC:
701 		break;
702 	case CEPH_OSD_OP_NOTIFY_ACK:
703 	case CEPH_OSD_OP_WATCH:
704 		dst->watch.cookie = cpu_to_le64(src->watch.cookie);
705 		dst->watch.ver = cpu_to_le64(src->watch.ver);
706 		dst->watch.flag = src->watch.flag;
707 		break;
708 	case CEPH_OSD_OP_SETALLOCHINT:
709 		dst->alloc_hint.expected_object_size =
710 		    cpu_to_le64(src->alloc_hint.expected_object_size);
711 		dst->alloc_hint.expected_write_size =
712 		    cpu_to_le64(src->alloc_hint.expected_write_size);
713 		break;
714 	case CEPH_OSD_OP_SETXATTR:
715 	case CEPH_OSD_OP_CMPXATTR:
716 		dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
717 		dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
718 		dst->xattr.cmp_op = src->xattr.cmp_op;
719 		dst->xattr.cmp_mode = src->xattr.cmp_mode;
720 		osd_data = &src->xattr.osd_data;
721 		ceph_osdc_msg_data_add(req->r_request, osd_data);
722 		request_data_len = osd_data->pagelist->length;
723 		break;
724 	case CEPH_OSD_OP_CREATE:
725 	case CEPH_OSD_OP_DELETE:
726 		break;
727 	default:
728 		pr_err("unsupported osd opcode %s\n",
729 			ceph_osd_op_name(src->op));
730 		WARN_ON(1);
731 
732 		return 0;
733 	}
734 
735 	dst->op = cpu_to_le16(src->op);
736 	dst->flags = cpu_to_le32(src->flags);
737 	dst->payload_len = cpu_to_le32(src->payload_len);
738 
739 	return request_data_len;
740 }
741 
742 /*
743  * build new request AND message, calculate layout, and adjust file
744  * extent as needed.
745  *
746  * if the file was recently truncated, we include information about its
747  * old and new size so that the object can be updated appropriately.  (we
748  * avoid synchronously deleting truncated objects because it's slow.)
749  *
750  * if @do_sync, include a 'startsync' command so that the osd will flush
751  * data quickly.
752  */
753 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
754 					       struct ceph_file_layout *layout,
755 					       struct ceph_vino vino,
756 					       u64 off, u64 *plen,
757 					       unsigned int which, int num_ops,
758 					       int opcode, int flags,
759 					       struct ceph_snap_context *snapc,
760 					       u32 truncate_seq,
761 					       u64 truncate_size,
762 					       bool use_mempool)
763 {
764 	struct ceph_osd_request *req;
765 	u64 objnum = 0;
766 	u64 objoff = 0;
767 	u64 objlen = 0;
768 	int r;
769 
770 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
771 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
772 	       opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
773 
774 	req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
775 					GFP_NOFS);
776 	if (!req)
777 		return ERR_PTR(-ENOMEM);
778 
779 	req->r_flags = flags;
780 
781 	/* calculate max write size */
782 	r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
783 	if (r < 0) {
784 		ceph_osdc_put_request(req);
785 		return ERR_PTR(r);
786 	}
787 
788 	if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
789 		osd_req_op_init(req, which, opcode);
790 	} else {
791 		u32 object_size = le32_to_cpu(layout->fl_object_size);
792 		u32 object_base = off - objoff;
793 		if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
794 			if (truncate_size <= object_base) {
795 				truncate_size = 0;
796 			} else {
797 				truncate_size -= object_base;
798 				if (truncate_size > object_size)
799 					truncate_size = object_size;
800 			}
801 		}
802 		osd_req_op_extent_init(req, which, opcode, objoff, objlen,
803 				       truncate_size, truncate_seq);
804 	}
805 
806 	req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
807 
808 	snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
809 		 "%llx.%08llx", vino.ino, objnum);
810 	req->r_base_oid.name_len = strlen(req->r_base_oid.name);
811 
812 	return req;
813 }
814 EXPORT_SYMBOL(ceph_osdc_new_request);
815 
816 /*
817  * We keep osd requests in an rbtree, sorted by ->r_tid.
818  */
819 static void __insert_request(struct ceph_osd_client *osdc,
820 			     struct ceph_osd_request *new)
821 {
822 	struct rb_node **p = &osdc->requests.rb_node;
823 	struct rb_node *parent = NULL;
824 	struct ceph_osd_request *req = NULL;
825 
826 	while (*p) {
827 		parent = *p;
828 		req = rb_entry(parent, struct ceph_osd_request, r_node);
829 		if (new->r_tid < req->r_tid)
830 			p = &(*p)->rb_left;
831 		else if (new->r_tid > req->r_tid)
832 			p = &(*p)->rb_right;
833 		else
834 			BUG();
835 	}
836 
837 	rb_link_node(&new->r_node, parent, p);
838 	rb_insert_color(&new->r_node, &osdc->requests);
839 }
840 
841 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
842 						 u64 tid)
843 {
844 	struct ceph_osd_request *req;
845 	struct rb_node *n = osdc->requests.rb_node;
846 
847 	while (n) {
848 		req = rb_entry(n, struct ceph_osd_request, r_node);
849 		if (tid < req->r_tid)
850 			n = n->rb_left;
851 		else if (tid > req->r_tid)
852 			n = n->rb_right;
853 		else
854 			return req;
855 	}
856 	return NULL;
857 }
858 
859 static struct ceph_osd_request *
860 __lookup_request_ge(struct ceph_osd_client *osdc,
861 		    u64 tid)
862 {
863 	struct ceph_osd_request *req;
864 	struct rb_node *n = osdc->requests.rb_node;
865 
866 	while (n) {
867 		req = rb_entry(n, struct ceph_osd_request, r_node);
868 		if (tid < req->r_tid) {
869 			if (!n->rb_left)
870 				return req;
871 			n = n->rb_left;
872 		} else if (tid > req->r_tid) {
873 			n = n->rb_right;
874 		} else {
875 			return req;
876 		}
877 	}
878 	return NULL;
879 }
880 
881 static void __kick_linger_request(struct ceph_osd_request *req)
882 {
883 	struct ceph_osd_client *osdc = req->r_osdc;
884 	struct ceph_osd *osd = req->r_osd;
885 
886 	/*
887 	 * Linger requests need to be resent with a new tid to avoid
888 	 * the dup op detection logic on the OSDs.  Achieve this with
889 	 * a re-register dance instead of open-coding.
890 	 */
891 	ceph_osdc_get_request(req);
892 	if (!list_empty(&req->r_linger_item))
893 		__unregister_linger_request(osdc, req);
894 	else
895 		__unregister_request(osdc, req);
896 	__register_request(osdc, req);
897 	ceph_osdc_put_request(req);
898 
899 	/*
900 	 * Unless request has been registered as both normal and
901 	 * lingering, __unregister{,_linger}_request clears r_osd.
902 	 * However, here we need to preserve r_osd to make sure we
903 	 * requeue on the same OSD.
904 	 */
905 	WARN_ON(req->r_osd || !osd);
906 	req->r_osd = osd;
907 
908 	dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
909 	__enqueue_request(req);
910 }
911 
912 /*
913  * Resubmit requests pending on the given osd.
914  */
915 static void __kick_osd_requests(struct ceph_osd_client *osdc,
916 				struct ceph_osd *osd)
917 {
918 	struct ceph_osd_request *req, *nreq;
919 	LIST_HEAD(resend);
920 	LIST_HEAD(resend_linger);
921 	int err;
922 
923 	dout("%s osd%d\n", __func__, osd->o_osd);
924 	err = __reset_osd(osdc, osd);
925 	if (err)
926 		return;
927 
928 	/*
929 	 * Build up a list of requests to resend by traversing the
930 	 * osd's list of requests.  Requests for a given object are
931 	 * sent in tid order, and that is also the order they're
932 	 * kept on this list.  Therefore all requests that are in
933 	 * flight will be found first, followed by all requests that
934 	 * have not yet been sent.  And to resend requests while
935 	 * preserving this order we will want to put any sent
936 	 * requests back on the front of the osd client's unsent
937 	 * list.
938 	 *
939 	 * So we build a separate ordered list of already-sent
940 	 * requests for the affected osd and splice it onto the
941 	 * front of the osd client's unsent list.  Once we've seen a
942 	 * request that has not yet been sent we're done.  Those
943 	 * requests are already sitting right where they belong.
944 	 */
945 	list_for_each_entry(req, &osd->o_requests, r_osd_item) {
946 		if (!req->r_sent)
947 			break;
948 
949 		if (!req->r_linger) {
950 			dout("%s requeueing %p tid %llu\n", __func__, req,
951 			     req->r_tid);
952 			list_move_tail(&req->r_req_lru_item, &resend);
953 			req->r_flags |= CEPH_OSD_FLAG_RETRY;
954 		} else {
955 			list_move_tail(&req->r_req_lru_item, &resend_linger);
956 		}
957 	}
958 	list_splice(&resend, &osdc->req_unsent);
959 
960 	/*
961 	 * Both registered and not yet registered linger requests are
962 	 * enqueued with a new tid on the same OSD.  We add/move them
963 	 * to req_unsent/o_requests at the end to keep things in tid
964 	 * order.
965 	 */
966 	list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
967 				 r_linger_osd_item) {
968 		WARN_ON(!list_empty(&req->r_req_lru_item));
969 		__kick_linger_request(req);
970 	}
971 
972 	list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
973 		__kick_linger_request(req);
974 }
975 
976 /*
977  * If the osd connection drops, we need to resubmit all requests.
978  */
979 static void osd_reset(struct ceph_connection *con)
980 {
981 	struct ceph_osd *osd = con->private;
982 	struct ceph_osd_client *osdc;
983 
984 	if (!osd)
985 		return;
986 	dout("osd_reset osd%d\n", osd->o_osd);
987 	osdc = osd->o_osdc;
988 	down_read(&osdc->map_sem);
989 	mutex_lock(&osdc->request_mutex);
990 	__kick_osd_requests(osdc, osd);
991 	__send_queued(osdc);
992 	mutex_unlock(&osdc->request_mutex);
993 	up_read(&osdc->map_sem);
994 }
995 
996 /*
997  * Track open sessions with osds.
998  */
999 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1000 {
1001 	struct ceph_osd *osd;
1002 
1003 	osd = kzalloc(sizeof(*osd), GFP_NOFS);
1004 	if (!osd)
1005 		return NULL;
1006 
1007 	atomic_set(&osd->o_ref, 1);
1008 	osd->o_osdc = osdc;
1009 	osd->o_osd = onum;
1010 	RB_CLEAR_NODE(&osd->o_node);
1011 	INIT_LIST_HEAD(&osd->o_requests);
1012 	INIT_LIST_HEAD(&osd->o_linger_requests);
1013 	INIT_LIST_HEAD(&osd->o_osd_lru);
1014 	osd->o_incarnation = 1;
1015 
1016 	ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1017 
1018 	INIT_LIST_HEAD(&osd->o_keepalive_item);
1019 	return osd;
1020 }
1021 
1022 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1023 {
1024 	if (atomic_inc_not_zero(&osd->o_ref)) {
1025 		dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1026 		     atomic_read(&osd->o_ref));
1027 		return osd;
1028 	} else {
1029 		dout("get_osd %p FAIL\n", osd);
1030 		return NULL;
1031 	}
1032 }
1033 
1034 static void put_osd(struct ceph_osd *osd)
1035 {
1036 	dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1037 	     atomic_read(&osd->o_ref) - 1);
1038 	if (atomic_dec_and_test(&osd->o_ref)) {
1039 		struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1040 
1041 		if (osd->o_auth.authorizer)
1042 			ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1043 		kfree(osd);
1044 	}
1045 }
1046 
1047 /*
1048  * remove an osd from our map
1049  */
1050 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1051 {
1052 	dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1053 	WARN_ON(!list_empty(&osd->o_requests));
1054 	WARN_ON(!list_empty(&osd->o_linger_requests));
1055 
1056 	list_del_init(&osd->o_osd_lru);
1057 	rb_erase(&osd->o_node, &osdc->osds);
1058 	RB_CLEAR_NODE(&osd->o_node);
1059 }
1060 
1061 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1062 {
1063 	dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1064 
1065 	if (!RB_EMPTY_NODE(&osd->o_node)) {
1066 		ceph_con_close(&osd->o_con);
1067 		__remove_osd(osdc, osd);
1068 		put_osd(osd);
1069 	}
1070 }
1071 
1072 static void remove_all_osds(struct ceph_osd_client *osdc)
1073 {
1074 	dout("%s %p\n", __func__, osdc);
1075 	mutex_lock(&osdc->request_mutex);
1076 	while (!RB_EMPTY_ROOT(&osdc->osds)) {
1077 		struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1078 						struct ceph_osd, o_node);
1079 		remove_osd(osdc, osd);
1080 	}
1081 	mutex_unlock(&osdc->request_mutex);
1082 }
1083 
1084 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1085 			      struct ceph_osd *osd)
1086 {
1087 	dout("%s %p\n", __func__, osd);
1088 	BUG_ON(!list_empty(&osd->o_osd_lru));
1089 
1090 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1091 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1092 }
1093 
1094 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1095 				  struct ceph_osd *osd)
1096 {
1097 	dout("%s %p\n", __func__, osd);
1098 
1099 	if (list_empty(&osd->o_requests) &&
1100 	    list_empty(&osd->o_linger_requests))
1101 		__move_osd_to_lru(osdc, osd);
1102 }
1103 
1104 static void __remove_osd_from_lru(struct ceph_osd *osd)
1105 {
1106 	dout("__remove_osd_from_lru %p\n", osd);
1107 	if (!list_empty(&osd->o_osd_lru))
1108 		list_del_init(&osd->o_osd_lru);
1109 }
1110 
1111 static void remove_old_osds(struct ceph_osd_client *osdc)
1112 {
1113 	struct ceph_osd *osd, *nosd;
1114 
1115 	dout("__remove_old_osds %p\n", osdc);
1116 	mutex_lock(&osdc->request_mutex);
1117 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1118 		if (time_before(jiffies, osd->lru_ttl))
1119 			break;
1120 		remove_osd(osdc, osd);
1121 	}
1122 	mutex_unlock(&osdc->request_mutex);
1123 }
1124 
1125 /*
1126  * reset osd connect
1127  */
1128 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1129 {
1130 	struct ceph_entity_addr *peer_addr;
1131 
1132 	dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1133 	if (list_empty(&osd->o_requests) &&
1134 	    list_empty(&osd->o_linger_requests)) {
1135 		remove_osd(osdc, osd);
1136 		return -ENODEV;
1137 	}
1138 
1139 	peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1140 	if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1141 			!ceph_con_opened(&osd->o_con)) {
1142 		struct ceph_osd_request *req;
1143 
1144 		dout("osd addr hasn't changed and connection never opened, "
1145 		     "letting msgr retry\n");
1146 		/* touch each r_stamp for handle_timeout()'s benfit */
1147 		list_for_each_entry(req, &osd->o_requests, r_osd_item)
1148 			req->r_stamp = jiffies;
1149 
1150 		return -EAGAIN;
1151 	}
1152 
1153 	ceph_con_close(&osd->o_con);
1154 	ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1155 	osd->o_incarnation++;
1156 
1157 	return 0;
1158 }
1159 
1160 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1161 {
1162 	struct rb_node **p = &osdc->osds.rb_node;
1163 	struct rb_node *parent = NULL;
1164 	struct ceph_osd *osd = NULL;
1165 
1166 	dout("__insert_osd %p osd%d\n", new, new->o_osd);
1167 	while (*p) {
1168 		parent = *p;
1169 		osd = rb_entry(parent, struct ceph_osd, o_node);
1170 		if (new->o_osd < osd->o_osd)
1171 			p = &(*p)->rb_left;
1172 		else if (new->o_osd > osd->o_osd)
1173 			p = &(*p)->rb_right;
1174 		else
1175 			BUG();
1176 	}
1177 
1178 	rb_link_node(&new->o_node, parent, p);
1179 	rb_insert_color(&new->o_node, &osdc->osds);
1180 }
1181 
1182 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1183 {
1184 	struct ceph_osd *osd;
1185 	struct rb_node *n = osdc->osds.rb_node;
1186 
1187 	while (n) {
1188 		osd = rb_entry(n, struct ceph_osd, o_node);
1189 		if (o < osd->o_osd)
1190 			n = n->rb_left;
1191 		else if (o > osd->o_osd)
1192 			n = n->rb_right;
1193 		else
1194 			return osd;
1195 	}
1196 	return NULL;
1197 }
1198 
1199 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1200 {
1201 	schedule_delayed_work(&osdc->timeout_work,
1202 			osdc->client->options->osd_keepalive_timeout * HZ);
1203 }
1204 
1205 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1206 {
1207 	cancel_delayed_work(&osdc->timeout_work);
1208 }
1209 
1210 /*
1211  * Register request, assign tid.  If this is the first request, set up
1212  * the timeout event.
1213  */
1214 static void __register_request(struct ceph_osd_client *osdc,
1215 			       struct ceph_osd_request *req)
1216 {
1217 	req->r_tid = ++osdc->last_tid;
1218 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1219 	dout("__register_request %p tid %lld\n", req, req->r_tid);
1220 	__insert_request(osdc, req);
1221 	ceph_osdc_get_request(req);
1222 	osdc->num_requests++;
1223 	if (osdc->num_requests == 1) {
1224 		dout(" first request, scheduling timeout\n");
1225 		__schedule_osd_timeout(osdc);
1226 	}
1227 }
1228 
1229 /*
1230  * called under osdc->request_mutex
1231  */
1232 static void __unregister_request(struct ceph_osd_client *osdc,
1233 				 struct ceph_osd_request *req)
1234 {
1235 	if (RB_EMPTY_NODE(&req->r_node)) {
1236 		dout("__unregister_request %p tid %lld not registered\n",
1237 			req, req->r_tid);
1238 		return;
1239 	}
1240 
1241 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1242 	rb_erase(&req->r_node, &osdc->requests);
1243 	RB_CLEAR_NODE(&req->r_node);
1244 	osdc->num_requests--;
1245 
1246 	if (req->r_osd) {
1247 		/* make sure the original request isn't in flight. */
1248 		ceph_msg_revoke(req->r_request);
1249 
1250 		list_del_init(&req->r_osd_item);
1251 		maybe_move_osd_to_lru(osdc, req->r_osd);
1252 		if (list_empty(&req->r_linger_osd_item))
1253 			req->r_osd = NULL;
1254 	}
1255 
1256 	list_del_init(&req->r_req_lru_item);
1257 	ceph_osdc_put_request(req);
1258 
1259 	if (osdc->num_requests == 0) {
1260 		dout(" no requests, canceling timeout\n");
1261 		__cancel_osd_timeout(osdc);
1262 	}
1263 }
1264 
1265 /*
1266  * Cancel a previously queued request message
1267  */
1268 static void __cancel_request(struct ceph_osd_request *req)
1269 {
1270 	if (req->r_sent && req->r_osd) {
1271 		ceph_msg_revoke(req->r_request);
1272 		req->r_sent = 0;
1273 	}
1274 }
1275 
1276 static void __register_linger_request(struct ceph_osd_client *osdc,
1277 				    struct ceph_osd_request *req)
1278 {
1279 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1280 	WARN_ON(!req->r_linger);
1281 
1282 	ceph_osdc_get_request(req);
1283 	list_add_tail(&req->r_linger_item, &osdc->req_linger);
1284 	if (req->r_osd)
1285 		list_add_tail(&req->r_linger_osd_item,
1286 			      &req->r_osd->o_linger_requests);
1287 }
1288 
1289 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1290 					struct ceph_osd_request *req)
1291 {
1292 	WARN_ON(!req->r_linger);
1293 
1294 	if (list_empty(&req->r_linger_item)) {
1295 		dout("%s %p tid %llu not registered\n", __func__, req,
1296 		     req->r_tid);
1297 		return;
1298 	}
1299 
1300 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1301 	list_del_init(&req->r_linger_item);
1302 
1303 	if (req->r_osd) {
1304 		list_del_init(&req->r_linger_osd_item);
1305 		maybe_move_osd_to_lru(osdc, req->r_osd);
1306 		if (list_empty(&req->r_osd_item))
1307 			req->r_osd = NULL;
1308 	}
1309 	ceph_osdc_put_request(req);
1310 }
1311 
1312 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1313 				  struct ceph_osd_request *req)
1314 {
1315 	if (!req->r_linger) {
1316 		dout("set_request_linger %p\n", req);
1317 		req->r_linger = 1;
1318 	}
1319 }
1320 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1321 
1322 /*
1323  * Returns whether a request should be blocked from being sent
1324  * based on the current osdmap and osd_client settings.
1325  *
1326  * Caller should hold map_sem for read.
1327  */
1328 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1329 				   struct ceph_osd_request *req)
1330 {
1331 	bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1332 	bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1333 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1334 	return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1335 		(req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1336 }
1337 
1338 /*
1339  * Calculate mapping of a request to a PG.  Takes tiering into account.
1340  */
1341 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1342 			     struct ceph_osd_request *req,
1343 			     struct ceph_pg *pg_out)
1344 {
1345 	bool need_check_tiering;
1346 
1347 	need_check_tiering = false;
1348 	if (req->r_target_oloc.pool == -1) {
1349 		req->r_target_oloc = req->r_base_oloc; /* struct */
1350 		need_check_tiering = true;
1351 	}
1352 	if (req->r_target_oid.name_len == 0) {
1353 		ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1354 		need_check_tiering = true;
1355 	}
1356 
1357 	if (need_check_tiering &&
1358 	    (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1359 		struct ceph_pg_pool_info *pi;
1360 
1361 		pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1362 		if (pi) {
1363 			if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1364 			    pi->read_tier >= 0)
1365 				req->r_target_oloc.pool = pi->read_tier;
1366 			if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1367 			    pi->write_tier >= 0)
1368 				req->r_target_oloc.pool = pi->write_tier;
1369 		}
1370 		/* !pi is caught in ceph_oloc_oid_to_pg() */
1371 	}
1372 
1373 	return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1374 				   &req->r_target_oid, pg_out);
1375 }
1376 
1377 static void __enqueue_request(struct ceph_osd_request *req)
1378 {
1379 	struct ceph_osd_client *osdc = req->r_osdc;
1380 
1381 	dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1382 	     req->r_osd ? req->r_osd->o_osd : -1);
1383 
1384 	if (req->r_osd) {
1385 		__remove_osd_from_lru(req->r_osd);
1386 		list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1387 		list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1388 	} else {
1389 		list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1390 	}
1391 }
1392 
1393 /*
1394  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1395  * (as needed), and set the request r_osd appropriately.  If there is
1396  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1397  * (unsent, homeless) or leave on in-flight lru.
1398  *
1399  * Return 0 if unchanged, 1 if changed, or negative on error.
1400  *
1401  * Caller should hold map_sem for read and request_mutex.
1402  */
1403 static int __map_request(struct ceph_osd_client *osdc,
1404 			 struct ceph_osd_request *req, int force_resend)
1405 {
1406 	struct ceph_pg pgid;
1407 	int acting[CEPH_PG_MAX_SIZE];
1408 	int num, o;
1409 	int err;
1410 	bool was_paused;
1411 
1412 	dout("map_request %p tid %lld\n", req, req->r_tid);
1413 
1414 	err = __calc_request_pg(osdc->osdmap, req, &pgid);
1415 	if (err) {
1416 		list_move(&req->r_req_lru_item, &osdc->req_notarget);
1417 		return err;
1418 	}
1419 	req->r_pgid = pgid;
1420 
1421 	num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1422 	if (num < 0)
1423 		num = 0;
1424 
1425 	was_paused = req->r_paused;
1426 	req->r_paused = __req_should_be_paused(osdc, req);
1427 	if (was_paused && !req->r_paused)
1428 		force_resend = 1;
1429 
1430 	if ((!force_resend &&
1431 	     req->r_osd && req->r_osd->o_osd == o &&
1432 	     req->r_sent >= req->r_osd->o_incarnation &&
1433 	     req->r_num_pg_osds == num &&
1434 	     memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1435 	    (req->r_osd == NULL && o == -1) ||
1436 	    req->r_paused)
1437 		return 0;  /* no change */
1438 
1439 	dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1440 	     req->r_tid, pgid.pool, pgid.seed, o,
1441 	     req->r_osd ? req->r_osd->o_osd : -1);
1442 
1443 	/* record full pg acting set */
1444 	memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1445 	req->r_num_pg_osds = num;
1446 
1447 	if (req->r_osd) {
1448 		__cancel_request(req);
1449 		list_del_init(&req->r_osd_item);
1450 		list_del_init(&req->r_linger_osd_item);
1451 		req->r_osd = NULL;
1452 	}
1453 
1454 	req->r_osd = __lookup_osd(osdc, o);
1455 	if (!req->r_osd && o >= 0) {
1456 		err = -ENOMEM;
1457 		req->r_osd = create_osd(osdc, o);
1458 		if (!req->r_osd) {
1459 			list_move(&req->r_req_lru_item, &osdc->req_notarget);
1460 			goto out;
1461 		}
1462 
1463 		dout("map_request osd %p is osd%d\n", req->r_osd, o);
1464 		__insert_osd(osdc, req->r_osd);
1465 
1466 		ceph_con_open(&req->r_osd->o_con,
1467 			      CEPH_ENTITY_TYPE_OSD, o,
1468 			      &osdc->osdmap->osd_addr[o]);
1469 	}
1470 
1471 	__enqueue_request(req);
1472 	err = 1;   /* osd or pg changed */
1473 
1474 out:
1475 	return err;
1476 }
1477 
1478 /*
1479  * caller should hold map_sem (for read) and request_mutex
1480  */
1481 static void __send_request(struct ceph_osd_client *osdc,
1482 			   struct ceph_osd_request *req)
1483 {
1484 	void *p;
1485 
1486 	dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1487 	     req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1488 	     (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1489 
1490 	/* fill in message content that changes each time we send it */
1491 	put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1492 	put_unaligned_le32(req->r_flags, req->r_request_flags);
1493 	put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1494 	p = req->r_request_pgid;
1495 	ceph_encode_64(&p, req->r_pgid.pool);
1496 	ceph_encode_32(&p, req->r_pgid.seed);
1497 	put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1498 	memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1499 	       sizeof(req->r_reassert_version));
1500 
1501 	req->r_stamp = jiffies;
1502 	list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1503 
1504 	ceph_msg_get(req->r_request); /* send consumes a ref */
1505 
1506 	req->r_sent = req->r_osd->o_incarnation;
1507 
1508 	ceph_con_send(&req->r_osd->o_con, req->r_request);
1509 }
1510 
1511 /*
1512  * Send any requests in the queue (req_unsent).
1513  */
1514 static void __send_queued(struct ceph_osd_client *osdc)
1515 {
1516 	struct ceph_osd_request *req, *tmp;
1517 
1518 	dout("__send_queued\n");
1519 	list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1520 		__send_request(osdc, req);
1521 }
1522 
1523 /*
1524  * Caller should hold map_sem for read and request_mutex.
1525  */
1526 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1527 				     struct ceph_osd_request *req,
1528 				     bool nofail)
1529 {
1530 	int rc;
1531 
1532 	__register_request(osdc, req);
1533 	req->r_sent = 0;
1534 	req->r_got_reply = 0;
1535 	rc = __map_request(osdc, req, 0);
1536 	if (rc < 0) {
1537 		if (nofail) {
1538 			dout("osdc_start_request failed map, "
1539 				" will retry %lld\n", req->r_tid);
1540 			rc = 0;
1541 		} else {
1542 			__unregister_request(osdc, req);
1543 		}
1544 		return rc;
1545 	}
1546 
1547 	if (req->r_osd == NULL) {
1548 		dout("send_request %p no up osds in pg\n", req);
1549 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1550 	} else {
1551 		__send_queued(osdc);
1552 	}
1553 
1554 	return 0;
1555 }
1556 
1557 /*
1558  * Timeout callback, called every N seconds when 1 or more osd
1559  * requests has been active for more than N seconds.  When this
1560  * happens, we ping all OSDs with requests who have timed out to
1561  * ensure any communications channel reset is detected.  Reset the
1562  * request timeouts another N seconds in the future as we go.
1563  * Reschedule the timeout event another N seconds in future (unless
1564  * there are no open requests).
1565  */
1566 static void handle_timeout(struct work_struct *work)
1567 {
1568 	struct ceph_osd_client *osdc =
1569 		container_of(work, struct ceph_osd_client, timeout_work.work);
1570 	struct ceph_osd_request *req;
1571 	struct ceph_osd *osd;
1572 	unsigned long keepalive =
1573 		osdc->client->options->osd_keepalive_timeout * HZ;
1574 	struct list_head slow_osds;
1575 	dout("timeout\n");
1576 	down_read(&osdc->map_sem);
1577 
1578 	ceph_monc_request_next_osdmap(&osdc->client->monc);
1579 
1580 	mutex_lock(&osdc->request_mutex);
1581 
1582 	/*
1583 	 * ping osds that are a bit slow.  this ensures that if there
1584 	 * is a break in the TCP connection we will notice, and reopen
1585 	 * a connection with that osd (from the fault callback).
1586 	 */
1587 	INIT_LIST_HEAD(&slow_osds);
1588 	list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1589 		if (time_before(jiffies, req->r_stamp + keepalive))
1590 			break;
1591 
1592 		osd = req->r_osd;
1593 		BUG_ON(!osd);
1594 		dout(" tid %llu is slow, will send keepalive on osd%d\n",
1595 		     req->r_tid, osd->o_osd);
1596 		list_move_tail(&osd->o_keepalive_item, &slow_osds);
1597 	}
1598 	while (!list_empty(&slow_osds)) {
1599 		osd = list_entry(slow_osds.next, struct ceph_osd,
1600 				 o_keepalive_item);
1601 		list_del_init(&osd->o_keepalive_item);
1602 		ceph_con_keepalive(&osd->o_con);
1603 	}
1604 
1605 	__schedule_osd_timeout(osdc);
1606 	__send_queued(osdc);
1607 	mutex_unlock(&osdc->request_mutex);
1608 	up_read(&osdc->map_sem);
1609 }
1610 
1611 static void handle_osds_timeout(struct work_struct *work)
1612 {
1613 	struct ceph_osd_client *osdc =
1614 		container_of(work, struct ceph_osd_client,
1615 			     osds_timeout_work.work);
1616 	unsigned long delay =
1617 		osdc->client->options->osd_idle_ttl * HZ >> 2;
1618 
1619 	dout("osds timeout\n");
1620 	down_read(&osdc->map_sem);
1621 	remove_old_osds(osdc);
1622 	up_read(&osdc->map_sem);
1623 
1624 	schedule_delayed_work(&osdc->osds_timeout_work,
1625 			      round_jiffies_relative(delay));
1626 }
1627 
1628 static int ceph_oloc_decode(void **p, void *end,
1629 			    struct ceph_object_locator *oloc)
1630 {
1631 	u8 struct_v, struct_cv;
1632 	u32 len;
1633 	void *struct_end;
1634 	int ret = 0;
1635 
1636 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1637 	struct_v = ceph_decode_8(p);
1638 	struct_cv = ceph_decode_8(p);
1639 	if (struct_v < 3) {
1640 		pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1641 			struct_v, struct_cv);
1642 		goto e_inval;
1643 	}
1644 	if (struct_cv > 6) {
1645 		pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1646 			struct_v, struct_cv);
1647 		goto e_inval;
1648 	}
1649 	len = ceph_decode_32(p);
1650 	ceph_decode_need(p, end, len, e_inval);
1651 	struct_end = *p + len;
1652 
1653 	oloc->pool = ceph_decode_64(p);
1654 	*p += 4; /* skip preferred */
1655 
1656 	len = ceph_decode_32(p);
1657 	if (len > 0) {
1658 		pr_warn("ceph_object_locator::key is set\n");
1659 		goto e_inval;
1660 	}
1661 
1662 	if (struct_v >= 5) {
1663 		len = ceph_decode_32(p);
1664 		if (len > 0) {
1665 			pr_warn("ceph_object_locator::nspace is set\n");
1666 			goto e_inval;
1667 		}
1668 	}
1669 
1670 	if (struct_v >= 6) {
1671 		s64 hash = ceph_decode_64(p);
1672 		if (hash != -1) {
1673 			pr_warn("ceph_object_locator::hash is set\n");
1674 			goto e_inval;
1675 		}
1676 	}
1677 
1678 	/* skip the rest */
1679 	*p = struct_end;
1680 out:
1681 	return ret;
1682 
1683 e_inval:
1684 	ret = -EINVAL;
1685 	goto out;
1686 }
1687 
1688 static int ceph_redirect_decode(void **p, void *end,
1689 				struct ceph_request_redirect *redir)
1690 {
1691 	u8 struct_v, struct_cv;
1692 	u32 len;
1693 	void *struct_end;
1694 	int ret;
1695 
1696 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1697 	struct_v = ceph_decode_8(p);
1698 	struct_cv = ceph_decode_8(p);
1699 	if (struct_cv > 1) {
1700 		pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1701 			struct_v, struct_cv);
1702 		goto e_inval;
1703 	}
1704 	len = ceph_decode_32(p);
1705 	ceph_decode_need(p, end, len, e_inval);
1706 	struct_end = *p + len;
1707 
1708 	ret = ceph_oloc_decode(p, end, &redir->oloc);
1709 	if (ret)
1710 		goto out;
1711 
1712 	len = ceph_decode_32(p);
1713 	if (len > 0) {
1714 		pr_warn("ceph_request_redirect::object_name is set\n");
1715 		goto e_inval;
1716 	}
1717 
1718 	len = ceph_decode_32(p);
1719 	*p += len; /* skip osd_instructions */
1720 
1721 	/* skip the rest */
1722 	*p = struct_end;
1723 out:
1724 	return ret;
1725 
1726 e_inval:
1727 	ret = -EINVAL;
1728 	goto out;
1729 }
1730 
1731 static void complete_request(struct ceph_osd_request *req)
1732 {
1733 	complete_all(&req->r_safe_completion);  /* fsync waiter */
1734 }
1735 
1736 /*
1737  * handle osd op reply.  either call the callback if it is specified,
1738  * or do the completion to wake up the waiting thread.
1739  */
1740 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1741 			 struct ceph_connection *con)
1742 {
1743 	void *p, *end;
1744 	struct ceph_osd_request *req;
1745 	struct ceph_request_redirect redir;
1746 	u64 tid;
1747 	int object_len;
1748 	unsigned int numops;
1749 	int payload_len, flags;
1750 	s32 result;
1751 	s32 retry_attempt;
1752 	struct ceph_pg pg;
1753 	int err;
1754 	u32 reassert_epoch;
1755 	u64 reassert_version;
1756 	u32 osdmap_epoch;
1757 	int already_completed;
1758 	u32 bytes;
1759 	unsigned int i;
1760 
1761 	tid = le64_to_cpu(msg->hdr.tid);
1762 	dout("handle_reply %p tid %llu\n", msg, tid);
1763 
1764 	p = msg->front.iov_base;
1765 	end = p + msg->front.iov_len;
1766 
1767 	ceph_decode_need(&p, end, 4, bad);
1768 	object_len = ceph_decode_32(&p);
1769 	ceph_decode_need(&p, end, object_len, bad);
1770 	p += object_len;
1771 
1772 	err = ceph_decode_pgid(&p, end, &pg);
1773 	if (err)
1774 		goto bad;
1775 
1776 	ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1777 	flags = ceph_decode_64(&p);
1778 	result = ceph_decode_32(&p);
1779 	reassert_epoch = ceph_decode_32(&p);
1780 	reassert_version = ceph_decode_64(&p);
1781 	osdmap_epoch = ceph_decode_32(&p);
1782 
1783 	/* lookup */
1784 	down_read(&osdc->map_sem);
1785 	mutex_lock(&osdc->request_mutex);
1786 	req = __lookup_request(osdc, tid);
1787 	if (req == NULL) {
1788 		dout("handle_reply tid %llu dne\n", tid);
1789 		goto bad_mutex;
1790 	}
1791 	ceph_osdc_get_request(req);
1792 
1793 	dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1794 	     req, result);
1795 
1796 	ceph_decode_need(&p, end, 4, bad_put);
1797 	numops = ceph_decode_32(&p);
1798 	if (numops > CEPH_OSD_MAX_OP)
1799 		goto bad_put;
1800 	if (numops != req->r_num_ops)
1801 		goto bad_put;
1802 	payload_len = 0;
1803 	ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1804 	for (i = 0; i < numops; i++) {
1805 		struct ceph_osd_op *op = p;
1806 		int len;
1807 
1808 		len = le32_to_cpu(op->payload_len);
1809 		req->r_reply_op_len[i] = len;
1810 		dout(" op %d has %d bytes\n", i, len);
1811 		payload_len += len;
1812 		p += sizeof(*op);
1813 	}
1814 	bytes = le32_to_cpu(msg->hdr.data_len);
1815 	if (payload_len != bytes) {
1816 		pr_warn("sum of op payload lens %d != data_len %d\n",
1817 			payload_len, bytes);
1818 		goto bad_put;
1819 	}
1820 
1821 	ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1822 	retry_attempt = ceph_decode_32(&p);
1823 	for (i = 0; i < numops; i++)
1824 		req->r_reply_op_result[i] = ceph_decode_32(&p);
1825 
1826 	if (le16_to_cpu(msg->hdr.version) >= 6) {
1827 		p += 8 + 4; /* skip replay_version */
1828 		p += 8; /* skip user_version */
1829 
1830 		err = ceph_redirect_decode(&p, end, &redir);
1831 		if (err)
1832 			goto bad_put;
1833 	} else {
1834 		redir.oloc.pool = -1;
1835 	}
1836 
1837 	if (redir.oloc.pool != -1) {
1838 		dout("redirect pool %lld\n", redir.oloc.pool);
1839 
1840 		__unregister_request(osdc, req);
1841 
1842 		req->r_target_oloc = redir.oloc; /* struct */
1843 
1844 		/*
1845 		 * Start redirect requests with nofail=true.  If
1846 		 * mapping fails, request will end up on the notarget
1847 		 * list, waiting for the new osdmap (which can take
1848 		 * a while), even though the original request mapped
1849 		 * successfully.  In the future we might want to follow
1850 		 * original request's nofail setting here.
1851 		 */
1852 		err = __ceph_osdc_start_request(osdc, req, true);
1853 		BUG_ON(err);
1854 
1855 		goto out_unlock;
1856 	}
1857 
1858 	already_completed = req->r_got_reply;
1859 	if (!req->r_got_reply) {
1860 		req->r_result = result;
1861 		dout("handle_reply result %d bytes %d\n", req->r_result,
1862 		     bytes);
1863 		if (req->r_result == 0)
1864 			req->r_result = bytes;
1865 
1866 		/* in case this is a write and we need to replay, */
1867 		req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1868 		req->r_reassert_version.version = cpu_to_le64(reassert_version);
1869 
1870 		req->r_got_reply = 1;
1871 	} else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1872 		dout("handle_reply tid %llu dup ack\n", tid);
1873 		goto out_unlock;
1874 	}
1875 
1876 	dout("handle_reply tid %llu flags %d\n", tid, flags);
1877 
1878 	if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1879 		__register_linger_request(osdc, req);
1880 
1881 	/* either this is a read, or we got the safe response */
1882 	if (result < 0 ||
1883 	    (flags & CEPH_OSD_FLAG_ONDISK) ||
1884 	    ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1885 		__unregister_request(osdc, req);
1886 
1887 	mutex_unlock(&osdc->request_mutex);
1888 	up_read(&osdc->map_sem);
1889 
1890 	if (!already_completed) {
1891 		if (req->r_unsafe_callback &&
1892 		    result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1893 			req->r_unsafe_callback(req, true);
1894 		if (req->r_callback)
1895 			req->r_callback(req, msg);
1896 		else
1897 			complete_all(&req->r_completion);
1898 	}
1899 
1900 	if (flags & CEPH_OSD_FLAG_ONDISK) {
1901 		if (req->r_unsafe_callback && already_completed)
1902 			req->r_unsafe_callback(req, false);
1903 		complete_request(req);
1904 	}
1905 
1906 out:
1907 	dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1908 	ceph_osdc_put_request(req);
1909 	return;
1910 out_unlock:
1911 	mutex_unlock(&osdc->request_mutex);
1912 	up_read(&osdc->map_sem);
1913 	goto out;
1914 
1915 bad_put:
1916 	req->r_result = -EIO;
1917 	__unregister_request(osdc, req);
1918 	if (req->r_callback)
1919 		req->r_callback(req, msg);
1920 	else
1921 		complete_all(&req->r_completion);
1922 	complete_request(req);
1923 	ceph_osdc_put_request(req);
1924 bad_mutex:
1925 	mutex_unlock(&osdc->request_mutex);
1926 	up_read(&osdc->map_sem);
1927 bad:
1928 	pr_err("corrupt osd_op_reply got %d %d\n",
1929 	       (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1930 	ceph_msg_dump(msg);
1931 }
1932 
1933 static void reset_changed_osds(struct ceph_osd_client *osdc)
1934 {
1935 	struct rb_node *p, *n;
1936 
1937 	dout("%s %p\n", __func__, osdc);
1938 	for (p = rb_first(&osdc->osds); p; p = n) {
1939 		struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1940 
1941 		n = rb_next(p);
1942 		if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1943 		    memcmp(&osd->o_con.peer_addr,
1944 			   ceph_osd_addr(osdc->osdmap,
1945 					 osd->o_osd),
1946 			   sizeof(struct ceph_entity_addr)) != 0)
1947 			__reset_osd(osdc, osd);
1948 	}
1949 }
1950 
1951 /*
1952  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1953  * no osd, request a new map.
1954  *
1955  * Caller should hold map_sem for read.
1956  */
1957 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1958 			  bool force_resend_writes)
1959 {
1960 	struct ceph_osd_request *req, *nreq;
1961 	struct rb_node *p;
1962 	int needmap = 0;
1963 	int err;
1964 	bool force_resend_req;
1965 
1966 	dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1967 		force_resend_writes ? " (force resend writes)" : "");
1968 	mutex_lock(&osdc->request_mutex);
1969 	for (p = rb_first(&osdc->requests); p; ) {
1970 		req = rb_entry(p, struct ceph_osd_request, r_node);
1971 		p = rb_next(p);
1972 
1973 		/*
1974 		 * For linger requests that have not yet been
1975 		 * registered, move them to the linger list; they'll
1976 		 * be sent to the osd in the loop below.  Unregister
1977 		 * the request before re-registering it as a linger
1978 		 * request to ensure the __map_request() below
1979 		 * will decide it needs to be sent.
1980 		 */
1981 		if (req->r_linger && list_empty(&req->r_linger_item)) {
1982 			dout("%p tid %llu restart on osd%d\n",
1983 			     req, req->r_tid,
1984 			     req->r_osd ? req->r_osd->o_osd : -1);
1985 			ceph_osdc_get_request(req);
1986 			__unregister_request(osdc, req);
1987 			__register_linger_request(osdc, req);
1988 			ceph_osdc_put_request(req);
1989 			continue;
1990 		}
1991 
1992 		force_resend_req = force_resend ||
1993 			(force_resend_writes &&
1994 				req->r_flags & CEPH_OSD_FLAG_WRITE);
1995 		err = __map_request(osdc, req, force_resend_req);
1996 		if (err < 0)
1997 			continue;  /* error */
1998 		if (req->r_osd == NULL) {
1999 			dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2000 			needmap++;  /* request a newer map */
2001 		} else if (err > 0) {
2002 			if (!req->r_linger) {
2003 				dout("%p tid %llu requeued on osd%d\n", req,
2004 				     req->r_tid,
2005 				     req->r_osd ? req->r_osd->o_osd : -1);
2006 				req->r_flags |= CEPH_OSD_FLAG_RETRY;
2007 			}
2008 		}
2009 	}
2010 
2011 	list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2012 				 r_linger_item) {
2013 		dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2014 
2015 		err = __map_request(osdc, req,
2016 				    force_resend || force_resend_writes);
2017 		dout("__map_request returned %d\n", err);
2018 		if (err < 0)
2019 			continue;  /* hrm! */
2020 		if (req->r_osd == NULL || err > 0) {
2021 			if (req->r_osd == NULL) {
2022 				dout("lingering %p tid %llu maps to no osd\n",
2023 				     req, req->r_tid);
2024 				/*
2025 				 * A homeless lingering request makes
2026 				 * no sense, as it's job is to keep
2027 				 * a particular OSD connection open.
2028 				 * Request a newer map and kick the
2029 				 * request, knowing that it won't be
2030 				 * resent until we actually get a map
2031 				 * that can tell us where to send it.
2032 				 */
2033 				needmap++;
2034 			}
2035 
2036 			dout("kicking lingering %p tid %llu osd%d\n", req,
2037 			     req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2038 			__register_request(osdc, req);
2039 			__unregister_linger_request(osdc, req);
2040 		}
2041 	}
2042 	reset_changed_osds(osdc);
2043 	mutex_unlock(&osdc->request_mutex);
2044 
2045 	if (needmap) {
2046 		dout("%d requests for down osds, need new map\n", needmap);
2047 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2048 	}
2049 }
2050 
2051 
2052 /*
2053  * Process updated osd map.
2054  *
2055  * The message contains any number of incremental and full maps, normally
2056  * indicating some sort of topology change in the cluster.  Kick requests
2057  * off to different OSDs as needed.
2058  */
2059 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2060 {
2061 	void *p, *end, *next;
2062 	u32 nr_maps, maplen;
2063 	u32 epoch;
2064 	struct ceph_osdmap *newmap = NULL, *oldmap;
2065 	int err;
2066 	struct ceph_fsid fsid;
2067 	bool was_full;
2068 
2069 	dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2070 	p = msg->front.iov_base;
2071 	end = p + msg->front.iov_len;
2072 
2073 	/* verify fsid */
2074 	ceph_decode_need(&p, end, sizeof(fsid), bad);
2075 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
2076 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
2077 		return;
2078 
2079 	down_write(&osdc->map_sem);
2080 
2081 	was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2082 
2083 	/* incremental maps */
2084 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2085 	dout(" %d inc maps\n", nr_maps);
2086 	while (nr_maps > 0) {
2087 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2088 		epoch = ceph_decode_32(&p);
2089 		maplen = ceph_decode_32(&p);
2090 		ceph_decode_need(&p, end, maplen, bad);
2091 		next = p + maplen;
2092 		if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2093 			dout("applying incremental map %u len %d\n",
2094 			     epoch, maplen);
2095 			newmap = osdmap_apply_incremental(&p, next,
2096 							  osdc->osdmap,
2097 							  &osdc->client->msgr);
2098 			if (IS_ERR(newmap)) {
2099 				err = PTR_ERR(newmap);
2100 				goto bad;
2101 			}
2102 			BUG_ON(!newmap);
2103 			if (newmap != osdc->osdmap) {
2104 				ceph_osdmap_destroy(osdc->osdmap);
2105 				osdc->osdmap = newmap;
2106 			}
2107 			was_full = was_full ||
2108 				ceph_osdmap_flag(osdc->osdmap,
2109 						 CEPH_OSDMAP_FULL);
2110 			kick_requests(osdc, 0, was_full);
2111 		} else {
2112 			dout("ignoring incremental map %u len %d\n",
2113 			     epoch, maplen);
2114 		}
2115 		p = next;
2116 		nr_maps--;
2117 	}
2118 	if (newmap)
2119 		goto done;
2120 
2121 	/* full maps */
2122 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2123 	dout(" %d full maps\n", nr_maps);
2124 	while (nr_maps) {
2125 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2126 		epoch = ceph_decode_32(&p);
2127 		maplen = ceph_decode_32(&p);
2128 		ceph_decode_need(&p, end, maplen, bad);
2129 		if (nr_maps > 1) {
2130 			dout("skipping non-latest full map %u len %d\n",
2131 			     epoch, maplen);
2132 		} else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2133 			dout("skipping full map %u len %d, "
2134 			     "older than our %u\n", epoch, maplen,
2135 			     osdc->osdmap->epoch);
2136 		} else {
2137 			int skipped_map = 0;
2138 
2139 			dout("taking full map %u len %d\n", epoch, maplen);
2140 			newmap = ceph_osdmap_decode(&p, p+maplen);
2141 			if (IS_ERR(newmap)) {
2142 				err = PTR_ERR(newmap);
2143 				goto bad;
2144 			}
2145 			BUG_ON(!newmap);
2146 			oldmap = osdc->osdmap;
2147 			osdc->osdmap = newmap;
2148 			if (oldmap) {
2149 				if (oldmap->epoch + 1 < newmap->epoch)
2150 					skipped_map = 1;
2151 				ceph_osdmap_destroy(oldmap);
2152 			}
2153 			was_full = was_full ||
2154 				ceph_osdmap_flag(osdc->osdmap,
2155 						 CEPH_OSDMAP_FULL);
2156 			kick_requests(osdc, skipped_map, was_full);
2157 		}
2158 		p += maplen;
2159 		nr_maps--;
2160 	}
2161 
2162 	if (!osdc->osdmap)
2163 		goto bad;
2164 done:
2165 	downgrade_write(&osdc->map_sem);
2166 	ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2167 
2168 	/*
2169 	 * subscribe to subsequent osdmap updates if full to ensure
2170 	 * we find out when we are no longer full and stop returning
2171 	 * ENOSPC.
2172 	 */
2173 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2174 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2175 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2176 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2177 
2178 	mutex_lock(&osdc->request_mutex);
2179 	__send_queued(osdc);
2180 	mutex_unlock(&osdc->request_mutex);
2181 	up_read(&osdc->map_sem);
2182 	wake_up_all(&osdc->client->auth_wq);
2183 	return;
2184 
2185 bad:
2186 	pr_err("osdc handle_map corrupt msg\n");
2187 	ceph_msg_dump(msg);
2188 	up_write(&osdc->map_sem);
2189 }
2190 
2191 /*
2192  * watch/notify callback event infrastructure
2193  *
2194  * These callbacks are used both for watch and notify operations.
2195  */
2196 static void __release_event(struct kref *kref)
2197 {
2198 	struct ceph_osd_event *event =
2199 		container_of(kref, struct ceph_osd_event, kref);
2200 
2201 	dout("__release_event %p\n", event);
2202 	kfree(event);
2203 }
2204 
2205 static void get_event(struct ceph_osd_event *event)
2206 {
2207 	kref_get(&event->kref);
2208 }
2209 
2210 void ceph_osdc_put_event(struct ceph_osd_event *event)
2211 {
2212 	kref_put(&event->kref, __release_event);
2213 }
2214 EXPORT_SYMBOL(ceph_osdc_put_event);
2215 
2216 static void __insert_event(struct ceph_osd_client *osdc,
2217 			     struct ceph_osd_event *new)
2218 {
2219 	struct rb_node **p = &osdc->event_tree.rb_node;
2220 	struct rb_node *parent = NULL;
2221 	struct ceph_osd_event *event = NULL;
2222 
2223 	while (*p) {
2224 		parent = *p;
2225 		event = rb_entry(parent, struct ceph_osd_event, node);
2226 		if (new->cookie < event->cookie)
2227 			p = &(*p)->rb_left;
2228 		else if (new->cookie > event->cookie)
2229 			p = &(*p)->rb_right;
2230 		else
2231 			BUG();
2232 	}
2233 
2234 	rb_link_node(&new->node, parent, p);
2235 	rb_insert_color(&new->node, &osdc->event_tree);
2236 }
2237 
2238 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2239 					        u64 cookie)
2240 {
2241 	struct rb_node **p = &osdc->event_tree.rb_node;
2242 	struct rb_node *parent = NULL;
2243 	struct ceph_osd_event *event = NULL;
2244 
2245 	while (*p) {
2246 		parent = *p;
2247 		event = rb_entry(parent, struct ceph_osd_event, node);
2248 		if (cookie < event->cookie)
2249 			p = &(*p)->rb_left;
2250 		else if (cookie > event->cookie)
2251 			p = &(*p)->rb_right;
2252 		else
2253 			return event;
2254 	}
2255 	return NULL;
2256 }
2257 
2258 static void __remove_event(struct ceph_osd_event *event)
2259 {
2260 	struct ceph_osd_client *osdc = event->osdc;
2261 
2262 	if (!RB_EMPTY_NODE(&event->node)) {
2263 		dout("__remove_event removed %p\n", event);
2264 		rb_erase(&event->node, &osdc->event_tree);
2265 		ceph_osdc_put_event(event);
2266 	} else {
2267 		dout("__remove_event didn't remove %p\n", event);
2268 	}
2269 }
2270 
2271 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2272 			   void (*event_cb)(u64, u64, u8, void *),
2273 			   void *data, struct ceph_osd_event **pevent)
2274 {
2275 	struct ceph_osd_event *event;
2276 
2277 	event = kmalloc(sizeof(*event), GFP_NOIO);
2278 	if (!event)
2279 		return -ENOMEM;
2280 
2281 	dout("create_event %p\n", event);
2282 	event->cb = event_cb;
2283 	event->one_shot = 0;
2284 	event->data = data;
2285 	event->osdc = osdc;
2286 	INIT_LIST_HEAD(&event->osd_node);
2287 	RB_CLEAR_NODE(&event->node);
2288 	kref_init(&event->kref);   /* one ref for us */
2289 	kref_get(&event->kref);    /* one ref for the caller */
2290 
2291 	spin_lock(&osdc->event_lock);
2292 	event->cookie = ++osdc->event_count;
2293 	__insert_event(osdc, event);
2294 	spin_unlock(&osdc->event_lock);
2295 
2296 	*pevent = event;
2297 	return 0;
2298 }
2299 EXPORT_SYMBOL(ceph_osdc_create_event);
2300 
2301 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2302 {
2303 	struct ceph_osd_client *osdc = event->osdc;
2304 
2305 	dout("cancel_event %p\n", event);
2306 	spin_lock(&osdc->event_lock);
2307 	__remove_event(event);
2308 	spin_unlock(&osdc->event_lock);
2309 	ceph_osdc_put_event(event); /* caller's */
2310 }
2311 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2312 
2313 
2314 static void do_event_work(struct work_struct *work)
2315 {
2316 	struct ceph_osd_event_work *event_work =
2317 		container_of(work, struct ceph_osd_event_work, work);
2318 	struct ceph_osd_event *event = event_work->event;
2319 	u64 ver = event_work->ver;
2320 	u64 notify_id = event_work->notify_id;
2321 	u8 opcode = event_work->opcode;
2322 
2323 	dout("do_event_work completing %p\n", event);
2324 	event->cb(ver, notify_id, opcode, event->data);
2325 	dout("do_event_work completed %p\n", event);
2326 	ceph_osdc_put_event(event);
2327 	kfree(event_work);
2328 }
2329 
2330 
2331 /*
2332  * Process osd watch notifications
2333  */
2334 static void handle_watch_notify(struct ceph_osd_client *osdc,
2335 				struct ceph_msg *msg)
2336 {
2337 	void *p, *end;
2338 	u8 proto_ver;
2339 	u64 cookie, ver, notify_id;
2340 	u8 opcode;
2341 	struct ceph_osd_event *event;
2342 	struct ceph_osd_event_work *event_work;
2343 
2344 	p = msg->front.iov_base;
2345 	end = p + msg->front.iov_len;
2346 
2347 	ceph_decode_8_safe(&p, end, proto_ver, bad);
2348 	ceph_decode_8_safe(&p, end, opcode, bad);
2349 	ceph_decode_64_safe(&p, end, cookie, bad);
2350 	ceph_decode_64_safe(&p, end, ver, bad);
2351 	ceph_decode_64_safe(&p, end, notify_id, bad);
2352 
2353 	spin_lock(&osdc->event_lock);
2354 	event = __find_event(osdc, cookie);
2355 	if (event) {
2356 		BUG_ON(event->one_shot);
2357 		get_event(event);
2358 	}
2359 	spin_unlock(&osdc->event_lock);
2360 	dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2361 	     cookie, ver, event);
2362 	if (event) {
2363 		event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2364 		if (!event_work) {
2365 			pr_err("couldn't allocate event_work\n");
2366 			ceph_osdc_put_event(event);
2367 			return;
2368 		}
2369 		INIT_WORK(&event_work->work, do_event_work);
2370 		event_work->event = event;
2371 		event_work->ver = ver;
2372 		event_work->notify_id = notify_id;
2373 		event_work->opcode = opcode;
2374 
2375 		queue_work(osdc->notify_wq, &event_work->work);
2376 	}
2377 
2378 	return;
2379 
2380 bad:
2381 	pr_err("osdc handle_watch_notify corrupt msg\n");
2382 }
2383 
2384 /*
2385  * build new request AND message
2386  *
2387  */
2388 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2389 				struct ceph_snap_context *snapc, u64 snap_id,
2390 				struct timespec *mtime)
2391 {
2392 	struct ceph_msg *msg = req->r_request;
2393 	void *p;
2394 	size_t msg_size;
2395 	int flags = req->r_flags;
2396 	u64 data_len;
2397 	unsigned int i;
2398 
2399 	req->r_snapid = snap_id;
2400 	req->r_snapc = ceph_get_snap_context(snapc);
2401 
2402 	/* encode request */
2403 	msg->hdr.version = cpu_to_le16(4);
2404 
2405 	p = msg->front.iov_base;
2406 	ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2407 	req->r_request_osdmap_epoch = p;
2408 	p += 4;
2409 	req->r_request_flags = p;
2410 	p += 4;
2411 	if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2412 		ceph_encode_timespec(p, mtime);
2413 	p += sizeof(struct ceph_timespec);
2414 	req->r_request_reassert_version = p;
2415 	p += sizeof(struct ceph_eversion); /* will get filled in */
2416 
2417 	/* oloc */
2418 	ceph_encode_8(&p, 4);
2419 	ceph_encode_8(&p, 4);
2420 	ceph_encode_32(&p, 8 + 4 + 4);
2421 	req->r_request_pool = p;
2422 	p += 8;
2423 	ceph_encode_32(&p, -1);  /* preferred */
2424 	ceph_encode_32(&p, 0);   /* key len */
2425 
2426 	ceph_encode_8(&p, 1);
2427 	req->r_request_pgid = p;
2428 	p += 8 + 4;
2429 	ceph_encode_32(&p, -1);  /* preferred */
2430 
2431 	/* oid */
2432 	ceph_encode_32(&p, req->r_base_oid.name_len);
2433 	memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2434 	dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2435 	     req->r_base_oid.name, req->r_base_oid.name_len);
2436 	p += req->r_base_oid.name_len;
2437 
2438 	/* ops--can imply data */
2439 	ceph_encode_16(&p, (u16)req->r_num_ops);
2440 	data_len = 0;
2441 	for (i = 0; i < req->r_num_ops; i++) {
2442 		data_len += osd_req_encode_op(req, p, i);
2443 		p += sizeof(struct ceph_osd_op);
2444 	}
2445 
2446 	/* snaps */
2447 	ceph_encode_64(&p, req->r_snapid);
2448 	ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2449 	ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2450 	if (req->r_snapc) {
2451 		for (i = 0; i < snapc->num_snaps; i++) {
2452 			ceph_encode_64(&p, req->r_snapc->snaps[i]);
2453 		}
2454 	}
2455 
2456 	req->r_request_attempts = p;
2457 	p += 4;
2458 
2459 	/* data */
2460 	if (flags & CEPH_OSD_FLAG_WRITE) {
2461 		u16 data_off;
2462 
2463 		/*
2464 		 * The header "data_off" is a hint to the receiver
2465 		 * allowing it to align received data into its
2466 		 * buffers such that there's no need to re-copy
2467 		 * it before writing it to disk (direct I/O).
2468 		 */
2469 		data_off = (u16) (off & 0xffff);
2470 		req->r_request->hdr.data_off = cpu_to_le16(data_off);
2471 	}
2472 	req->r_request->hdr.data_len = cpu_to_le32(data_len);
2473 
2474 	BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2475 	msg_size = p - msg->front.iov_base;
2476 	msg->front.iov_len = msg_size;
2477 	msg->hdr.front_len = cpu_to_le32(msg_size);
2478 
2479 	dout("build_request msg_size was %d\n", (int)msg_size);
2480 }
2481 EXPORT_SYMBOL(ceph_osdc_build_request);
2482 
2483 /*
2484  * Register request, send initial attempt.
2485  */
2486 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2487 			    struct ceph_osd_request *req,
2488 			    bool nofail)
2489 {
2490 	int rc;
2491 
2492 	down_read(&osdc->map_sem);
2493 	mutex_lock(&osdc->request_mutex);
2494 
2495 	rc = __ceph_osdc_start_request(osdc, req, nofail);
2496 
2497 	mutex_unlock(&osdc->request_mutex);
2498 	up_read(&osdc->map_sem);
2499 
2500 	return rc;
2501 }
2502 EXPORT_SYMBOL(ceph_osdc_start_request);
2503 
2504 /*
2505  * Unregister a registered request.  The request is not completed (i.e.
2506  * no callbacks or wakeups) - higher layers are supposed to know what
2507  * they are canceling.
2508  */
2509 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2510 {
2511 	struct ceph_osd_client *osdc = req->r_osdc;
2512 
2513 	mutex_lock(&osdc->request_mutex);
2514 	if (req->r_linger)
2515 		__unregister_linger_request(osdc, req);
2516 	__unregister_request(osdc, req);
2517 	mutex_unlock(&osdc->request_mutex);
2518 
2519 	dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2520 }
2521 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2522 
2523 /*
2524  * wait for a request to complete
2525  */
2526 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2527 			   struct ceph_osd_request *req)
2528 {
2529 	int rc;
2530 
2531 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2532 
2533 	rc = wait_for_completion_interruptible(&req->r_completion);
2534 	if (rc < 0) {
2535 		dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2536 		ceph_osdc_cancel_request(req);
2537 		complete_request(req);
2538 		return rc;
2539 	}
2540 
2541 	dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2542 	     req->r_result);
2543 	return req->r_result;
2544 }
2545 EXPORT_SYMBOL(ceph_osdc_wait_request);
2546 
2547 /*
2548  * sync - wait for all in-flight requests to flush.  avoid starvation.
2549  */
2550 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2551 {
2552 	struct ceph_osd_request *req;
2553 	u64 last_tid, next_tid = 0;
2554 
2555 	mutex_lock(&osdc->request_mutex);
2556 	last_tid = osdc->last_tid;
2557 	while (1) {
2558 		req = __lookup_request_ge(osdc, next_tid);
2559 		if (!req)
2560 			break;
2561 		if (req->r_tid > last_tid)
2562 			break;
2563 
2564 		next_tid = req->r_tid + 1;
2565 		if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2566 			continue;
2567 
2568 		ceph_osdc_get_request(req);
2569 		mutex_unlock(&osdc->request_mutex);
2570 		dout("sync waiting on tid %llu (last is %llu)\n",
2571 		     req->r_tid, last_tid);
2572 		wait_for_completion(&req->r_safe_completion);
2573 		mutex_lock(&osdc->request_mutex);
2574 		ceph_osdc_put_request(req);
2575 	}
2576 	mutex_unlock(&osdc->request_mutex);
2577 	dout("sync done (thru tid %llu)\n", last_tid);
2578 }
2579 EXPORT_SYMBOL(ceph_osdc_sync);
2580 
2581 /*
2582  * Call all pending notify callbacks - for use after a watch is
2583  * unregistered, to make sure no more callbacks for it will be invoked
2584  */
2585 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2586 {
2587 	flush_workqueue(osdc->notify_wq);
2588 }
2589 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2590 
2591 
2592 /*
2593  * init, shutdown
2594  */
2595 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2596 {
2597 	int err;
2598 
2599 	dout("init\n");
2600 	osdc->client = client;
2601 	osdc->osdmap = NULL;
2602 	init_rwsem(&osdc->map_sem);
2603 	init_completion(&osdc->map_waiters);
2604 	osdc->last_requested_map = 0;
2605 	mutex_init(&osdc->request_mutex);
2606 	osdc->last_tid = 0;
2607 	osdc->osds = RB_ROOT;
2608 	INIT_LIST_HEAD(&osdc->osd_lru);
2609 	osdc->requests = RB_ROOT;
2610 	INIT_LIST_HEAD(&osdc->req_lru);
2611 	INIT_LIST_HEAD(&osdc->req_unsent);
2612 	INIT_LIST_HEAD(&osdc->req_notarget);
2613 	INIT_LIST_HEAD(&osdc->req_linger);
2614 	osdc->num_requests = 0;
2615 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2616 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2617 	spin_lock_init(&osdc->event_lock);
2618 	osdc->event_tree = RB_ROOT;
2619 	osdc->event_count = 0;
2620 
2621 	schedule_delayed_work(&osdc->osds_timeout_work,
2622 	   round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2623 
2624 	err = -ENOMEM;
2625 	osdc->req_mempool = mempool_create_kmalloc_pool(10,
2626 					sizeof(struct ceph_osd_request));
2627 	if (!osdc->req_mempool)
2628 		goto out;
2629 
2630 	err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2631 				OSD_OP_FRONT_LEN, 10, true,
2632 				"osd_op");
2633 	if (err < 0)
2634 		goto out_mempool;
2635 	err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2636 				OSD_OPREPLY_FRONT_LEN, 10, true,
2637 				"osd_op_reply");
2638 	if (err < 0)
2639 		goto out_msgpool;
2640 
2641 	err = -ENOMEM;
2642 	osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2643 	if (!osdc->notify_wq)
2644 		goto out_msgpool_reply;
2645 
2646 	return 0;
2647 
2648 out_msgpool_reply:
2649 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2650 out_msgpool:
2651 	ceph_msgpool_destroy(&osdc->msgpool_op);
2652 out_mempool:
2653 	mempool_destroy(osdc->req_mempool);
2654 out:
2655 	return err;
2656 }
2657 
2658 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2659 {
2660 	flush_workqueue(osdc->notify_wq);
2661 	destroy_workqueue(osdc->notify_wq);
2662 	cancel_delayed_work_sync(&osdc->timeout_work);
2663 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
2664 	if (osdc->osdmap) {
2665 		ceph_osdmap_destroy(osdc->osdmap);
2666 		osdc->osdmap = NULL;
2667 	}
2668 	remove_all_osds(osdc);
2669 	mempool_destroy(osdc->req_mempool);
2670 	ceph_msgpool_destroy(&osdc->msgpool_op);
2671 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2672 }
2673 
2674 /*
2675  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2676  * *plen.  Return number of bytes read, or error.
2677  */
2678 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2679 			struct ceph_vino vino, struct ceph_file_layout *layout,
2680 			u64 off, u64 *plen,
2681 			u32 truncate_seq, u64 truncate_size,
2682 			struct page **pages, int num_pages, int page_align)
2683 {
2684 	struct ceph_osd_request *req;
2685 	int rc = 0;
2686 
2687 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2688 	     vino.snap, off, *plen);
2689 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2690 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2691 				    NULL, truncate_seq, truncate_size,
2692 				    false);
2693 	if (IS_ERR(req))
2694 		return PTR_ERR(req);
2695 
2696 	/* it may be a short read due to an object boundary */
2697 
2698 	osd_req_op_extent_osd_data_pages(req, 0,
2699 				pages, *plen, page_align, false, false);
2700 
2701 	dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2702 	     off, *plen, *plen, page_align);
2703 
2704 	ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2705 
2706 	rc = ceph_osdc_start_request(osdc, req, false);
2707 	if (!rc)
2708 		rc = ceph_osdc_wait_request(osdc, req);
2709 
2710 	ceph_osdc_put_request(req);
2711 	dout("readpages result %d\n", rc);
2712 	return rc;
2713 }
2714 EXPORT_SYMBOL(ceph_osdc_readpages);
2715 
2716 /*
2717  * do a synchronous write on N pages
2718  */
2719 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2720 			 struct ceph_file_layout *layout,
2721 			 struct ceph_snap_context *snapc,
2722 			 u64 off, u64 len,
2723 			 u32 truncate_seq, u64 truncate_size,
2724 			 struct timespec *mtime,
2725 			 struct page **pages, int num_pages)
2726 {
2727 	struct ceph_osd_request *req;
2728 	int rc = 0;
2729 	int page_align = off & ~PAGE_MASK;
2730 
2731 	BUG_ON(vino.snap != CEPH_NOSNAP);	/* snapshots aren't writeable */
2732 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2733 				    CEPH_OSD_OP_WRITE,
2734 				    CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2735 				    snapc, truncate_seq, truncate_size,
2736 				    true);
2737 	if (IS_ERR(req))
2738 		return PTR_ERR(req);
2739 
2740 	/* it may be a short write due to an object boundary */
2741 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2742 				false, false);
2743 	dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2744 
2745 	ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2746 
2747 	rc = ceph_osdc_start_request(osdc, req, true);
2748 	if (!rc)
2749 		rc = ceph_osdc_wait_request(osdc, req);
2750 
2751 	ceph_osdc_put_request(req);
2752 	if (rc == 0)
2753 		rc = len;
2754 	dout("writepages result %d\n", rc);
2755 	return rc;
2756 }
2757 EXPORT_SYMBOL(ceph_osdc_writepages);
2758 
2759 int ceph_osdc_setup(void)
2760 {
2761 	BUG_ON(ceph_osd_request_cache);
2762 	ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2763 					sizeof (struct ceph_osd_request),
2764 					__alignof__(struct ceph_osd_request),
2765 					0, NULL);
2766 
2767 	return ceph_osd_request_cache ? 0 : -ENOMEM;
2768 }
2769 EXPORT_SYMBOL(ceph_osdc_setup);
2770 
2771 void ceph_osdc_cleanup(void)
2772 {
2773 	BUG_ON(!ceph_osd_request_cache);
2774 	kmem_cache_destroy(ceph_osd_request_cache);
2775 	ceph_osd_request_cache = NULL;
2776 }
2777 EXPORT_SYMBOL(ceph_osdc_cleanup);
2778 
2779 /*
2780  * handle incoming message
2781  */
2782 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2783 {
2784 	struct ceph_osd *osd = con->private;
2785 	struct ceph_osd_client *osdc;
2786 	int type = le16_to_cpu(msg->hdr.type);
2787 
2788 	if (!osd)
2789 		goto out;
2790 	osdc = osd->o_osdc;
2791 
2792 	switch (type) {
2793 	case CEPH_MSG_OSD_MAP:
2794 		ceph_osdc_handle_map(osdc, msg);
2795 		break;
2796 	case CEPH_MSG_OSD_OPREPLY:
2797 		handle_reply(osdc, msg, con);
2798 		break;
2799 	case CEPH_MSG_WATCH_NOTIFY:
2800 		handle_watch_notify(osdc, msg);
2801 		break;
2802 
2803 	default:
2804 		pr_err("received unknown message type %d %s\n", type,
2805 		       ceph_msg_type_name(type));
2806 	}
2807 out:
2808 	ceph_msg_put(msg);
2809 }
2810 
2811 /*
2812  * lookup and return message for incoming reply.  set up reply message
2813  * pages.
2814  */
2815 static struct ceph_msg *get_reply(struct ceph_connection *con,
2816 				  struct ceph_msg_header *hdr,
2817 				  int *skip)
2818 {
2819 	struct ceph_osd *osd = con->private;
2820 	struct ceph_osd_client *osdc = osd->o_osdc;
2821 	struct ceph_msg *m;
2822 	struct ceph_osd_request *req;
2823 	int front_len = le32_to_cpu(hdr->front_len);
2824 	int data_len = le32_to_cpu(hdr->data_len);
2825 	u64 tid;
2826 
2827 	tid = le64_to_cpu(hdr->tid);
2828 	mutex_lock(&osdc->request_mutex);
2829 	req = __lookup_request(osdc, tid);
2830 	if (!req) {
2831 		*skip = 1;
2832 		m = NULL;
2833 		dout("get_reply unknown tid %llu from osd%d\n", tid,
2834 		     osd->o_osd);
2835 		goto out;
2836 	}
2837 
2838 	if (req->r_reply->con)
2839 		dout("%s revoking msg %p from old con %p\n", __func__,
2840 		     req->r_reply, req->r_reply->con);
2841 	ceph_msg_revoke_incoming(req->r_reply);
2842 
2843 	if (front_len > req->r_reply->front_alloc_len) {
2844 		pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2845 			front_len, req->r_reply->front_alloc_len,
2846 			(unsigned int)con->peer_name.type,
2847 			le64_to_cpu(con->peer_name.num));
2848 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2849 				 false);
2850 		if (!m)
2851 			goto out;
2852 		ceph_msg_put(req->r_reply);
2853 		req->r_reply = m;
2854 	}
2855 	m = ceph_msg_get(req->r_reply);
2856 
2857 	if (data_len > 0) {
2858 		struct ceph_osd_data *osd_data;
2859 
2860 		/*
2861 		 * XXX This is assuming there is only one op containing
2862 		 * XXX page data.  Probably OK for reads, but this
2863 		 * XXX ought to be done more generally.
2864 		 */
2865 		osd_data = osd_req_op_extent_osd_data(req, 0);
2866 		if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2867 			if (osd_data->pages &&
2868 				unlikely(osd_data->length < data_len)) {
2869 
2870 				pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2871 					tid, data_len, osd_data->length);
2872 				*skip = 1;
2873 				ceph_msg_put(m);
2874 				m = NULL;
2875 				goto out;
2876 			}
2877 		}
2878 	}
2879 	*skip = 0;
2880 	dout("get_reply tid %lld %p\n", tid, m);
2881 
2882 out:
2883 	mutex_unlock(&osdc->request_mutex);
2884 	return m;
2885 
2886 }
2887 
2888 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2889 				  struct ceph_msg_header *hdr,
2890 				  int *skip)
2891 {
2892 	struct ceph_osd *osd = con->private;
2893 	int type = le16_to_cpu(hdr->type);
2894 	int front = le32_to_cpu(hdr->front_len);
2895 
2896 	*skip = 0;
2897 	switch (type) {
2898 	case CEPH_MSG_OSD_MAP:
2899 	case CEPH_MSG_WATCH_NOTIFY:
2900 		return ceph_msg_new(type, front, GFP_NOFS, false);
2901 	case CEPH_MSG_OSD_OPREPLY:
2902 		return get_reply(con, hdr, skip);
2903 	default:
2904 		pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2905 			osd->o_osd);
2906 		*skip = 1;
2907 		return NULL;
2908 	}
2909 }
2910 
2911 /*
2912  * Wrappers to refcount containing ceph_osd struct
2913  */
2914 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2915 {
2916 	struct ceph_osd *osd = con->private;
2917 	if (get_osd(osd))
2918 		return con;
2919 	return NULL;
2920 }
2921 
2922 static void put_osd_con(struct ceph_connection *con)
2923 {
2924 	struct ceph_osd *osd = con->private;
2925 	put_osd(osd);
2926 }
2927 
2928 /*
2929  * authentication
2930  */
2931 /*
2932  * Note: returned pointer is the address of a structure that's
2933  * managed separately.  Caller must *not* attempt to free it.
2934  */
2935 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2936 					int *proto, int force_new)
2937 {
2938 	struct ceph_osd *o = con->private;
2939 	struct ceph_osd_client *osdc = o->o_osdc;
2940 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2941 	struct ceph_auth_handshake *auth = &o->o_auth;
2942 
2943 	if (force_new && auth->authorizer) {
2944 		ceph_auth_destroy_authorizer(ac, auth->authorizer);
2945 		auth->authorizer = NULL;
2946 	}
2947 	if (!auth->authorizer) {
2948 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2949 						      auth);
2950 		if (ret)
2951 			return ERR_PTR(ret);
2952 	} else {
2953 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2954 						     auth);
2955 		if (ret)
2956 			return ERR_PTR(ret);
2957 	}
2958 	*proto = ac->protocol;
2959 
2960 	return auth;
2961 }
2962 
2963 
2964 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2965 {
2966 	struct ceph_osd *o = con->private;
2967 	struct ceph_osd_client *osdc = o->o_osdc;
2968 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2969 
2970 	return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2971 }
2972 
2973 static int invalidate_authorizer(struct ceph_connection *con)
2974 {
2975 	struct ceph_osd *o = con->private;
2976 	struct ceph_osd_client *osdc = o->o_osdc;
2977 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2978 
2979 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2980 	return ceph_monc_validate_auth(&osdc->client->monc);
2981 }
2982 
2983 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
2984 {
2985 	struct ceph_osd *o = con->private;
2986 	struct ceph_auth_handshake *auth = &o->o_auth;
2987 	return ceph_auth_sign_message(auth, msg);
2988 }
2989 
2990 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
2991 {
2992 	struct ceph_osd *o = con->private;
2993 	struct ceph_auth_handshake *auth = &o->o_auth;
2994 	return ceph_auth_check_message_signature(auth, msg);
2995 }
2996 
2997 static const struct ceph_connection_operations osd_con_ops = {
2998 	.get = get_osd_con,
2999 	.put = put_osd_con,
3000 	.dispatch = dispatch,
3001 	.get_authorizer = get_authorizer,
3002 	.verify_authorizer_reply = verify_authorizer_reply,
3003 	.invalidate_authorizer = invalidate_authorizer,
3004 	.alloc_msg = alloc_msg,
3005 	.sign_message = sign_message,
3006 	.check_message_signature = check_message_signature,
3007 	.fault = osd_reset,
3008 };
3009