net/ceph/osd_client.c

   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         struct ceph_osd_request *__oreq = (oreq);                       \
 126         unsigned int __whch = (whch);                                   \
 127         BUG_ON(__whch >= __oreq->r_num_ops);                            \
 128         &__oreq->r_ops[__whch].typ.fld;                                 \
 129 })
 130
 131 static struct ceph_osd_data *
 132 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
 133 {
 134         BUG_ON(which >= osd_req->r_num_ops);
 135
 136         return &osd_req->r_ops[which].raw_data_in;
 137 }
 138
 139 struct ceph_osd_data *
 140 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
 141                         unsigned int which)
 142 {
 143         return osd_req_op_data(osd_req, which, extent, osd_data);
 144 }
 145 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
 146
 147 struct ceph_osd_data *
 148 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
 149                         unsigned int which)
 150 {
 151         return osd_req_op_data(osd_req, which, cls, response_data);
 152 }
 153 EXPORT_SYMBOL(osd_req_op_cls_response_data);    /* ??? */
 154
 155 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
 156                         unsigned int which, struct page **pages,
 157                         u64 length, u32 alignment,
 158                         bool pages_from_pool, bool own_pages)
 159 {
 160         struct ceph_osd_data *osd_data;
 161
 162         osd_data = osd_req_op_raw_data_in(osd_req, which);
 163         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 164                                 pages_from_pool, own_pages);
 165 }
 166 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
 167
 168 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
 169                         unsigned int which, struct page **pages,
 170                         u64 length, u32 alignment,
 171                         bool pages_from_pool, bool own_pages)
 172 {
 173         struct ceph_osd_data *osd_data;
 174
 175         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 176         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 177                                 pages_from_pool, own_pages);
 178 }
 179 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
 180
 181 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
 182                         unsigned int which, struct ceph_pagelist *pagelist)
 183 {
 184         struct ceph_osd_data *osd_data;
 185
 186         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 187         ceph_osd_data_pagelist_init(osd_data, pagelist);
 188 }
 189 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
 190
 191 #ifdef CONFIG_BLOCK
 192 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
 193                         unsigned int which, struct bio *bio, size_t bio_length)
 194 {
 195         struct ceph_osd_data *osd_data;
 196
 197         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 198         ceph_osd_data_bio_init(osd_data, bio, bio_length);
 199 }
 200 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
 201 #endif /* CONFIG_BLOCK */
 202
 203 static void osd_req_op_cls_request_info_pagelist(
 204                         struct ceph_osd_request *osd_req,
 205                         unsigned int which, struct ceph_pagelist *pagelist)
 206 {
 207         struct ceph_osd_data *osd_data;
 208
 209         osd_data = osd_req_op_data(osd_req, which, cls, request_info);
 210         ceph_osd_data_pagelist_init(osd_data, pagelist);
 211 }
 212
 213 void osd_req_op_cls_request_data_pagelist(
 214                         struct ceph_osd_request *osd_req,
 215                         unsigned int which, struct ceph_pagelist *pagelist)
 216 {
 217         struct ceph_osd_data *osd_data;
 218
 219         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 220         ceph_osd_data_pagelist_init(osd_data, pagelist);
 221 }
 222 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
 223
 224 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
 225                         unsigned int which, struct page **pages, u64 length,
 226                         u32 alignment, bool pages_from_pool, bool own_pages)
 227 {
 228         struct ceph_osd_data *osd_data;
 229
 230         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 231         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 232                                 pages_from_pool, own_pages);
 233 }
 234 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
 235
 236 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
 237                         unsigned int which, struct page **pages, u64 length,
 238                         u32 alignment, bool pages_from_pool, bool own_pages)
 239 {
 240         struct ceph_osd_data *osd_data;
 241
 242         osd_data = osd_req_op_data(osd_req, which, cls, response_data);
 243         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 244                                 pages_from_pool, own_pages);
 245 }
 246 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
 247
 248 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
 249 {
 250         switch (osd_data->type) {
 251         case CEPH_OSD_DATA_TYPE_NONE:
 252                 return 0;
 253         case CEPH_OSD_DATA_TYPE_PAGES:
 254                 return osd_data->length;
 255         case CEPH_OSD_DATA_TYPE_PAGELIST:
 256                 return (u64)osd_data->pagelist->length;
 257 #ifdef CONFIG_BLOCK
 258         case CEPH_OSD_DATA_TYPE_BIO:
 259                 return (u64)osd_data->bio_length;
 260 #endif /* CONFIG_BLOCK */
 261         default:
 262                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
 263                 return 0;
 264         }
 265 }
 266
 267 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
 268 {
 269         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
 270                 int num_pages;
 271
 272                 num_pages = calc_pages_for((u64)osd_data->alignment,
 273                                                 (u64)osd_data->length);
 274                 ceph_release_page_vector(osd_data->pages, num_pages);
 275         }
 276         ceph_osd_data_init(osd_data);
 277 }
 278
 279 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
 280                         unsigned int which)
 281 {
 282         struct ceph_osd_req_op *op;
 283
 284         BUG_ON(which >= osd_req->r_num_ops);
 285         op = &osd_req->r_ops[which];
 286
 287         switch (op->op) {
 288         case CEPH_OSD_OP_READ:
 289         case CEPH_OSD_OP_WRITE:
 290         case CEPH_OSD_OP_WRITEFULL:
 291                 ceph_osd_data_release(&op->extent.osd_data);
 292                 break;
 293         case CEPH_OSD_OP_CALL:
 294                 ceph_osd_data_release(&op->cls.request_info);
 295                 ceph_osd_data_release(&op->cls.request_data);
 296                 ceph_osd_data_release(&op->cls.response_data);
 297                 break;
 298         case CEPH_OSD_OP_SETXATTR:
 299         case CEPH_OSD_OP_CMPXATTR:
 300                 ceph_osd_data_release(&op->xattr.osd_data);
 301                 break;
 302         case CEPH_OSD_OP_STAT:
 303                 ceph_osd_data_release(&op->raw_data_in);
 304                 break;
 305         default:
 306                 break;
 307         }
 308 }
 309
 310 /*
 311  * requests
 312  */
 313 static void ceph_osdc_release_request(struct kref *kref)
 314 {
 315         struct ceph_osd_request *req = container_of(kref,
 316                                             struct ceph_osd_request, r_kref);
 317         unsigned int which;
 318
 319         dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
 320              req->r_request, req->r_reply);
 321         WARN_ON(!RB_EMPTY_NODE(&req->r_node));
 322         WARN_ON(!list_empty(&req->r_req_lru_item));
 323         WARN_ON(!list_empty(&req->r_osd_item));
 324         WARN_ON(!list_empty(&req->r_linger_item));
 325         WARN_ON(!list_empty(&req->r_linger_osd_item));
 326         WARN_ON(req->r_osd);
 327
 328         if (req->r_request)
 329                 ceph_msg_put(req->r_request);
 330         if (req->r_reply) {
 331                 ceph_msg_revoke_incoming(req->r_reply);
 332                 ceph_msg_put(req->r_reply);
 333         }
 334
 335         for (which = 0; which < req->r_num_ops; which++)
 336                 osd_req_op_data_release(req, which);
 337
 338         ceph_put_snap_context(req->r_snapc);
 339         if (req->r_mempool)
 340                 mempool_free(req, req->r_osdc->req_mempool);
 341         else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS)
 342                 kmem_cache_free(ceph_osd_request_cache, req);
 343         else
 344                 kfree(req);
 345 }
 346
 347 void ceph_osdc_get_request(struct ceph_osd_request *req)
 348 {
 349         dout("%s %p (was %d)\n", __func__, req,
 350              atomic_read(&req->r_kref.refcount));
 351         kref_get(&req->r_kref);
 352 }
 353 EXPORT_SYMBOL(ceph_osdc_get_request);
 354
 355 void ceph_osdc_put_request(struct ceph_osd_request *req)
 356 {
 357         dout("%s %p (was %d)\n", __func__, req,
 358              atomic_read(&req->r_kref.refcount));
 359         kref_put(&req->r_kref, ceph_osdc_release_request);
 360 }
 361 EXPORT_SYMBOL(ceph_osdc_put_request);
 362
 363 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 364                                                struct ceph_snap_context *snapc,
 365                                                unsigned int num_ops,
 366                                                bool use_mempool,
 367                                                gfp_t gfp_flags)
 368 {
 369         struct ceph_osd_request *req;
 370         struct ceph_msg *msg;
 371         size_t msg_size;
 372
 373         if (use_mempool) {
 374                 BUG_ON(num_ops > CEPH_OSD_SLAB_OPS);
 375                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 376         } else if (num_ops <= CEPH_OSD_SLAB_OPS) {
 377                 req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags);
 378         } else {
 379                 BUG_ON(num_ops > CEPH_OSD_MAX_OPS);
 380                 req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]),
 381                               gfp_flags);
 382         }
 383         if (unlikely(!req))
 384                 return NULL;
 385
 386         /* req only, each op is zeroed in _osd_req_op_init() */
 387         memset(req, 0, sizeof(*req));
 388
 389         req->r_osdc = osdc;
 390         req->r_mempool = use_mempool;
 391         req->r_num_ops = num_ops;
 392
 393         kref_init(&req->r_kref);
 394         init_completion(&req->r_completion);
 395         init_completion(&req->r_safe_completion);
 396         RB_CLEAR_NODE(&req->r_node);
 397         INIT_LIST_HEAD(&req->r_unsafe_item);
 398         INIT_LIST_HEAD(&req->r_linger_item);
 399         INIT_LIST_HEAD(&req->r_linger_osd_item);
 400         INIT_LIST_HEAD(&req->r_req_lru_item);
 401         INIT_LIST_HEAD(&req->r_osd_item);
 402
 403         req->r_base_oloc.pool = -1;
 404         req->r_target_oloc.pool = -1;
 405
 406         msg_size = OSD_OPREPLY_FRONT_LEN;
 407         if (num_ops > CEPH_OSD_SLAB_OPS) {
 408                 /* ceph_osd_op and rval */
 409                 msg_size += (num_ops - CEPH_OSD_SLAB_OPS) *
 410                             (sizeof(struct ceph_osd_op) + 4);
 411         }
 412
 413         /* create reply message */
 414         if (use_mempool)
 415                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 416         else
 417                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size,
 418                                    gfp_flags, true);
 419         if (!msg) {
 420                 ceph_osdc_put_request(req);
 421                 return NULL;
 422         }
 423         req->r_reply = msg;
 424
 425         msg_size = 4 + 4 + 4; /* client_inc, osdmap_epoch, flags */
 426         msg_size += 4 + 4 + 4 + 8; /* mtime, reassert_version */
 427         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
 428         msg_size += 1 + 8 + 4 + 4; /* pgid */
 429         msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
 430         msg_size += 2 + num_ops * sizeof(struct ceph_osd_op);
 431         msg_size += 8; /* snapid */
 432         msg_size += 8; /* snap_seq */
 433         msg_size += 4 + 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
 434         msg_size += 4; /* retry_attempt */
 435
 436         /* create request message; allow space for oid */
 437         if (use_mempool)
 438                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 439         else
 440                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 441         if (!msg) {
 442                 ceph_osdc_put_request(req);
 443                 return NULL;
 444         }
 445
 446         memset(msg->front.iov_base, 0, msg->front.iov_len);
 447
 448         req->r_request = msg;
 449
 450         return req;
 451 }
 452 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 453
 454 static bool osd_req_opcode_valid(u16 opcode)
 455 {
 456         switch (opcode) {
 457 #define GENERATE_CASE(op, opcode, str)  case CEPH_OSD_OP_##op: return true;
 458 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
 459 #undef GENERATE_CASE
 460         default:
 461                 return false;
 462         }
 463 }
 464
 465 /*
 466  * This is an osd op init function for opcodes that have no data or
 467  * other information associated with them.  It also serves as a
 468  * common init routine for all the other init functions, below.
 469  */
 470 static struct ceph_osd_req_op *
 471 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
 472                  u16 opcode, u32 flags)
 473 {
 474         struct ceph_osd_req_op *op;
 475
 476         BUG_ON(which >= osd_req->r_num_ops);
 477         BUG_ON(!osd_req_opcode_valid(opcode));
 478
 479         op = &osd_req->r_ops[which];
 480         memset(op, 0, sizeof (*op));
 481         op->op = opcode;
 482         op->flags = flags;
 483
 484         return op;
 485 }
 486
 487 void osd_req_op_init(struct ceph_osd_request *osd_req,
 488                      unsigned int which, u16 opcode, u32 flags)
 489 {
 490         (void)_osd_req_op_init(osd_req, which, opcode, flags);
 491 }
 492 EXPORT_SYMBOL(osd_req_op_init);
 493
 494 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 495                                 unsigned int which, u16 opcode,
 496                                 u64 offset, u64 length,
 497                                 u64 truncate_size, u32 truncate_seq)
 498 {
 499         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 500                                                       opcode, 0);
 501         size_t payload_len = 0;
 502
 503         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 504                opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO &&
 505                opcode != CEPH_OSD_OP_TRUNCATE);
 506
 507         op->extent.offset = offset;
 508         op->extent.length = length;
 509         op->extent.truncate_size = truncate_size;
 510         op->extent.truncate_seq = truncate_seq;
 511         if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL)
 512                 payload_len += length;
 513
 514         op->indata_len = payload_len;
 515 }
 516 EXPORT_SYMBOL(osd_req_op_extent_init);
 517
 518 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 519                                 unsigned int which, u64 length)
 520 {
 521         struct ceph_osd_req_op *op;
 522         u64 previous;
 523
 524         BUG_ON(which >= osd_req->r_num_ops);
 525         op = &osd_req->r_ops[which];
 526         previous = op->extent.length;
 527
 528         if (length == previous)
 529                 return;         /* Nothing to do */
 530         BUG_ON(length > previous);
 531
 532         op->extent.length = length;
 533         op->indata_len -= previous - length;
 534 }
 535 EXPORT_SYMBOL(osd_req_op_extent_update);
 536
 537 void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 538                                 unsigned int which, u64 offset_inc)
 539 {
 540         struct ceph_osd_req_op *op, *prev_op;
 541
 542         BUG_ON(which + 1 >= osd_req->r_num_ops);
 543
 544         prev_op = &osd_req->r_ops[which];
 545         op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags);
 546         /* dup previous one */
 547         op->indata_len = prev_op->indata_len;
 548         op->outdata_len = prev_op->outdata_len;
 549         op->extent = prev_op->extent;
 550         /* adjust offset */
 551         op->extent.offset += offset_inc;
 552         op->extent.length -= offset_inc;
 553
 554         if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
 555                 op->indata_len -= offset_inc;
 556 }
 557 EXPORT_SYMBOL(osd_req_op_extent_dup_last);
 558
 559 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
 560                         u16 opcode, const char *class, const char *method)
 561 {
 562         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 563                                                       opcode, 0);
 564         struct ceph_pagelist *pagelist;
 565         size_t payload_len = 0;
 566         size_t size;
 567
 568         BUG_ON(opcode != CEPH_OSD_OP_CALL);
 569
 570         pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
 571         BUG_ON(!pagelist);
 572         ceph_pagelist_init(pagelist);
 573
 574         op->cls.class_name = class;
 575         size = strlen(class);
 576         BUG_ON(size > (size_t) U8_MAX);
 577         op->cls.class_len = size;
 578         ceph_pagelist_append(pagelist, class, size);
 579         payload_len += size;
 580
 581         op->cls.method_name = method;
 582         size = strlen(method);
 583         BUG_ON(size > (size_t) U8_MAX);
 584         op->cls.method_len = size;
 585         ceph_pagelist_append(pagelist, method, size);
 586         payload_len += size;
 587
 588         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
 589
 590         op->cls.argc = 0;       /* currently unused */
 591
 592         op->indata_len = payload_len;
 593 }
 594 EXPORT_SYMBOL(osd_req_op_cls_init);
 595
 596 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 597                           u16 opcode, const char *name, const void *value,
 598                           size_t size, u8 cmp_op, u8 cmp_mode)
 599 {
 600         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 601                                                       opcode, 0);
 602         struct ceph_pagelist *pagelist;
 603         size_t payload_len;
 604
 605         BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
 606
 607         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
 608         if (!pagelist)
 609                 return -ENOMEM;
 610
 611         ceph_pagelist_init(pagelist);
 612
 613         payload_len = strlen(name);
 614         op->xattr.name_len = payload_len;
 615         ceph_pagelist_append(pagelist, name, payload_len);
 616
 617         op->xattr.value_len = size;
 618         ceph_pagelist_append(pagelist, value, size);
 619         payload_len += size;
 620
 621         op->xattr.cmp_op = cmp_op;
 622         op->xattr.cmp_mode = cmp_mode;
 623
 624         ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
 625         op->indata_len = payload_len;
 626         return 0;
 627 }
 628 EXPORT_SYMBOL(osd_req_op_xattr_init);
 629
 630 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
 631                                 unsigned int which, u16 opcode,
 632                                 u64 cookie, u64 version, int flag)
 633 {
 634         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 635                                                       opcode, 0);
 636
 637         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
 638
 639         op->watch.cookie = cookie;
 640         op->watch.ver = version;
 641         if (opcode == CEPH_OSD_OP_WATCH && flag)
 642                 op->watch.flag = (u8)1;
 643 }
 644 EXPORT_SYMBOL(osd_req_op_watch_init);
 645
 646 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 647                                 unsigned int which,
 648                                 u64 expected_object_size,
 649                                 u64 expected_write_size)
 650 {
 651         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 652                                                       CEPH_OSD_OP_SETALLOCHINT,
 653                                                       0);
 654
 655         op->alloc_hint.expected_object_size = expected_object_size;
 656         op->alloc_hint.expected_write_size = expected_write_size;
 657
 658         /*
 659          * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
 660          * not worth a feature bit.  Set FAILOK per-op flag to make
 661          * sure older osds don't trip over an unsupported opcode.
 662          */
 663         op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
 664 }
 665 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
 666
 667 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
 668                                 struct ceph_osd_data *osd_data)
 669 {
 670         u64 length = ceph_osd_data_length(osd_data);
 671
 672         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 673                 BUG_ON(length > (u64) SIZE_MAX);
 674                 if (length)
 675                         ceph_msg_data_add_pages(msg, osd_data->pages,
 676                                         length, osd_data->alignment);
 677         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 678                 BUG_ON(!length);
 679                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
 680 #ifdef CONFIG_BLOCK
 681         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
 682                 ceph_msg_data_add_bio(msg, osd_data->bio, length);
 683 #endif
 684         } else {
 685                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
 686         }
 687 }
 688
 689 static u64 osd_req_encode_op(struct ceph_osd_request *req,
 690                               struct ceph_osd_op *dst, unsigned int which)
 691 {
 692         struct ceph_osd_req_op *src;
 693         struct ceph_osd_data *osd_data;
 694         u64 request_data_len = 0;
 695         u64 data_length;
 696
 697         BUG_ON(which >= req->r_num_ops);
 698         src = &req->r_ops[which];
 699         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
 700                 pr_err("unrecognized osd opcode %d\n", src->op);
 701
 702                 return 0;
 703         }
 704
 705         switch (src->op) {
 706         case CEPH_OSD_OP_STAT:
 707                 osd_data = &src->raw_data_in;
 708                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 709                 break;
 710         case CEPH_OSD_OP_READ:
 711         case CEPH_OSD_OP_WRITE:
 712         case CEPH_OSD_OP_WRITEFULL:
 713         case CEPH_OSD_OP_ZERO:
 714         case CEPH_OSD_OP_TRUNCATE:
 715                 if (src->op == CEPH_OSD_OP_WRITE ||
 716                     src->op == CEPH_OSD_OP_WRITEFULL)
 717                         request_data_len = src->extent.length;
 718                 dst->extent.offset = cpu_to_le64(src->extent.offset);
 719                 dst->extent.length = cpu_to_le64(src->extent.length);
 720                 dst->extent.truncate_size =
 721                         cpu_to_le64(src->extent.truncate_size);
 722                 dst->extent.truncate_seq =
 723                         cpu_to_le32(src->extent.truncate_seq);
 724                 osd_data = &src->extent.osd_data;
 725                 if (src->op == CEPH_OSD_OP_WRITE ||
 726                     src->op == CEPH_OSD_OP_WRITEFULL)
 727                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 728                 else
 729                         ceph_osdc_msg_data_add(req->r_reply, osd_data);
 730                 break;
 731         case CEPH_OSD_OP_CALL:
 732                 dst->cls.class_len = src->cls.class_len;
 733                 dst->cls.method_len = src->cls.method_len;
 734                 osd_data = &src->cls.request_info;
 735                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 736                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
 737                 request_data_len = osd_data->pagelist->length;
 738
 739                 osd_data = &src->cls.request_data;
 740                 data_length = ceph_osd_data_length(osd_data);
 741                 if (data_length) {
 742                         BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
 743                         dst->cls.indata_len = cpu_to_le32(data_length);
 744                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 745                         src->indata_len += data_length;
 746                         request_data_len += data_length;
 747                 }
 748                 osd_data = &src->cls.response_data;
 749                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 750                 break;
 751         case CEPH_OSD_OP_STARTSYNC:
 752                 break;
 753         case CEPH_OSD_OP_NOTIFY_ACK:
 754         case CEPH_OSD_OP_WATCH:
 755                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 756                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 757                 dst->watch.flag = src->watch.flag;
 758                 break;
 759         case CEPH_OSD_OP_SETALLOCHINT:
 760                 dst->alloc_hint.expected_object_size =
 761                     cpu_to_le64(src->alloc_hint.expected_object_size);
 762                 dst->alloc_hint.expected_write_size =
 763                     cpu_to_le64(src->alloc_hint.expected_write_size);
 764                 break;
 765         case CEPH_OSD_OP_SETXATTR:
 766         case CEPH_OSD_OP_CMPXATTR:
 767                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 768                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 769                 dst->xattr.cmp_op = src->xattr.cmp_op;
 770                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 771                 osd_data = &src->xattr.osd_data;
 772                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 773                 request_data_len = osd_data->pagelist->length;
 774                 break;
 775         case CEPH_OSD_OP_CREATE:
 776         case CEPH_OSD_OP_DELETE:
 777                 break;
 778         default:
 779                 pr_err("unsupported osd opcode %s\n",
 780                         ceph_osd_op_name(src->op));
 781                 WARN_ON(1);
 782
 783                 return 0;
 784         }
 785
 786         dst->op = cpu_to_le16(src->op);
 787         dst->flags = cpu_to_le32(src->flags);
 788         dst->payload_len = cpu_to_le32(src->indata_len);
 789
 790         return request_data_len;
 791 }
 792
 793 /*
 794  * build new request AND message, calculate layout, and adjust file
 795  * extent as needed.
 796  *
 797  * if the file was recently truncated, we include information about its
 798  * old and new size so that the object can be updated appropriately.  (we
 799  * avoid synchronously deleting truncated objects because it's slow.)
 800  *
 801  * if @do_sync, include a 'startsync' command so that the osd will flush
 802  * data quickly.
 803  */
 804 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 805                                                struct ceph_file_layout *layout,
 806                                                struct ceph_vino vino,
 807                                                u64 off, u64 *plen,
 808                                                unsigned int which, int num_ops,
 809                                                int opcode, int flags,
 810                                                struct ceph_snap_context *snapc,
 811                                                u32 truncate_seq,
 812                                                u64 truncate_size,
 813                                                bool use_mempool)
 814 {
 815         struct ceph_osd_request *req;
 816         u64 objnum = 0;
 817         u64 objoff = 0;
 818         u64 objlen = 0;
 819         int r;
 820
 821         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 822                opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
 823                opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
 824
 825         req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
 826                                         GFP_NOFS);
 827         if (!req)
 828                 return ERR_PTR(-ENOMEM);
 829
 830         req->r_flags = flags;
 831
 832         /* calculate max write size */
 833         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
 834         if (r < 0) {
 835                 ceph_osdc_put_request(req);
 836                 return ERR_PTR(r);
 837         }
 838
 839         if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
 840                 osd_req_op_init(req, which, opcode, 0);
 841         } else {
 842                 u32 object_size = le32_to_cpu(layout->fl_object_size);
 843                 u32 object_base = off - objoff;
 844                 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
 845                         if (truncate_size <= object_base) {
 846                                 truncate_size = 0;
 847                         } else {
 848                                 truncate_size -= object_base;
 849                                 if (truncate_size > object_size)
 850                                         truncate_size = object_size;
 851                         }
 852                 }
 853                 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
 854                                        truncate_size, truncate_seq);
 855         }
 856
 857         req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
 858
 859         snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
 860                  "%llx.%08llx", vino.ino, objnum);
 861         req->r_base_oid.name_len = strlen(req->r_base_oid.name);
 862
 863         return req;
 864 }
 865 EXPORT_SYMBOL(ceph_osdc_new_request);
 866
 867 /*
 868  * We keep osd requests in an rbtree, sorted by ->r_tid.
 869  */
 870 static void __insert_request(struct ceph_osd_client *osdc,
 871                              struct ceph_osd_request *new)
 872 {
 873         struct rb_node **p = &osdc->requests.rb_node;
 874         struct rb_node *parent = NULL;
 875         struct ceph_osd_request *req = NULL;
 876
 877         while (*p) {
 878                 parent = *p;
 879                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 880                 if (new->r_tid < req->r_tid)
 881                         p = &(*p)->rb_left;
 882                 else if (new->r_tid > req->r_tid)
 883                         p = &(*p)->rb_right;
 884                 else
 885                         BUG();
 886         }
 887
 888         rb_link_node(&new->r_node, parent, p);
 889         rb_insert_color(&new->r_node, &osdc->requests);
 890 }
 891
 892 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 893                                                  u64 tid)
 894 {
 895         struct ceph_osd_request *req;
 896         struct rb_node *n = osdc->requests.rb_node;
 897
 898         while (n) {
 899                 req = rb_entry(n, struct ceph_osd_request, r_node);
 900                 if (tid < req->r_tid)
 901                         n = n->rb_left;
 902                 else if (tid > req->r_tid)
 903                         n = n->rb_right;
 904                 else
 905                         return req;
 906         }
 907         return NULL;
 908 }
 909
 910 static struct ceph_osd_request *
 911 __lookup_request_ge(struct ceph_osd_client *osdc,
 912                     u64 tid)
 913 {
 914         struct ceph_osd_request *req;
 915         struct rb_node *n = osdc->requests.rb_node;
 916
 917         while (n) {
 918                 req = rb_entry(n, struct ceph_osd_request, r_node);
 919                 if (tid < req->r_tid) {
 920                         if (!n->rb_left)
 921                                 return req;
 922                         n = n->rb_left;
 923                 } else if (tid > req->r_tid) {
 924                         n = n->rb_right;
 925                 } else {
 926                         return req;
 927                 }
 928         }
 929         return NULL;
 930 }
 931
 932 static void __kick_linger_request(struct ceph_osd_request *req)
 933 {
 934         struct ceph_osd_client *osdc = req->r_osdc;
 935         struct ceph_osd *osd = req->r_osd;
 936
 937         /*
 938          * Linger requests need to be resent with a new tid to avoid
 939          * the dup op detection logic on the OSDs.  Achieve this with
 940          * a re-register dance instead of open-coding.
 941          */
 942         ceph_osdc_get_request(req);
 943         if (!list_empty(&req->r_linger_item))
 944                 __unregister_linger_request(osdc, req);
 945         else
 946                 __unregister_request(osdc, req);
 947         __register_request(osdc, req);
 948         ceph_osdc_put_request(req);
 949
 950         /*
 951          * Unless request has been registered as both normal and
 952          * lingering, __unregister{,_linger}_request clears r_osd.
 953          * However, here we need to preserve r_osd to make sure we
 954          * requeue on the same OSD.
 955          */
 956         WARN_ON(req->r_osd || !osd);
 957         req->r_osd = osd;
 958
 959         dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
 960         __enqueue_request(req);
 961 }
 962
 963 /*
 964  * Resubmit requests pending on the given osd.
 965  */
 966 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 967                                 struct ceph_osd *osd)
 968 {
 969         struct ceph_osd_request *req, *nreq;
 970         LIST_HEAD(resend);
 971         LIST_HEAD(resend_linger);
 972         int err;
 973
 974         dout("%s osd%d\n", __func__, osd->o_osd);
 975         err = __reset_osd(osdc, osd);
 976         if (err)
 977                 return;
 978
 979         /*
 980          * Build up a list of requests to resend by traversing the
 981          * osd's list of requests.  Requests for a given object are
 982          * sent in tid order, and that is also the order they're
 983          * kept on this list.  Therefore all requests that are in
 984          * flight will be found first, followed by all requests that
 985          * have not yet been sent.  And to resend requests while
 986          * preserving this order we will want to put any sent
 987          * requests back on the front of the osd client's unsent
 988          * list.
 989          *
 990          * So we build a separate ordered list of already-sent
 991          * requests for the affected osd and splice it onto the
 992          * front of the osd client's unsent list.  Once we've seen a
 993          * request that has not yet been sent we're done.  Those
 994          * requests are already sitting right where they belong.
 995          */
 996         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 997                 if (!req->r_sent)
 998                         break;
 999
1000                 if (!req->r_linger) {
1001                         dout("%s requeueing %p tid %llu\n", __func__, req,
1002                              req->r_tid);
1003                         list_move_tail(&req->r_req_lru_item, &resend);
1004                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
1005                 } else {
1006                         list_move_tail(&req->r_req_lru_item, &resend_linger);
1007                 }
1008         }
1009         list_splice(&resend, &osdc->req_unsent);
1010
1011         /*
1012          * Both registered and not yet registered linger requests are
1013          * enqueued with a new tid on the same OSD.  We add/move them
1014          * to req_unsent/o_requests at the end to keep things in tid
1015          * order.
1016          */
1017         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
1018                                  r_linger_osd_item) {
1019                 WARN_ON(!list_empty(&req->r_req_lru_item));
1020                 __kick_linger_request(req);
1021         }
1022
1023         list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
1024                 __kick_linger_request(req);
1025 }
1026
1027 /*
1028  * If the osd connection drops, we need to resubmit all requests.
1029  */
1030 static void osd_reset(struct ceph_connection *con)
1031 {
1032         struct ceph_osd *osd = con->private;
1033         struct ceph_osd_client *osdc;
1034
1035         if (!osd)
1036                 return;
1037         dout("osd_reset osd%d\n", osd->o_osd);
1038         osdc = osd->o_osdc;
1039         down_read(&osdc->map_sem);
1040         mutex_lock(&osdc->request_mutex);
1041         __kick_osd_requests(osdc, osd);
1042         __send_queued(osdc);
1043         mutex_unlock(&osdc->request_mutex);
1044         up_read(&osdc->map_sem);
1045 }
1046
1047 /*
1048  * Track open sessions with osds.
1049  */
1050 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1051 {
1052         struct ceph_osd *osd;
1053
1054         osd = kzalloc(sizeof(*osd), GFP_NOFS);
1055         if (!osd)
1056                 return NULL;
1057
1058         atomic_set(&osd->o_ref, 1);
1059         osd->o_osdc = osdc;
1060         osd->o_osd = onum;
1061         RB_CLEAR_NODE(&osd->o_node);
1062         INIT_LIST_HEAD(&osd->o_requests);
1063         INIT_LIST_HEAD(&osd->o_linger_requests);
1064         INIT_LIST_HEAD(&osd->o_osd_lru);
1065         osd->o_incarnation = 1;
1066
1067         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1068
1069         INIT_LIST_HEAD(&osd->o_keepalive_item);
1070         return osd;
1071 }
1072
1073 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1074 {
1075         if (atomic_inc_not_zero(&osd->o_ref)) {
1076                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1077                      atomic_read(&osd->o_ref));
1078                 return osd;
1079         } else {
1080                 dout("get_osd %p FAIL\n", osd);
1081                 return NULL;
1082         }
1083 }
1084
1085 static void put_osd(struct ceph_osd *osd)
1086 {
1087         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1088              atomic_read(&osd->o_ref) - 1);
1089         if (atomic_dec_and_test(&osd->o_ref)) {
1090                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1091
1092                 if (osd->o_auth.authorizer)
1093                         ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1094                 kfree(osd);
1095         }
1096 }
1097
1098 /*
1099  * remove an osd from our map
1100  */
1101 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1102 {
1103         dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1104         WARN_ON(!list_empty(&osd->o_requests));
1105         WARN_ON(!list_empty(&osd->o_linger_requests));
1106
1107         list_del_init(&osd->o_osd_lru);
1108         rb_erase(&osd->o_node, &osdc->osds);
1109         RB_CLEAR_NODE(&osd->o_node);
1110 }
1111
1112 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1113 {
1114         dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1115
1116         if (!RB_EMPTY_NODE(&osd->o_node)) {
1117                 ceph_con_close(&osd->o_con);
1118                 __remove_osd(osdc, osd);
1119                 put_osd(osd);
1120         }
1121 }
1122
1123 static void remove_all_osds(struct ceph_osd_client *osdc)
1124 {
1125         dout("%s %p\n", __func__, osdc);
1126         mutex_lock(&osdc->request_mutex);
1127         while (!RB_EMPTY_ROOT(&osdc->osds)) {
1128                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1129                                                 struct ceph_osd, o_node);
1130                 remove_osd(osdc, osd);
1131         }
1132         mutex_unlock(&osdc->request_mutex);
1133 }
1134
1135 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1136                               struct ceph_osd *osd)
1137 {
1138         dout("%s %p\n", __func__, osd);
1139         BUG_ON(!list_empty(&osd->o_osd_lru));
1140
1141         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1142         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1143 }
1144
1145 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1146                                   struct ceph_osd *osd)
1147 {
1148         dout("%s %p\n", __func__, osd);
1149
1150         if (list_empty(&osd->o_requests) &&
1151             list_empty(&osd->o_linger_requests))
1152                 __move_osd_to_lru(osdc, osd);
1153 }
1154
1155 static void __remove_osd_from_lru(struct ceph_osd *osd)
1156 {
1157         dout("__remove_osd_from_lru %p\n", osd);
1158         if (!list_empty(&osd->o_osd_lru))
1159                 list_del_init(&osd->o_osd_lru);
1160 }
1161
1162 static void remove_old_osds(struct ceph_osd_client *osdc)
1163 {
1164         struct ceph_osd *osd, *nosd;
1165
1166         dout("__remove_old_osds %p\n", osdc);
1167         mutex_lock(&osdc->request_mutex);
1168         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1169                 if (time_before(jiffies, osd->lru_ttl))
1170                         break;
1171                 remove_osd(osdc, osd);
1172         }
1173         mutex_unlock(&osdc->request_mutex);
1174 }
1175
1176 /*
1177  * reset osd connect
1178  */
1179 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1180 {
1181         struct ceph_entity_addr *peer_addr;
1182
1183         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1184         if (list_empty(&osd->o_requests) &&
1185             list_empty(&osd->o_linger_requests)) {
1186                 remove_osd(osdc, osd);
1187                 return -ENODEV;
1188         }
1189
1190         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1191         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1192                         !ceph_con_opened(&osd->o_con)) {
1193                 struct ceph_osd_request *req;
1194
1195                 dout("osd addr hasn't changed and connection never opened, "
1196                      "letting msgr retry\n");
1197                 /* touch each r_stamp for handle_timeout()'s benfit */
1198                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1199                         req->r_stamp = jiffies;
1200
1201                 return -EAGAIN;
1202         }
1203
1204         ceph_con_close(&osd->o_con);
1205         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1206         osd->o_incarnation++;
1207
1208         return 0;
1209 }
1210
1211 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1212 {
1213         struct rb_node **p = &osdc->osds.rb_node;
1214         struct rb_node *parent = NULL;
1215         struct ceph_osd *osd = NULL;
1216
1217         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1218         while (*p) {
1219                 parent = *p;
1220                 osd = rb_entry(parent, struct ceph_osd, o_node);
1221                 if (new->o_osd < osd->o_osd)
1222                         p = &(*p)->rb_left;
1223                 else if (new->o_osd > osd->o_osd)
1224                         p = &(*p)->rb_right;
1225                 else
1226                         BUG();
1227         }
1228
1229         rb_link_node(&new->o_node, parent, p);
1230         rb_insert_color(&new->o_node, &osdc->osds);
1231 }
1232
1233 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1234 {
1235         struct ceph_osd *osd;
1236         struct rb_node *n = osdc->osds.rb_node;
1237
1238         while (n) {
1239                 osd = rb_entry(n, struct ceph_osd, o_node);
1240                 if (o < osd->o_osd)
1241                         n = n->rb_left;
1242                 else if (o > osd->o_osd)
1243                         n = n->rb_right;
1244                 else
1245                         return osd;
1246         }
1247         return NULL;
1248 }
1249
1250 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1251 {
1252         schedule_delayed_work(&osdc->timeout_work,
1253                               osdc->client->options->osd_keepalive_timeout);
1254 }
1255
1256 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1257 {
1258         cancel_delayed_work(&osdc->timeout_work);
1259 }
1260
1261 /*
1262  * Register request, assign tid.  If this is the first request, set up
1263  * the timeout event.
1264  */
1265 static void __register_request(struct ceph_osd_client *osdc,
1266                                struct ceph_osd_request *req)
1267 {
1268         req->r_tid = ++osdc->last_tid;
1269         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1270         dout("__register_request %p tid %lld\n", req, req->r_tid);
1271         __insert_request(osdc, req);
1272         ceph_osdc_get_request(req);
1273         osdc->num_requests++;
1274         if (osdc->num_requests == 1) {
1275                 dout(" first request, scheduling timeout\n");
1276                 __schedule_osd_timeout(osdc);
1277         }
1278 }
1279
1280 /*
1281  * called under osdc->request_mutex
1282  */
1283 static void __unregister_request(struct ceph_osd_client *osdc,
1284                                  struct ceph_osd_request *req)
1285 {
1286         if (RB_EMPTY_NODE(&req->r_node)) {
1287                 dout("__unregister_request %p tid %lld not registered\n",
1288                         req, req->r_tid);
1289                 return;
1290         }
1291
1292         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1293         rb_erase(&req->r_node, &osdc->requests);
1294         RB_CLEAR_NODE(&req->r_node);
1295         osdc->num_requests--;
1296
1297         if (req->r_osd) {
1298                 /* make sure the original request isn't in flight. */
1299                 ceph_msg_revoke(req->r_request);
1300
1301                 list_del_init(&req->r_osd_item);
1302                 maybe_move_osd_to_lru(osdc, req->r_osd);
1303                 if (list_empty(&req->r_linger_osd_item))
1304                         req->r_osd = NULL;
1305         }
1306
1307         list_del_init(&req->r_req_lru_item);
1308         ceph_osdc_put_request(req);
1309
1310         if (osdc->num_requests == 0) {
1311                 dout(" no requests, canceling timeout\n");
1312                 __cancel_osd_timeout(osdc);
1313         }
1314 }
1315
1316 /*
1317  * Cancel a previously queued request message
1318  */
1319 static void __cancel_request(struct ceph_osd_request *req)
1320 {
1321         if (req->r_sent && req->r_osd) {
1322                 ceph_msg_revoke(req->r_request);
1323                 req->r_sent = 0;
1324         }
1325 }
1326
1327 static void __register_linger_request(struct ceph_osd_client *osdc,
1328                                     struct ceph_osd_request *req)
1329 {
1330         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1331         WARN_ON(!req->r_linger);
1332
1333         ceph_osdc_get_request(req);
1334         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1335         if (req->r_osd)
1336                 list_add_tail(&req->r_linger_osd_item,
1337                               &req->r_osd->o_linger_requests);
1338 }
1339
1340 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1341                                         struct ceph_osd_request *req)
1342 {
1343         WARN_ON(!req->r_linger);
1344
1345         if (list_empty(&req->r_linger_item)) {
1346                 dout("%s %p tid %llu not registered\n", __func__, req,
1347                      req->r_tid);
1348                 return;
1349         }
1350
1351         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1352         list_del_init(&req->r_linger_item);
1353
1354         if (req->r_osd) {
1355                 list_del_init(&req->r_linger_osd_item);
1356                 maybe_move_osd_to_lru(osdc, req->r_osd);
1357                 if (list_empty(&req->r_osd_item))
1358                         req->r_osd = NULL;
1359         }
1360         ceph_osdc_put_request(req);
1361 }
1362
1363 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1364                                   struct ceph_osd_request *req)
1365 {
1366         if (!req->r_linger) {
1367                 dout("set_request_linger %p\n", req);
1368                 req->r_linger = 1;
1369         }
1370 }
1371 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1372
1373 /*
1374  * Returns whether a request should be blocked from being sent
1375  * based on the current osdmap and osd_client settings.
1376  *
1377  * Caller should hold map_sem for read.
1378  */
1379 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1380                                    struct ceph_osd_request *req)
1381 {
1382         bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1383         bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1384                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1385         return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1386                 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1387 }
1388
1389 /*
1390  * Calculate mapping of a request to a PG.  Takes tiering into account.
1391  */
1392 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1393                              struct ceph_osd_request *req,
1394                              struct ceph_pg *pg_out)
1395 {
1396         bool need_check_tiering;
1397
1398         need_check_tiering = false;
1399         if (req->r_target_oloc.pool == -1) {
1400                 req->r_target_oloc = req->r_base_oloc; /* struct */
1401                 need_check_tiering = true;
1402         }
1403         if (req->r_target_oid.name_len == 0) {
1404                 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1405                 need_check_tiering = true;
1406         }
1407
1408         if (need_check_tiering &&
1409             (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1410                 struct ceph_pg_pool_info *pi;
1411
1412                 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1413                 if (pi) {
1414                         if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1415                             pi->read_tier >= 0)
1416                                 req->r_target_oloc.pool = pi->read_tier;
1417                         if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1418                             pi->write_tier >= 0)
1419                                 req->r_target_oloc.pool = pi->write_tier;
1420                 }
1421                 /* !pi is caught in ceph_oloc_oid_to_pg() */
1422         }
1423
1424         return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1425                                    &req->r_target_oid, pg_out);
1426 }
1427
1428 static void __enqueue_request(struct ceph_osd_request *req)
1429 {
1430         struct ceph_osd_client *osdc = req->r_osdc;
1431
1432         dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1433              req->r_osd ? req->r_osd->o_osd : -1);
1434
1435         if (req->r_osd) {
1436                 __remove_osd_from_lru(req->r_osd);
1437                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1438                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1439         } else {
1440                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1441         }
1442 }
1443
1444 /*
1445  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1446  * (as needed), and set the request r_osd appropriately.  If there is
1447  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1448  * (unsent, homeless) or leave on in-flight lru.
1449  *
1450  * Return 0 if unchanged, 1 if changed, or negative on error.
1451  *
1452  * Caller should hold map_sem for read and request_mutex.
1453  */
1454 static int __map_request(struct ceph_osd_client *osdc,
1455                          struct ceph_osd_request *req, int force_resend)
1456 {
1457         struct ceph_pg pgid;
1458         int acting[CEPH_PG_MAX_SIZE];
1459         int num, o;
1460         int err;
1461         bool was_paused;
1462
1463         dout("map_request %p tid %lld\n", req, req->r_tid);
1464
1465         err = __calc_request_pg(osdc->osdmap, req, &pgid);
1466         if (err) {
1467                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1468                 return err;
1469         }
1470         req->r_pgid = pgid;
1471
1472         num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1473         if (num < 0)
1474                 num = 0;
1475
1476         was_paused = req->r_paused;
1477         req->r_paused = __req_should_be_paused(osdc, req);
1478         if (was_paused && !req->r_paused)
1479                 force_resend = 1;
1480
1481         if ((!force_resend &&
1482              req->r_osd && req->r_osd->o_osd == o &&
1483              req->r_sent >= req->r_osd->o_incarnation &&
1484              req->r_num_pg_osds == num &&
1485              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1486             (req->r_osd == NULL && o == -1) ||
1487             req->r_paused)
1488                 return 0;  /* no change */
1489
1490         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1491              req->r_tid, pgid.pool, pgid.seed, o,
1492              req->r_osd ? req->r_osd->o_osd : -1);
1493
1494         /* record full pg acting set */
1495         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1496         req->r_num_pg_osds = num;
1497
1498         if (req->r_osd) {
1499                 __cancel_request(req);
1500                 list_del_init(&req->r_osd_item);
1501                 list_del_init(&req->r_linger_osd_item);
1502                 req->r_osd = NULL;
1503         }
1504
1505         req->r_osd = __lookup_osd(osdc, o);
1506         if (!req->r_osd && o >= 0) {
1507                 err = -ENOMEM;
1508                 req->r_osd = create_osd(osdc, o);
1509                 if (!req->r_osd) {
1510                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1511                         goto out;
1512                 }
1513
1514                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1515                 __insert_osd(osdc, req->r_osd);
1516
1517                 ceph_con_open(&req->r_osd->o_con,
1518                               CEPH_ENTITY_TYPE_OSD, o,
1519                               &osdc->osdmap->osd_addr[o]);
1520         }
1521
1522         __enqueue_request(req);
1523         err = 1;   /* osd or pg changed */
1524
1525 out:
1526         return err;
1527 }
1528
1529 /*
1530  * caller should hold map_sem (for read) and request_mutex
1531  */
1532 static void __send_request(struct ceph_osd_client *osdc,
1533                            struct ceph_osd_request *req)
1534 {
1535         void *p;
1536
1537         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1538              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1539              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1540
1541         /* fill in message content that changes each time we send it */
1542         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1543         put_unaligned_le32(req->r_flags, req->r_request_flags);
1544         put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1545         p = req->r_request_pgid;
1546         ceph_encode_64(&p, req->r_pgid.pool);
1547         ceph_encode_32(&p, req->r_pgid.seed);
1548         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1549         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1550                sizeof(req->r_reassert_version));
1551
1552         req->r_stamp = jiffies;
1553         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1554
1555         ceph_msg_get(req->r_request); /* send consumes a ref */
1556
1557         req->r_sent = req->r_osd->o_incarnation;
1558
1559         ceph_con_send(&req->r_osd->o_con, req->r_request);
1560 }
1561
1562 /*
1563  * Send any requests in the queue (req_unsent).
1564  */
1565 static void __send_queued(struct ceph_osd_client *osdc)
1566 {
1567         struct ceph_osd_request *req, *tmp;
1568
1569         dout("__send_queued\n");
1570         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1571                 __send_request(osdc, req);
1572 }
1573
1574 /*
1575  * Caller should hold map_sem for read and request_mutex.
1576  */
1577 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1578                                      struct ceph_osd_request *req,
1579                                      bool nofail)
1580 {
1581         int rc;
1582
1583         __register_request(osdc, req);
1584         req->r_sent = 0;
1585         req->r_got_reply = 0;
1586         rc = __map_request(osdc, req, 0);
1587         if (rc < 0) {
1588                 if (nofail) {
1589                         dout("osdc_start_request failed map, "
1590                                 " will retry %lld\n", req->r_tid);
1591                         rc = 0;
1592                 } else {
1593                         __unregister_request(osdc, req);
1594                 }
1595                 return rc;
1596         }
1597
1598         if (req->r_osd == NULL) {
1599                 dout("send_request %p no up osds in pg\n", req);
1600                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1601         } else {
1602                 __send_queued(osdc);
1603         }
1604
1605         return 0;
1606 }
1607
1608 /*
1609  * Timeout callback, called every N seconds when 1 or more osd
1610  * requests has been active for more than N seconds.  When this
1611  * happens, we ping all OSDs with requests who have timed out to
1612  * ensure any communications channel reset is detected.  Reset the
1613  * request timeouts another N seconds in the future as we go.
1614  * Reschedule the timeout event another N seconds in future (unless
1615  * there are no open requests).
1616  */
1617 static void handle_timeout(struct work_struct *work)
1618 {
1619         struct ceph_osd_client *osdc =
1620                 container_of(work, struct ceph_osd_client, timeout_work.work);
1621         struct ceph_options *opts = osdc->client->options;
1622         struct ceph_osd_request *req;
1623         struct ceph_osd *osd;
1624         struct list_head slow_osds;
1625         dout("timeout\n");
1626         down_read(&osdc->map_sem);
1627
1628         ceph_monc_request_next_osdmap(&osdc->client->monc);
1629
1630         mutex_lock(&osdc->request_mutex);
1631
1632         /*
1633          * ping osds that are a bit slow.  this ensures that if there
1634          * is a break in the TCP connection we will notice, and reopen
1635          * a connection with that osd (from the fault callback).
1636          */
1637         INIT_LIST_HEAD(&slow_osds);
1638         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1639                 if (time_before(jiffies,
1640                                 req->r_stamp + opts->osd_keepalive_timeout))
1641                         break;
1642
1643                 osd = req->r_osd;
1644                 BUG_ON(!osd);
1645                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1646                      req->r_tid, osd->o_osd);
1647                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1648         }
1649         while (!list_empty(&slow_osds)) {
1650                 osd = list_entry(slow_osds.next, struct ceph_osd,
1651                                  o_keepalive_item);
1652                 list_del_init(&osd->o_keepalive_item);
1653                 ceph_con_keepalive(&osd->o_con);
1654         }
1655
1656         __schedule_osd_timeout(osdc);
1657         __send_queued(osdc);
1658         mutex_unlock(&osdc->request_mutex);
1659         up_read(&osdc->map_sem);
1660 }
1661
1662 static void handle_osds_timeout(struct work_struct *work)
1663 {
1664         struct ceph_osd_client *osdc =
1665                 container_of(work, struct ceph_osd_client,
1666                              osds_timeout_work.work);
1667         unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
1668
1669         dout("osds timeout\n");
1670         down_read(&osdc->map_sem);
1671         remove_old_osds(osdc);
1672         up_read(&osdc->map_sem);
1673
1674         schedule_delayed_work(&osdc->osds_timeout_work,
1675                               round_jiffies_relative(delay));
1676 }
1677
1678 static int ceph_oloc_decode(void **p, void *end,
1679                             struct ceph_object_locator *oloc)
1680 {
1681         u8 struct_v, struct_cv;
1682         u32 len;
1683         void *struct_end;
1684         int ret = 0;
1685
1686         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1687         struct_v = ceph_decode_8(p);
1688         struct_cv = ceph_decode_8(p);
1689         if (struct_v < 3) {
1690                 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1691                         struct_v, struct_cv);
1692                 goto e_inval;
1693         }
1694         if (struct_cv > 6) {
1695                 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1696                         struct_v, struct_cv);
1697                 goto e_inval;
1698         }
1699         len = ceph_decode_32(p);
1700         ceph_decode_need(p, end, len, e_inval);
1701         struct_end = *p + len;
1702
1703         oloc->pool = ceph_decode_64(p);
1704         *p += 4; /* skip preferred */
1705
1706         len = ceph_decode_32(p);
1707         if (len > 0) {
1708                 pr_warn("ceph_object_locator::key is set\n");
1709                 goto e_inval;
1710         }
1711
1712         if (struct_v >= 5) {
1713                 len = ceph_decode_32(p);
1714                 if (len > 0) {
1715                         pr_warn("ceph_object_locator::nspace is set\n");
1716                         goto e_inval;
1717                 }
1718         }
1719
1720         if (struct_v >= 6) {
1721                 s64 hash = ceph_decode_64(p);
1722                 if (hash != -1) {
1723                         pr_warn("ceph_object_locator::hash is set\n");
1724                         goto e_inval;
1725                 }
1726         }
1727
1728         /* skip the rest */
1729         *p = struct_end;
1730 out:
1731         return ret;
1732
1733 e_inval:
1734         ret = -EINVAL;
1735         goto out;
1736 }
1737
1738 static int ceph_redirect_decode(void **p, void *end,
1739                                 struct ceph_request_redirect *redir)
1740 {
1741         u8 struct_v, struct_cv;
1742         u32 len;
1743         void *struct_end;
1744         int ret;
1745
1746         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1747         struct_v = ceph_decode_8(p);
1748         struct_cv = ceph_decode_8(p);
1749         if (struct_cv > 1) {
1750                 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1751                         struct_v, struct_cv);
1752                 goto e_inval;
1753         }
1754         len = ceph_decode_32(p);
1755         ceph_decode_need(p, end, len, e_inval);
1756         struct_end = *p + len;
1757
1758         ret = ceph_oloc_decode(p, end, &redir->oloc);
1759         if (ret)
1760                 goto out;
1761
1762         len = ceph_decode_32(p);
1763         if (len > 0) {
1764                 pr_warn("ceph_request_redirect::object_name is set\n");
1765                 goto e_inval;
1766         }
1767
1768         len = ceph_decode_32(p);
1769         *p += len; /* skip osd_instructions */
1770
1771         /* skip the rest */
1772         *p = struct_end;
1773 out:
1774         return ret;
1775
1776 e_inval:
1777         ret = -EINVAL;
1778         goto out;
1779 }
1780
1781 static void complete_request(struct ceph_osd_request *req)
1782 {
1783         complete_all(&req->r_safe_completion);  /* fsync waiter */
1784 }
1785
1786 /*
1787  * handle osd op reply.  either call the callback if it is specified,
1788  * or do the completion to wake up the waiting thread.
1789  */
1790 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1791 {
1792         void *p, *end;
1793         struct ceph_osd_request *req;
1794         struct ceph_request_redirect redir;
1795         u64 tid;
1796         int object_len;
1797         unsigned int numops;
1798         int payload_len, flags;
1799         s32 result;
1800         s32 retry_attempt;
1801         struct ceph_pg pg;
1802         int err;
1803         u32 reassert_epoch;
1804         u64 reassert_version;
1805         u32 osdmap_epoch;
1806         int already_completed;
1807         u32 bytes;
1808         u8 decode_redir;
1809         unsigned int i;
1810
1811         tid = le64_to_cpu(msg->hdr.tid);
1812         dout("handle_reply %p tid %llu\n", msg, tid);
1813
1814         p = msg->front.iov_base;
1815         end = p + msg->front.iov_len;
1816
1817         ceph_decode_need(&p, end, 4, bad);
1818         object_len = ceph_decode_32(&p);
1819         ceph_decode_need(&p, end, object_len, bad);
1820         p += object_len;
1821
1822         err = ceph_decode_pgid(&p, end, &pg);
1823         if (err)
1824                 goto bad;
1825
1826         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1827         flags = ceph_decode_64(&p);
1828         result = ceph_decode_32(&p);
1829         reassert_epoch = ceph_decode_32(&p);
1830         reassert_version = ceph_decode_64(&p);
1831         osdmap_epoch = ceph_decode_32(&p);
1832
1833         /* lookup */
1834         down_read(&osdc->map_sem);
1835         mutex_lock(&osdc->request_mutex);
1836         req = __lookup_request(osdc, tid);
1837         if (req == NULL) {
1838                 dout("handle_reply tid %llu dne\n", tid);
1839                 goto bad_mutex;
1840         }
1841         ceph_osdc_get_request(req);
1842
1843         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1844              req, result);
1845
1846         ceph_decode_need(&p, end, 4, bad_put);
1847         numops = ceph_decode_32(&p);
1848         if (numops > CEPH_OSD_MAX_OPS)
1849                 goto bad_put;
1850         if (numops != req->r_num_ops)
1851                 goto bad_put;
1852         payload_len = 0;
1853         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1854         for (i = 0; i < numops; i++) {
1855                 struct ceph_osd_op *op = p;
1856                 int len;
1857
1858                 len = le32_to_cpu(op->payload_len);
1859                 req->r_ops[i].outdata_len = len;
1860                 dout(" op %d has %d bytes\n", i, len);
1861                 payload_len += len;
1862                 p += sizeof(*op);
1863         }
1864         bytes = le32_to_cpu(msg->hdr.data_len);
1865         if (payload_len != bytes) {
1866                 pr_warn("sum of op payload lens %d != data_len %d\n",
1867                         payload_len, bytes);
1868                 goto bad_put;
1869         }
1870
1871         ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1872         retry_attempt = ceph_decode_32(&p);
1873         for (i = 0; i < numops; i++)
1874                 req->r_ops[i].rval = ceph_decode_32(&p);
1875
1876         if (le16_to_cpu(msg->hdr.version) >= 6) {
1877                 p += 8 + 4; /* skip replay_version */
1878                 p += 8; /* skip user_version */
1879
1880                 if (le16_to_cpu(msg->hdr.version) >= 7)
1881                         ceph_decode_8_safe(&p, end, decode_redir, bad_put);
1882                 else
1883                         decode_redir = 1;
1884         } else {
1885                 decode_redir = 0;
1886         }
1887
1888         if (decode_redir) {
1889                 err = ceph_redirect_decode(&p, end, &redir);
1890                 if (err)
1891                         goto bad_put;
1892         } else {
1893                 redir.oloc.pool = -1;
1894         }
1895
1896         if (redir.oloc.pool != -1) {
1897                 dout("redirect pool %lld\n", redir.oloc.pool);
1898
1899                 __unregister_request(osdc, req);
1900
1901                 req->r_target_oloc = redir.oloc; /* struct */
1902
1903                 /*
1904                  * Start redirect requests with nofail=true.  If
1905                  * mapping fails, request will end up on the notarget
1906                  * list, waiting for the new osdmap (which can take
1907                  * a while), even though the original request mapped
1908                  * successfully.  In the future we might want to follow
1909                  * original request's nofail setting here.
1910                  */
1911                 err = __ceph_osdc_start_request(osdc, req, true);
1912                 BUG_ON(err);
1913
1914                 goto out_unlock;
1915         }
1916
1917         already_completed = req->r_got_reply;
1918         if (!req->r_got_reply) {
1919                 req->r_result = result;
1920                 dout("handle_reply result %d bytes %d\n", req->r_result,
1921                      bytes);
1922                 if (req->r_result == 0)
1923                         req->r_result = bytes;
1924
1925                 /* in case this is a write and we need to replay, */
1926                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1927                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1928
1929                 req->r_got_reply = 1;
1930         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1931                 dout("handle_reply tid %llu dup ack\n", tid);
1932                 goto out_unlock;
1933         }
1934
1935         dout("handle_reply tid %llu flags %d\n", tid, flags);
1936
1937         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1938                 __register_linger_request(osdc, req);
1939
1940         /* either this is a read, or we got the safe response */
1941         if (result < 0 ||
1942             (flags & CEPH_OSD_FLAG_ONDISK) ||
1943             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1944                 __unregister_request(osdc, req);
1945
1946         mutex_unlock(&osdc->request_mutex);
1947         up_read(&osdc->map_sem);
1948
1949         if (!already_completed) {
1950                 if (req->r_unsafe_callback &&
1951                     result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1952                         req->r_unsafe_callback(req, true);
1953                 if (req->r_callback)
1954                         req->r_callback(req, msg);
1955                 else
1956                         complete_all(&req->r_completion);
1957         }
1958
1959         if (flags & CEPH_OSD_FLAG_ONDISK) {
1960                 if (req->r_unsafe_callback && already_completed)
1961                         req->r_unsafe_callback(req, false);
1962                 complete_request(req);
1963         }
1964
1965 out:
1966         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1967         ceph_osdc_put_request(req);
1968         return;
1969 out_unlock:
1970         mutex_unlock(&osdc->request_mutex);
1971         up_read(&osdc->map_sem);
1972         goto out;
1973
1974 bad_put:
1975         req->r_result = -EIO;
1976         __unregister_request(osdc, req);
1977         if (req->r_callback)
1978                 req->r_callback(req, msg);
1979         else
1980                 complete_all(&req->r_completion);
1981         complete_request(req);
1982         ceph_osdc_put_request(req);
1983 bad_mutex:
1984         mutex_unlock(&osdc->request_mutex);
1985         up_read(&osdc->map_sem);
1986 bad:
1987         pr_err("corrupt osd_op_reply got %d %d\n",
1988                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1989         ceph_msg_dump(msg);
1990 }
1991
1992 static void reset_changed_osds(struct ceph_osd_client *osdc)
1993 {
1994         struct rb_node *p, *n;
1995
1996         dout("%s %p\n", __func__, osdc);
1997         for (p = rb_first(&osdc->osds); p; p = n) {
1998                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1999
2000                 n = rb_next(p);
2001                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
2002                     memcmp(&osd->o_con.peer_addr,
2003                            ceph_osd_addr(osdc->osdmap,
2004                                          osd->o_osd),
2005                            sizeof(struct ceph_entity_addr)) != 0)
2006                         __reset_osd(osdc, osd);
2007         }
2008 }
2009
2010 /*
2011  * Requeue requests whose mapping to an OSD has changed.  If requests map to
2012  * no osd, request a new map.
2013  *
2014  * Caller should hold map_sem for read.
2015  */
2016 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
2017                           bool force_resend_writes)
2018 {
2019         struct ceph_osd_request *req, *nreq;
2020         struct rb_node *p;
2021         int needmap = 0;
2022         int err;
2023         bool force_resend_req;
2024
2025         dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
2026                 force_resend_writes ? " (force resend writes)" : "");
2027         mutex_lock(&osdc->request_mutex);
2028         for (p = rb_first(&osdc->requests); p; ) {
2029                 req = rb_entry(p, struct ceph_osd_request, r_node);
2030                 p = rb_next(p);
2031
2032                 /*
2033                  * For linger requests that have not yet been
2034                  * registered, move them to the linger list; they'll
2035                  * be sent to the osd in the loop below.  Unregister
2036                  * the request before re-registering it as a linger
2037                  * request to ensure the __map_request() below
2038                  * will decide it needs to be sent.
2039                  */
2040                 if (req->r_linger && list_empty(&req->r_linger_item)) {
2041                         dout("%p tid %llu restart on osd%d\n",
2042                              req, req->r_tid,
2043                              req->r_osd ? req->r_osd->o_osd : -1);
2044                         ceph_osdc_get_request(req);
2045                         __unregister_request(osdc, req);
2046                         __register_linger_request(osdc, req);
2047                         ceph_osdc_put_request(req);
2048                         continue;
2049                 }
2050
2051                 force_resend_req = force_resend ||
2052                         (force_resend_writes &&
2053                                 req->r_flags & CEPH_OSD_FLAG_WRITE);
2054                 err = __map_request(osdc, req, force_resend_req);
2055                 if (err < 0)
2056                         continue;  /* error */
2057                 if (req->r_osd == NULL) {
2058                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2059                         needmap++;  /* request a newer map */
2060                 } else if (err > 0) {
2061                         if (!req->r_linger) {
2062                                 dout("%p tid %llu requeued on osd%d\n", req,
2063                                      req->r_tid,
2064                                      req->r_osd ? req->r_osd->o_osd : -1);
2065                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2066                         }
2067                 }
2068         }
2069
2070         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2071                                  r_linger_item) {
2072                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2073
2074                 err = __map_request(osdc, req,
2075                                     force_resend || force_resend_writes);
2076                 dout("__map_request returned %d\n", err);
2077                 if (err < 0)
2078                         continue;  /* hrm! */
2079                 if (req->r_osd == NULL || err > 0) {
2080                         if (req->r_osd == NULL) {
2081                                 dout("lingering %p tid %llu maps to no osd\n",
2082                                      req, req->r_tid);
2083                                 /*
2084                                  * A homeless lingering request makes
2085                                  * no sense, as it's job is to keep
2086                                  * a particular OSD connection open.
2087                                  * Request a newer map and kick the
2088                                  * request, knowing that it won't be
2089                                  * resent until we actually get a map
2090                                  * that can tell us where to send it.
2091                                  */
2092                                 needmap++;
2093                         }
2094
2095                         dout("kicking lingering %p tid %llu osd%d\n", req,
2096                              req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2097                         __register_request(osdc, req);
2098                         __unregister_linger_request(osdc, req);
2099                 }
2100         }
2101         reset_changed_osds(osdc);
2102         mutex_unlock(&osdc->request_mutex);
2103
2104         if (needmap) {
2105                 dout("%d requests for down osds, need new map\n", needmap);
2106                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2107         }
2108 }
2109
2110
2111 /*
2112  * Process updated osd map.
2113  *
2114  * The message contains any number of incremental and full maps, normally
2115  * indicating some sort of topology change in the cluster.  Kick requests
2116  * off to different OSDs as needed.
2117  */
2118 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2119 {
2120         void *p, *end, *next;
2121         u32 nr_maps, maplen;
2122         u32 epoch;
2123         struct ceph_osdmap *newmap = NULL, *oldmap;
2124         int err;
2125         struct ceph_fsid fsid;
2126         bool was_full;
2127
2128         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2129         p = msg->front.iov_base;
2130         end = p + msg->front.iov_len;
2131
2132         /* verify fsid */
2133         ceph_decode_need(&p, end, sizeof(fsid), bad);
2134         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2135         if (ceph_check_fsid(osdc->client, &fsid) < 0)
2136                 return;
2137
2138         down_write(&osdc->map_sem);
2139
2140         was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2141
2142         /* incremental maps */
2143         ceph_decode_32_safe(&p, end, nr_maps, bad);
2144         dout(" %d inc maps\n", nr_maps);
2145         while (nr_maps > 0) {
2146                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2147                 epoch = ceph_decode_32(&p);
2148                 maplen = ceph_decode_32(&p);
2149                 ceph_decode_need(&p, end, maplen, bad);
2150                 next = p + maplen;
2151                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2152                         dout("applying incremental map %u len %d\n",
2153                              epoch, maplen);
2154                         newmap = osdmap_apply_incremental(&p, next,
2155                                                           osdc->osdmap,
2156                                                           &osdc->client->msgr);
2157                         if (IS_ERR(newmap)) {
2158                                 err = PTR_ERR(newmap);
2159                                 goto bad;
2160                         }
2161                         BUG_ON(!newmap);
2162                         if (newmap != osdc->osdmap) {
2163                                 ceph_osdmap_destroy(osdc->osdmap);
2164                                 osdc->osdmap = newmap;
2165                         }
2166                         was_full = was_full ||
2167                                 ceph_osdmap_flag(osdc->osdmap,
2168                                                  CEPH_OSDMAP_FULL);
2169                         kick_requests(osdc, 0, was_full);
2170                 } else {
2171                         dout("ignoring incremental map %u len %d\n",
2172                              epoch, maplen);
2173                 }
2174                 p = next;
2175                 nr_maps--;
2176         }
2177         if (newmap)
2178                 goto done;
2179
2180         /* full maps */
2181         ceph_decode_32_safe(&p, end, nr_maps, bad);
2182         dout(" %d full maps\n", nr_maps);
2183         while (nr_maps) {
2184                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2185                 epoch = ceph_decode_32(&p);
2186                 maplen = ceph_decode_32(&p);
2187                 ceph_decode_need(&p, end, maplen, bad);
2188                 if (nr_maps > 1) {
2189                         dout("skipping non-latest full map %u len %d\n",
2190                              epoch, maplen);
2191                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2192                         dout("skipping full map %u len %d, "
2193                              "older than our %u\n", epoch, maplen,
2194                              osdc->osdmap->epoch);
2195                 } else {
2196                         int skipped_map = 0;
2197
2198                         dout("taking full map %u len %d\n", epoch, maplen);
2199                         newmap = ceph_osdmap_decode(&p, p+maplen);
2200                         if (IS_ERR(newmap)) {
2201                                 err = PTR_ERR(newmap);
2202                                 goto bad;
2203                         }
2204                         BUG_ON(!newmap);
2205                         oldmap = osdc->osdmap;
2206                         osdc->osdmap = newmap;
2207                         if (oldmap) {
2208                                 if (oldmap->epoch + 1 < newmap->epoch)
2209                                         skipped_map = 1;
2210                                 ceph_osdmap_destroy(oldmap);
2211                         }
2212                         was_full = was_full ||
2213                                 ceph_osdmap_flag(osdc->osdmap,
2214                                                  CEPH_OSDMAP_FULL);
2215                         kick_requests(osdc, skipped_map, was_full);
2216                 }
2217                 p += maplen;
2218                 nr_maps--;
2219         }
2220
2221         if (!osdc->osdmap)
2222                 goto bad;
2223 done:
2224         downgrade_write(&osdc->map_sem);
2225         ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
2226                           osdc->osdmap->epoch);
2227
2228         /*
2229          * subscribe to subsequent osdmap updates if full to ensure
2230          * we find out when we are no longer full and stop returning
2231          * ENOSPC.
2232          */
2233         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2234                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2235                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2236                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2237
2238         mutex_lock(&osdc->request_mutex);
2239         __send_queued(osdc);
2240         mutex_unlock(&osdc->request_mutex);
2241         up_read(&osdc->map_sem);
2242         wake_up_all(&osdc->client->auth_wq);
2243         return;
2244
2245 bad:
2246         pr_err("osdc handle_map corrupt msg\n");
2247         ceph_msg_dump(msg);
2248         up_write(&osdc->map_sem);
2249 }
2250
2251 /*
2252  * watch/notify callback event infrastructure
2253  *
2254  * These callbacks are used both for watch and notify operations.
2255  */
2256 static void __release_event(struct kref *kref)
2257 {
2258         struct ceph_osd_event *event =
2259                 container_of(kref, struct ceph_osd_event, kref);
2260
2261         dout("__release_event %p\n", event);
2262         kfree(event);
2263 }
2264
2265 static void get_event(struct ceph_osd_event *event)
2266 {
2267         kref_get(&event->kref);
2268 }
2269
2270 void ceph_osdc_put_event(struct ceph_osd_event *event)
2271 {
2272         kref_put(&event->kref, __release_event);
2273 }
2274 EXPORT_SYMBOL(ceph_osdc_put_event);
2275
2276 static void __insert_event(struct ceph_osd_client *osdc,
2277                              struct ceph_osd_event *new)
2278 {
2279         struct rb_node **p = &osdc->event_tree.rb_node;
2280         struct rb_node *parent = NULL;
2281         struct ceph_osd_event *event = NULL;
2282
2283         while (*p) {
2284                 parent = *p;
2285                 event = rb_entry(parent, struct ceph_osd_event, node);
2286                 if (new->cookie < event->cookie)
2287                         p = &(*p)->rb_left;
2288                 else if (new->cookie > event->cookie)
2289                         p = &(*p)->rb_right;
2290                 else
2291                         BUG();
2292         }
2293
2294         rb_link_node(&new->node, parent, p);
2295         rb_insert_color(&new->node, &osdc->event_tree);
2296 }
2297
2298 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2299                                                 u64 cookie)
2300 {
2301         struct rb_node **p = &osdc->event_tree.rb_node;
2302         struct rb_node *parent = NULL;
2303         struct ceph_osd_event *event = NULL;
2304
2305         while (*p) {
2306                 parent = *p;
2307                 event = rb_entry(parent, struct ceph_osd_event, node);
2308                 if (cookie < event->cookie)
2309                         p = &(*p)->rb_left;
2310                 else if (cookie > event->cookie)
2311                         p = &(*p)->rb_right;
2312                 else
2313                         return event;
2314         }
2315         return NULL;
2316 }
2317
2318 static void __remove_event(struct ceph_osd_event *event)
2319 {
2320         struct ceph_osd_client *osdc = event->osdc;
2321
2322         if (!RB_EMPTY_NODE(&event->node)) {
2323                 dout("__remove_event removed %p\n", event);
2324                 rb_erase(&event->node, &osdc->event_tree);
2325                 ceph_osdc_put_event(event);
2326         } else {
2327                 dout("__remove_event didn't remove %p\n", event);
2328         }
2329 }
2330
2331 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2332                            void (*event_cb)(u64, u64, u8, void *),
2333                            void *data, struct ceph_osd_event **pevent)
2334 {
2335         struct ceph_osd_event *event;
2336
2337         event = kmalloc(sizeof(*event), GFP_NOIO);
2338         if (!event)
2339                 return -ENOMEM;
2340
2341         dout("create_event %p\n", event);
2342         event->cb = event_cb;
2343         event->one_shot = 0;
2344         event->data = data;
2345         event->osdc = osdc;
2346         INIT_LIST_HEAD(&event->osd_node);
2347         RB_CLEAR_NODE(&event->node);
2348         kref_init(&event->kref);   /* one ref for us */
2349         kref_get(&event->kref);    /* one ref for the caller */
2350
2351         spin_lock(&osdc->event_lock);
2352         event->cookie = ++osdc->event_count;
2353         __insert_event(osdc, event);
2354         spin_unlock(&osdc->event_lock);
2355
2356         *pevent = event;
2357         return 0;
2358 }
2359 EXPORT_SYMBOL(ceph_osdc_create_event);
2360
2361 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2362 {
2363         struct ceph_osd_client *osdc = event->osdc;
2364
2365         dout("cancel_event %p\n", event);
2366         spin_lock(&osdc->event_lock);
2367         __remove_event(event);
2368         spin_unlock(&osdc->event_lock);
2369         ceph_osdc_put_event(event); /* caller's */
2370 }
2371 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2372
2373
2374 static void do_event_work(struct work_struct *work)
2375 {
2376         struct ceph_osd_event_work *event_work =
2377                 container_of(work, struct ceph_osd_event_work, work);
2378         struct ceph_osd_event *event = event_work->event;
2379         u64 ver = event_work->ver;
2380         u64 notify_id = event_work->notify_id;
2381         u8 opcode = event_work->opcode;
2382
2383         dout("do_event_work completing %p\n", event);
2384         event->cb(ver, notify_id, opcode, event->data);
2385         dout("do_event_work completed %p\n", event);
2386         ceph_osdc_put_event(event);
2387         kfree(event_work);
2388 }
2389
2390
2391 /*
2392  * Process osd watch notifications
2393  */
2394 static void handle_watch_notify(struct ceph_osd_client *osdc,
2395                                 struct ceph_msg *msg)
2396 {
2397         void *p, *end;
2398         u8 proto_ver;
2399         u64 cookie, ver, notify_id;
2400         u8 opcode;
2401         struct ceph_osd_event *event;
2402         struct ceph_osd_event_work *event_work;
2403
2404         p = msg->front.iov_base;
2405         end = p + msg->front.iov_len;
2406
2407         ceph_decode_8_safe(&p, end, proto_ver, bad);
2408         ceph_decode_8_safe(&p, end, opcode, bad);
2409         ceph_decode_64_safe(&p, end, cookie, bad);
2410         ceph_decode_64_safe(&p, end, ver, bad);
2411         ceph_decode_64_safe(&p, end, notify_id, bad);
2412
2413         spin_lock(&osdc->event_lock);
2414         event = __find_event(osdc, cookie);
2415         if (event) {
2416                 BUG_ON(event->one_shot);
2417                 get_event(event);
2418         }
2419         spin_unlock(&osdc->event_lock);
2420         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2421              cookie, ver, event);
2422         if (event) {
2423                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2424                 if (!event_work) {
2425                         pr_err("couldn't allocate event_work\n");
2426                         ceph_osdc_put_event(event);
2427                         return;
2428                 }
2429                 INIT_WORK(&event_work->work, do_event_work);
2430                 event_work->event = event;
2431                 event_work->ver = ver;
2432                 event_work->notify_id = notify_id;
2433                 event_work->opcode = opcode;
2434
2435                 queue_work(osdc->notify_wq, &event_work->work);
2436         }
2437
2438         return;
2439
2440 bad:
2441         pr_err("osdc handle_watch_notify corrupt msg\n");
2442 }
2443
2444 /*
2445  * build new request AND message
2446  *
2447  */
2448 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2449                                 struct ceph_snap_context *snapc, u64 snap_id,
2450                                 struct timespec *mtime)
2451 {
2452         struct ceph_msg *msg = req->r_request;
2453         void *p;
2454         size_t msg_size;
2455         int flags = req->r_flags;
2456         u64 data_len;
2457         unsigned int i;
2458
2459         req->r_snapid = snap_id;
2460         req->r_snapc = ceph_get_snap_context(snapc);
2461
2462         /* encode request */
2463         msg->hdr.version = cpu_to_le16(4);
2464
2465         p = msg->front.iov_base;
2466         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2467         req->r_request_osdmap_epoch = p;
2468         p += 4;
2469         req->r_request_flags = p;
2470         p += 4;
2471         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2472                 ceph_encode_timespec(p, mtime);
2473         p += sizeof(struct ceph_timespec);
2474         req->r_request_reassert_version = p;
2475         p += sizeof(struct ceph_eversion); /* will get filled in */
2476
2477         /* oloc */
2478         ceph_encode_8(&p, 4);
2479         ceph_encode_8(&p, 4);
2480         ceph_encode_32(&p, 8 + 4 + 4);
2481         req->r_request_pool = p;
2482         p += 8;
2483         ceph_encode_32(&p, -1);  /* preferred */
2484         ceph_encode_32(&p, 0);   /* key len */
2485
2486         ceph_encode_8(&p, 1);
2487         req->r_request_pgid = p;
2488         p += 8 + 4;
2489         ceph_encode_32(&p, -1);  /* preferred */
2490
2491         /* oid */
2492         ceph_encode_32(&p, req->r_base_oid.name_len);
2493         memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2494         dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2495              req->r_base_oid.name, req->r_base_oid.name_len);
2496         p += req->r_base_oid.name_len;
2497
2498         /* ops--can imply data */
2499         ceph_encode_16(&p, (u16)req->r_num_ops);
2500         data_len = 0;
2501         for (i = 0; i < req->r_num_ops; i++) {
2502                 data_len += osd_req_encode_op(req, p, i);
2503                 p += sizeof(struct ceph_osd_op);
2504         }
2505
2506         /* snaps */
2507         ceph_encode_64(&p, req->r_snapid);
2508         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2509         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2510         if (req->r_snapc) {
2511                 for (i = 0; i < snapc->num_snaps; i++) {
2512                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2513                 }
2514         }
2515
2516         req->r_request_attempts = p;
2517         p += 4;
2518
2519         /* data */
2520         if (flags & CEPH_OSD_FLAG_WRITE) {
2521                 u16 data_off;
2522
2523                 /*
2524                  * The header "data_off" is a hint to the receiver
2525                  * allowing it to align received data into its
2526                  * buffers such that there's no need to re-copy
2527                  * it before writing it to disk (direct I/O).
2528                  */
2529                 data_off = (u16) (off & 0xffff);
2530                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2531         }
2532         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2533
2534         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2535         msg_size = p - msg->front.iov_base;
2536         msg->front.iov_len = msg_size;
2537         msg->hdr.front_len = cpu_to_le32(msg_size);
2538
2539         dout("build_request msg_size was %d\n", (int)msg_size);
2540 }
2541 EXPORT_SYMBOL(ceph_osdc_build_request);
2542
2543 /*
2544  * Register request, send initial attempt.
2545  */
2546 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2547                             struct ceph_osd_request *req,
2548                             bool nofail)
2549 {
2550         int rc;
2551
2552         down_read(&osdc->map_sem);
2553         mutex_lock(&osdc->request_mutex);
2554
2555         rc = __ceph_osdc_start_request(osdc, req, nofail);
2556
2557         mutex_unlock(&osdc->request_mutex);
2558         up_read(&osdc->map_sem);
2559
2560         return rc;
2561 }
2562 EXPORT_SYMBOL(ceph_osdc_start_request);
2563
2564 /*
2565  * Unregister a registered request.  The request is not completed (i.e.
2566  * no callbacks or wakeups) - higher layers are supposed to know what
2567  * they are canceling.
2568  */
2569 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2570 {
2571         struct ceph_osd_client *osdc = req->r_osdc;
2572
2573         mutex_lock(&osdc->request_mutex);
2574         if (req->r_linger)
2575                 __unregister_linger_request(osdc, req);
2576         __unregister_request(osdc, req);
2577         mutex_unlock(&osdc->request_mutex);
2578
2579         dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2580 }
2581 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2582
2583 /*
2584  * wait for a request to complete
2585  */
2586 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2587                            struct ceph_osd_request *req)
2588 {
2589         int rc;
2590
2591         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2592
2593         rc = wait_for_completion_interruptible(&req->r_completion);
2594         if (rc < 0) {
2595                 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2596                 ceph_osdc_cancel_request(req);
2597                 complete_request(req);
2598                 return rc;
2599         }
2600
2601         dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2602              req->r_result);
2603         return req->r_result;
2604 }
2605 EXPORT_SYMBOL(ceph_osdc_wait_request);
2606
2607 /*
2608  * sync - wait for all in-flight requests to flush.  avoid starvation.
2609  */
2610 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2611 {
2612         struct ceph_osd_request *req;
2613         u64 last_tid, next_tid = 0;
2614
2615         mutex_lock(&osdc->request_mutex);
2616         last_tid = osdc->last_tid;
2617         while (1) {
2618                 req = __lookup_request_ge(osdc, next_tid);
2619                 if (!req)
2620                         break;
2621                 if (req->r_tid > last_tid)
2622                         break;
2623
2624                 next_tid = req->r_tid + 1;
2625                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2626                         continue;
2627
2628                 ceph_osdc_get_request(req);
2629                 mutex_unlock(&osdc->request_mutex);
2630                 dout("sync waiting on tid %llu (last is %llu)\n",
2631                      req->r_tid, last_tid);
2632                 wait_for_completion(&req->r_safe_completion);
2633                 mutex_lock(&osdc->request_mutex);
2634                 ceph_osdc_put_request(req);
2635         }
2636         mutex_unlock(&osdc->request_mutex);
2637         dout("sync done (thru tid %llu)\n", last_tid);
2638 }
2639 EXPORT_SYMBOL(ceph_osdc_sync);
2640
2641 /*
2642  * Call all pending notify callbacks - for use after a watch is
2643  * unregistered, to make sure no more callbacks for it will be invoked
2644  */
2645 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2646 {
2647         flush_workqueue(osdc->notify_wq);
2648 }
2649 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2650
2651
2652 /*
2653  * init, shutdown
2654  */
2655 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2656 {
2657         int err;
2658
2659         dout("init\n");
2660         osdc->client = client;
2661         osdc->osdmap = NULL;
2662         init_rwsem(&osdc->map_sem);
2663         init_completion(&osdc->map_waiters);
2664         osdc->last_requested_map = 0;
2665         mutex_init(&osdc->request_mutex);
2666         osdc->last_tid = 0;
2667         osdc->osds = RB_ROOT;
2668         INIT_LIST_HEAD(&osdc->osd_lru);
2669         osdc->requests = RB_ROOT;
2670         INIT_LIST_HEAD(&osdc->req_lru);
2671         INIT_LIST_HEAD(&osdc->req_unsent);
2672         INIT_LIST_HEAD(&osdc->req_notarget);
2673         INIT_LIST_HEAD(&osdc->req_linger);
2674         osdc->num_requests = 0;
2675         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2676         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2677         spin_lock_init(&osdc->event_lock);
2678         osdc->event_tree = RB_ROOT;
2679         osdc->event_count = 0;
2680
2681         schedule_delayed_work(&osdc->osds_timeout_work,
2682             round_jiffies_relative(osdc->client->options->osd_idle_ttl));
2683
2684         err = -ENOMEM;
2685         osdc->req_mempool = mempool_create_slab_pool(10,
2686                                                      ceph_osd_request_cache);
2687         if (!osdc->req_mempool)
2688                 goto out;
2689
2690         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2691                                 OSD_OP_FRONT_LEN, 10, true,
2692                                 "osd_op");
2693         if (err < 0)
2694                 goto out_mempool;
2695         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2696                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2697                                 "osd_op_reply");
2698         if (err < 0)
2699                 goto out_msgpool;
2700
2701         err = -ENOMEM;
2702         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2703         if (!osdc->notify_wq)
2704                 goto out_msgpool_reply;
2705
2706         return 0;
2707
2708 out_msgpool_reply:
2709         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2710 out_msgpool:
2711         ceph_msgpool_destroy(&osdc->msgpool_op);
2712 out_mempool:
2713         mempool_destroy(osdc->req_mempool);
2714 out:
2715         return err;
2716 }
2717
2718 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2719 {
2720         flush_workqueue(osdc->notify_wq);
2721         destroy_workqueue(osdc->notify_wq);
2722         cancel_delayed_work_sync(&osdc->timeout_work);
2723         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2724         if (osdc->osdmap) {
2725                 ceph_osdmap_destroy(osdc->osdmap);
2726                 osdc->osdmap = NULL;
2727         }
2728         remove_all_osds(osdc);
2729         mempool_destroy(osdc->req_mempool);
2730         ceph_msgpool_destroy(&osdc->msgpool_op);
2731         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2732 }
2733
2734 /*
2735  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2736  * *plen.  Return number of bytes read, or error.
2737  */
2738 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2739                         struct ceph_vino vino, struct ceph_file_layout *layout,
2740                         u64 off, u64 *plen,
2741                         u32 truncate_seq, u64 truncate_size,
2742                         struct page **pages, int num_pages, int page_align)
2743 {
2744         struct ceph_osd_request *req;
2745         int rc = 0;
2746
2747         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2748              vino.snap, off, *plen);
2749         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2750                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2751                                     NULL, truncate_seq, truncate_size,
2752                                     false);
2753         if (IS_ERR(req))
2754                 return PTR_ERR(req);
2755
2756         /* it may be a short read due to an object boundary */
2757
2758         osd_req_op_extent_osd_data_pages(req, 0,
2759                                 pages, *plen, page_align, false, false);
2760
2761         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2762              off, *plen, *plen, page_align);
2763
2764         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2765
2766         rc = ceph_osdc_start_request(osdc, req, false);
2767         if (!rc)
2768                 rc = ceph_osdc_wait_request(osdc, req);
2769
2770         ceph_osdc_put_request(req);
2771         dout("readpages result %d\n", rc);
2772         return rc;
2773 }
2774 EXPORT_SYMBOL(ceph_osdc_readpages);
2775
2776 /*
2777  * do a synchronous write on N pages
2778  */
2779 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2780                          struct ceph_file_layout *layout,
2781                          struct ceph_snap_context *snapc,
2782                          u64 off, u64 len,
2783                          u32 truncate_seq, u64 truncate_size,
2784                          struct timespec *mtime,
2785                          struct page **pages, int num_pages)
2786 {
2787         struct ceph_osd_request *req;
2788         int rc = 0;
2789         int page_align = off & ~PAGE_MASK;
2790
2791         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2792         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2793                                     CEPH_OSD_OP_WRITE,
2794                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2795                                     snapc, truncate_seq, truncate_size,
2796                                     true);
2797         if (IS_ERR(req))
2798                 return PTR_ERR(req);
2799
2800         /* it may be a short write due to an object boundary */
2801         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2802                                 false, false);
2803         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2804
2805         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2806
2807         rc = ceph_osdc_start_request(osdc, req, true);
2808         if (!rc)
2809                 rc = ceph_osdc_wait_request(osdc, req);
2810
2811         ceph_osdc_put_request(req);
2812         if (rc == 0)
2813                 rc = len;
2814         dout("writepages result %d\n", rc);
2815         return rc;
2816 }
2817 EXPORT_SYMBOL(ceph_osdc_writepages);
2818
2819 int ceph_osdc_setup(void)
2820 {
2821         size_t size = sizeof(struct ceph_osd_request) +
2822             CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op);
2823
2824         BUG_ON(ceph_osd_request_cache);
2825         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size,
2826                                                    0, 0, NULL);
2827
2828         return ceph_osd_request_cache ? 0 : -ENOMEM;
2829 }
2830 EXPORT_SYMBOL(ceph_osdc_setup);
2831
2832 void ceph_osdc_cleanup(void)
2833 {
2834         BUG_ON(!ceph_osd_request_cache);
2835         kmem_cache_destroy(ceph_osd_request_cache);
2836         ceph_osd_request_cache = NULL;
2837 }
2838 EXPORT_SYMBOL(ceph_osdc_cleanup);
2839
2840 /*
2841  * handle incoming message
2842  */
2843 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2844 {
2845         struct ceph_osd *osd = con->private;
2846         struct ceph_osd_client *osdc;
2847         int type = le16_to_cpu(msg->hdr.type);
2848
2849         if (!osd)
2850                 goto out;
2851         osdc = osd->o_osdc;
2852
2853         switch (type) {
2854         case CEPH_MSG_OSD_MAP:
2855                 ceph_osdc_handle_map(osdc, msg);
2856                 break;
2857         case CEPH_MSG_OSD_OPREPLY:
2858                 handle_reply(osdc, msg);
2859                 break;
2860         case CEPH_MSG_WATCH_NOTIFY:
2861                 handle_watch_notify(osdc, msg);
2862                 break;
2863
2864         default:
2865                 pr_err("received unknown message type %d %s\n", type,
2866                        ceph_msg_type_name(type));
2867         }
2868 out:
2869         ceph_msg_put(msg);
2870 }
2871
2872 /*
2873  * Lookup and return message for incoming reply.  Don't try to do
2874  * anything about a larger than preallocated data portion of the
2875  * message at the moment - for now, just skip the message.
2876  */
2877 static struct ceph_msg *get_reply(struct ceph_connection *con,
2878                                   struct ceph_msg_header *hdr,
2879                                   int *skip)
2880 {
2881         struct ceph_osd *osd = con->private;
2882         struct ceph_osd_client *osdc = osd->o_osdc;
2883         struct ceph_msg *m;
2884         struct ceph_osd_request *req;
2885         int front_len = le32_to_cpu(hdr->front_len);
2886         int data_len = le32_to_cpu(hdr->data_len);
2887         u64 tid;
2888
2889         tid = le64_to_cpu(hdr->tid);
2890         mutex_lock(&osdc->request_mutex);
2891         req = __lookup_request(osdc, tid);
2892         if (!req) {
2893                 dout("%s osd%d tid %llu unknown, skipping\n", __func__,
2894                      osd->o_osd, tid);
2895                 m = NULL;
2896                 *skip = 1;
2897                 goto out;
2898         }
2899
2900         ceph_msg_revoke_incoming(req->r_reply);
2901
2902         if (front_len > req->r_reply->front_alloc_len) {
2903                 pr_warn("%s osd%d tid %llu front %d > preallocated %d\n",
2904                         __func__, osd->o_osd, req->r_tid, front_len,
2905                         req->r_reply->front_alloc_len);
2906                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2907                                  false);
2908                 if (!m)
2909                         goto out;
2910                 ceph_msg_put(req->r_reply);
2911                 req->r_reply = m;
2912         }
2913
2914         if (data_len > req->r_reply->data_length) {
2915                 pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n",
2916                         __func__, osd->o_osd, req->r_tid, data_len,
2917                         req->r_reply->data_length);
2918                 m = NULL;
2919                 *skip = 1;
2920                 goto out;
2921         }
2922
2923         m = ceph_msg_get(req->r_reply);
2924         dout("get_reply tid %lld %p\n", tid, m);
2925
2926 out:
2927         mutex_unlock(&osdc->request_mutex);
2928         return m;
2929 }
2930
2931 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2932                                   struct ceph_msg_header *hdr,
2933                                   int *skip)
2934 {
2935         struct ceph_osd *osd = con->private;
2936         int type = le16_to_cpu(hdr->type);
2937         int front = le32_to_cpu(hdr->front_len);
2938
2939         *skip = 0;
2940         switch (type) {
2941         case CEPH_MSG_OSD_MAP:
2942         case CEPH_MSG_WATCH_NOTIFY:
2943                 return ceph_msg_new(type, front, GFP_NOFS, false);
2944         case CEPH_MSG_OSD_OPREPLY:
2945                 return get_reply(con, hdr, skip);
2946         default:
2947                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2948                         osd->o_osd);
2949                 *skip = 1;
2950                 return NULL;
2951         }
2952 }
2953
2954 /*
2955  * Wrappers to refcount containing ceph_osd struct
2956  */
2957 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2958 {
2959         struct ceph_osd *osd = con->private;
2960         if (get_osd(osd))
2961                 return con;
2962         return NULL;
2963 }
2964
2965 static void put_osd_con(struct ceph_connection *con)
2966 {
2967         struct ceph_osd *osd = con->private;
2968         put_osd(osd);
2969 }
2970
2971 /*
2972  * authentication
2973  */
2974 /*
2975  * Note: returned pointer is the address of a structure that's
2976  * managed separately.  Caller must *not* attempt to free it.
2977  */
2978 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2979                                         int *proto, int force_new)
2980 {
2981         struct ceph_osd *o = con->private;
2982         struct ceph_osd_client *osdc = o->o_osdc;
2983         struct ceph_auth_client *ac = osdc->client->monc.auth;
2984         struct ceph_auth_handshake *auth = &o->o_auth;
2985
2986         if (force_new && auth->authorizer) {
2987                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2988                 auth->authorizer = NULL;
2989         }
2990         if (!auth->authorizer) {
2991                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2992                                                       auth);
2993                 if (ret)
2994                         return ERR_PTR(ret);
2995         } else {
2996                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2997                                                      auth);
2998                 if (ret)
2999                         return ERR_PTR(ret);
3000         }
3001         *proto = ac->protocol;
3002
3003         return auth;
3004 }
3005
3006
3007 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3008 {
3009         struct ceph_osd *o = con->private;
3010         struct ceph_osd_client *osdc = o->o_osdc;
3011         struct ceph_auth_client *ac = osdc->client->monc.auth;
3012
3013         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
3014 }
3015
3016 static int invalidate_authorizer(struct ceph_connection *con)
3017 {
3018         struct ceph_osd *o = con->private;
3019         struct ceph_osd_client *osdc = o->o_osdc;
3020         struct ceph_auth_client *ac = osdc->client->monc.auth;
3021
3022         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
3023         return ceph_monc_validate_auth(&osdc->client->monc);
3024 }
3025
3026 static int osd_sign_message(struct ceph_msg *msg)
3027 {
3028         struct ceph_osd *o = msg->con->private;
3029         struct ceph_auth_handshake *auth = &o->o_auth;
3030
3031         return ceph_auth_sign_message(auth, msg);
3032 }
3033
3034 static int osd_check_message_signature(struct ceph_msg *msg)
3035 {
3036         struct ceph_osd *o = msg->con->private;
3037         struct ceph_auth_handshake *auth = &o->o_auth;
3038
3039         return ceph_auth_check_message_signature(auth, msg);
3040 }
3041
3042 static const struct ceph_connection_operations osd_con_ops = {
3043         .get = get_osd_con,
3044         .put = put_osd_con,
3045         .dispatch = dispatch,
3046         .get_authorizer = get_authorizer,
3047         .verify_authorizer_reply = verify_authorizer_reply,
3048         .invalidate_authorizer = invalidate_authorizer,
3049         .alloc_msg = alloc_msg,
3050         .sign_message = osd_sign_message,
3051         .check_message_signature = osd_check_message_signature,
3052         .fault = osd_reset,
3053 };