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