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_size <= object_base) {
 737                 truncate_size = 0;
 738         } else {
 739                 truncate_size -= object_base;
 740                 if (truncate_size > object_size)
 741                         truncate_size = object_size;
 742         }
 743
 744         osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
 745                                 truncate_size, truncate_seq);
 746
 747         /*
 748          * A second op in the ops array means the caller wants to
 749          * also issue a include a 'startsync' command so that the
 750          * osd will flush data quickly.
 751          */
 752         if (num_ops > 1)
 753                 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
 754
 755         req->r_file_layout = *layout;  /* keep a copy */
 756
 757         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx",
 758                 vino.ino, objnum);
 759         req->r_oid_len = strlen(req->r_oid);
 760
 761         return req;
 762 }
 763 EXPORT_SYMBOL(ceph_osdc_new_request);
 764
 765 /*
 766  * We keep osd requests in an rbtree, sorted by ->r_tid.
 767  */
 768 static void __insert_request(struct ceph_osd_client *osdc,
 769                              struct ceph_osd_request *new)
 770 {
 771         struct rb_node **p = &osdc->requests.rb_node;
 772         struct rb_node *parent = NULL;
 773         struct ceph_osd_request *req = NULL;
 774
 775         while (*p) {
 776                 parent = *p;
 777                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 778                 if (new->r_tid < req->r_tid)
 779                         p = &(*p)->rb_left;
 780                 else if (new->r_tid > req->r_tid)
 781                         p = &(*p)->rb_right;
 782                 else
 783                         BUG();
 784         }
 785
 786         rb_link_node(&new->r_node, parent, p);
 787         rb_insert_color(&new->r_node, &osdc->requests);
 788 }
 789
 790 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 791                                                  u64 tid)
 792 {
 793         struct ceph_osd_request *req;
 794         struct rb_node *n = osdc->requests.rb_node;
 795
 796         while (n) {
 797                 req = rb_entry(n, struct ceph_osd_request, r_node);
 798                 if (tid < req->r_tid)
 799                         n = n->rb_left;
 800                 else if (tid > req->r_tid)
 801                         n = n->rb_right;
 802                 else
 803                         return req;
 804         }
 805         return NULL;
 806 }
 807
 808 static struct ceph_osd_request *
 809 __lookup_request_ge(struct ceph_osd_client *osdc,
 810                     u64 tid)
 811 {
 812         struct ceph_osd_request *req;
 813         struct rb_node *n = osdc->requests.rb_node;
 814
 815         while (n) {
 816                 req = rb_entry(n, struct ceph_osd_request, r_node);
 817                 if (tid < req->r_tid) {
 818                         if (!n->rb_left)
 819                                 return req;
 820                         n = n->rb_left;
 821                 } else if (tid > req->r_tid) {
 822                         n = n->rb_right;
 823                 } else {
 824                         return req;
 825                 }
 826         }
 827         return NULL;
 828 }
 829
 830 /*
 831  * Resubmit requests pending on the given osd.
 832  */
 833 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 834                                 struct ceph_osd *osd)
 835 {
 836         struct ceph_osd_request *req, *nreq;
 837         LIST_HEAD(resend);
 838         int err;
 839
 840         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 841         err = __reset_osd(osdc, osd);
 842         if (err)
 843                 return;
 844         /*
 845          * Build up a list of requests to resend by traversing the
 846          * osd's list of requests.  Requests for a given object are
 847          * sent in tid order, and that is also the order they're
 848          * kept on this list.  Therefore all requests that are in
 849          * flight will be found first, followed by all requests that
 850          * have not yet been sent.  And to resend requests while
 851          * preserving this order we will want to put any sent
 852          * requests back on the front of the osd client's unsent
 853          * list.
 854          *
 855          * So we build a separate ordered list of already-sent
 856          * requests for the affected osd and splice it onto the
 857          * front of the osd client's unsent list.  Once we've seen a
 858          * request that has not yet been sent we're done.  Those
 859          * requests are already sitting right where they belong.
 860          */
 861         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 862                 if (!req->r_sent)
 863                         break;
 864                 list_move_tail(&req->r_req_lru_item, &resend);
 865                 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
 866                      osd->o_osd);
 867                 if (!req->r_linger)
 868                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 869         }
 870         list_splice(&resend, &osdc->req_unsent);
 871
 872         /*
 873          * Linger requests are re-registered before sending, which
 874          * sets up a new tid for each.  We add them to the unsent
 875          * list at the end to keep things in tid order.
 876          */
 877         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 878                                  r_linger_osd) {
 879                 /*
 880                  * reregister request prior to unregistering linger so
 881                  * that r_osd is preserved.
 882                  */
 883                 BUG_ON(!list_empty(&req->r_req_lru_item));
 884                 __register_request(osdc, req);
 885                 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
 886                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
 887                 __unregister_linger_request(osdc, req);
 888                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 889                      osd->o_osd);
 890         }
 891 }
 892
 893 /*
 894  * If the osd connection drops, we need to resubmit all requests.
 895  */
 896 static void osd_reset(struct ceph_connection *con)
 897 {
 898         struct ceph_osd *osd = con->private;
 899         struct ceph_osd_client *osdc;
 900
 901         if (!osd)
 902                 return;
 903         dout("osd_reset osd%d\n", osd->o_osd);
 904         osdc = osd->o_osdc;
 905         down_read(&osdc->map_sem);
 906         mutex_lock(&osdc->request_mutex);
 907         __kick_osd_requests(osdc, osd);
 908         __send_queued(osdc);
 909         mutex_unlock(&osdc->request_mutex);
 910         up_read(&osdc->map_sem);
 911 }
 912
 913 /*
 914  * Track open sessions with osds.
 915  */
 916 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 917 {
 918         struct ceph_osd *osd;
 919
 920         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 921         if (!osd)
 922                 return NULL;
 923
 924         atomic_set(&osd->o_ref, 1);
 925         osd->o_osdc = osdc;
 926         osd->o_osd = onum;
 927         RB_CLEAR_NODE(&osd->o_node);
 928         INIT_LIST_HEAD(&osd->o_requests);
 929         INIT_LIST_HEAD(&osd->o_linger_requests);
 930         INIT_LIST_HEAD(&osd->o_osd_lru);
 931         osd->o_incarnation = 1;
 932
 933         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 934
 935         INIT_LIST_HEAD(&osd->o_keepalive_item);
 936         return osd;
 937 }
 938
 939 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 940 {
 941         if (atomic_inc_not_zero(&osd->o_ref)) {
 942                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 943                      atomic_read(&osd->o_ref));
 944                 return osd;
 945         } else {
 946                 dout("get_osd %p FAIL\n", osd);
 947                 return NULL;
 948         }
 949 }
 950
 951 static void put_osd(struct ceph_osd *osd)
 952 {
 953         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 954              atomic_read(&osd->o_ref) - 1);
 955         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 956                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 957
 958                 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
 959                 kfree(osd);
 960         }
 961 }
 962
 963 /*
 964  * remove an osd from our map
 965  */
 966 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 967 {
 968         dout("__remove_osd %p\n", osd);
 969         BUG_ON(!list_empty(&osd->o_requests));
 970         rb_erase(&osd->o_node, &osdc->osds);
 971         list_del_init(&osd->o_osd_lru);
 972         ceph_con_close(&osd->o_con);
 973         put_osd(osd);
 974 }
 975
 976 static void remove_all_osds(struct ceph_osd_client *osdc)
 977 {
 978         dout("%s %p\n", __func__, osdc);
 979         mutex_lock(&osdc->request_mutex);
 980         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 981                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 982                                                 struct ceph_osd, o_node);
 983                 __remove_osd(osdc, osd);
 984         }
 985         mutex_unlock(&osdc->request_mutex);
 986 }
 987
 988 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 989                               struct ceph_osd *osd)
 990 {
 991         dout("__move_osd_to_lru %p\n", osd);
 992         BUG_ON(!list_empty(&osd->o_osd_lru));
 993         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 994         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 995 }
 996
 997 static void __remove_osd_from_lru(struct ceph_osd *osd)
 998 {
 999         dout("__remove_osd_from_lru %p\n", osd);
1000         if (!list_empty(&osd->o_osd_lru))
1001                 list_del_init(&osd->o_osd_lru);
1002 }
1003
1004 static void remove_old_osds(struct ceph_osd_client *osdc)
1005 {
1006         struct ceph_osd *osd, *nosd;
1007
1008         dout("__remove_old_osds %p\n", osdc);
1009         mutex_lock(&osdc->request_mutex);
1010         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1011                 if (time_before(jiffies, osd->lru_ttl))
1012                         break;
1013                 __remove_osd(osdc, osd);
1014         }
1015         mutex_unlock(&osdc->request_mutex);
1016 }
1017
1018 /*
1019  * reset osd connect
1020  */
1021 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1022 {
1023         struct ceph_entity_addr *peer_addr;
1024
1025         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1026         if (list_empty(&osd->o_requests) &&
1027             list_empty(&osd->o_linger_requests)) {
1028                 __remove_osd(osdc, osd);
1029
1030                 return -ENODEV;
1031         }
1032
1033         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1034         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1035                         !ceph_con_opened(&osd->o_con)) {
1036                 struct ceph_osd_request *req;
1037
1038                 dout(" osd addr hasn't changed and connection never opened,"
1039                      " letting msgr retry");
1040                 /* touch each r_stamp for handle_timeout()'s benfit */
1041                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1042                         req->r_stamp = jiffies;
1043
1044                 return -EAGAIN;
1045         }
1046
1047         ceph_con_close(&osd->o_con);
1048         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1049         osd->o_incarnation++;
1050
1051         return 0;
1052 }
1053
1054 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1055 {
1056         struct rb_node **p = &osdc->osds.rb_node;
1057         struct rb_node *parent = NULL;
1058         struct ceph_osd *osd = NULL;
1059
1060         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1061         while (*p) {
1062                 parent = *p;
1063                 osd = rb_entry(parent, struct ceph_osd, o_node);
1064                 if (new->o_osd < osd->o_osd)
1065                         p = &(*p)->rb_left;
1066                 else if (new->o_osd > osd->o_osd)
1067                         p = &(*p)->rb_right;
1068                 else
1069                         BUG();
1070         }
1071
1072         rb_link_node(&new->o_node, parent, p);
1073         rb_insert_color(&new->o_node, &osdc->osds);
1074 }
1075
1076 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1077 {
1078         struct ceph_osd *osd;
1079         struct rb_node *n = osdc->osds.rb_node;
1080
1081         while (n) {
1082                 osd = rb_entry(n, struct ceph_osd, o_node);
1083                 if (o < osd->o_osd)
1084                         n = n->rb_left;
1085                 else if (o > osd->o_osd)
1086                         n = n->rb_right;
1087                 else
1088                         return osd;
1089         }
1090         return NULL;
1091 }
1092
1093 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1094 {
1095         schedule_delayed_work(&osdc->timeout_work,
1096                         osdc->client->options->osd_keepalive_timeout * HZ);
1097 }
1098
1099 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1100 {
1101         cancel_delayed_work(&osdc->timeout_work);
1102 }
1103
1104 /*
1105  * Register request, assign tid.  If this is the first request, set up
1106  * the timeout event.
1107  */
1108 static void __register_request(struct ceph_osd_client *osdc,
1109                                struct ceph_osd_request *req)
1110 {
1111         req->r_tid = ++osdc->last_tid;
1112         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1113         dout("__register_request %p tid %lld\n", req, req->r_tid);
1114         __insert_request(osdc, req);
1115         ceph_osdc_get_request(req);
1116         osdc->num_requests++;
1117         if (osdc->num_requests == 1) {
1118                 dout(" first request, scheduling timeout\n");
1119                 __schedule_osd_timeout(osdc);
1120         }
1121 }
1122
1123 /*
1124  * called under osdc->request_mutex
1125  */
1126 static void __unregister_request(struct ceph_osd_client *osdc,
1127                                  struct ceph_osd_request *req)
1128 {
1129         if (RB_EMPTY_NODE(&req->r_node)) {
1130                 dout("__unregister_request %p tid %lld not registered\n",
1131                         req, req->r_tid);
1132                 return;
1133         }
1134
1135         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1136         rb_erase(&req->r_node, &osdc->requests);
1137         osdc->num_requests--;
1138
1139         if (req->r_osd) {
1140                 /* make sure the original request isn't in flight. */
1141                 ceph_msg_revoke(req->r_request);
1142
1143                 list_del_init(&req->r_osd_item);
1144                 if (list_empty(&req->r_osd->o_requests) &&
1145                     list_empty(&req->r_osd->o_linger_requests)) {
1146                         dout("moving osd to %p lru\n", req->r_osd);
1147                         __move_osd_to_lru(osdc, req->r_osd);
1148                 }
1149                 if (list_empty(&req->r_linger_item))
1150                         req->r_osd = NULL;
1151         }
1152
1153         list_del_init(&req->r_req_lru_item);
1154         ceph_osdc_put_request(req);
1155
1156         if (osdc->num_requests == 0) {
1157                 dout(" no requests, canceling timeout\n");
1158                 __cancel_osd_timeout(osdc);
1159         }
1160 }
1161
1162 /*
1163  * Cancel a previously queued request message
1164  */
1165 static void __cancel_request(struct ceph_osd_request *req)
1166 {
1167         if (req->r_sent && req->r_osd) {
1168                 ceph_msg_revoke(req->r_request);
1169                 req->r_sent = 0;
1170         }
1171 }
1172
1173 static void __register_linger_request(struct ceph_osd_client *osdc,
1174                                     struct ceph_osd_request *req)
1175 {
1176         dout("__register_linger_request %p\n", req);
1177         ceph_osdc_get_request(req);
1178         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1179         if (req->r_osd)
1180                 list_add_tail(&req->r_linger_osd,
1181                               &req->r_osd->o_linger_requests);
1182 }
1183
1184 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1185                                         struct ceph_osd_request *req)
1186 {
1187         dout("__unregister_linger_request %p\n", req);
1188         list_del_init(&req->r_linger_item);
1189         if (req->r_osd) {
1190                 list_del_init(&req->r_linger_osd);
1191
1192                 if (list_empty(&req->r_osd->o_requests) &&
1193                     list_empty(&req->r_osd->o_linger_requests)) {
1194                         dout("moving osd to %p lru\n", req->r_osd);
1195                         __move_osd_to_lru(osdc, req->r_osd);
1196                 }
1197                 if (list_empty(&req->r_osd_item))
1198                         req->r_osd = NULL;
1199         }
1200         ceph_osdc_put_request(req);
1201 }
1202
1203 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1204                                          struct ceph_osd_request *req)
1205 {
1206         mutex_lock(&osdc->request_mutex);
1207         if (req->r_linger) {
1208                 req->r_linger = 0;
1209                 __unregister_linger_request(osdc, req);
1210         }
1211         mutex_unlock(&osdc->request_mutex);
1212 }
1213 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1214
1215 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1216                                   struct ceph_osd_request *req)
1217 {
1218         if (!req->r_linger) {
1219                 dout("set_request_linger %p\n", req);
1220                 req->r_linger = 1;
1221         }
1222 }
1223 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1224
1225 /*
1226  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1227  * (as needed), and set the request r_osd appropriately.  If there is
1228  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1229  * (unsent, homeless) or leave on in-flight lru.
1230  *
1231  * Return 0 if unchanged, 1 if changed, or negative on error.
1232  *
1233  * Caller should hold map_sem for read and request_mutex.
1234  */
1235 static int __map_request(struct ceph_osd_client *osdc,
1236                          struct ceph_osd_request *req, int force_resend)
1237 {
1238         struct ceph_pg pgid;
1239         int acting[CEPH_PG_MAX_SIZE];
1240         int o = -1, num = 0;
1241         int err;
1242
1243         dout("map_request %p tid %lld\n", req, req->r_tid);
1244         err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap,
1245                                 ceph_file_layout_pg_pool(req->r_file_layout));
1246         if (err) {
1247                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1248                 return err;
1249         }
1250         req->r_pgid = pgid;
1251
1252         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
1253         if (err > 0) {
1254                 o = acting[0];
1255                 num = err;
1256         }
1257
1258         if ((!force_resend &&
1259              req->r_osd && req->r_osd->o_osd == o &&
1260              req->r_sent >= req->r_osd->o_incarnation &&
1261              req->r_num_pg_osds == num &&
1262              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1263             (req->r_osd == NULL && o == -1))
1264                 return 0;  /* no change */
1265
1266         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1267              req->r_tid, pgid.pool, pgid.seed, o,
1268              req->r_osd ? req->r_osd->o_osd : -1);
1269
1270         /* record full pg acting set */
1271         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1272         req->r_num_pg_osds = num;
1273
1274         if (req->r_osd) {
1275                 __cancel_request(req);
1276                 list_del_init(&req->r_osd_item);
1277                 req->r_osd = NULL;
1278         }
1279
1280         req->r_osd = __lookup_osd(osdc, o);
1281         if (!req->r_osd && o >= 0) {
1282                 err = -ENOMEM;
1283                 req->r_osd = create_osd(osdc, o);
1284                 if (!req->r_osd) {
1285                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1286                         goto out;
1287                 }
1288
1289                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1290                 __insert_osd(osdc, req->r_osd);
1291
1292                 ceph_con_open(&req->r_osd->o_con,
1293                               CEPH_ENTITY_TYPE_OSD, o,
1294                               &osdc->osdmap->osd_addr[o]);
1295         }
1296
1297         if (req->r_osd) {
1298                 __remove_osd_from_lru(req->r_osd);
1299                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1300                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1301         } else {
1302                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1303         }
1304         err = 1;   /* osd or pg changed */
1305
1306 out:
1307         return err;
1308 }
1309
1310 /*
1311  * caller should hold map_sem (for read) and request_mutex
1312  */
1313 static void __send_request(struct ceph_osd_client *osdc,
1314                            struct ceph_osd_request *req)
1315 {
1316         void *p;
1317
1318         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1319              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1320              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1321
1322         /* fill in message content that changes each time we send it */
1323         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1324         put_unaligned_le32(req->r_flags, req->r_request_flags);
1325         put_unaligned_le64(req->r_pgid.pool, req->r_request_pool);
1326         p = req->r_request_pgid;
1327         ceph_encode_64(&p, req->r_pgid.pool);
1328         ceph_encode_32(&p, req->r_pgid.seed);
1329         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1330         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1331                sizeof(req->r_reassert_version));
1332
1333         req->r_stamp = jiffies;
1334         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1335
1336         ceph_msg_get(req->r_request); /* send consumes a ref */
1337
1338         /* Mark the request unsafe if this is the first timet's being sent. */
1339
1340         if (!req->r_sent && req->r_unsafe_callback)
1341                 req->r_unsafe_callback(req, true);
1342         req->r_sent = req->r_osd->o_incarnation;
1343
1344         ceph_con_send(&req->r_osd->o_con, req->r_request);
1345 }
1346
1347 /*
1348  * Send any requests in the queue (req_unsent).
1349  */
1350 static void __send_queued(struct ceph_osd_client *osdc)
1351 {
1352         struct ceph_osd_request *req, *tmp;
1353
1354         dout("__send_queued\n");
1355         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1356                 __send_request(osdc, req);
1357 }
1358
1359 /*
1360  * Timeout callback, called every N seconds when 1 or more osd
1361  * requests has been active for more than N seconds.  When this
1362  * happens, we ping all OSDs with requests who have timed out to
1363  * ensure any communications channel reset is detected.  Reset the
1364  * request timeouts another N seconds in the future as we go.
1365  * Reschedule the timeout event another N seconds in future (unless
1366  * there are no open requests).
1367  */
1368 static void handle_timeout(struct work_struct *work)
1369 {
1370         struct ceph_osd_client *osdc =
1371                 container_of(work, struct ceph_osd_client, timeout_work.work);
1372         struct ceph_osd_request *req;
1373         struct ceph_osd *osd;
1374         unsigned long keepalive =
1375                 osdc->client->options->osd_keepalive_timeout * HZ;
1376         struct list_head slow_osds;
1377         dout("timeout\n");
1378         down_read(&osdc->map_sem);
1379
1380         ceph_monc_request_next_osdmap(&osdc->client->monc);
1381
1382         mutex_lock(&osdc->request_mutex);
1383
1384         /*
1385          * ping osds that are a bit slow.  this ensures that if there
1386          * is a break in the TCP connection we will notice, and reopen
1387          * a connection with that osd (from the fault callback).
1388          */
1389         INIT_LIST_HEAD(&slow_osds);
1390         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1391                 if (time_before(jiffies, req->r_stamp + keepalive))
1392                         break;
1393
1394                 osd = req->r_osd;
1395                 BUG_ON(!osd);
1396                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1397                      req->r_tid, osd->o_osd);
1398                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1399         }
1400         while (!list_empty(&slow_osds)) {
1401                 osd = list_entry(slow_osds.next, struct ceph_osd,
1402                                  o_keepalive_item);
1403                 list_del_init(&osd->o_keepalive_item);
1404                 ceph_con_keepalive(&osd->o_con);
1405         }
1406
1407         __schedule_osd_timeout(osdc);
1408         __send_queued(osdc);
1409         mutex_unlock(&osdc->request_mutex);
1410         up_read(&osdc->map_sem);
1411 }
1412
1413 static void handle_osds_timeout(struct work_struct *work)
1414 {
1415         struct ceph_osd_client *osdc =
1416                 container_of(work, struct ceph_osd_client,
1417                              osds_timeout_work.work);
1418         unsigned long delay =
1419                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1420
1421         dout("osds timeout\n");
1422         down_read(&osdc->map_sem);
1423         remove_old_osds(osdc);
1424         up_read(&osdc->map_sem);
1425
1426         schedule_delayed_work(&osdc->osds_timeout_work,
1427                               round_jiffies_relative(delay));
1428 }
1429
1430 static void complete_request(struct ceph_osd_request *req)
1431 {
1432         if (req->r_unsafe_callback)
1433                 req->r_unsafe_callback(req, false);
1434         complete_all(&req->r_safe_completion);  /* fsync waiter */
1435 }
1436
1437 /*
1438  * handle osd op reply.  either call the callback if it is specified,
1439  * or do the completion to wake up the waiting thread.
1440  */
1441 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1442                          struct ceph_connection *con)
1443 {
1444         void *p, *end;
1445         struct ceph_osd_request *req;
1446         u64 tid;
1447         int object_len;
1448         unsigned int numops;
1449         int payload_len, flags;
1450         s32 result;
1451         s32 retry_attempt;
1452         struct ceph_pg pg;
1453         int err;
1454         u32 reassert_epoch;
1455         u64 reassert_version;
1456         u32 osdmap_epoch;
1457         int already_completed;
1458         u32 bytes;
1459         unsigned int i;
1460
1461         tid = le64_to_cpu(msg->hdr.tid);
1462         dout("handle_reply %p tid %llu\n", msg, tid);
1463
1464         p = msg->front.iov_base;
1465         end = p + msg->front.iov_len;
1466
1467         ceph_decode_need(&p, end, 4, bad);
1468         object_len = ceph_decode_32(&p);
1469         ceph_decode_need(&p, end, object_len, bad);
1470         p += object_len;
1471
1472         err = ceph_decode_pgid(&p, end, &pg);
1473         if (err)
1474                 goto bad;
1475
1476         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1477         flags = ceph_decode_64(&p);
1478         result = ceph_decode_32(&p);
1479         reassert_epoch = ceph_decode_32(&p);
1480         reassert_version = ceph_decode_64(&p);
1481         osdmap_epoch = ceph_decode_32(&p);
1482
1483         /* lookup */
1484         mutex_lock(&osdc->request_mutex);
1485         req = __lookup_request(osdc, tid);
1486         if (req == NULL) {
1487                 dout("handle_reply tid %llu dne\n", tid);
1488                 goto bad_mutex;
1489         }
1490         ceph_osdc_get_request(req);
1491
1492         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1493              req, result);
1494
1495         ceph_decode_need(&p, end, 4, bad);
1496         numops = ceph_decode_32(&p);
1497         if (numops > CEPH_OSD_MAX_OP)
1498                 goto bad_put;
1499         if (numops != req->r_num_ops)
1500                 goto bad_put;
1501         payload_len = 0;
1502         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad);
1503         for (i = 0; i < numops; i++) {
1504                 struct ceph_osd_op *op = p;
1505                 int len;
1506
1507                 len = le32_to_cpu(op->payload_len);
1508                 req->r_reply_op_len[i] = len;
1509                 dout(" op %d has %d bytes\n", i, len);
1510                 payload_len += len;
1511                 p += sizeof(*op);
1512         }
1513         bytes = le32_to_cpu(msg->hdr.data_len);
1514         if (payload_len != bytes) {
1515                 pr_warning("sum of op payload lens %d != data_len %d",
1516                            payload_len, bytes);
1517                 goto bad_put;
1518         }
1519
1520         ceph_decode_need(&p, end, 4 + numops * 4, bad);
1521         retry_attempt = ceph_decode_32(&p);
1522         for (i = 0; i < numops; i++)
1523                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1524
1525         if (!req->r_got_reply) {
1526
1527                 req->r_result = result;
1528                 dout("handle_reply result %d bytes %d\n", req->r_result,
1529                      bytes);
1530                 if (req->r_result == 0)
1531                         req->r_result = bytes;
1532
1533                 /* in case this is a write and we need to replay, */
1534                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1535                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1536
1537                 req->r_got_reply = 1;
1538         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1539                 dout("handle_reply tid %llu dup ack\n", tid);
1540                 mutex_unlock(&osdc->request_mutex);
1541                 goto done;
1542         }
1543
1544         dout("handle_reply tid %llu flags %d\n", tid, flags);
1545
1546         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1547                 __register_linger_request(osdc, req);
1548
1549         /* either this is a read, or we got the safe response */
1550         if (result < 0 ||
1551             (flags & CEPH_OSD_FLAG_ONDISK) ||
1552             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1553                 __unregister_request(osdc, req);
1554
1555         already_completed = req->r_completed;
1556         req->r_completed = 1;
1557         mutex_unlock(&osdc->request_mutex);
1558         if (already_completed)
1559                 goto done;
1560
1561         if (req->r_callback)
1562                 req->r_callback(req, msg);
1563         else
1564                 complete_all(&req->r_completion);
1565
1566         if (flags & CEPH_OSD_FLAG_ONDISK)
1567                 complete_request(req);
1568
1569 done:
1570         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1571         ceph_osdc_put_request(req);
1572         return;
1573
1574 bad_put:
1575         ceph_osdc_put_request(req);
1576 bad_mutex:
1577         mutex_unlock(&osdc->request_mutex);
1578 bad:
1579         pr_err("corrupt osd_op_reply got %d %d\n",
1580                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1581         ceph_msg_dump(msg);
1582 }
1583
1584 static void reset_changed_osds(struct ceph_osd_client *osdc)
1585 {
1586         struct rb_node *p, *n;
1587
1588         for (p = rb_first(&osdc->osds); p; p = n) {
1589                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1590
1591                 n = rb_next(p);
1592                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1593                     memcmp(&osd->o_con.peer_addr,
1594                            ceph_osd_addr(osdc->osdmap,
1595                                          osd->o_osd),
1596                            sizeof(struct ceph_entity_addr)) != 0)
1597                         __reset_osd(osdc, osd);
1598         }
1599 }
1600
1601 /*
1602  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1603  * no osd, request a new map.
1604  *
1605  * Caller should hold map_sem for read.
1606  */
1607 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1608 {
1609         struct ceph_osd_request *req, *nreq;
1610         struct rb_node *p;
1611         int needmap = 0;
1612         int err;
1613
1614         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1615         mutex_lock(&osdc->request_mutex);
1616         for (p = rb_first(&osdc->requests); p; ) {
1617                 req = rb_entry(p, struct ceph_osd_request, r_node);
1618                 p = rb_next(p);
1619
1620                 /*
1621                  * For linger requests that have not yet been
1622                  * registered, move them to the linger list; they'll
1623                  * be sent to the osd in the loop below.  Unregister
1624                  * the request before re-registering it as a linger
1625                  * request to ensure the __map_request() below
1626                  * will decide it needs to be sent.
1627                  */
1628                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1629                         dout("%p tid %llu restart on osd%d\n",
1630                              req, req->r_tid,
1631                              req->r_osd ? req->r_osd->o_osd : -1);
1632                         ceph_osdc_get_request(req);
1633                         __unregister_request(osdc, req);
1634                         __register_linger_request(osdc, req);
1635                         ceph_osdc_put_request(req);
1636                         continue;
1637                 }
1638
1639                 err = __map_request(osdc, req, force_resend);
1640                 if (err < 0)
1641                         continue;  /* error */
1642                 if (req->r_osd == NULL) {
1643                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1644                         needmap++;  /* request a newer map */
1645                 } else if (err > 0) {
1646                         if (!req->r_linger) {
1647                                 dout("%p tid %llu requeued on osd%d\n", req,
1648                                      req->r_tid,
1649                                      req->r_osd ? req->r_osd->o_osd : -1);
1650                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1651                         }
1652                 }
1653         }
1654
1655         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1656                                  r_linger_item) {
1657                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1658
1659                 err = __map_request(osdc, req, force_resend);
1660                 dout("__map_request returned %d\n", err);
1661                 if (err == 0)
1662                         continue;  /* no change and no osd was specified */
1663                 if (err < 0)
1664                         continue;  /* hrm! */
1665                 if (req->r_osd == NULL) {
1666                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1667                         needmap++;  /* request a newer map */
1668                         continue;
1669                 }
1670
1671                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1672                      req->r_osd ? req->r_osd->o_osd : -1);
1673                 __register_request(osdc, req);
1674                 __unregister_linger_request(osdc, req);
1675         }
1676         reset_changed_osds(osdc);
1677         mutex_unlock(&osdc->request_mutex);
1678
1679         if (needmap) {
1680                 dout("%d requests for down osds, need new map\n", needmap);
1681                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1682         }
1683 }
1684
1685
1686 /*
1687  * Process updated osd map.
1688  *
1689  * The message contains any number of incremental and full maps, normally
1690  * indicating some sort of topology change in the cluster.  Kick requests
1691  * off to different OSDs as needed.
1692  */
1693 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1694 {
1695         void *p, *end, *next;
1696         u32 nr_maps, maplen;
1697         u32 epoch;
1698         struct ceph_osdmap *newmap = NULL, *oldmap;
1699         int err;
1700         struct ceph_fsid fsid;
1701
1702         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1703         p = msg->front.iov_base;
1704         end = p + msg->front.iov_len;
1705
1706         /* verify fsid */
1707         ceph_decode_need(&p, end, sizeof(fsid), bad);
1708         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1709         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1710                 return;
1711
1712         down_write(&osdc->map_sem);
1713
1714         /* incremental maps */
1715         ceph_decode_32_safe(&p, end, nr_maps, bad);
1716         dout(" %d inc maps\n", nr_maps);
1717         while (nr_maps > 0) {
1718                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1719                 epoch = ceph_decode_32(&p);
1720                 maplen = ceph_decode_32(&p);
1721                 ceph_decode_need(&p, end, maplen, bad);
1722                 next = p + maplen;
1723                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1724                         dout("applying incremental map %u len %d\n",
1725                              epoch, maplen);
1726                         newmap = osdmap_apply_incremental(&p, next,
1727                                                           osdc->osdmap,
1728                                                           &osdc->client->msgr);
1729                         if (IS_ERR(newmap)) {
1730                                 err = PTR_ERR(newmap);
1731                                 goto bad;
1732                         }
1733                         BUG_ON(!newmap);
1734                         if (newmap != osdc->osdmap) {
1735                                 ceph_osdmap_destroy(osdc->osdmap);
1736                                 osdc->osdmap = newmap;
1737                         }
1738                         kick_requests(osdc, 0);
1739                 } else {
1740                         dout("ignoring incremental map %u len %d\n",
1741                              epoch, maplen);
1742                 }
1743                 p = next;
1744                 nr_maps--;
1745         }
1746         if (newmap)
1747                 goto done;
1748
1749         /* full maps */
1750         ceph_decode_32_safe(&p, end, nr_maps, bad);
1751         dout(" %d full maps\n", nr_maps);
1752         while (nr_maps) {
1753                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1754                 epoch = ceph_decode_32(&p);
1755                 maplen = ceph_decode_32(&p);
1756                 ceph_decode_need(&p, end, maplen, bad);
1757                 if (nr_maps > 1) {
1758                         dout("skipping non-latest full map %u len %d\n",
1759                              epoch, maplen);
1760                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1761                         dout("skipping full map %u len %d, "
1762                              "older than our %u\n", epoch, maplen,
1763                              osdc->osdmap->epoch);
1764                 } else {
1765                         int skipped_map = 0;
1766
1767                         dout("taking full map %u len %d\n", epoch, maplen);
1768                         newmap = osdmap_decode(&p, p+maplen);
1769                         if (IS_ERR(newmap)) {
1770                                 err = PTR_ERR(newmap);
1771                                 goto bad;
1772                         }
1773                         BUG_ON(!newmap);
1774                         oldmap = osdc->osdmap;
1775                         osdc->osdmap = newmap;
1776                         if (oldmap) {
1777                                 if (oldmap->epoch + 1 < newmap->epoch)
1778                                         skipped_map = 1;
1779                                 ceph_osdmap_destroy(oldmap);
1780                         }
1781                         kick_requests(osdc, skipped_map);
1782                 }
1783                 p += maplen;
1784                 nr_maps--;
1785         }
1786
1787 done:
1788         downgrade_write(&osdc->map_sem);
1789         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1790
1791         /*
1792          * subscribe to subsequent osdmap updates if full to ensure
1793          * we find out when we are no longer full and stop returning
1794          * ENOSPC.
1795          */
1796         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1797                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1798
1799         mutex_lock(&osdc->request_mutex);
1800         __send_queued(osdc);
1801         mutex_unlock(&osdc->request_mutex);
1802         up_read(&osdc->map_sem);
1803         wake_up_all(&osdc->client->auth_wq);
1804         return;
1805
1806 bad:
1807         pr_err("osdc handle_map corrupt msg\n");
1808         ceph_msg_dump(msg);
1809         up_write(&osdc->map_sem);
1810         return;
1811 }
1812
1813 /*
1814  * watch/notify callback event infrastructure
1815  *
1816  * These callbacks are used both for watch and notify operations.
1817  */
1818 static void __release_event(struct kref *kref)
1819 {
1820         struct ceph_osd_event *event =
1821                 container_of(kref, struct ceph_osd_event, kref);
1822
1823         dout("__release_event %p\n", event);
1824         kfree(event);
1825 }
1826
1827 static void get_event(struct ceph_osd_event *event)
1828 {
1829         kref_get(&event->kref);
1830 }
1831
1832 void ceph_osdc_put_event(struct ceph_osd_event *event)
1833 {
1834         kref_put(&event->kref, __release_event);
1835 }
1836 EXPORT_SYMBOL(ceph_osdc_put_event);
1837
1838 static void __insert_event(struct ceph_osd_client *osdc,
1839                              struct ceph_osd_event *new)
1840 {
1841         struct rb_node **p = &osdc->event_tree.rb_node;
1842         struct rb_node *parent = NULL;
1843         struct ceph_osd_event *event = NULL;
1844
1845         while (*p) {
1846                 parent = *p;
1847                 event = rb_entry(parent, struct ceph_osd_event, node);
1848                 if (new->cookie < event->cookie)
1849                         p = &(*p)->rb_left;
1850                 else if (new->cookie > event->cookie)
1851                         p = &(*p)->rb_right;
1852                 else
1853                         BUG();
1854         }
1855
1856         rb_link_node(&new->node, parent, p);
1857         rb_insert_color(&new->node, &osdc->event_tree);
1858 }
1859
1860 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1861                                                 u64 cookie)
1862 {
1863         struct rb_node **p = &osdc->event_tree.rb_node;
1864         struct rb_node *parent = NULL;
1865         struct ceph_osd_event *event = NULL;
1866
1867         while (*p) {
1868                 parent = *p;
1869                 event = rb_entry(parent, struct ceph_osd_event, node);
1870                 if (cookie < event->cookie)
1871                         p = &(*p)->rb_left;
1872                 else if (cookie > event->cookie)
1873                         p = &(*p)->rb_right;
1874                 else
1875                         return event;
1876         }
1877         return NULL;
1878 }
1879
1880 static void __remove_event(struct ceph_osd_event *event)
1881 {
1882         struct ceph_osd_client *osdc = event->osdc;
1883
1884         if (!RB_EMPTY_NODE(&event->node)) {
1885                 dout("__remove_event removed %p\n", event);
1886                 rb_erase(&event->node, &osdc->event_tree);
1887                 ceph_osdc_put_event(event);
1888         } else {
1889                 dout("__remove_event didn't remove %p\n", event);
1890         }
1891 }
1892
1893 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1894                            void (*event_cb)(u64, u64, u8, void *),
1895                            void *data, struct ceph_osd_event **pevent)
1896 {
1897         struct ceph_osd_event *event;
1898
1899         event = kmalloc(sizeof(*event), GFP_NOIO);
1900         if (!event)
1901                 return -ENOMEM;
1902
1903         dout("create_event %p\n", event);
1904         event->cb = event_cb;
1905         event->one_shot = 0;
1906         event->data = data;
1907         event->osdc = osdc;
1908         INIT_LIST_HEAD(&event->osd_node);
1909         RB_CLEAR_NODE(&event->node);
1910         kref_init(&event->kref);   /* one ref for us */
1911         kref_get(&event->kref);    /* one ref for the caller */
1912
1913         spin_lock(&osdc->event_lock);
1914         event->cookie = ++osdc->event_count;
1915         __insert_event(osdc, event);
1916         spin_unlock(&osdc->event_lock);
1917
1918         *pevent = event;
1919         return 0;
1920 }
1921 EXPORT_SYMBOL(ceph_osdc_create_event);
1922
1923 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1924 {
1925         struct ceph_osd_client *osdc = event->osdc;
1926
1927         dout("cancel_event %p\n", event);
1928         spin_lock(&osdc->event_lock);
1929         __remove_event(event);
1930         spin_unlock(&osdc->event_lock);
1931         ceph_osdc_put_event(event); /* caller's */
1932 }
1933 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1934
1935
1936 static void do_event_work(struct work_struct *work)
1937 {
1938         struct ceph_osd_event_work *event_work =
1939                 container_of(work, struct ceph_osd_event_work, work);
1940         struct ceph_osd_event *event = event_work->event;
1941         u64 ver = event_work->ver;
1942         u64 notify_id = event_work->notify_id;
1943         u8 opcode = event_work->opcode;
1944
1945         dout("do_event_work completing %p\n", event);
1946         event->cb(ver, notify_id, opcode, event->data);
1947         dout("do_event_work completed %p\n", event);
1948         ceph_osdc_put_event(event);
1949         kfree(event_work);
1950 }
1951
1952
1953 /*
1954  * Process osd watch notifications
1955  */
1956 static void handle_watch_notify(struct ceph_osd_client *osdc,
1957                                 struct ceph_msg *msg)
1958 {
1959         void *p, *end;
1960         u8 proto_ver;
1961         u64 cookie, ver, notify_id;
1962         u8 opcode;
1963         struct ceph_osd_event *event;
1964         struct ceph_osd_event_work *event_work;
1965
1966         p = msg->front.iov_base;
1967         end = p + msg->front.iov_len;
1968
1969         ceph_decode_8_safe(&p, end, proto_ver, bad);
1970         ceph_decode_8_safe(&p, end, opcode, bad);
1971         ceph_decode_64_safe(&p, end, cookie, bad);
1972         ceph_decode_64_safe(&p, end, ver, bad);
1973         ceph_decode_64_safe(&p, end, notify_id, bad);
1974
1975         spin_lock(&osdc->event_lock);
1976         event = __find_event(osdc, cookie);
1977         if (event) {
1978                 BUG_ON(event->one_shot);
1979                 get_event(event);
1980         }
1981         spin_unlock(&osdc->event_lock);
1982         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1983              cookie, ver, event);
1984         if (event) {
1985                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1986                 if (!event_work) {
1987                         dout("ERROR: could not allocate event_work\n");
1988                         goto done_err;
1989                 }
1990                 INIT_WORK(&event_work->work, do_event_work);
1991                 event_work->event = event;
1992                 event_work->ver = ver;
1993                 event_work->notify_id = notify_id;
1994                 event_work->opcode = opcode;
1995                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1996                         dout("WARNING: failed to queue notify event work\n");
1997                         goto done_err;
1998                 }
1999         }
2000
2001         return;
2002
2003 done_err:
2004         ceph_osdc_put_event(event);
2005         return;
2006
2007 bad:
2008         pr_err("osdc handle_watch_notify corrupt msg\n");
2009         return;
2010 }
2011
2012 /*
2013  * build new request AND message
2014  *
2015  */
2016 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2017                                 struct ceph_snap_context *snapc, u64 snap_id,
2018                                 struct timespec *mtime)
2019 {
2020         struct ceph_msg *msg = req->r_request;
2021         void *p;
2022         size_t msg_size;
2023         int flags = req->r_flags;
2024         u64 data_len;
2025         unsigned int i;
2026
2027         req->r_snapid = snap_id;
2028         req->r_snapc = ceph_get_snap_context(snapc);
2029
2030         /* encode request */
2031         msg->hdr.version = cpu_to_le16(4);
2032
2033         p = msg->front.iov_base;
2034         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2035         req->r_request_osdmap_epoch = p;
2036         p += 4;
2037         req->r_request_flags = p;
2038         p += 4;
2039         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2040                 ceph_encode_timespec(p, mtime);
2041         p += sizeof(struct ceph_timespec);
2042         req->r_request_reassert_version = p;
2043         p += sizeof(struct ceph_eversion); /* will get filled in */
2044
2045         /* oloc */
2046         ceph_encode_8(&p, 4);
2047         ceph_encode_8(&p, 4);
2048         ceph_encode_32(&p, 8 + 4 + 4);
2049         req->r_request_pool = p;
2050         p += 8;
2051         ceph_encode_32(&p, -1);  /* preferred */
2052         ceph_encode_32(&p, 0);   /* key len */
2053
2054         ceph_encode_8(&p, 1);
2055         req->r_request_pgid = p;
2056         p += 8 + 4;
2057         ceph_encode_32(&p, -1);  /* preferred */
2058
2059         /* oid */
2060         ceph_encode_32(&p, req->r_oid_len);
2061         memcpy(p, req->r_oid, req->r_oid_len);
2062         dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len);
2063         p += req->r_oid_len;
2064
2065         /* ops--can imply data */
2066         ceph_encode_16(&p, (u16)req->r_num_ops);
2067         data_len = 0;
2068         for (i = 0; i < req->r_num_ops; i++) {
2069                 data_len += osd_req_encode_op(req, p, i);
2070                 p += sizeof(struct ceph_osd_op);
2071         }
2072
2073         /* snaps */
2074         ceph_encode_64(&p, req->r_snapid);
2075         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2076         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2077         if (req->r_snapc) {
2078                 for (i = 0; i < snapc->num_snaps; i++) {
2079                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2080                 }
2081         }
2082
2083         req->r_request_attempts = p;
2084         p += 4;
2085
2086         /* data */
2087         if (flags & CEPH_OSD_FLAG_WRITE) {
2088                 u16 data_off;
2089
2090                 /*
2091                  * The header "data_off" is a hint to the receiver
2092                  * allowing it to align received data into its
2093                  * buffers such that there's no need to re-copy
2094                  * it before writing it to disk (direct I/O).
2095                  */
2096                 data_off = (u16) (off & 0xffff);
2097                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2098         }
2099         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2100
2101         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2102         msg_size = p - msg->front.iov_base;
2103         msg->front.iov_len = msg_size;
2104         msg->hdr.front_len = cpu_to_le32(msg_size);
2105
2106         dout("build_request msg_size was %d\n", (int)msg_size);
2107 }
2108 EXPORT_SYMBOL(ceph_osdc_build_request);
2109
2110 /*
2111  * Register request, send initial attempt.
2112  */
2113 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2114                             struct ceph_osd_request *req,
2115                             bool nofail)
2116 {
2117         int rc = 0;
2118
2119         down_read(&osdc->map_sem);
2120         mutex_lock(&osdc->request_mutex);
2121         __register_request(osdc, req);
2122         req->r_sent = 0;
2123         req->r_got_reply = 0;
2124         req->r_completed = 0;
2125         rc = __map_request(osdc, req, 0);
2126         if (rc < 0) {
2127                 if (nofail) {
2128                         dout("osdc_start_request failed map, "
2129                                 " will retry %lld\n", req->r_tid);
2130                         rc = 0;
2131                 }
2132                 goto out_unlock;
2133         }
2134         if (req->r_osd == NULL) {
2135                 dout("send_request %p no up osds in pg\n", req);
2136                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2137         } else {
2138                 __send_queued(osdc);
2139         }
2140         rc = 0;
2141 out_unlock:
2142         mutex_unlock(&osdc->request_mutex);
2143         up_read(&osdc->map_sem);
2144         return rc;
2145 }
2146 EXPORT_SYMBOL(ceph_osdc_start_request);
2147
2148 /*
2149  * wait for a request to complete
2150  */
2151 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2152                            struct ceph_osd_request *req)
2153 {
2154         int rc;
2155
2156         rc = wait_for_completion_interruptible(&req->r_completion);
2157         if (rc < 0) {
2158                 mutex_lock(&osdc->request_mutex);
2159                 __cancel_request(req);
2160                 __unregister_request(osdc, req);
2161                 mutex_unlock(&osdc->request_mutex);
2162                 complete_request(req);
2163                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
2164                 return rc;
2165         }
2166
2167         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
2168         return req->r_result;
2169 }
2170 EXPORT_SYMBOL(ceph_osdc_wait_request);
2171
2172 /*
2173  * sync - wait for all in-flight requests to flush.  avoid starvation.
2174  */
2175 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2176 {
2177         struct ceph_osd_request *req;
2178         u64 last_tid, next_tid = 0;
2179
2180         mutex_lock(&osdc->request_mutex);
2181         last_tid = osdc->last_tid;
2182         while (1) {
2183                 req = __lookup_request_ge(osdc, next_tid);
2184                 if (!req)
2185                         break;
2186                 if (req->r_tid > last_tid)
2187                         break;
2188
2189                 next_tid = req->r_tid + 1;
2190                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2191                         continue;
2192
2193                 ceph_osdc_get_request(req);
2194                 mutex_unlock(&osdc->request_mutex);
2195                 dout("sync waiting on tid %llu (last is %llu)\n",
2196                      req->r_tid, last_tid);
2197                 wait_for_completion(&req->r_safe_completion);
2198                 mutex_lock(&osdc->request_mutex);
2199                 ceph_osdc_put_request(req);
2200         }
2201         mutex_unlock(&osdc->request_mutex);
2202         dout("sync done (thru tid %llu)\n", last_tid);
2203 }
2204 EXPORT_SYMBOL(ceph_osdc_sync);
2205
2206 /*
2207  * init, shutdown
2208  */
2209 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2210 {
2211         int err;
2212
2213         dout("init\n");
2214         osdc->client = client;
2215         osdc->osdmap = NULL;
2216         init_rwsem(&osdc->map_sem);
2217         init_completion(&osdc->map_waiters);
2218         osdc->last_requested_map = 0;
2219         mutex_init(&osdc->request_mutex);
2220         osdc->last_tid = 0;
2221         osdc->osds = RB_ROOT;
2222         INIT_LIST_HEAD(&osdc->osd_lru);
2223         osdc->requests = RB_ROOT;
2224         INIT_LIST_HEAD(&osdc->req_lru);
2225         INIT_LIST_HEAD(&osdc->req_unsent);
2226         INIT_LIST_HEAD(&osdc->req_notarget);
2227         INIT_LIST_HEAD(&osdc->req_linger);
2228         osdc->num_requests = 0;
2229         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2230         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2231         spin_lock_init(&osdc->event_lock);
2232         osdc->event_tree = RB_ROOT;
2233         osdc->event_count = 0;
2234
2235         schedule_delayed_work(&osdc->osds_timeout_work,
2236            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2237
2238         err = -ENOMEM;
2239         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2240                                         sizeof(struct ceph_osd_request));
2241         if (!osdc->req_mempool)
2242                 goto out;
2243
2244         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2245                                 OSD_OP_FRONT_LEN, 10, true,
2246                                 "osd_op");
2247         if (err < 0)
2248                 goto out_mempool;
2249         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2250                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2251                                 "osd_op_reply");
2252         if (err < 0)
2253                 goto out_msgpool;
2254
2255         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2256         if (IS_ERR(osdc->notify_wq)) {
2257                 err = PTR_ERR(osdc->notify_wq);
2258                 osdc->notify_wq = NULL;
2259                 goto out_msgpool;
2260         }
2261         return 0;
2262
2263 out_msgpool:
2264         ceph_msgpool_destroy(&osdc->msgpool_op);
2265 out_mempool:
2266         mempool_destroy(osdc->req_mempool);
2267 out:
2268         return err;
2269 }
2270
2271 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2272 {
2273         flush_workqueue(osdc->notify_wq);
2274         destroy_workqueue(osdc->notify_wq);
2275         cancel_delayed_work_sync(&osdc->timeout_work);
2276         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2277         if (osdc->osdmap) {
2278                 ceph_osdmap_destroy(osdc->osdmap);
2279                 osdc->osdmap = NULL;
2280         }
2281         remove_all_osds(osdc);
2282         mempool_destroy(osdc->req_mempool);
2283         ceph_msgpool_destroy(&osdc->msgpool_op);
2284         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2285 }
2286
2287 /*
2288  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2289  * *plen.  Return number of bytes read, or error.
2290  */
2291 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2292                         struct ceph_vino vino, struct ceph_file_layout *layout,
2293                         u64 off, u64 *plen,
2294                         u32 truncate_seq, u64 truncate_size,
2295                         struct page **pages, int num_pages, int page_align)
2296 {
2297         struct ceph_osd_request *req;
2298         int rc = 0;
2299
2300         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2301              vino.snap, off, *plen);
2302         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2303                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2304                                     NULL, truncate_seq, truncate_size,
2305                                     false);
2306         if (IS_ERR(req))
2307                 return PTR_ERR(req);
2308
2309         /* it may be a short read due to an object boundary */
2310
2311         osd_req_op_extent_osd_data_pages(req, 0,
2312                                 pages, *plen, page_align, false, false);
2313
2314         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2315              off, *plen, *plen, page_align);
2316
2317         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2318
2319         rc = ceph_osdc_start_request(osdc, req, false);
2320         if (!rc)
2321                 rc = ceph_osdc_wait_request(osdc, req);
2322
2323         ceph_osdc_put_request(req);
2324         dout("readpages result %d\n", rc);
2325         return rc;
2326 }
2327 EXPORT_SYMBOL(ceph_osdc_readpages);
2328
2329 /*
2330  * do a synchronous write on N pages
2331  */
2332 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2333                          struct ceph_file_layout *layout,
2334                          struct ceph_snap_context *snapc,
2335                          u64 off, u64 len,
2336                          u32 truncate_seq, u64 truncate_size,
2337                          struct timespec *mtime,
2338                          struct page **pages, int num_pages)
2339 {
2340         struct ceph_osd_request *req;
2341         int rc = 0;
2342         int page_align = off & ~PAGE_MASK;
2343
2344         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2345         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2346                                     CEPH_OSD_OP_WRITE,
2347                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2348                                     snapc, truncate_seq, truncate_size,
2349                                     true);
2350         if (IS_ERR(req))
2351                 return PTR_ERR(req);
2352
2353         /* it may be a short write due to an object boundary */
2354         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2355                                 false, false);
2356         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2357
2358         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2359
2360         rc = ceph_osdc_start_request(osdc, req, true);
2361         if (!rc)
2362                 rc = ceph_osdc_wait_request(osdc, req);
2363
2364         ceph_osdc_put_request(req);
2365         if (rc == 0)
2366                 rc = len;
2367         dout("writepages result %d\n", rc);
2368         return rc;
2369 }
2370 EXPORT_SYMBOL(ceph_osdc_writepages);
2371
2372 int ceph_osdc_setup(void)
2373 {
2374         BUG_ON(ceph_osd_request_cache);
2375         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2376                                         sizeof (struct ceph_osd_request),
2377                                         __alignof__(struct ceph_osd_request),
2378                                         0, NULL);
2379
2380         return ceph_osd_request_cache ? 0 : -ENOMEM;
2381 }
2382 EXPORT_SYMBOL(ceph_osdc_setup);
2383
2384 void ceph_osdc_cleanup(void)
2385 {
2386         BUG_ON(!ceph_osd_request_cache);
2387         kmem_cache_destroy(ceph_osd_request_cache);
2388         ceph_osd_request_cache = NULL;
2389 }
2390 EXPORT_SYMBOL(ceph_osdc_cleanup);
2391
2392 /*
2393  * handle incoming message
2394  */
2395 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2396 {
2397         struct ceph_osd *osd = con->private;
2398         struct ceph_osd_client *osdc;
2399         int type = le16_to_cpu(msg->hdr.type);
2400
2401         if (!osd)
2402                 goto out;
2403         osdc = osd->o_osdc;
2404
2405         switch (type) {
2406         case CEPH_MSG_OSD_MAP:
2407                 ceph_osdc_handle_map(osdc, msg);
2408                 break;
2409         case CEPH_MSG_OSD_OPREPLY:
2410                 handle_reply(osdc, msg, con);
2411                 break;
2412         case CEPH_MSG_WATCH_NOTIFY:
2413                 handle_watch_notify(osdc, msg);
2414                 break;
2415
2416         default:
2417                 pr_err("received unknown message type %d %s\n", type,
2418                        ceph_msg_type_name(type));
2419         }
2420 out:
2421         ceph_msg_put(msg);
2422 }
2423
2424 /*
2425  * lookup and return message for incoming reply.  set up reply message
2426  * pages.
2427  */
2428 static struct ceph_msg *get_reply(struct ceph_connection *con,
2429                                   struct ceph_msg_header *hdr,
2430                                   int *skip)
2431 {
2432         struct ceph_osd *osd = con->private;
2433         struct ceph_osd_client *osdc = osd->o_osdc;
2434         struct ceph_msg *m;
2435         struct ceph_osd_request *req;
2436         int front = le32_to_cpu(hdr->front_len);
2437         int data_len = le32_to_cpu(hdr->data_len);
2438         u64 tid;
2439
2440         tid = le64_to_cpu(hdr->tid);
2441         mutex_lock(&osdc->request_mutex);
2442         req = __lookup_request(osdc, tid);
2443         if (!req) {
2444                 *skip = 1;
2445                 m = NULL;
2446                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2447                      osd->o_osd);
2448                 goto out;
2449         }
2450
2451         if (req->r_reply->con)
2452                 dout("%s revoking msg %p from old con %p\n", __func__,
2453                      req->r_reply, req->r_reply->con);
2454         ceph_msg_revoke_incoming(req->r_reply);
2455
2456         if (front > req->r_reply->front.iov_len) {
2457                 pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n",
2458                            front, (int)req->r_reply->front.iov_len,
2459                            (unsigned int)con->peer_name.type,
2460                            le64_to_cpu(con->peer_name.num));
2461                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2462                 if (!m)
2463                         goto out;
2464                 ceph_msg_put(req->r_reply);
2465                 req->r_reply = m;
2466         }
2467         m = ceph_msg_get(req->r_reply);
2468
2469         if (data_len > 0) {
2470                 struct ceph_osd_data *osd_data;
2471
2472                 /*
2473                  * XXX This is assuming there is only one op containing
2474                  * XXX page data.  Probably OK for reads, but this
2475                  * XXX ought to be done more generally.
2476                  */
2477                 osd_data = osd_req_op_extent_osd_data(req, 0);
2478                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2479                         if (osd_data->pages &&
2480                                 unlikely(osd_data->length < data_len)) {
2481
2482                                 pr_warning("tid %lld reply has %d bytes "
2483                                         "we had only %llu bytes ready\n",
2484                                         tid, data_len, osd_data->length);
2485                                 *skip = 1;
2486                                 ceph_msg_put(m);
2487                                 m = NULL;
2488                                 goto out;
2489                         }
2490                 }
2491         }
2492         *skip = 0;
2493         dout("get_reply tid %lld %p\n", tid, m);
2494
2495 out:
2496         mutex_unlock(&osdc->request_mutex);
2497         return m;
2498
2499 }
2500
2501 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2502                                   struct ceph_msg_header *hdr,
2503                                   int *skip)
2504 {
2505         struct ceph_osd *osd = con->private;
2506         int type = le16_to_cpu(hdr->type);
2507         int front = le32_to_cpu(hdr->front_len);
2508
2509         *skip = 0;
2510         switch (type) {
2511         case CEPH_MSG_OSD_MAP:
2512         case CEPH_MSG_WATCH_NOTIFY:
2513                 return ceph_msg_new(type, front, GFP_NOFS, false);
2514         case CEPH_MSG_OSD_OPREPLY:
2515                 return get_reply(con, hdr, skip);
2516         default:
2517                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2518                         osd->o_osd);
2519                 *skip = 1;
2520                 return NULL;
2521         }
2522 }
2523
2524 /*
2525  * Wrappers to refcount containing ceph_osd struct
2526  */
2527 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2528 {
2529         struct ceph_osd *osd = con->private;
2530         if (get_osd(osd))
2531                 return con;
2532         return NULL;
2533 }
2534
2535 static void put_osd_con(struct ceph_connection *con)
2536 {
2537         struct ceph_osd *osd = con->private;
2538         put_osd(osd);
2539 }
2540
2541 /*
2542  * authentication
2543  */
2544 /*
2545  * Note: returned pointer is the address of a structure that's
2546  * managed separately.  Caller must *not* attempt to free it.
2547  */
2548 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2549                                         int *proto, int force_new)
2550 {
2551         struct ceph_osd *o = con->private;
2552         struct ceph_osd_client *osdc = o->o_osdc;
2553         struct ceph_auth_client *ac = osdc->client->monc.auth;
2554         struct ceph_auth_handshake *auth = &o->o_auth;
2555
2556         if (force_new && auth->authorizer) {
2557                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2558                 auth->authorizer = NULL;
2559         }
2560         if (!auth->authorizer) {
2561                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2562                                                       auth);
2563                 if (ret)
2564                         return ERR_PTR(ret);
2565         } else {
2566                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2567                                                      auth);
2568                 if (ret)
2569                         return ERR_PTR(ret);
2570         }
2571         *proto = ac->protocol;
2572
2573         return auth;
2574 }
2575
2576
2577 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2578 {
2579         struct ceph_osd *o = con->private;
2580         struct ceph_osd_client *osdc = o->o_osdc;
2581         struct ceph_auth_client *ac = osdc->client->monc.auth;
2582
2583         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2584 }
2585
2586 static int invalidate_authorizer(struct ceph_connection *con)
2587 {
2588         struct ceph_osd *o = con->private;
2589         struct ceph_osd_client *osdc = o->o_osdc;
2590         struct ceph_auth_client *ac = osdc->client->monc.auth;
2591
2592         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2593         return ceph_monc_validate_auth(&osdc->client->monc);
2594 }
2595
2596 static const struct ceph_connection_operations osd_con_ops = {
2597         .get = get_osd_con,
2598         .put = put_osd_con,
2599         .dispatch = dispatch,
2600         .get_authorizer = get_authorizer,
2601         .verify_authorizer_reply = verify_authorizer_reply,
2602         .invalidate_authorizer = invalidate_authorizer,
2603         .alloc_msg = alloc_msg,
2604         .fault = osd_reset,
2605 };