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