net/ceph/osd_client.c

   1
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/module.h>
   5 #include <linux/err.h>
   6 #include <linux/highmem.h>
   7 #include <linux/mm.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/slab.h>
  10 #include <linux/uaccess.h>
  11 #ifdef CONFIG_BLOCK
  12 #include <linux/bio.h>
  13 #endif
  14
  15 #include <linux/ceph/libceph.h>
  16 #include <linux/ceph/osd_client.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/auth.h>
  20 #include <linux/ceph/pagelist.h>
  21
  22 #define OSD_OP_FRONT_LEN        4096
  23 #define OSD_OPREPLY_FRONT_LEN   512
  24
  25 static struct kmem_cache        *ceph_osd_request_cache;
  26
  27 static const struct ceph_connection_operations osd_con_ops;
  28
  29 static void __send_queued(struct ceph_osd_client *osdc);
  30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  31 static void __register_request(struct ceph_osd_client *osdc,
  32                                struct ceph_osd_request *req);
  33 static void __unregister_request(struct ceph_osd_client *osdc,
  34                                  struct ceph_osd_request *req);
  35 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  36                                         struct ceph_osd_request *req);
  37 static void __enqueue_request(struct ceph_osd_request *req);
  38 static void __send_request(struct ceph_osd_client *osdc,
  39                            struct ceph_osd_request *req);
  40
  41 /*
  42  * Implement client access to distributed object storage cluster.
  43  *
  44  * All data objects are stored within a cluster/cloud of OSDs, or
  45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  47  * remote daemons serving up and coordinating consistent and safe
  48  * access to storage.
  49  *
  50  * Cluster membership and the mapping of data objects onto storage devices
  51  * are described by the osd map.
  52  *
  53  * We keep track of pending OSD requests (read, write), resubmit
  54  * requests to different OSDs when the cluster topology/data layout
  55  * change, or retry the affected requests when the communications
  56  * channel with an OSD is reset.
  57  */
  58
  59 /*
  60  * calculate the mapping of a file extent onto an object, and fill out the
  61  * request accordingly.  shorten extent as necessary if it crosses an
  62  * object boundary.
  63  *
  64  * fill osd op in request message.
  65  */
  66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
  67                         u64 *objnum, u64 *objoff, u64 *objlen)
  68 {
  69         u64 orig_len = *plen;
  70         int r;
  71
  72         /* object extent? */
  73         r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
  74                                           objoff, objlen);
  75         if (r < 0)
  76                 return r;
  77         if (*objlen < orig_len) {
  78                 *plen = *objlen;
  79                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  80                      orig_len - *plen, off, *plen);
  81         }
  82
  83         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
  84
  85         return 0;
  86 }
  87
  88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
  89 {
  90         memset(osd_data, 0, sizeof (*osd_data));
  91         osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
  92 }
  93
  94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
  95                         struct page **pages, u64 length, u32 alignment,
  96                         bool pages_from_pool, bool own_pages)
  97 {
  98         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
  99         osd_data->pages = pages;
 100         osd_data->length = length;
 101         osd_data->alignment = alignment;
 102         osd_data->pages_from_pool = pages_from_pool;
 103         osd_data->own_pages = own_pages;
 104 }
 105
 106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
 107                         struct ceph_pagelist *pagelist)
 108 {
 109         osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
 110         osd_data->pagelist = pagelist;
 111 }
 112
 113 #ifdef CONFIG_BLOCK
 114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
 115                         struct bio *bio, size_t bio_length)
 116 {
 117         osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
 118         osd_data->bio = bio;
 119         osd_data->bio_length = bio_length;
 120 }
 121 #endif /* CONFIG_BLOCK */
 122
 123 #define osd_req_op_data(oreq, whch, typ, fld)   \
 124         ({                                              \
 125                 BUG_ON(whch >= (oreq)->r_num_ops);      \
 126                 &(oreq)->r_ops[whch].typ.fld;           \
 127         })
 128
 129 static struct ceph_osd_data *
 130 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
 131 {
 132         BUG_ON(which >= osd_req->r_num_ops);
 133
 134         return &osd_req->r_ops[which].raw_data_in;
 135 }
 136
 137 struct ceph_osd_data *
 138 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
 139                         unsigned int which)
 140 {
 141         return osd_req_op_data(osd_req, which, extent, osd_data);
 142 }
 143 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
 144
 145 struct ceph_osd_data *
 146 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
 147                         unsigned int which)
 148 {
 149         return osd_req_op_data(osd_req, which, cls, response_data);
 150 }
 151 EXPORT_SYMBOL(osd_req_op_cls_response_data);    /* ??? */
 152
 153 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
 154                         unsigned int which, struct page **pages,
 155                         u64 length, u32 alignment,
 156                         bool pages_from_pool, bool own_pages)
 157 {
 158         struct ceph_osd_data *osd_data;
 159
 160         osd_data = osd_req_op_raw_data_in(osd_req, which);
 161         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 162                                 pages_from_pool, own_pages);
 163 }
 164 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
 165
 166 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
 167                         unsigned int which, struct page **pages,
 168                         u64 length, u32 alignment,
 169                         bool pages_from_pool, bool own_pages)
 170 {
 171         struct ceph_osd_data *osd_data;
 172
 173         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 174         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 175                                 pages_from_pool, own_pages);
 176 }
 177 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
 178
 179 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
 180                         unsigned int which, struct ceph_pagelist *pagelist)
 181 {
 182         struct ceph_osd_data *osd_data;
 183
 184         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 185         ceph_osd_data_pagelist_init(osd_data, pagelist);
 186 }
 187 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
 188
 189 #ifdef CONFIG_BLOCK
 190 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
 191                         unsigned int which, struct bio *bio, size_t bio_length)
 192 {
 193         struct ceph_osd_data *osd_data;
 194
 195         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 196         ceph_osd_data_bio_init(osd_data, bio, bio_length);
 197 }
 198 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
 199 #endif /* CONFIG_BLOCK */
 200
 201 static void osd_req_op_cls_request_info_pagelist(
 202                         struct ceph_osd_request *osd_req,
 203                         unsigned int which, struct ceph_pagelist *pagelist)
 204 {
 205         struct ceph_osd_data *osd_data;
 206
 207         osd_data = osd_req_op_data(osd_req, which, cls, request_info);
 208         ceph_osd_data_pagelist_init(osd_data, pagelist);
 209 }
 210
 211 void osd_req_op_cls_request_data_pagelist(
 212                         struct ceph_osd_request *osd_req,
 213                         unsigned int which, struct ceph_pagelist *pagelist)
 214 {
 215         struct ceph_osd_data *osd_data;
 216
 217         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 218         ceph_osd_data_pagelist_init(osd_data, pagelist);
 219 }
 220 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
 221
 222 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
 223                         unsigned int which, struct page **pages, u64 length,
 224                         u32 alignment, bool pages_from_pool, bool own_pages)
 225 {
 226         struct ceph_osd_data *osd_data;
 227
 228         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 229         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 230                                 pages_from_pool, own_pages);
 231 }
 232 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
 233
 234 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
 235                         unsigned int which, struct page **pages, u64 length,
 236                         u32 alignment, bool pages_from_pool, bool own_pages)
 237 {
 238         struct ceph_osd_data *osd_data;
 239
 240         osd_data = osd_req_op_data(osd_req, which, cls, response_data);
 241         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 242                                 pages_from_pool, own_pages);
 243 }
 244 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
 245
 246 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
 247 {
 248         switch (osd_data->type) {
 249         case CEPH_OSD_DATA_TYPE_NONE:
 250                 return 0;
 251         case CEPH_OSD_DATA_TYPE_PAGES:
 252                 return osd_data->length;
 253         case CEPH_OSD_DATA_TYPE_PAGELIST:
 254                 return (u64)osd_data->pagelist->length;
 255 #ifdef CONFIG_BLOCK
 256         case CEPH_OSD_DATA_TYPE_BIO:
 257                 return (u64)osd_data->bio_length;
 258 #endif /* CONFIG_BLOCK */
 259         default:
 260                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
 261                 return 0;
 262         }
 263 }
 264
 265 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
 266 {
 267         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
 268                 int num_pages;
 269
 270                 num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                                 (u64)osd_data->length);
 272                 ceph_release_page_vector(osd_data->pages, num_pages);
 273         }
 274         ceph_osd_data_init(osd_data);
 275 }
 276
 277 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
 278                         unsigned int which)
 279 {
 280         struct ceph_osd_req_op *op;
 281
 282         BUG_ON(which >= osd_req->r_num_ops);
 283         op = &osd_req->r_ops[which];
 284
 285         switch (op->op) {
 286         case CEPH_OSD_OP_READ:
 287         case CEPH_OSD_OP_WRITE:
 288                 ceph_osd_data_release(&op->extent.osd_data);
 289                 break;
 290         case CEPH_OSD_OP_CALL:
 291                 ceph_osd_data_release(&op->cls.request_info);
 292                 ceph_osd_data_release(&op->cls.request_data);
 293                 ceph_osd_data_release(&op->cls.response_data);
 294                 break;
 295         case CEPH_OSD_OP_SETXATTR:
 296         case CEPH_OSD_OP_CMPXATTR:
 297                 ceph_osd_data_release(&op->xattr.osd_data);
 298                 break;
 299         case CEPH_OSD_OP_STAT:
 300                 ceph_osd_data_release(&op->raw_data_in);
 301                 break;
 302         default:
 303                 break;
 304         }
 305 }
 306
 307 /*
 308  * requests
 309  */
 310 static void ceph_osdc_release_request(struct kref *kref)
 311 {
 312         struct ceph_osd_request *req = container_of(kref,
 313                                             struct ceph_osd_request, r_kref);
 314         unsigned int which;
 315
 316         dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
 317              req->r_request, req->r_reply);
 318         WARN_ON(!RB_EMPTY_NODE(&req->r_node));
 319         WARN_ON(!list_empty(&req->r_req_lru_item));
 320         WARN_ON(!list_empty(&req->r_osd_item));
 321         WARN_ON(!list_empty(&req->r_linger_item));
 322         WARN_ON(!list_empty(&req->r_linger_osd_item));
 323         WARN_ON(req->r_osd);
 324
 325         if (req->r_request)
 326                 ceph_msg_put(req->r_request);
 327         if (req->r_reply) {
 328                 ceph_msg_revoke_incoming(req->r_reply);
 329                 ceph_msg_put(req->r_reply);
 330         }
 331
 332         for (which = 0; which < req->r_num_ops; which++)
 333                 osd_req_op_data_release(req, which);
 334
 335         ceph_put_snap_context(req->r_snapc);
 336         if (req->r_mempool)
 337                 mempool_free(req, req->r_osdc->req_mempool);
 338         else
 339                 kmem_cache_free(ceph_osd_request_cache, req);
 340
 341 }
 342
 343 void ceph_osdc_get_request(struct ceph_osd_request *req)
 344 {
 345         dout("%s %p (was %d)\n", __func__, req,
 346              atomic_read(&req->r_kref.refcount));
 347         kref_get(&req->r_kref);
 348 }
 349 EXPORT_SYMBOL(ceph_osdc_get_request);
 350
 351 void ceph_osdc_put_request(struct ceph_osd_request *req)
 352 {
 353         dout("%s %p (was %d)\n", __func__, req,
 354              atomic_read(&req->r_kref.refcount));
 355         kref_put(&req->r_kref, ceph_osdc_release_request);
 356 }
 357 EXPORT_SYMBOL(ceph_osdc_put_request);
 358
 359 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 360                                                struct ceph_snap_context *snapc,
 361                                                unsigned int num_ops,
 362                                                bool use_mempool,
 363                                                gfp_t gfp_flags)
 364 {
 365         struct ceph_osd_request *req;
 366         struct ceph_msg *msg;
 367         size_t msg_size;
 368
 369         BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
 370         BUG_ON(num_ops > CEPH_OSD_MAX_OP);
 371
 372         msg_size = 4 + 4 + 8 + 8 + 4+8;
 373         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
 374         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
 375         msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
 376         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
 377         msg_size += 8;  /* snapid */
 378         msg_size += 8;  /* snap_seq */
 379         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
 380         msg_size += 4;
 381
 382         if (use_mempool) {
 383                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 384                 memset(req, 0, sizeof(*req));
 385         } else {
 386                 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
 387         }
 388         if (req == NULL)
 389                 return NULL;
 390
 391         req->r_osdc = osdc;
 392         req->r_mempool = use_mempool;
 393         req->r_num_ops = num_ops;
 394
 395         kref_init(&req->r_kref);
 396         init_completion(&req->r_completion);
 397         init_completion(&req->r_safe_completion);
 398         RB_CLEAR_NODE(&req->r_node);
 399         INIT_LIST_HEAD(&req->r_unsafe_item);
 400         INIT_LIST_HEAD(&req->r_linger_item);
 401         INIT_LIST_HEAD(&req->r_linger_osd_item);
 402         INIT_LIST_HEAD(&req->r_req_lru_item);
 403         INIT_LIST_HEAD(&req->r_osd_item);
 404
 405         req->r_base_oloc.pool = -1;
 406         req->r_target_oloc.pool = -1;
 407
 408         /* create reply message */
 409         if (use_mempool)
 410                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 411         else
 412                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 413                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 414         if (!msg) {
 415                 ceph_osdc_put_request(req);
 416                 return NULL;
 417         }
 418         req->r_reply = msg;
 419
 420         /* create request message; allow space for oid */
 421         if (use_mempool)
 422                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 423         else
 424                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 425         if (!msg) {
 426                 ceph_osdc_put_request(req);
 427                 return NULL;
 428         }
 429
 430         memset(msg->front.iov_base, 0, msg->front.iov_len);
 431
 432         req->r_request = msg;
 433
 434         return req;
 435 }
 436 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 437
 438 static bool osd_req_opcode_valid(u16 opcode)
 439 {
 440         switch (opcode) {
 441 #define GENERATE_CASE(op, opcode, str)  case CEPH_OSD_OP_##op: return true;
 442 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
 443 #undef GENERATE_CASE
 444         default:
 445                 return false;
 446         }
 447 }
 448
 449 /*
 450  * This is an osd op init function for opcodes that have no data or
 451  * other information associated with them.  It also serves as a
 452  * common init routine for all the other init functions, below.
 453  */
 454 static struct ceph_osd_req_op *
 455 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
 456                  u16 opcode, u32 flags)
 457 {
 458         struct ceph_osd_req_op *op;
 459
 460         BUG_ON(which >= osd_req->r_num_ops);
 461         BUG_ON(!osd_req_opcode_valid(opcode));
 462
 463         op = &osd_req->r_ops[which];
 464         memset(op, 0, sizeof (*op));
 465         op->op = opcode;
 466         op->flags = flags;
 467
 468         return op;
 469 }
 470
 471 void osd_req_op_init(struct ceph_osd_request *osd_req,
 472                      unsigned int which, u16 opcode, u32 flags)
 473 {
 474         (void)_osd_req_op_init(osd_req, which, opcode, flags);
 475 }
 476 EXPORT_SYMBOL(osd_req_op_init);
 477
 478 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 479                                 unsigned int which, u16 opcode,
 480                                 u64 offset, u64 length,
 481                                 u64 truncate_size, u32 truncate_seq)
 482 {
 483         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 484                                                       opcode, 0);
 485         size_t payload_len = 0;
 486
 487         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 488                opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE);
 489
 490         op->extent.offset = offset;
 491         op->extent.length = length;
 492         op->extent.truncate_size = truncate_size;
 493         op->extent.truncate_seq = truncate_seq;
 494         if (opcode == CEPH_OSD_OP_WRITE)
 495                 payload_len += length;
 496
 497         op->payload_len = payload_len;
 498 }
 499 EXPORT_SYMBOL(osd_req_op_extent_init);
 500
 501 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 502                                 unsigned int which, u64 length)
 503 {
 504         struct ceph_osd_req_op *op;
 505         u64 previous;
 506
 507         BUG_ON(which >= osd_req->r_num_ops);
 508         op = &osd_req->r_ops[which];
 509         previous = op->extent.length;
 510
 511         if (length == previous)
 512                 return;         /* Nothing to do */
 513         BUG_ON(length > previous);
 514
 515         op->extent.length = length;
 516         op->payload_len -= previous - length;
 517 }
 518 EXPORT_SYMBOL(osd_req_op_extent_update);
 519
 520 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
 521                         u16 opcode, const char *class, const char *method)
 522 {
 523         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 524                                                       opcode, 0);
 525         struct ceph_pagelist *pagelist;
 526         size_t payload_len = 0;
 527         size_t size;
 528
 529         BUG_ON(opcode != CEPH_OSD_OP_CALL);
 530
 531         pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
 532         BUG_ON(!pagelist);
 533         ceph_pagelist_init(pagelist);
 534
 535         op->cls.class_name = class;
 536         size = strlen(class);
 537         BUG_ON(size > (size_t) U8_MAX);
 538         op->cls.class_len = size;
 539         ceph_pagelist_append(pagelist, class, size);
 540         payload_len += size;
 541
 542         op->cls.method_name = method;
 543         size = strlen(method);
 544         BUG_ON(size > (size_t) U8_MAX);
 545         op->cls.method_len = size;
 546         ceph_pagelist_append(pagelist, method, size);
 547         payload_len += size;
 548
 549         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
 550
 551         op->cls.argc = 0;       /* currently unused */
 552
 553         op->payload_len = payload_len;
 554 }
 555 EXPORT_SYMBOL(osd_req_op_cls_init);
 556
 557 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 558                           u16 opcode, const char *name, const void *value,
 559                           size_t size, u8 cmp_op, u8 cmp_mode)
 560 {
 561         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 562                                                       opcode, 0);
 563         struct ceph_pagelist *pagelist;
 564         size_t payload_len;
 565
 566         BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
 567
 568         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
 569         if (!pagelist)
 570                 return -ENOMEM;
 571
 572         ceph_pagelist_init(pagelist);
 573
 574         payload_len = strlen(name);
 575         op->xattr.name_len = payload_len;
 576         ceph_pagelist_append(pagelist, name, payload_len);
 577
 578         op->xattr.value_len = size;
 579         ceph_pagelist_append(pagelist, value, size);
 580         payload_len += size;
 581
 582         op->xattr.cmp_op = cmp_op;
 583         op->xattr.cmp_mode = cmp_mode;
 584
 585         ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
 586         op->payload_len = payload_len;
 587         return 0;
 588 }
 589 EXPORT_SYMBOL(osd_req_op_xattr_init);
 590
 591 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
 592                                 unsigned int which, u16 opcode,
 593                                 u64 cookie, u64 version, int flag)
 594 {
 595         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 596                                                       opcode, 0);
 597
 598         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
 599
 600         op->watch.cookie = cookie;
 601         op->watch.ver = version;
 602         if (opcode == CEPH_OSD_OP_WATCH && flag)
 603                 op->watch.flag = (u8)1;
 604 }
 605 EXPORT_SYMBOL(osd_req_op_watch_init);
 606
 607 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 608                                 unsigned int which,
 609                                 u64 expected_object_size,
 610                                 u64 expected_write_size)
 611 {
 612         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 613                                                       CEPH_OSD_OP_SETALLOCHINT,
 614                                                       0);
 615
 616         op->alloc_hint.expected_object_size = expected_object_size;
 617         op->alloc_hint.expected_write_size = expected_write_size;
 618
 619         /*
 620          * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
 621          * not worth a feature bit.  Set FAILOK per-op flag to make
 622          * sure older osds don't trip over an unsupported opcode.
 623          */
 624         op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
 625 }
 626 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
 627
 628 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
 629                                 struct ceph_osd_data *osd_data)
 630 {
 631         u64 length = ceph_osd_data_length(osd_data);
 632
 633         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 634                 BUG_ON(length > (u64) SIZE_MAX);
 635                 if (length)
 636                         ceph_msg_data_add_pages(msg, osd_data->pages,
 637                                         length, osd_data->alignment);
 638         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 639                 BUG_ON(!length);
 640                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
 641 #ifdef CONFIG_BLOCK
 642         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
 643                 ceph_msg_data_add_bio(msg, osd_data->bio, length);
 644 #endif
 645         } else {
 646                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
 647         }
 648 }
 649
 650 static u64 osd_req_encode_op(struct ceph_osd_request *req,
 651                               struct ceph_osd_op *dst, unsigned int which)
 652 {
 653         struct ceph_osd_req_op *src;
 654         struct ceph_osd_data *osd_data;
 655         u64 request_data_len = 0;
 656         u64 data_length;
 657
 658         BUG_ON(which >= req->r_num_ops);
 659         src = &req->r_ops[which];
 660         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
 661                 pr_err("unrecognized osd opcode %d\n", src->op);
 662
 663                 return 0;
 664         }
 665
 666         switch (src->op) {
 667         case CEPH_OSD_OP_STAT:
 668                 osd_data = &src->raw_data_in;
 669                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 670                 break;
 671         case CEPH_OSD_OP_READ:
 672         case CEPH_OSD_OP_WRITE:
 673         case CEPH_OSD_OP_ZERO:
 674         case CEPH_OSD_OP_TRUNCATE:
 675                 if (src->op == CEPH_OSD_OP_WRITE)
 676                         request_data_len = src->extent.length;
 677                 dst->extent.offset = cpu_to_le64(src->extent.offset);
 678                 dst->extent.length = cpu_to_le64(src->extent.length);
 679                 dst->extent.truncate_size =
 680                         cpu_to_le64(src->extent.truncate_size);
 681                 dst->extent.truncate_seq =
 682                         cpu_to_le32(src->extent.truncate_seq);
 683                 osd_data = &src->extent.osd_data;
 684                 if (src->op == CEPH_OSD_OP_WRITE)
 685                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 686                 else
 687                         ceph_osdc_msg_data_add(req->r_reply, osd_data);
 688                 break;
 689         case CEPH_OSD_OP_CALL:
 690                 dst->cls.class_len = src->cls.class_len;
 691                 dst->cls.method_len = src->cls.method_len;
 692                 osd_data = &src->cls.request_info;
 693                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 694                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
 695                 request_data_len = osd_data->pagelist->length;
 696
 697                 osd_data = &src->cls.request_data;
 698                 data_length = ceph_osd_data_length(osd_data);
 699                 if (data_length) {
 700                         BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
 701                         dst->cls.indata_len = cpu_to_le32(data_length);
 702                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 703                         src->payload_len += data_length;
 704                         request_data_len += data_length;
 705                 }
 706                 osd_data = &src->cls.response_data;
 707                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 708                 break;
 709         case CEPH_OSD_OP_STARTSYNC:
 710                 break;
 711         case CEPH_OSD_OP_NOTIFY_ACK:
 712         case CEPH_OSD_OP_WATCH:
 713                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 714                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 715                 dst->watch.flag = src->watch.flag;
 716                 break;
 717         case CEPH_OSD_OP_SETALLOCHINT:
 718                 dst->alloc_hint.expected_object_size =
 719                     cpu_to_le64(src->alloc_hint.expected_object_size);
 720                 dst->alloc_hint.expected_write_size =
 721                     cpu_to_le64(src->alloc_hint.expected_write_size);
 722                 break;
 723         case CEPH_OSD_OP_SETXATTR:
 724         case CEPH_OSD_OP_CMPXATTR:
 725                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 726                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 727                 dst->xattr.cmp_op = src->xattr.cmp_op;
 728                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 729                 osd_data = &src->xattr.osd_data;
 730                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 731                 request_data_len = osd_data->pagelist->length;
 732                 break;
 733         case CEPH_OSD_OP_CREATE:
 734         case CEPH_OSD_OP_DELETE:
 735                 break;
 736         default:
 737                 pr_err("unsupported osd opcode %s\n",
 738                         ceph_osd_op_name(src->op));
 739                 WARN_ON(1);
 740
 741                 return 0;
 742         }
 743
 744         dst->op = cpu_to_le16(src->op);
 745         dst->flags = cpu_to_le32(src->flags);
 746         dst->payload_len = cpu_to_le32(src->payload_len);
 747
 748         return request_data_len;
 749 }
 750
 751 /*
 752  * build new request AND message, calculate layout, and adjust file
 753  * extent as needed.
 754  *
 755  * if the file was recently truncated, we include information about its
 756  * old and new size so that the object can be updated appropriately.  (we
 757  * avoid synchronously deleting truncated objects because it's slow.)
 758  *
 759  * if @do_sync, include a 'startsync' command so that the osd will flush
 760  * data quickly.
 761  */
 762 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 763                                                struct ceph_file_layout *layout,
 764                                                struct ceph_vino vino,
 765                                                u64 off, u64 *plen,
 766                                                unsigned int which, int num_ops,
 767                                                int opcode, int flags,
 768                                                struct ceph_snap_context *snapc,
 769                                                u32 truncate_seq,
 770                                                u64 truncate_size,
 771                                                bool use_mempool)
 772 {
 773         struct ceph_osd_request *req;
 774         u64 objnum = 0;
 775         u64 objoff = 0;
 776         u64 objlen = 0;
 777         int r;
 778
 779         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 780                opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
 781                opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
 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         if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
 798                 osd_req_op_init(req, which, opcode, 0);
 799         } else {
 800                 u32 object_size = le32_to_cpu(layout->fl_object_size);
 801                 u32 object_base = off - objoff;
 802                 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
 803                         if (truncate_size <= object_base) {
 804                                 truncate_size = 0;
 805                         } else {
 806                                 truncate_size -= object_base;
 807                                 if (truncate_size > object_size)
 808                                         truncate_size = object_size;
 809                         }
 810                 }
 811                 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
 812                                        truncate_size, truncate_seq);
 813         }
 814
 815         req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
 816
 817         snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
 818                  "%llx.%08llx", vino.ino, objnum);
 819         req->r_base_oid.name_len = strlen(req->r_base_oid.name);
 820
 821         return req;
 822 }
 823 EXPORT_SYMBOL(ceph_osdc_new_request);
 824
 825 /*
 826  * We keep osd requests in an rbtree, sorted by ->r_tid.
 827  */
 828 static void __insert_request(struct ceph_osd_client *osdc,
 829                              struct ceph_osd_request *new)
 830 {
 831         struct rb_node **p = &osdc->requests.rb_node;
 832         struct rb_node *parent = NULL;
 833         struct ceph_osd_request *req = NULL;
 834
 835         while (*p) {
 836                 parent = *p;
 837                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 838                 if (new->r_tid < req->r_tid)
 839                         p = &(*p)->rb_left;
 840                 else if (new->r_tid > req->r_tid)
 841                         p = &(*p)->rb_right;
 842                 else
 843                         BUG();
 844         }
 845
 846         rb_link_node(&new->r_node, parent, p);
 847         rb_insert_color(&new->r_node, &osdc->requests);
 848 }
 849
 850 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 851                                                  u64 tid)
 852 {
 853         struct ceph_osd_request *req;
 854         struct rb_node *n = osdc->requests.rb_node;
 855
 856         while (n) {
 857                 req = rb_entry(n, struct ceph_osd_request, r_node);
 858                 if (tid < req->r_tid)
 859                         n = n->rb_left;
 860                 else if (tid > req->r_tid)
 861                         n = n->rb_right;
 862                 else
 863                         return req;
 864         }
 865         return NULL;
 866 }
 867
 868 static struct ceph_osd_request *
 869 __lookup_request_ge(struct ceph_osd_client *osdc,
 870                     u64 tid)
 871 {
 872         struct ceph_osd_request *req;
 873         struct rb_node *n = osdc->requests.rb_node;
 874
 875         while (n) {
 876                 req = rb_entry(n, struct ceph_osd_request, r_node);
 877                 if (tid < req->r_tid) {
 878                         if (!n->rb_left)
 879                                 return req;
 880                         n = n->rb_left;
 881                 } else if (tid > req->r_tid) {
 882                         n = n->rb_right;
 883                 } else {
 884                         return req;
 885                 }
 886         }
 887         return NULL;
 888 }
 889
 890 static void __kick_linger_request(struct ceph_osd_request *req)
 891 {
 892         struct ceph_osd_client *osdc = req->r_osdc;
 893         struct ceph_osd *osd = req->r_osd;
 894
 895         /*
 896          * Linger requests need to be resent with a new tid to avoid
 897          * the dup op detection logic on the OSDs.  Achieve this with
 898          * a re-register dance instead of open-coding.
 899          */
 900         ceph_osdc_get_request(req);
 901         if (!list_empty(&req->r_linger_item))
 902                 __unregister_linger_request(osdc, req);
 903         else
 904                 __unregister_request(osdc, req);
 905         __register_request(osdc, req);
 906         ceph_osdc_put_request(req);
 907
 908         /*
 909          * Unless request has been registered as both normal and
 910          * lingering, __unregister{,_linger}_request clears r_osd.
 911          * However, here we need to preserve r_osd to make sure we
 912          * requeue on the same OSD.
 913          */
 914         WARN_ON(req->r_osd || !osd);
 915         req->r_osd = osd;
 916
 917         dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
 918         __enqueue_request(req);
 919 }
 920
 921 /*
 922  * Resubmit requests pending on the given osd.
 923  */
 924 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 925                                 struct ceph_osd *osd)
 926 {
 927         struct ceph_osd_request *req, *nreq;
 928         LIST_HEAD(resend);
 929         LIST_HEAD(resend_linger);
 930         int err;
 931
 932         dout("%s osd%d\n", __func__, osd->o_osd);
 933         err = __reset_osd(osdc, osd);
 934         if (err)
 935                 return;
 936
 937         /*
 938          * Build up a list of requests to resend by traversing the
 939          * osd's list of requests.  Requests for a given object are
 940          * sent in tid order, and that is also the order they're
 941          * kept on this list.  Therefore all requests that are in
 942          * flight will be found first, followed by all requests that
 943          * have not yet been sent.  And to resend requests while
 944          * preserving this order we will want to put any sent
 945          * requests back on the front of the osd client's unsent
 946          * list.
 947          *
 948          * So we build a separate ordered list of already-sent
 949          * requests for the affected osd and splice it onto the
 950          * front of the osd client's unsent list.  Once we've seen a
 951          * request that has not yet been sent we're done.  Those
 952          * requests are already sitting right where they belong.
 953          */
 954         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 955                 if (!req->r_sent)
 956                         break;
 957
 958                 if (!req->r_linger) {
 959                         dout("%s requeueing %p tid %llu\n", __func__, req,
 960                              req->r_tid);
 961                         list_move_tail(&req->r_req_lru_item, &resend);
 962                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 963                 } else {
 964                         list_move_tail(&req->r_req_lru_item, &resend_linger);
 965                 }
 966         }
 967         list_splice(&resend, &osdc->req_unsent);
 968
 969         /*
 970          * Both registered and not yet registered linger requests are
 971          * enqueued with a new tid on the same OSD.  We add/move them
 972          * to req_unsent/o_requests at the end to keep things in tid
 973          * order.
 974          */
 975         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 976                                  r_linger_osd_item) {
 977                 WARN_ON(!list_empty(&req->r_req_lru_item));
 978                 __kick_linger_request(req);
 979         }
 980
 981         list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
 982                 __kick_linger_request(req);
 983 }
 984
 985 /*
 986  * If the osd connection drops, we need to resubmit all requests.
 987  */
 988 static void osd_reset(struct ceph_connection *con)
 989 {
 990         struct ceph_osd *osd = con->private;
 991         struct ceph_osd_client *osdc;
 992
 993         if (!osd)
 994                 return;
 995         dout("osd_reset osd%d\n", osd->o_osd);
 996         osdc = osd->o_osdc;
 997         down_read(&osdc->map_sem);
 998         mutex_lock(&osdc->request_mutex);
 999         __kick_osd_requests(osdc, osd);
1000         __send_queued(osdc);
1001         mutex_unlock(&osdc->request_mutex);
1002         up_read(&osdc->map_sem);
1003 }
1004
1005 /*
1006  * Track open sessions with osds.
1007  */
1008 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1009 {
1010         struct ceph_osd *osd;
1011
1012         osd = kzalloc(sizeof(*osd), GFP_NOFS);
1013         if (!osd)
1014                 return NULL;
1015
1016         atomic_set(&osd->o_ref, 1);
1017         osd->o_osdc = osdc;
1018         osd->o_osd = onum;
1019         RB_CLEAR_NODE(&osd->o_node);
1020         INIT_LIST_HEAD(&osd->o_requests);
1021         INIT_LIST_HEAD(&osd->o_linger_requests);
1022         INIT_LIST_HEAD(&osd->o_osd_lru);
1023         osd->o_incarnation = 1;
1024
1025         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1026
1027         INIT_LIST_HEAD(&osd->o_keepalive_item);
1028         return osd;
1029 }
1030
1031 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1032 {
1033         if (atomic_inc_not_zero(&osd->o_ref)) {
1034                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1035                      atomic_read(&osd->o_ref));
1036                 return osd;
1037         } else {
1038                 dout("get_osd %p FAIL\n", osd);
1039                 return NULL;
1040         }
1041 }
1042
1043 static void put_osd(struct ceph_osd *osd)
1044 {
1045         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1046              atomic_read(&osd->o_ref) - 1);
1047         if (atomic_dec_and_test(&osd->o_ref)) {
1048                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1049
1050                 if (osd->o_auth.authorizer)
1051                         ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1052                 kfree(osd);
1053         }
1054 }
1055
1056 /*
1057  * remove an osd from our map
1058  */
1059 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1060 {
1061         dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1062         WARN_ON(!list_empty(&osd->o_requests));
1063         WARN_ON(!list_empty(&osd->o_linger_requests));
1064
1065         list_del_init(&osd->o_osd_lru);
1066         rb_erase(&osd->o_node, &osdc->osds);
1067         RB_CLEAR_NODE(&osd->o_node);
1068 }
1069
1070 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1071 {
1072         dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1073
1074         if (!RB_EMPTY_NODE(&osd->o_node)) {
1075                 ceph_con_close(&osd->o_con);
1076                 __remove_osd(osdc, osd);
1077                 put_osd(osd);
1078         }
1079 }
1080
1081 static void remove_all_osds(struct ceph_osd_client *osdc)
1082 {
1083         dout("%s %p\n", __func__, osdc);
1084         mutex_lock(&osdc->request_mutex);
1085         while (!RB_EMPTY_ROOT(&osdc->osds)) {
1086                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1087                                                 struct ceph_osd, o_node);
1088                 remove_osd(osdc, osd);
1089         }
1090         mutex_unlock(&osdc->request_mutex);
1091 }
1092
1093 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1094                               struct ceph_osd *osd)
1095 {
1096         dout("%s %p\n", __func__, osd);
1097         BUG_ON(!list_empty(&osd->o_osd_lru));
1098
1099         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1100         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1101 }
1102
1103 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1104                                   struct ceph_osd *osd)
1105 {
1106         dout("%s %p\n", __func__, osd);
1107
1108         if (list_empty(&osd->o_requests) &&
1109             list_empty(&osd->o_linger_requests))
1110                 __move_osd_to_lru(osdc, osd);
1111 }
1112
1113 static void __remove_osd_from_lru(struct ceph_osd *osd)
1114 {
1115         dout("__remove_osd_from_lru %p\n", osd);
1116         if (!list_empty(&osd->o_osd_lru))
1117                 list_del_init(&osd->o_osd_lru);
1118 }
1119
1120 static void remove_old_osds(struct ceph_osd_client *osdc)
1121 {
1122         struct ceph_osd *osd, *nosd;
1123
1124         dout("__remove_old_osds %p\n", osdc);
1125         mutex_lock(&osdc->request_mutex);
1126         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1127                 if (time_before(jiffies, osd->lru_ttl))
1128                         break;
1129                 remove_osd(osdc, osd);
1130         }
1131         mutex_unlock(&osdc->request_mutex);
1132 }
1133
1134 /*
1135  * reset osd connect
1136  */
1137 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1138 {
1139         struct ceph_entity_addr *peer_addr;
1140
1141         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1142         if (list_empty(&osd->o_requests) &&
1143             list_empty(&osd->o_linger_requests)) {
1144                 remove_osd(osdc, osd);
1145                 return -ENODEV;
1146         }
1147
1148         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1149         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1150                         !ceph_con_opened(&osd->o_con)) {
1151                 struct ceph_osd_request *req;
1152
1153                 dout("osd addr hasn't changed and connection never opened, "
1154                      "letting msgr retry\n");
1155                 /* touch each r_stamp for handle_timeout()'s benfit */
1156                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1157                         req->r_stamp = jiffies;
1158
1159                 return -EAGAIN;
1160         }
1161
1162         ceph_con_close(&osd->o_con);
1163         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1164         osd->o_incarnation++;
1165
1166         return 0;
1167 }
1168
1169 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1170 {
1171         struct rb_node **p = &osdc->osds.rb_node;
1172         struct rb_node *parent = NULL;
1173         struct ceph_osd *osd = NULL;
1174
1175         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1176         while (*p) {
1177                 parent = *p;
1178                 osd = rb_entry(parent, struct ceph_osd, o_node);
1179                 if (new->o_osd < osd->o_osd)
1180                         p = &(*p)->rb_left;
1181                 else if (new->o_osd > osd->o_osd)
1182                         p = &(*p)->rb_right;
1183                 else
1184                         BUG();
1185         }
1186
1187         rb_link_node(&new->o_node, parent, p);
1188         rb_insert_color(&new->o_node, &osdc->osds);
1189 }
1190
1191 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1192 {
1193         struct ceph_osd *osd;
1194         struct rb_node *n = osdc->osds.rb_node;
1195
1196         while (n) {
1197                 osd = rb_entry(n, struct ceph_osd, o_node);
1198                 if (o < osd->o_osd)
1199                         n = n->rb_left;
1200                 else if (o > osd->o_osd)
1201                         n = n->rb_right;
1202                 else
1203                         return osd;
1204         }
1205         return NULL;
1206 }
1207
1208 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1209 {
1210         schedule_delayed_work(&osdc->timeout_work,
1211                               osdc->client->options->osd_keepalive_timeout);
1212 }
1213
1214 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1215 {
1216         cancel_delayed_work(&osdc->timeout_work);
1217 }
1218
1219 /*
1220  * Register request, assign tid.  If this is the first request, set up
1221  * the timeout event.
1222  */
1223 static void __register_request(struct ceph_osd_client *osdc,
1224                                struct ceph_osd_request *req)
1225 {
1226         req->r_tid = ++osdc->last_tid;
1227         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1228         dout("__register_request %p tid %lld\n", req, req->r_tid);
1229         __insert_request(osdc, req);
1230         ceph_osdc_get_request(req);
1231         osdc->num_requests++;
1232         if (osdc->num_requests == 1) {
1233                 dout(" first request, scheduling timeout\n");
1234                 __schedule_osd_timeout(osdc);
1235         }
1236 }
1237
1238 /*
1239  * called under osdc->request_mutex
1240  */
1241 static void __unregister_request(struct ceph_osd_client *osdc,
1242                                  struct ceph_osd_request *req)
1243 {
1244         if (RB_EMPTY_NODE(&req->r_node)) {
1245                 dout("__unregister_request %p tid %lld not registered\n",
1246                         req, req->r_tid);
1247                 return;
1248         }
1249
1250         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1251         rb_erase(&req->r_node, &osdc->requests);
1252         RB_CLEAR_NODE(&req->r_node);
1253         osdc->num_requests--;
1254
1255         if (req->r_osd) {
1256                 /* make sure the original request isn't in flight. */
1257                 ceph_msg_revoke(req->r_request);
1258
1259                 list_del_init(&req->r_osd_item);
1260                 maybe_move_osd_to_lru(osdc, req->r_osd);
1261                 if (list_empty(&req->r_linger_osd_item))
1262                         req->r_osd = NULL;
1263         }
1264
1265         list_del_init(&req->r_req_lru_item);
1266         ceph_osdc_put_request(req);
1267
1268         if (osdc->num_requests == 0) {
1269                 dout(" no requests, canceling timeout\n");
1270                 __cancel_osd_timeout(osdc);
1271         }
1272 }
1273
1274 /*
1275  * Cancel a previously queued request message
1276  */
1277 static void __cancel_request(struct ceph_osd_request *req)
1278 {
1279         if (req->r_sent && req->r_osd) {
1280                 ceph_msg_revoke(req->r_request);
1281                 req->r_sent = 0;
1282         }
1283 }
1284
1285 static void __register_linger_request(struct ceph_osd_client *osdc,
1286                                     struct ceph_osd_request *req)
1287 {
1288         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1289         WARN_ON(!req->r_linger);
1290
1291         ceph_osdc_get_request(req);
1292         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1293         if (req->r_osd)
1294                 list_add_tail(&req->r_linger_osd_item,
1295                               &req->r_osd->o_linger_requests);
1296 }
1297
1298 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1299                                         struct ceph_osd_request *req)
1300 {
1301         WARN_ON(!req->r_linger);
1302
1303         if (list_empty(&req->r_linger_item)) {
1304                 dout("%s %p tid %llu not registered\n", __func__, req,
1305                      req->r_tid);
1306                 return;
1307         }
1308
1309         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1310         list_del_init(&req->r_linger_item);
1311
1312         if (req->r_osd) {
1313                 list_del_init(&req->r_linger_osd_item);
1314                 maybe_move_osd_to_lru(osdc, req->r_osd);
1315                 if (list_empty(&req->r_osd_item))
1316                         req->r_osd = NULL;
1317         }
1318         ceph_osdc_put_request(req);
1319 }
1320
1321 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1322                                   struct ceph_osd_request *req)
1323 {
1324         if (!req->r_linger) {
1325                 dout("set_request_linger %p\n", req);
1326                 req->r_linger = 1;
1327         }
1328 }
1329 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1330
1331 /*
1332  * Returns whether a request should be blocked from being sent
1333  * based on the current osdmap and osd_client settings.
1334  *
1335  * Caller should hold map_sem for read.
1336  */
1337 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1338                                    struct ceph_osd_request *req)
1339 {
1340         bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1341         bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1342                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1343         return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1344                 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1345 }
1346
1347 /*
1348  * Calculate mapping of a request to a PG.  Takes tiering into account.
1349  */
1350 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1351                              struct ceph_osd_request *req,
1352                              struct ceph_pg *pg_out)
1353 {
1354         bool need_check_tiering;
1355
1356         need_check_tiering = false;
1357         if (req->r_target_oloc.pool == -1) {
1358                 req->r_target_oloc = req->r_base_oloc; /* struct */
1359                 need_check_tiering = true;
1360         }
1361         if (req->r_target_oid.name_len == 0) {
1362                 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1363                 need_check_tiering = true;
1364         }
1365
1366         if (need_check_tiering &&
1367             (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1368                 struct ceph_pg_pool_info *pi;
1369
1370                 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1371                 if (pi) {
1372                         if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1373                             pi->read_tier >= 0)
1374                                 req->r_target_oloc.pool = pi->read_tier;
1375                         if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1376                             pi->write_tier >= 0)
1377                                 req->r_target_oloc.pool = pi->write_tier;
1378                 }
1379                 /* !pi is caught in ceph_oloc_oid_to_pg() */
1380         }
1381
1382         return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1383                                    &req->r_target_oid, pg_out);
1384 }
1385
1386 static void __enqueue_request(struct ceph_osd_request *req)
1387 {
1388         struct ceph_osd_client *osdc = req->r_osdc;
1389
1390         dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1391              req->r_osd ? req->r_osd->o_osd : -1);
1392
1393         if (req->r_osd) {
1394                 __remove_osd_from_lru(req->r_osd);
1395                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1396                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1397         } else {
1398                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1399         }
1400 }
1401
1402 /*
1403  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1404  * (as needed), and set the request r_osd appropriately.  If there is
1405  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1406  * (unsent, homeless) or leave on in-flight lru.
1407  *
1408  * Return 0 if unchanged, 1 if changed, or negative on error.
1409  *
1410  * Caller should hold map_sem for read and request_mutex.
1411  */
1412 static int __map_request(struct ceph_osd_client *osdc,
1413                          struct ceph_osd_request *req, int force_resend)
1414 {
1415         struct ceph_pg pgid;
1416         int acting[CEPH_PG_MAX_SIZE];
1417         int num, o;
1418         int err;
1419         bool was_paused;
1420
1421         dout("map_request %p tid %lld\n", req, req->r_tid);
1422
1423         err = __calc_request_pg(osdc->osdmap, req, &pgid);
1424         if (err) {
1425                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1426                 return err;
1427         }
1428         req->r_pgid = pgid;
1429
1430         num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1431         if (num < 0)
1432                 num = 0;
1433
1434         was_paused = req->r_paused;
1435         req->r_paused = __req_should_be_paused(osdc, req);
1436         if (was_paused && !req->r_paused)
1437                 force_resend = 1;
1438
1439         if ((!force_resend &&
1440              req->r_osd && req->r_osd->o_osd == o &&
1441              req->r_sent >= req->r_osd->o_incarnation &&
1442              req->r_num_pg_osds == num &&
1443              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1444             (req->r_osd == NULL && o == -1) ||
1445             req->r_paused)
1446                 return 0;  /* no change */
1447
1448         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1449              req->r_tid, pgid.pool, pgid.seed, o,
1450              req->r_osd ? req->r_osd->o_osd : -1);
1451
1452         /* record full pg acting set */
1453         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1454         req->r_num_pg_osds = num;
1455
1456         if (req->r_osd) {
1457                 __cancel_request(req);
1458                 list_del_init(&req->r_osd_item);
1459                 list_del_init(&req->r_linger_osd_item);
1460                 req->r_osd = NULL;
1461         }
1462
1463         req->r_osd = __lookup_osd(osdc, o);
1464         if (!req->r_osd && o >= 0) {
1465                 err = -ENOMEM;
1466                 req->r_osd = create_osd(osdc, o);
1467                 if (!req->r_osd) {
1468                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1469                         goto out;
1470                 }
1471
1472                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1473                 __insert_osd(osdc, req->r_osd);
1474
1475                 ceph_con_open(&req->r_osd->o_con,
1476                               CEPH_ENTITY_TYPE_OSD, o,
1477                               &osdc->osdmap->osd_addr[o]);
1478         }
1479
1480         __enqueue_request(req);
1481         err = 1;   /* osd or pg changed */
1482
1483 out:
1484         return err;
1485 }
1486
1487 /*
1488  * caller should hold map_sem (for read) and request_mutex
1489  */
1490 static void __send_request(struct ceph_osd_client *osdc,
1491                            struct ceph_osd_request *req)
1492 {
1493         void *p;
1494
1495         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1496              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1497              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1498
1499         /* fill in message content that changes each time we send it */
1500         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1501         put_unaligned_le32(req->r_flags, req->r_request_flags);
1502         put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1503         p = req->r_request_pgid;
1504         ceph_encode_64(&p, req->r_pgid.pool);
1505         ceph_encode_32(&p, req->r_pgid.seed);
1506         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1507         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1508                sizeof(req->r_reassert_version));
1509
1510         req->r_stamp = jiffies;
1511         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1512
1513         ceph_msg_get(req->r_request); /* send consumes a ref */
1514
1515         req->r_sent = req->r_osd->o_incarnation;
1516
1517         ceph_con_send(&req->r_osd->o_con, req->r_request);
1518 }
1519
1520 /*
1521  * Send any requests in the queue (req_unsent).
1522  */
1523 static void __send_queued(struct ceph_osd_client *osdc)
1524 {
1525         struct ceph_osd_request *req, *tmp;
1526
1527         dout("__send_queued\n");
1528         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1529                 __send_request(osdc, req);
1530 }
1531
1532 /*
1533  * Caller should hold map_sem for read and request_mutex.
1534  */
1535 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1536                                      struct ceph_osd_request *req,
1537                                      bool nofail)
1538 {
1539         int rc;
1540
1541         __register_request(osdc, req);
1542         req->r_sent = 0;
1543         req->r_got_reply = 0;
1544         rc = __map_request(osdc, req, 0);
1545         if (rc < 0) {
1546                 if (nofail) {
1547                         dout("osdc_start_request failed map, "
1548                                 " will retry %lld\n", req->r_tid);
1549                         rc = 0;
1550                 } else {
1551                         __unregister_request(osdc, req);
1552                 }
1553                 return rc;
1554         }
1555
1556         if (req->r_osd == NULL) {
1557                 dout("send_request %p no up osds in pg\n", req);
1558                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1559         } else {
1560                 __send_queued(osdc);
1561         }
1562
1563         return 0;
1564 }
1565
1566 /*
1567  * Timeout callback, called every N seconds when 1 or more osd
1568  * requests has been active for more than N seconds.  When this
1569  * happens, we ping all OSDs with requests who have timed out to
1570  * ensure any communications channel reset is detected.  Reset the
1571  * request timeouts another N seconds in the future as we go.
1572  * Reschedule the timeout event another N seconds in future (unless
1573  * there are no open requests).
1574  */
1575 static void handle_timeout(struct work_struct *work)
1576 {
1577         struct ceph_osd_client *osdc =
1578                 container_of(work, struct ceph_osd_client, timeout_work.work);
1579         struct ceph_options *opts = osdc->client->options;
1580         struct ceph_osd_request *req;
1581         struct ceph_osd *osd;
1582         struct list_head slow_osds;
1583         dout("timeout\n");
1584         down_read(&osdc->map_sem);
1585
1586         ceph_monc_request_next_osdmap(&osdc->client->monc);
1587
1588         mutex_lock(&osdc->request_mutex);
1589
1590         /*
1591          * ping osds that are a bit slow.  this ensures that if there
1592          * is a break in the TCP connection we will notice, and reopen
1593          * a connection with that osd (from the fault callback).
1594          */
1595         INIT_LIST_HEAD(&slow_osds);
1596         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1597                 if (time_before(jiffies,
1598                                 req->r_stamp + opts->osd_keepalive_timeout))
1599                         break;
1600
1601                 osd = req->r_osd;
1602                 BUG_ON(!osd);
1603                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1604                      req->r_tid, osd->o_osd);
1605                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1606         }
1607         while (!list_empty(&slow_osds)) {
1608                 osd = list_entry(slow_osds.next, struct ceph_osd,
1609                                  o_keepalive_item);
1610                 list_del_init(&osd->o_keepalive_item);
1611                 ceph_con_keepalive(&osd->o_con);
1612         }
1613
1614         __schedule_osd_timeout(osdc);
1615         __send_queued(osdc);
1616         mutex_unlock(&osdc->request_mutex);
1617         up_read(&osdc->map_sem);
1618 }
1619
1620 static void handle_osds_timeout(struct work_struct *work)
1621 {
1622         struct ceph_osd_client *osdc =
1623                 container_of(work, struct ceph_osd_client,
1624                              osds_timeout_work.work);
1625         unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
1626
1627         dout("osds timeout\n");
1628         down_read(&osdc->map_sem);
1629         remove_old_osds(osdc);
1630         up_read(&osdc->map_sem);
1631
1632         schedule_delayed_work(&osdc->osds_timeout_work,
1633                               round_jiffies_relative(delay));
1634 }
1635
1636 static int ceph_oloc_decode(void **p, void *end,
1637                             struct ceph_object_locator *oloc)
1638 {
1639         u8 struct_v, struct_cv;
1640         u32 len;
1641         void *struct_end;
1642         int ret = 0;
1643
1644         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1645         struct_v = ceph_decode_8(p);
1646         struct_cv = ceph_decode_8(p);
1647         if (struct_v < 3) {
1648                 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1649                         struct_v, struct_cv);
1650                 goto e_inval;
1651         }
1652         if (struct_cv > 6) {
1653                 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1654                         struct_v, struct_cv);
1655                 goto e_inval;
1656         }
1657         len = ceph_decode_32(p);
1658         ceph_decode_need(p, end, len, e_inval);
1659         struct_end = *p + len;
1660
1661         oloc->pool = ceph_decode_64(p);
1662         *p += 4; /* skip preferred */
1663
1664         len = ceph_decode_32(p);
1665         if (len > 0) {
1666                 pr_warn("ceph_object_locator::key is set\n");
1667                 goto e_inval;
1668         }
1669
1670         if (struct_v >= 5) {
1671                 len = ceph_decode_32(p);
1672                 if (len > 0) {
1673                         pr_warn("ceph_object_locator::nspace is set\n");
1674                         goto e_inval;
1675                 }
1676         }
1677
1678         if (struct_v >= 6) {
1679                 s64 hash = ceph_decode_64(p);
1680                 if (hash != -1) {
1681                         pr_warn("ceph_object_locator::hash is set\n");
1682                         goto e_inval;
1683                 }
1684         }
1685
1686         /* skip the rest */
1687         *p = struct_end;
1688 out:
1689         return ret;
1690
1691 e_inval:
1692         ret = -EINVAL;
1693         goto out;
1694 }
1695
1696 static int ceph_redirect_decode(void **p, void *end,
1697                                 struct ceph_request_redirect *redir)
1698 {
1699         u8 struct_v, struct_cv;
1700         u32 len;
1701         void *struct_end;
1702         int ret;
1703
1704         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1705         struct_v = ceph_decode_8(p);
1706         struct_cv = ceph_decode_8(p);
1707         if (struct_cv > 1) {
1708                 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1709                         struct_v, struct_cv);
1710                 goto e_inval;
1711         }
1712         len = ceph_decode_32(p);
1713         ceph_decode_need(p, end, len, e_inval);
1714         struct_end = *p + len;
1715
1716         ret = ceph_oloc_decode(p, end, &redir->oloc);
1717         if (ret)
1718                 goto out;
1719
1720         len = ceph_decode_32(p);
1721         if (len > 0) {
1722                 pr_warn("ceph_request_redirect::object_name is set\n");
1723                 goto e_inval;
1724         }
1725
1726         len = ceph_decode_32(p);
1727         *p += len; /* skip osd_instructions */
1728
1729         /* skip the rest */
1730         *p = struct_end;
1731 out:
1732         return ret;
1733
1734 e_inval:
1735         ret = -EINVAL;
1736         goto out;
1737 }
1738
1739 static void complete_request(struct ceph_osd_request *req)
1740 {
1741         complete_all(&req->r_safe_completion);  /* fsync waiter */
1742 }
1743
1744 /*
1745  * handle osd op reply.  either call the callback if it is specified,
1746  * or do the completion to wake up the waiting thread.
1747  */
1748 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1749                          struct ceph_connection *con)
1750 {
1751         void *p, *end;
1752         struct ceph_osd_request *req;
1753         struct ceph_request_redirect redir;
1754         u64 tid;
1755         int object_len;
1756         unsigned int numops;
1757         int payload_len, flags;
1758         s32 result;
1759         s32 retry_attempt;
1760         struct ceph_pg pg;
1761         int err;
1762         u32 reassert_epoch;
1763         u64 reassert_version;
1764         u32 osdmap_epoch;
1765         int already_completed;
1766         u32 bytes;
1767         unsigned int i;
1768
1769         tid = le64_to_cpu(msg->hdr.tid);
1770         dout("handle_reply %p tid %llu\n", msg, tid);
1771
1772         p = msg->front.iov_base;
1773         end = p + msg->front.iov_len;
1774
1775         ceph_decode_need(&p, end, 4, bad);
1776         object_len = ceph_decode_32(&p);
1777         ceph_decode_need(&p, end, object_len, bad);
1778         p += object_len;
1779
1780         err = ceph_decode_pgid(&p, end, &pg);
1781         if (err)
1782                 goto bad;
1783
1784         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1785         flags = ceph_decode_64(&p);
1786         result = ceph_decode_32(&p);
1787         reassert_epoch = ceph_decode_32(&p);
1788         reassert_version = ceph_decode_64(&p);
1789         osdmap_epoch = ceph_decode_32(&p);
1790
1791         /* lookup */
1792         down_read(&osdc->map_sem);
1793         mutex_lock(&osdc->request_mutex);
1794         req = __lookup_request(osdc, tid);
1795         if (req == NULL) {
1796                 dout("handle_reply tid %llu dne\n", tid);
1797                 goto bad_mutex;
1798         }
1799         ceph_osdc_get_request(req);
1800
1801         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1802              req, result);
1803
1804         ceph_decode_need(&p, end, 4, bad_put);
1805         numops = ceph_decode_32(&p);
1806         if (numops > CEPH_OSD_MAX_OP)
1807                 goto bad_put;
1808         if (numops != req->r_num_ops)
1809                 goto bad_put;
1810         payload_len = 0;
1811         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1812         for (i = 0; i < numops; i++) {
1813                 struct ceph_osd_op *op = p;
1814                 int len;
1815
1816                 len = le32_to_cpu(op->payload_len);
1817                 req->r_reply_op_len[i] = len;
1818                 dout(" op %d has %d bytes\n", i, len);
1819                 payload_len += len;
1820                 p += sizeof(*op);
1821         }
1822         bytes = le32_to_cpu(msg->hdr.data_len);
1823         if (payload_len != bytes) {
1824                 pr_warn("sum of op payload lens %d != data_len %d\n",
1825                         payload_len, bytes);
1826                 goto bad_put;
1827         }
1828
1829         ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1830         retry_attempt = ceph_decode_32(&p);
1831         for (i = 0; i < numops; i++)
1832                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1833
1834         if (le16_to_cpu(msg->hdr.version) >= 6) {
1835                 p += 8 + 4; /* skip replay_version */
1836                 p += 8; /* skip user_version */
1837
1838                 err = ceph_redirect_decode(&p, end, &redir);
1839                 if (err)
1840                         goto bad_put;
1841         } else {
1842                 redir.oloc.pool = -1;
1843         }
1844
1845         if (redir.oloc.pool != -1) {
1846                 dout("redirect pool %lld\n", redir.oloc.pool);
1847
1848                 __unregister_request(osdc, req);
1849
1850                 req->r_target_oloc = redir.oloc; /* struct */
1851
1852                 /*
1853                  * Start redirect requests with nofail=true.  If
1854                  * mapping fails, request will end up on the notarget
1855                  * list, waiting for the new osdmap (which can take
1856                  * a while), even though the original request mapped
1857                  * successfully.  In the future we might want to follow
1858                  * original request's nofail setting here.
1859                  */
1860                 err = __ceph_osdc_start_request(osdc, req, true);
1861                 BUG_ON(err);
1862
1863                 goto out_unlock;
1864         }
1865
1866         already_completed = req->r_got_reply;
1867         if (!req->r_got_reply) {
1868                 req->r_result = result;
1869                 dout("handle_reply result %d bytes %d\n", req->r_result,
1870                      bytes);
1871                 if (req->r_result == 0)
1872                         req->r_result = bytes;
1873
1874                 /* in case this is a write and we need to replay, */
1875                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1876                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1877
1878                 req->r_got_reply = 1;
1879         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1880                 dout("handle_reply tid %llu dup ack\n", tid);
1881                 goto out_unlock;
1882         }
1883
1884         dout("handle_reply tid %llu flags %d\n", tid, flags);
1885
1886         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1887                 __register_linger_request(osdc, req);
1888
1889         /* either this is a read, or we got the safe response */
1890         if (result < 0 ||
1891             (flags & CEPH_OSD_FLAG_ONDISK) ||
1892             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1893                 __unregister_request(osdc, req);
1894
1895         mutex_unlock(&osdc->request_mutex);
1896         up_read(&osdc->map_sem);
1897
1898         if (!already_completed) {
1899                 if (req->r_unsafe_callback &&
1900                     result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1901                         req->r_unsafe_callback(req, true);
1902                 if (req->r_callback)
1903                         req->r_callback(req, msg);
1904                 else
1905                         complete_all(&req->r_completion);
1906         }
1907
1908         if (flags & CEPH_OSD_FLAG_ONDISK) {
1909                 if (req->r_unsafe_callback && already_completed)
1910                         req->r_unsafe_callback(req, false);
1911                 complete_request(req);
1912         }
1913
1914 out:
1915         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1916         ceph_osdc_put_request(req);
1917         return;
1918 out_unlock:
1919         mutex_unlock(&osdc->request_mutex);
1920         up_read(&osdc->map_sem);
1921         goto out;
1922
1923 bad_put:
1924         req->r_result = -EIO;
1925         __unregister_request(osdc, req);
1926         if (req->r_callback)
1927                 req->r_callback(req, msg);
1928         else
1929                 complete_all(&req->r_completion);
1930         complete_request(req);
1931         ceph_osdc_put_request(req);
1932 bad_mutex:
1933         mutex_unlock(&osdc->request_mutex);
1934         up_read(&osdc->map_sem);
1935 bad:
1936         pr_err("corrupt osd_op_reply got %d %d\n",
1937                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1938         ceph_msg_dump(msg);
1939 }
1940
1941 static void reset_changed_osds(struct ceph_osd_client *osdc)
1942 {
1943         struct rb_node *p, *n;
1944
1945         dout("%s %p\n", __func__, osdc);
1946         for (p = rb_first(&osdc->osds); p; p = n) {
1947                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1948
1949                 n = rb_next(p);
1950                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1951                     memcmp(&osd->o_con.peer_addr,
1952                            ceph_osd_addr(osdc->osdmap,
1953                                          osd->o_osd),
1954                            sizeof(struct ceph_entity_addr)) != 0)
1955                         __reset_osd(osdc, osd);
1956         }
1957 }
1958
1959 /*
1960  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1961  * no osd, request a new map.
1962  *
1963  * Caller should hold map_sem for read.
1964  */
1965 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1966                           bool force_resend_writes)
1967 {
1968         struct ceph_osd_request *req, *nreq;
1969         struct rb_node *p;
1970         int needmap = 0;
1971         int err;
1972         bool force_resend_req;
1973
1974         dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1975                 force_resend_writes ? " (force resend writes)" : "");
1976         mutex_lock(&osdc->request_mutex);
1977         for (p = rb_first(&osdc->requests); p; ) {
1978                 req = rb_entry(p, struct ceph_osd_request, r_node);
1979                 p = rb_next(p);
1980
1981                 /*
1982                  * For linger requests that have not yet been
1983                  * registered, move them to the linger list; they'll
1984                  * be sent to the osd in the loop below.  Unregister
1985                  * the request before re-registering it as a linger
1986                  * request to ensure the __map_request() below
1987                  * will decide it needs to be sent.
1988                  */
1989                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1990                         dout("%p tid %llu restart on osd%d\n",
1991                              req, req->r_tid,
1992                              req->r_osd ? req->r_osd->o_osd : -1);
1993                         ceph_osdc_get_request(req);
1994                         __unregister_request(osdc, req);
1995                         __register_linger_request(osdc, req);
1996                         ceph_osdc_put_request(req);
1997                         continue;
1998                 }
1999
2000                 force_resend_req = force_resend ||
2001                         (force_resend_writes &&
2002                                 req->r_flags & CEPH_OSD_FLAG_WRITE);
2003                 err = __map_request(osdc, req, force_resend_req);
2004                 if (err < 0)
2005                         continue;  /* error */
2006                 if (req->r_osd == NULL) {
2007                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2008                         needmap++;  /* request a newer map */
2009                 } else if (err > 0) {
2010                         if (!req->r_linger) {
2011                                 dout("%p tid %llu requeued on osd%d\n", req,
2012                                      req->r_tid,
2013                                      req->r_osd ? req->r_osd->o_osd : -1);
2014                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2015                         }
2016                 }
2017         }
2018
2019         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2020                                  r_linger_item) {
2021                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2022
2023                 err = __map_request(osdc, req,
2024                                     force_resend || force_resend_writes);
2025                 dout("__map_request returned %d\n", err);
2026                 if (err < 0)
2027                         continue;  /* hrm! */
2028                 if (req->r_osd == NULL || err > 0) {
2029                         if (req->r_osd == NULL) {
2030                                 dout("lingering %p tid %llu maps to no osd\n",
2031                                      req, req->r_tid);
2032                                 /*
2033                                  * A homeless lingering request makes
2034                                  * no sense, as it's job is to keep
2035                                  * a particular OSD connection open.
2036                                  * Request a newer map and kick the
2037                                  * request, knowing that it won't be
2038                                  * resent until we actually get a map
2039                                  * that can tell us where to send it.
2040                                  */
2041                                 needmap++;
2042                         }
2043
2044                         dout("kicking lingering %p tid %llu osd%d\n", req,
2045                              req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2046                         __register_request(osdc, req);
2047                         __unregister_linger_request(osdc, req);
2048                 }
2049         }
2050         reset_changed_osds(osdc);
2051         mutex_unlock(&osdc->request_mutex);
2052
2053         if (needmap) {
2054                 dout("%d requests for down osds, need new map\n", needmap);
2055                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2056         }
2057 }
2058
2059
2060 /*
2061  * Process updated osd map.
2062  *
2063  * The message contains any number of incremental and full maps, normally
2064  * indicating some sort of topology change in the cluster.  Kick requests
2065  * off to different OSDs as needed.
2066  */
2067 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2068 {
2069         void *p, *end, *next;
2070         u32 nr_maps, maplen;
2071         u32 epoch;
2072         struct ceph_osdmap *newmap = NULL, *oldmap;
2073         int err;
2074         struct ceph_fsid fsid;
2075         bool was_full;
2076
2077         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2078         p = msg->front.iov_base;
2079         end = p + msg->front.iov_len;
2080
2081         /* verify fsid */
2082         ceph_decode_need(&p, end, sizeof(fsid), bad);
2083         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2084         if (ceph_check_fsid(osdc->client, &fsid) < 0)
2085                 return;
2086
2087         down_write(&osdc->map_sem);
2088
2089         was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2090
2091         /* incremental maps */
2092         ceph_decode_32_safe(&p, end, nr_maps, bad);
2093         dout(" %d inc maps\n", nr_maps);
2094         while (nr_maps > 0) {
2095                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2096                 epoch = ceph_decode_32(&p);
2097                 maplen = ceph_decode_32(&p);
2098                 ceph_decode_need(&p, end, maplen, bad);
2099                 next = p + maplen;
2100                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2101                         dout("applying incremental map %u len %d\n",
2102                              epoch, maplen);
2103                         newmap = osdmap_apply_incremental(&p, next,
2104                                                           osdc->osdmap,
2105                                                           &osdc->client->msgr);
2106                         if (IS_ERR(newmap)) {
2107                                 err = PTR_ERR(newmap);
2108                                 goto bad;
2109                         }
2110                         BUG_ON(!newmap);
2111                         if (newmap != osdc->osdmap) {
2112                                 ceph_osdmap_destroy(osdc->osdmap);
2113                                 osdc->osdmap = newmap;
2114                         }
2115                         was_full = was_full ||
2116                                 ceph_osdmap_flag(osdc->osdmap,
2117                                                  CEPH_OSDMAP_FULL);
2118                         kick_requests(osdc, 0, was_full);
2119                 } else {
2120                         dout("ignoring incremental map %u len %d\n",
2121                              epoch, maplen);
2122                 }
2123                 p = next;
2124                 nr_maps--;
2125         }
2126         if (newmap)
2127                 goto done;
2128
2129         /* full maps */
2130         ceph_decode_32_safe(&p, end, nr_maps, bad);
2131         dout(" %d full maps\n", nr_maps);
2132         while (nr_maps) {
2133                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2134                 epoch = ceph_decode_32(&p);
2135                 maplen = ceph_decode_32(&p);
2136                 ceph_decode_need(&p, end, maplen, bad);
2137                 if (nr_maps > 1) {
2138                         dout("skipping non-latest full map %u len %d\n",
2139                              epoch, maplen);
2140                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2141                         dout("skipping full map %u len %d, "
2142                              "older than our %u\n", epoch, maplen,
2143                              osdc->osdmap->epoch);
2144                 } else {
2145                         int skipped_map = 0;
2146
2147                         dout("taking full map %u len %d\n", epoch, maplen);
2148                         newmap = ceph_osdmap_decode(&p, p+maplen);
2149                         if (IS_ERR(newmap)) {
2150                                 err = PTR_ERR(newmap);
2151                                 goto bad;
2152                         }
2153                         BUG_ON(!newmap);
2154                         oldmap = osdc->osdmap;
2155                         osdc->osdmap = newmap;
2156                         if (oldmap) {
2157                                 if (oldmap->epoch + 1 < newmap->epoch)
2158                                         skipped_map = 1;
2159                                 ceph_osdmap_destroy(oldmap);
2160                         }
2161                         was_full = was_full ||
2162                                 ceph_osdmap_flag(osdc->osdmap,
2163                                                  CEPH_OSDMAP_FULL);
2164                         kick_requests(osdc, skipped_map, was_full);
2165                 }
2166                 p += maplen;
2167                 nr_maps--;
2168         }
2169
2170         if (!osdc->osdmap)
2171                 goto bad;
2172 done:
2173         downgrade_write(&osdc->map_sem);
2174         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2175
2176         /*
2177          * subscribe to subsequent osdmap updates if full to ensure
2178          * we find out when we are no longer full and stop returning
2179          * ENOSPC.
2180          */
2181         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2182                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2183                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2184                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2185
2186         mutex_lock(&osdc->request_mutex);
2187         __send_queued(osdc);
2188         mutex_unlock(&osdc->request_mutex);
2189         up_read(&osdc->map_sem);
2190         wake_up_all(&osdc->client->auth_wq);
2191         return;
2192
2193 bad:
2194         pr_err("osdc handle_map corrupt msg\n");
2195         ceph_msg_dump(msg);
2196         up_write(&osdc->map_sem);
2197 }
2198
2199 /*
2200  * watch/notify callback event infrastructure
2201  *
2202  * These callbacks are used both for watch and notify operations.
2203  */
2204 static void __release_event(struct kref *kref)
2205 {
2206         struct ceph_osd_event *event =
2207                 container_of(kref, struct ceph_osd_event, kref);
2208
2209         dout("__release_event %p\n", event);
2210         kfree(event);
2211 }
2212
2213 static void get_event(struct ceph_osd_event *event)
2214 {
2215         kref_get(&event->kref);
2216 }
2217
2218 void ceph_osdc_put_event(struct ceph_osd_event *event)
2219 {
2220         kref_put(&event->kref, __release_event);
2221 }
2222 EXPORT_SYMBOL(ceph_osdc_put_event);
2223
2224 static void __insert_event(struct ceph_osd_client *osdc,
2225                              struct ceph_osd_event *new)
2226 {
2227         struct rb_node **p = &osdc->event_tree.rb_node;
2228         struct rb_node *parent = NULL;
2229         struct ceph_osd_event *event = NULL;
2230
2231         while (*p) {
2232                 parent = *p;
2233                 event = rb_entry(parent, struct ceph_osd_event, node);
2234                 if (new->cookie < event->cookie)
2235                         p = &(*p)->rb_left;
2236                 else if (new->cookie > event->cookie)
2237                         p = &(*p)->rb_right;
2238                 else
2239                         BUG();
2240         }
2241
2242         rb_link_node(&new->node, parent, p);
2243         rb_insert_color(&new->node, &osdc->event_tree);
2244 }
2245
2246 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2247                                                 u64 cookie)
2248 {
2249         struct rb_node **p = &osdc->event_tree.rb_node;
2250         struct rb_node *parent = NULL;
2251         struct ceph_osd_event *event = NULL;
2252
2253         while (*p) {
2254                 parent = *p;
2255                 event = rb_entry(parent, struct ceph_osd_event, node);
2256                 if (cookie < event->cookie)
2257                         p = &(*p)->rb_left;
2258                 else if (cookie > event->cookie)
2259                         p = &(*p)->rb_right;
2260                 else
2261                         return event;
2262         }
2263         return NULL;
2264 }
2265
2266 static void __remove_event(struct ceph_osd_event *event)
2267 {
2268         struct ceph_osd_client *osdc = event->osdc;
2269
2270         if (!RB_EMPTY_NODE(&event->node)) {
2271                 dout("__remove_event removed %p\n", event);
2272                 rb_erase(&event->node, &osdc->event_tree);
2273                 ceph_osdc_put_event(event);
2274         } else {
2275                 dout("__remove_event didn't remove %p\n", event);
2276         }
2277 }
2278
2279 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2280                            void (*event_cb)(u64, u64, u8, void *),
2281                            void *data, struct ceph_osd_event **pevent)
2282 {
2283         struct ceph_osd_event *event;
2284
2285         event = kmalloc(sizeof(*event), GFP_NOIO);
2286         if (!event)
2287                 return -ENOMEM;
2288
2289         dout("create_event %p\n", event);
2290         event->cb = event_cb;
2291         event->one_shot = 0;
2292         event->data = data;
2293         event->osdc = osdc;
2294         INIT_LIST_HEAD(&event->osd_node);
2295         RB_CLEAR_NODE(&event->node);
2296         kref_init(&event->kref);   /* one ref for us */
2297         kref_get(&event->kref);    /* one ref for the caller */
2298
2299         spin_lock(&osdc->event_lock);
2300         event->cookie = ++osdc->event_count;
2301         __insert_event(osdc, event);
2302         spin_unlock(&osdc->event_lock);
2303
2304         *pevent = event;
2305         return 0;
2306 }
2307 EXPORT_SYMBOL(ceph_osdc_create_event);
2308
2309 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2310 {
2311         struct ceph_osd_client *osdc = event->osdc;
2312
2313         dout("cancel_event %p\n", event);
2314         spin_lock(&osdc->event_lock);
2315         __remove_event(event);
2316         spin_unlock(&osdc->event_lock);
2317         ceph_osdc_put_event(event); /* caller's */
2318 }
2319 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2320
2321
2322 static void do_event_work(struct work_struct *work)
2323 {
2324         struct ceph_osd_event_work *event_work =
2325                 container_of(work, struct ceph_osd_event_work, work);
2326         struct ceph_osd_event *event = event_work->event;
2327         u64 ver = event_work->ver;
2328         u64 notify_id = event_work->notify_id;
2329         u8 opcode = event_work->opcode;
2330
2331         dout("do_event_work completing %p\n", event);
2332         event->cb(ver, notify_id, opcode, event->data);
2333         dout("do_event_work completed %p\n", event);
2334         ceph_osdc_put_event(event);
2335         kfree(event_work);
2336 }
2337
2338
2339 /*
2340  * Process osd watch notifications
2341  */
2342 static void handle_watch_notify(struct ceph_osd_client *osdc,
2343                                 struct ceph_msg *msg)
2344 {
2345         void *p, *end;
2346         u8 proto_ver;
2347         u64 cookie, ver, notify_id;
2348         u8 opcode;
2349         struct ceph_osd_event *event;
2350         struct ceph_osd_event_work *event_work;
2351
2352         p = msg->front.iov_base;
2353         end = p + msg->front.iov_len;
2354
2355         ceph_decode_8_safe(&p, end, proto_ver, bad);
2356         ceph_decode_8_safe(&p, end, opcode, bad);
2357         ceph_decode_64_safe(&p, end, cookie, bad);
2358         ceph_decode_64_safe(&p, end, ver, bad);
2359         ceph_decode_64_safe(&p, end, notify_id, bad);
2360
2361         spin_lock(&osdc->event_lock);
2362         event = __find_event(osdc, cookie);
2363         if (event) {
2364                 BUG_ON(event->one_shot);
2365                 get_event(event);
2366         }
2367         spin_unlock(&osdc->event_lock);
2368         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2369              cookie, ver, event);
2370         if (event) {
2371                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2372                 if (!event_work) {
2373                         pr_err("couldn't allocate event_work\n");
2374                         ceph_osdc_put_event(event);
2375                         return;
2376                 }
2377                 INIT_WORK(&event_work->work, do_event_work);
2378                 event_work->event = event;
2379                 event_work->ver = ver;
2380                 event_work->notify_id = notify_id;
2381                 event_work->opcode = opcode;
2382
2383                 queue_work(osdc->notify_wq, &event_work->work);
2384         }
2385
2386         return;
2387
2388 bad:
2389         pr_err("osdc handle_watch_notify corrupt msg\n");
2390 }
2391
2392 /*
2393  * build new request AND message
2394  *
2395  */
2396 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2397                                 struct ceph_snap_context *snapc, u64 snap_id,
2398                                 struct timespec *mtime)
2399 {
2400         struct ceph_msg *msg = req->r_request;
2401         void *p;
2402         size_t msg_size;
2403         int flags = req->r_flags;
2404         u64 data_len;
2405         unsigned int i;
2406
2407         req->r_snapid = snap_id;
2408         req->r_snapc = ceph_get_snap_context(snapc);
2409
2410         /* encode request */
2411         msg->hdr.version = cpu_to_le16(4);
2412
2413         p = msg->front.iov_base;
2414         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2415         req->r_request_osdmap_epoch = p;
2416         p += 4;
2417         req->r_request_flags = p;
2418         p += 4;
2419         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2420                 ceph_encode_timespec(p, mtime);
2421         p += sizeof(struct ceph_timespec);
2422         req->r_request_reassert_version = p;
2423         p += sizeof(struct ceph_eversion); /* will get filled in */
2424
2425         /* oloc */
2426         ceph_encode_8(&p, 4);
2427         ceph_encode_8(&p, 4);
2428         ceph_encode_32(&p, 8 + 4 + 4);
2429         req->r_request_pool = p;
2430         p += 8;
2431         ceph_encode_32(&p, -1);  /* preferred */
2432         ceph_encode_32(&p, 0);   /* key len */
2433
2434         ceph_encode_8(&p, 1);
2435         req->r_request_pgid = p;
2436         p += 8 + 4;
2437         ceph_encode_32(&p, -1);  /* preferred */
2438
2439         /* oid */
2440         ceph_encode_32(&p, req->r_base_oid.name_len);
2441         memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2442         dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2443              req->r_base_oid.name, req->r_base_oid.name_len);
2444         p += req->r_base_oid.name_len;
2445
2446         /* ops--can imply data */
2447         ceph_encode_16(&p, (u16)req->r_num_ops);
2448         data_len = 0;
2449         for (i = 0; i < req->r_num_ops; i++) {
2450                 data_len += osd_req_encode_op(req, p, i);
2451                 p += sizeof(struct ceph_osd_op);
2452         }
2453
2454         /* snaps */
2455         ceph_encode_64(&p, req->r_snapid);
2456         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2457         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2458         if (req->r_snapc) {
2459                 for (i = 0; i < snapc->num_snaps; i++) {
2460                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2461                 }
2462         }
2463
2464         req->r_request_attempts = p;
2465         p += 4;
2466
2467         /* data */
2468         if (flags & CEPH_OSD_FLAG_WRITE) {
2469                 u16 data_off;
2470
2471                 /*
2472                  * The header "data_off" is a hint to the receiver
2473                  * allowing it to align received data into its
2474                  * buffers such that there's no need to re-copy
2475                  * it before writing it to disk (direct I/O).
2476                  */
2477                 data_off = (u16) (off & 0xffff);
2478                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2479         }
2480         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2481
2482         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2483         msg_size = p - msg->front.iov_base;
2484         msg->front.iov_len = msg_size;
2485         msg->hdr.front_len = cpu_to_le32(msg_size);
2486
2487         dout("build_request msg_size was %d\n", (int)msg_size);
2488 }
2489 EXPORT_SYMBOL(ceph_osdc_build_request);
2490
2491 /*
2492  * Register request, send initial attempt.
2493  */
2494 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2495                             struct ceph_osd_request *req,
2496                             bool nofail)
2497 {
2498         int rc;
2499
2500         down_read(&osdc->map_sem);
2501         mutex_lock(&osdc->request_mutex);
2502
2503         rc = __ceph_osdc_start_request(osdc, req, nofail);
2504
2505         mutex_unlock(&osdc->request_mutex);
2506         up_read(&osdc->map_sem);
2507
2508         return rc;
2509 }
2510 EXPORT_SYMBOL(ceph_osdc_start_request);
2511
2512 /*
2513  * Unregister a registered request.  The request is not completed (i.e.
2514  * no callbacks or wakeups) - higher layers are supposed to know what
2515  * they are canceling.
2516  */
2517 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2518 {
2519         struct ceph_osd_client *osdc = req->r_osdc;
2520
2521         mutex_lock(&osdc->request_mutex);
2522         if (req->r_linger)
2523                 __unregister_linger_request(osdc, req);
2524         __unregister_request(osdc, req);
2525         mutex_unlock(&osdc->request_mutex);
2526
2527         dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2528 }
2529 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2530
2531 /*
2532  * wait for a request to complete
2533  */
2534 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2535                            struct ceph_osd_request *req)
2536 {
2537         int rc;
2538
2539         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2540
2541         rc = wait_for_completion_interruptible(&req->r_completion);
2542         if (rc < 0) {
2543                 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2544                 ceph_osdc_cancel_request(req);
2545                 complete_request(req);
2546                 return rc;
2547         }
2548
2549         dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2550              req->r_result);
2551         return req->r_result;
2552 }
2553 EXPORT_SYMBOL(ceph_osdc_wait_request);
2554
2555 /*
2556  * sync - wait for all in-flight requests to flush.  avoid starvation.
2557  */
2558 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2559 {
2560         struct ceph_osd_request *req;
2561         u64 last_tid, next_tid = 0;
2562
2563         mutex_lock(&osdc->request_mutex);
2564         last_tid = osdc->last_tid;
2565         while (1) {
2566                 req = __lookup_request_ge(osdc, next_tid);
2567                 if (!req)
2568                         break;
2569                 if (req->r_tid > last_tid)
2570                         break;
2571
2572                 next_tid = req->r_tid + 1;
2573                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2574                         continue;
2575
2576                 ceph_osdc_get_request(req);
2577                 mutex_unlock(&osdc->request_mutex);
2578                 dout("sync waiting on tid %llu (last is %llu)\n",
2579                      req->r_tid, last_tid);
2580                 wait_for_completion(&req->r_safe_completion);
2581                 mutex_lock(&osdc->request_mutex);
2582                 ceph_osdc_put_request(req);
2583         }
2584         mutex_unlock(&osdc->request_mutex);
2585         dout("sync done (thru tid %llu)\n", last_tid);
2586 }
2587 EXPORT_SYMBOL(ceph_osdc_sync);
2588
2589 /*
2590  * Call all pending notify callbacks - for use after a watch is
2591  * unregistered, to make sure no more callbacks for it will be invoked
2592  */
2593 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2594 {
2595         flush_workqueue(osdc->notify_wq);
2596 }
2597 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2598
2599
2600 /*
2601  * init, shutdown
2602  */
2603 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2604 {
2605         int err;
2606
2607         dout("init\n");
2608         osdc->client = client;
2609         osdc->osdmap = NULL;
2610         init_rwsem(&osdc->map_sem);
2611         init_completion(&osdc->map_waiters);
2612         osdc->last_requested_map = 0;
2613         mutex_init(&osdc->request_mutex);
2614         osdc->last_tid = 0;
2615         osdc->osds = RB_ROOT;
2616         INIT_LIST_HEAD(&osdc->osd_lru);
2617         osdc->requests = RB_ROOT;
2618         INIT_LIST_HEAD(&osdc->req_lru);
2619         INIT_LIST_HEAD(&osdc->req_unsent);
2620         INIT_LIST_HEAD(&osdc->req_notarget);
2621         INIT_LIST_HEAD(&osdc->req_linger);
2622         osdc->num_requests = 0;
2623         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2624         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2625         spin_lock_init(&osdc->event_lock);
2626         osdc->event_tree = RB_ROOT;
2627         osdc->event_count = 0;
2628
2629         schedule_delayed_work(&osdc->osds_timeout_work,
2630             round_jiffies_relative(osdc->client->options->osd_idle_ttl));
2631
2632         err = -ENOMEM;
2633         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2634                                         sizeof(struct ceph_osd_request));
2635         if (!osdc->req_mempool)
2636                 goto out;
2637
2638         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2639                                 OSD_OP_FRONT_LEN, 10, true,
2640                                 "osd_op");
2641         if (err < 0)
2642                 goto out_mempool;
2643         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2644                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2645                                 "osd_op_reply");
2646         if (err < 0)
2647                 goto out_msgpool;
2648
2649         err = -ENOMEM;
2650         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2651         if (!osdc->notify_wq)
2652                 goto out_msgpool_reply;
2653
2654         return 0;
2655
2656 out_msgpool_reply:
2657         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2658 out_msgpool:
2659         ceph_msgpool_destroy(&osdc->msgpool_op);
2660 out_mempool:
2661         mempool_destroy(osdc->req_mempool);
2662 out:
2663         return err;
2664 }
2665
2666 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2667 {
2668         flush_workqueue(osdc->notify_wq);
2669         destroy_workqueue(osdc->notify_wq);
2670         cancel_delayed_work_sync(&osdc->timeout_work);
2671         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2672         if (osdc->osdmap) {
2673                 ceph_osdmap_destroy(osdc->osdmap);
2674                 osdc->osdmap = NULL;
2675         }
2676         remove_all_osds(osdc);
2677         mempool_destroy(osdc->req_mempool);
2678         ceph_msgpool_destroy(&osdc->msgpool_op);
2679         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2680 }
2681
2682 /*
2683  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2684  * *plen.  Return number of bytes read, or error.
2685  */
2686 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2687                         struct ceph_vino vino, struct ceph_file_layout *layout,
2688                         u64 off, u64 *plen,
2689                         u32 truncate_seq, u64 truncate_size,
2690                         struct page **pages, int num_pages, int page_align)
2691 {
2692         struct ceph_osd_request *req;
2693         int rc = 0;
2694
2695         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2696              vino.snap, off, *plen);
2697         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2698                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2699                                     NULL, truncate_seq, truncate_size,
2700                                     false);
2701         if (IS_ERR(req))
2702                 return PTR_ERR(req);
2703
2704         /* it may be a short read due to an object boundary */
2705
2706         osd_req_op_extent_osd_data_pages(req, 0,
2707                                 pages, *plen, page_align, false, false);
2708
2709         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2710              off, *plen, *plen, page_align);
2711
2712         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2713
2714         rc = ceph_osdc_start_request(osdc, req, false);
2715         if (!rc)
2716                 rc = ceph_osdc_wait_request(osdc, req);
2717
2718         ceph_osdc_put_request(req);
2719         dout("readpages result %d\n", rc);
2720         return rc;
2721 }
2722 EXPORT_SYMBOL(ceph_osdc_readpages);
2723
2724 /*
2725  * do a synchronous write on N pages
2726  */
2727 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2728                          struct ceph_file_layout *layout,
2729                          struct ceph_snap_context *snapc,
2730                          u64 off, u64 len,
2731                          u32 truncate_seq, u64 truncate_size,
2732                          struct timespec *mtime,
2733                          struct page **pages, int num_pages)
2734 {
2735         struct ceph_osd_request *req;
2736         int rc = 0;
2737         int page_align = off & ~PAGE_MASK;
2738
2739         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2740         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2741                                     CEPH_OSD_OP_WRITE,
2742                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2743                                     snapc, truncate_seq, truncate_size,
2744                                     true);
2745         if (IS_ERR(req))
2746                 return PTR_ERR(req);
2747
2748         /* it may be a short write due to an object boundary */
2749         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2750                                 false, false);
2751         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2752
2753         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2754
2755         rc = ceph_osdc_start_request(osdc, req, true);
2756         if (!rc)
2757                 rc = ceph_osdc_wait_request(osdc, req);
2758
2759         ceph_osdc_put_request(req);
2760         if (rc == 0)
2761                 rc = len;
2762         dout("writepages result %d\n", rc);
2763         return rc;
2764 }
2765 EXPORT_SYMBOL(ceph_osdc_writepages);
2766
2767 int ceph_osdc_setup(void)
2768 {
2769         BUG_ON(ceph_osd_request_cache);
2770         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2771                                         sizeof (struct ceph_osd_request),
2772                                         __alignof__(struct ceph_osd_request),
2773                                         0, NULL);
2774
2775         return ceph_osd_request_cache ? 0 : -ENOMEM;
2776 }
2777 EXPORT_SYMBOL(ceph_osdc_setup);
2778
2779 void ceph_osdc_cleanup(void)
2780 {
2781         BUG_ON(!ceph_osd_request_cache);
2782         kmem_cache_destroy(ceph_osd_request_cache);
2783         ceph_osd_request_cache = NULL;
2784 }
2785 EXPORT_SYMBOL(ceph_osdc_cleanup);
2786
2787 /*
2788  * handle incoming message
2789  */
2790 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2791 {
2792         struct ceph_osd *osd = con->private;
2793         struct ceph_osd_client *osdc;
2794         int type = le16_to_cpu(msg->hdr.type);
2795
2796         if (!osd)
2797                 goto out;
2798         osdc = osd->o_osdc;
2799
2800         switch (type) {
2801         case CEPH_MSG_OSD_MAP:
2802                 ceph_osdc_handle_map(osdc, msg);
2803                 break;
2804         case CEPH_MSG_OSD_OPREPLY:
2805                 handle_reply(osdc, msg, con);
2806                 break;
2807         case CEPH_MSG_WATCH_NOTIFY:
2808                 handle_watch_notify(osdc, msg);
2809                 break;
2810
2811         default:
2812                 pr_err("received unknown message type %d %s\n", type,
2813                        ceph_msg_type_name(type));
2814         }
2815 out:
2816         ceph_msg_put(msg);
2817 }
2818
2819 /*
2820  * lookup and return message for incoming reply.  set up reply message
2821  * pages.
2822  */
2823 static struct ceph_msg *get_reply(struct ceph_connection *con,
2824                                   struct ceph_msg_header *hdr,
2825                                   int *skip)
2826 {
2827         struct ceph_osd *osd = con->private;
2828         struct ceph_osd_client *osdc = osd->o_osdc;
2829         struct ceph_msg *m;
2830         struct ceph_osd_request *req;
2831         int front_len = le32_to_cpu(hdr->front_len);
2832         int data_len = le32_to_cpu(hdr->data_len);
2833         u64 tid;
2834
2835         tid = le64_to_cpu(hdr->tid);
2836         mutex_lock(&osdc->request_mutex);
2837         req = __lookup_request(osdc, tid);
2838         if (!req) {
2839                 *skip = 1;
2840                 m = NULL;
2841                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2842                      osd->o_osd);
2843                 goto out;
2844         }
2845
2846         if (req->r_reply->con)
2847                 dout("%s revoking msg %p from old con %p\n", __func__,
2848                      req->r_reply, req->r_reply->con);
2849         ceph_msg_revoke_incoming(req->r_reply);
2850
2851         if (front_len > req->r_reply->front_alloc_len) {
2852                 pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2853                         front_len, req->r_reply->front_alloc_len,
2854                         (unsigned int)con->peer_name.type,
2855                         le64_to_cpu(con->peer_name.num));
2856                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2857                                  false);
2858                 if (!m)
2859                         goto out;
2860                 ceph_msg_put(req->r_reply);
2861                 req->r_reply = m;
2862         }
2863         m = ceph_msg_get(req->r_reply);
2864
2865         if (data_len > 0) {
2866                 struct ceph_osd_data *osd_data;
2867
2868                 /*
2869                  * XXX This is assuming there is only one op containing
2870                  * XXX page data.  Probably OK for reads, but this
2871                  * XXX ought to be done more generally.
2872                  */
2873                 osd_data = osd_req_op_extent_osd_data(req, 0);
2874                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2875                         if (osd_data->pages &&
2876                                 unlikely(osd_data->length < data_len)) {
2877
2878                                 pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2879                                         tid, data_len, osd_data->length);
2880                                 *skip = 1;
2881                                 ceph_msg_put(m);
2882                                 m = NULL;
2883                                 goto out;
2884                         }
2885                 }
2886         }
2887         *skip = 0;
2888         dout("get_reply tid %lld %p\n", tid, m);
2889
2890 out:
2891         mutex_unlock(&osdc->request_mutex);
2892         return m;
2893
2894 }
2895
2896 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2897                                   struct ceph_msg_header *hdr,
2898                                   int *skip)
2899 {
2900         struct ceph_osd *osd = con->private;
2901         int type = le16_to_cpu(hdr->type);
2902         int front = le32_to_cpu(hdr->front_len);
2903
2904         *skip = 0;
2905         switch (type) {
2906         case CEPH_MSG_OSD_MAP:
2907         case CEPH_MSG_WATCH_NOTIFY:
2908                 return ceph_msg_new(type, front, GFP_NOFS, false);
2909         case CEPH_MSG_OSD_OPREPLY:
2910                 return get_reply(con, hdr, skip);
2911         default:
2912                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2913                         osd->o_osd);
2914                 *skip = 1;
2915                 return NULL;
2916         }
2917 }
2918
2919 /*
2920  * Wrappers to refcount containing ceph_osd struct
2921  */
2922 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2923 {
2924         struct ceph_osd *osd = con->private;
2925         if (get_osd(osd))
2926                 return con;
2927         return NULL;
2928 }
2929
2930 static void put_osd_con(struct ceph_connection *con)
2931 {
2932         struct ceph_osd *osd = con->private;
2933         put_osd(osd);
2934 }
2935
2936 /*
2937  * authentication
2938  */
2939 /*
2940  * Note: returned pointer is the address of a structure that's
2941  * managed separately.  Caller must *not* attempt to free it.
2942  */
2943 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2944                                         int *proto, int force_new)
2945 {
2946         struct ceph_osd *o = con->private;
2947         struct ceph_osd_client *osdc = o->o_osdc;
2948         struct ceph_auth_client *ac = osdc->client->monc.auth;
2949         struct ceph_auth_handshake *auth = &o->o_auth;
2950
2951         if (force_new && auth->authorizer) {
2952                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2953                 auth->authorizer = NULL;
2954         }
2955         if (!auth->authorizer) {
2956                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2957                                                       auth);
2958                 if (ret)
2959                         return ERR_PTR(ret);
2960         } else {
2961                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2962                                                      auth);
2963                 if (ret)
2964                         return ERR_PTR(ret);
2965         }
2966         *proto = ac->protocol;
2967
2968         return auth;
2969 }
2970
2971
2972 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2973 {
2974         struct ceph_osd *o = con->private;
2975         struct ceph_osd_client *osdc = o->o_osdc;
2976         struct ceph_auth_client *ac = osdc->client->monc.auth;
2977
2978         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2979 }
2980
2981 static int invalidate_authorizer(struct ceph_connection *con)
2982 {
2983         struct ceph_osd *o = con->private;
2984         struct ceph_osd_client *osdc = o->o_osdc;
2985         struct ceph_auth_client *ac = osdc->client->monc.auth;
2986
2987         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2988         return ceph_monc_validate_auth(&osdc->client->monc);
2989 }
2990
2991 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
2992 {
2993         struct ceph_osd *o = con->private;
2994         struct ceph_auth_handshake *auth = &o->o_auth;
2995         return ceph_auth_sign_message(auth, msg);
2996 }
2997
2998 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
2999 {
3000         struct ceph_osd *o = con->private;
3001         struct ceph_auth_handshake *auth = &o->o_auth;
3002         return ceph_auth_check_message_signature(auth, msg);
3003 }
3004
3005 static const struct ceph_connection_operations osd_con_ops = {
3006         .get = get_osd_con,
3007         .put = put_osd_con,
3008         .dispatch = dispatch,
3009         .get_authorizer = get_authorizer,
3010         .verify_authorizer_reply = verify_authorizer_reply,
3011         .invalidate_authorizer = invalidate_authorizer,
3012         .alloc_msg = alloc_msg,
3013         .sign_message = sign_message,
3014         .check_message_signature = check_message_signature,
3015         .fault = osd_reset,
3016 };