lustre/mdt/mdt_batch.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2020, DDN Storage Corporation.
  24  */
  25 /*
  26  * This file is part of Lustre, http://www.lustre.org/
  27  */
  28 /*
  29  * lustre/mdt/mdt_batch.c
  30  *
  31  * Batch Metadata Updating on the server (MDT)
  32  *
  33  * Author: Qian Yingjin <qian@ddn.com>
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_MDS
  37
  38 #include <linux/module.h>
  39
  40 #include <lustre_mds.h>
  41 #include "mdt_internal.h"
  42
  43 static struct ldlm_callback_suite mdt_dlm_cbs = {
  44         .lcs_completion = ldlm_server_completion_ast,
  45         .lcs_blocking   = tgt_blocking_ast,
  46         .lcs_glimpse    = ldlm_server_glimpse_ast
  47 };
  48
  49 static int mdt_batch_unpack(struct mdt_thread_info *info, __u32 opc)
  50 {
  51         int rc = 0;
  52
  53         switch (opc) {
  54         case BUT_GETATTR:
  55                 info->mti_dlm_req = req_capsule_client_get(info->mti_pill,
  56                                                            &RMF_DLM_REQ);
  57                 if (info->mti_dlm_req == NULL)
  58                         RETURN(-EFAULT);
  59                 break;
  60         default:
  61                 rc = -EOPNOTSUPP;
  62                 CERROR("%s: Unexpected opcode %d: rc = %d\n",
  63                        mdt_obd_name(info->mti_mdt), opc, rc);
  64                 break;
  65         }
  66
  67         RETURN(rc);
  68 }
  69
  70 static int mdt_batch_pack_repmsg(struct mdt_thread_info *info)
  71 {
  72         return 0;
  73 }
  74
  75 typedef int (*mdt_batch_reconstructor)(struct tgt_session_info *tsi);
  76
  77 static mdt_batch_reconstructor reconstructors[BUT_LAST_OPC];
  78
  79 static int mdt_batch_reconstruct(struct tgt_session_info *tsi, long opc)
  80 {
  81         mdt_batch_reconstructor reconst;
  82         int rc;
  83
  84         ENTRY;
  85
  86         if (opc >= BUT_LAST_OPC)
  87                 RETURN(-EOPNOTSUPP);
  88
  89         reconst = reconstructors[opc];
  90         LASSERT(reconst != NULL);
  91         rc = reconst(tsi);
  92         RETURN(rc);
  93 }
  94
  95 static int mdt_batch_getattr(struct tgt_session_info *tsi)
  96 {
  97         struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
  98         struct req_capsule *pill = &info->mti_sub_pill;
  99         int rc;
 100
 101         ENTRY;
 102
 103         rc = ldlm_handle_enqueue(info->mti_exp->exp_obd->obd_namespace,
 104                                  pill, info->mti_dlm_req, &mdt_dlm_cbs);
 105
 106         RETURN(rc);
 107 }
 108
 109 /* Batch UpdaTe Request with a format known in advance */
 110 #define TGT_BUT_HDL(flags, opc, fn)                     \
 111 [opc - BUT_FIRST_OPC] = {                               \
 112         .th_name        = #opc,                         \
 113         .th_fail_id     = 0,                            \
 114         .th_opc         = opc,                          \
 115         .th_flags       = flags,                        \
 116         .th_act         = fn,                           \
 117         .th_fmt         = &RQF_ ## opc,                 \
 118         .th_version     = LUSTRE_MDS_VERSION,           \
 119         .th_hp          = NULL,                         \
 120 }
 121
 122 static struct tgt_handler mdt_batch_handlers[] = {
 123 TGT_BUT_HDL(HAS_KEY | HAS_REPLY,        BUT_GETATTR,    mdt_batch_getattr),
 124 };
 125
 126 static struct tgt_handler *mdt_batch_handler_find(__u32 opc)
 127 {
 128         struct tgt_handler *h;
 129
 130         h = NULL;
 131         if (opc >= BUT_FIRST_OPC && opc < BUT_LAST_OPC) {
 132                 h = &mdt_batch_handlers[opc - BUT_FIRST_OPC];
 133                 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 134                          h->th_opc, opc);
 135         } else {
 136                 h = NULL; /* unsupported opc */
 137         }
 138         return h;
 139 }
 140
 141 int mdt_batch(struct tgt_session_info *tsi)
 142 {
 143         struct mdt_thread_info *info = tsi2mdt_info(tsi);
 144         struct req_capsule *pill = &info->mti_sub_pill;
 145         struct ptlrpc_request *req = tgt_ses_req(tsi);
 146         struct but_update_header *buh;
 147         struct but_update_buffer *bub = NULL;
 148         struct batch_update_reply *reply = NULL;
 149         struct ptlrpc_bulk_desc *desc = NULL;
 150         struct tg_reply_data *trd = NULL;
 151         struct lustre_msg *repmsg = NULL;
 152         bool need_reconstruct;
 153         __u32 handled_update_count = 0;
 154         __u32 update_buf_count;
 155         __u32 packed_replen;
 156         void **update_bufs;
 157         int buh_size;
 158         int rc;
 159         int i;
 160
 161         ENTRY;
 162
 163         buh_size = req_capsule_get_size(&req->rq_pill, &RMF_BUT_HEADER,
 164                                         RCL_CLIENT);
 165         if (buh_size <= 0)
 166                 RETURN(err_serious(-EPROTO));
 167
 168         buh = req_capsule_client_get(&req->rq_pill, &RMF_BUT_HEADER);
 169         if (buh == NULL)
 170                 RETURN(err_serious(-EPROTO));
 171
 172         if (buh->buh_magic != BUT_HEADER_MAGIC) {
 173                 CERROR("%s: invalid update header magic %x expect %x: "
 174                        "rc = %d\n", tgt_name(tsi->tsi_tgt), buh->buh_magic,
 175                        BUT_HEADER_MAGIC, -EPROTO);
 176                 RETURN(err_serious(-EPROTO));
 177         }
 178
 179         update_buf_count = buh->buh_count;
 180         if (update_buf_count == 0)
 181                 RETURN(err_serious(-EPROTO));
 182
 183         OBD_ALLOC_PTR_ARRAY(update_bufs, update_buf_count);
 184         if (update_bufs == NULL)
 185                 RETURN(err_serious(-ENOMEM));
 186
 187         if (buh->buh_inline_length > 0) {
 188                 update_bufs[0] = buh->buh_inline_data;
 189         } else {
 190                 struct but_update_buffer *tmp;
 191                 int page_count = 0;
 192
 193                 bub = req_capsule_client_get(&req->rq_pill, &RMF_BUT_BUF);
 194                 if (bub == NULL)
 195                         GOTO(out, rc = err_serious(-EPROTO));
 196
 197                 for (i = 0; i < update_buf_count; i++)
 198                         /* First *and* last might be partial pages, hence +1 */
 199                         page_count += DIV_ROUND_UP(bub[i].bub_size,
 200                                                    PAGE_SIZE) + 1;
 201
 202                 desc = ptlrpc_prep_bulk_exp(req, page_count,
 203                                             PTLRPC_BULK_OPS_COUNT,
 204                                             PTLRPC_BULK_GET_SINK,
 205                                             MDS_BULK_PORTAL,
 206                                             &ptlrpc_bulk_kiov_nopin_ops);
 207                 if (desc == NULL)
 208                         GOTO(out, rc = err_serious(-ENOMEM));
 209
 210                 tmp = bub;
 211                 for (i = 0; i < update_buf_count; i++, tmp++) {
 212                         if (tmp->bub_size >= OUT_MAXREQSIZE)
 213                                 GOTO(out, rc = err_serious(-EPROTO));
 214
 215                         OBD_ALLOC_LARGE(update_bufs[i], tmp->bub_size);
 216                         if (update_bufs[i] == NULL)
 217                                 GOTO(out, rc = err_serious(-ENOMEM));
 218
 219                         desc->bd_frag_ops->add_iov_frag(desc, update_bufs[i],
 220                                                         tmp->bub_size);
 221                 }
 222
 223                 req->rq_bulk_write = 1;
 224                 rc = sptlrpc_svc_prep_bulk(req, desc);
 225                 if (rc != 0)
 226                         GOTO(out, rc = err_serious(rc));
 227
 228                 rc = target_bulk_io(req->rq_export, desc);
 229                 if (rc < 0)
 230                         GOTO(out, rc = err_serious(rc));
 231         }
 232
 233         req_capsule_set_size(&req->rq_pill, &RMF_BUT_REPLY, RCL_SERVER,
 234                              buh->buh_reply_size);
 235         rc = req_capsule_server_pack(&req->rq_pill);
 236         if (rc != 0) {
 237                 DEBUG_REQ(D_ERROR, req, "%s: Can't pack response: rc = %d\n",
 238                        tgt_name(tsi->tsi_tgt), rc);
 239                 GOTO(out, rc);
 240         }
 241
 242         /* Prepare the update reply buffer */
 243         reply = req_capsule_server_get(&req->rq_pill, &RMF_BUT_REPLY);
 244         if (reply == NULL)
 245                 GOTO(out, rc = -EPROTO);
 246
 247         reply->burp_magic = BUT_REPLY_MAGIC;
 248         packed_replen = sizeof(*reply);
 249         info->mti_batch_env = 1;
 250         info->mti_pill = pill;
 251         tsi->tsi_batch_env = true;
 252
 253         OBD_ALLOC_PTR(trd);
 254         if (trd == NULL)
 255                 GOTO(out, rc = -ENOMEM);
 256
 257         need_reconstruct = tgt_check_resent(req, trd);
 258         /* Walk through sub requests in the batch request to execute them. */
 259         for (i = 0; i < update_buf_count; i++) {
 260                 struct batch_update_request *bur;
 261                 struct lustre_msg *reqmsg = NULL;
 262                 struct tgt_handler *h;
 263                 int update_count;
 264                 int j;
 265
 266                 bur = update_bufs[i];
 267                 update_count = bur->burq_count;
 268                 for (j = 0; j < update_count; j++) {
 269                         __u32 replen;
 270
 271                         reqmsg = batch_update_reqmsg_next(bur, reqmsg);
 272                         repmsg = batch_update_repmsg_next(reply, repmsg);
 273
 274                         if (handled_update_count > buh->buh_update_count)
 275                                 GOTO(out, rc = -EOVERFLOW);
 276
 277                         LASSERT(reqmsg != NULL && repmsg != NULL);
 278                         LASSERTF(reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V2,
 279                                  "Invalid reqmsg magic %x expected %x\n",
 280                                  reqmsg->lm_magic, LUSTRE_MSG_MAGIC_V2);
 281
 282                         h = mdt_batch_handler_find(reqmsg->lm_opc);
 283                         if (unlikely(h == NULL)) {
 284                                 CERROR("%s: unsupported opc: 0x%x\n",
 285                                        tgt_name(tsi->tsi_tgt), reqmsg->lm_opc);
 286                                 GOTO(out, rc = -ENOTSUPP);
 287                         }
 288
 289                         LASSERT(h->th_fmt != NULL);
 290                         req_capsule_subreq_init(pill, h->th_fmt, req,
 291                                                 reqmsg, repmsg, RCL_SERVER);
 292
 293                         rc = mdt_batch_unpack(info, reqmsg->lm_opc);
 294                         if (rc) {
 295                                 CERROR("%s: Can't unpack subreq, rc = %d\n",
 296                                        mdt_obd_name(info->mti_mdt), rc);
 297                                 GOTO(out, rc);
 298                         }
 299
 300                         rc = mdt_batch_pack_repmsg(info);
 301                         if (rc)
 302                                 GOTO(out, rc);
 303
 304                         /* Need to reconstruct the reply for committed sub
 305                          * requests in a batched RPC.
 306                          * It only calls reconstruct for modification sub
 307                          * requests.
 308                          * For uncommitted or read-only sub requests, the server
 309                          * should re-execute them via the ->th_act() below.
 310                          */
 311                         if ((h->th_flags & IS_MUTABLE) && need_reconstruct &&
 312                             handled_update_count <=
 313                             trd->trd_reply.lrd_batch_idx) {
 314                                 rc = mdt_batch_reconstruct(tsi, reqmsg->lm_opc);
 315                                 if (rc)
 316                                         GOTO(out, rc);
 317                                 GOTO(next, rc);
 318                         }
 319
 320                         tsi->tsi_batch_idx = handled_update_count;
 321                         rc = h->th_act(tsi);
 322                         if (rc)
 323                                 GOTO(out, rc);
 324 next:
 325                         /*
 326                          * As @repmsg may be changed if the reply buffer is
 327                          * too small to grow, thus it needs to reload it here.
 328                          */
 329                         repmsg = pill->rc_repmsg;
 330                         repmsg->lm_result = rc;
 331                         mdt_thread_info_reset(info);
 332
 333                         replen = lustre_packed_msg_size(repmsg);
 334                         packed_replen += replen;
 335                         handled_update_count++;
 336                 }
 337         }
 338
 339         CDEBUG(D_INFO, "reply size %u packed replen %u\n",
 340                buh->buh_reply_size, packed_replen);
 341         if (buh->buh_reply_size > packed_replen)
 342                 req_capsule_shrink(&req->rq_pill, &RMF_BUT_REPLY,
 343                                    packed_replen, RCL_SERVER);
 344 out:
 345         if (reply != NULL)
 346                 reply->burp_count = handled_update_count;
 347
 348         if (update_bufs != NULL) {
 349                 if (bub != NULL) {
 350                         for (i = 0; i < update_buf_count; i++, bub++) {
 351                                 if (update_bufs[i] != NULL)
 352                                         OBD_FREE_LARGE(update_bufs[i],
 353                                                        bub->bub_size);
 354                         }
 355                 }
 356
 357                 OBD_FREE_PTR_ARRAY(update_bufs, update_buf_count);
 358         }
 359
 360         if (trd)
 361                 OBD_FREE_PTR(trd);
 362
 363         if (desc != NULL)
 364                 ptlrpc_free_bulk(desc);
 365
 366         mdt_thread_info_fini(info);
 367         tsi->tsi_reply_fail_id = OBD_FAIL_BUT_UPDATE_NET_REP;
 368         RETURN(rc);
 369 }
 370