bulk-checkin.c

   1 /*
   2  * Copyright (c) 2011, Google Inc.
   3  */
   4
   5 #define USE_THE_REPOSITORY_VARIABLE
   6
   7 #include "git-compat-util.h"
   8 #include "bulk-checkin.h"
   9 #include "environment.h"
  10 #include "gettext.h"
  11 #include "hex.h"
  12 #include "lockfile.h"
  13 #include "repository.h"
  14 #include "csum-file.h"
  15 #include "pack.h"
  16 #include "strbuf.h"
  17 #include "tmp-objdir.h"
  18 #include "packfile.h"
  19 #include "object-file.h"
  20 #include "object-store-ll.h"
  21
  22 static int odb_transaction_nesting;
  23
  24 static struct tmp_objdir *bulk_fsync_objdir;
  25
  26 static struct bulk_checkin_packfile {
  27         char *pack_tmp_name;
  28         struct hashfile *f;
  29         off_t offset;
  30         struct pack_idx_option pack_idx_opts;
  31
  32         struct pack_idx_entry **written;
  33         uint32_t alloc_written;
  34         uint32_t nr_written;
  35 } bulk_checkin_packfile;
  36
  37 static void finish_tmp_packfile(struct strbuf *basename,
  38                                 const char *pack_tmp_name,
  39                                 struct pack_idx_entry **written_list,
  40                                 uint32_t nr_written,
  41                                 struct pack_idx_option *pack_idx_opts,
  42                                 unsigned char hash[])
  43 {
  44         char *idx_tmp_name = NULL;
  45
  46         stage_tmp_packfiles(basename, pack_tmp_name, written_list, nr_written,
  47                             NULL, pack_idx_opts, hash, &idx_tmp_name);
  48         rename_tmp_packfile_idx(basename, &idx_tmp_name);
  49
  50         free(idx_tmp_name);
  51 }
  52
  53 static void flush_bulk_checkin_packfile(struct bulk_checkin_packfile *state)
  54 {
  55         unsigned char hash[GIT_MAX_RAWSZ];
  56         struct strbuf packname = STRBUF_INIT;
  57         int i;
  58
  59         if (!state->f)
  60                 return;
  61
  62         if (state->nr_written == 0) {
  63                 close(state->f->fd);
  64                 free_hashfile(state->f);
  65                 unlink(state->pack_tmp_name);
  66                 goto clear_exit;
  67         } else if (state->nr_written == 1) {
  68                 finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK,
  69                                   CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
  70         } else {
  71                 int fd = finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK, 0);
  72                 fixup_pack_header_footer(fd, hash, state->pack_tmp_name,
  73                                          state->nr_written, hash,
  74                                          state->offset);
  75                 close(fd);
  76         }
  77
  78         strbuf_addf(&packname, "%s/pack/pack-%s.", repo_get_object_directory(the_repository),
  79                     hash_to_hex(hash));
  80         finish_tmp_packfile(&packname, state->pack_tmp_name,
  81                             state->written, state->nr_written,
  82                             &state->pack_idx_opts, hash);
  83         for (i = 0; i < state->nr_written; i++)
  84                 free(state->written[i]);
  85
  86 clear_exit:
  87         free(state->pack_tmp_name);
  88         free(state->written);
  89         memset(state, 0, sizeof(*state));
  90
  91         strbuf_release(&packname);
  92         /* Make objects we just wrote available to ourselves */
  93         reprepare_packed_git(the_repository);
  94 }
  95
  96 /*
  97  * Cleanup after batch-mode fsync_object_files.
  98  */
  99 static void flush_batch_fsync(void)
 100 {
 101         struct strbuf temp_path = STRBUF_INIT;
 102         struct tempfile *temp;
 103
 104         if (!bulk_fsync_objdir)
 105                 return;
 106
 107         /*
 108          * Issue a full hardware flush against a temporary file to ensure
 109          * that all objects are durable before any renames occur. The code in
 110          * fsync_loose_object_bulk_checkin has already issued a writeout
 111          * request, but it has not flushed any writeback cache in the storage
 112          * hardware or any filesystem logs. This fsync call acts as a barrier
 113          * to ensure that the data in each new object file is durable before
 114          * the final name is visible.
 115          */
 116         strbuf_addf(&temp_path, "%s/bulk_fsync_XXXXXX", repo_get_object_directory(the_repository));
 117         temp = xmks_tempfile(temp_path.buf);
 118         fsync_or_die(get_tempfile_fd(temp), get_tempfile_path(temp));
 119         delete_tempfile(&temp);
 120         strbuf_release(&temp_path);
 121
 122         /*
 123          * Make the object files visible in the primary ODB after their data is
 124          * fully durable.
 125          */
 126         tmp_objdir_migrate(bulk_fsync_objdir);
 127         bulk_fsync_objdir = NULL;
 128 }
 129
 130 static int already_written(struct bulk_checkin_packfile *state, struct object_id *oid)
 131 {
 132         int i;
 133
 134         /* The object may already exist in the repository */
 135         if (repo_has_object_file(the_repository, oid))
 136                 return 1;
 137
 138         /* Might want to keep the list sorted */
 139         for (i = 0; i < state->nr_written; i++)
 140                 if (oideq(&state->written[i]->oid, oid))
 141                         return 1;
 142
 143         /* This is a new object we need to keep */
 144         return 0;
 145 }
 146
 147 /*
 148  * Read the contents from fd for size bytes, streaming it to the
 149  * packfile in state while updating the hash in ctx. Signal a failure
 150  * by returning a negative value when the resulting pack would exceed
 151  * the pack size limit and this is not the first object in the pack,
 152  * so that the caller can discard what we wrote from the current pack
 153  * by truncating it and opening a new one. The caller will then call
 154  * us again after rewinding the input fd.
 155  *
 156  * The already_hashed_to pointer is kept untouched by the caller to
 157  * make sure we do not hash the same byte when we are called
 158  * again. This way, the caller does not have to checkpoint its hash
 159  * status before calling us just in case we ask it to call us again
 160  * with a new pack.
 161  */
 162 static int stream_blob_to_pack(struct bulk_checkin_packfile *state,
 163                                git_hash_ctx *ctx, off_t *already_hashed_to,
 164                                int fd, size_t size, const char *path,
 165                                unsigned flags)
 166 {
 167         git_zstream s;
 168         unsigned char ibuf[16384];
 169         unsigned char obuf[16384];
 170         unsigned hdrlen;
 171         int status = Z_OK;
 172         int write_object = (flags & HASH_WRITE_OBJECT);
 173         off_t offset = 0;
 174
 175         git_deflate_init(&s, pack_compression_level);
 176
 177         hdrlen = encode_in_pack_object_header(obuf, sizeof(obuf), OBJ_BLOB, size);
 178         s.next_out = obuf + hdrlen;
 179         s.avail_out = sizeof(obuf) - hdrlen;
 180
 181         while (status != Z_STREAM_END) {
 182                 if (size && !s.avail_in) {
 183                         ssize_t rsize = size < sizeof(ibuf) ? size : sizeof(ibuf);
 184                         ssize_t read_result = read_in_full(fd, ibuf, rsize);
 185                         if (read_result < 0)
 186                                 die_errno("failed to read from '%s'", path);
 187                         if (read_result != rsize)
 188                                 die("failed to read %d bytes from '%s'",
 189                                     (int)rsize, path);
 190                         offset += rsize;
 191                         if (*already_hashed_to < offset) {
 192                                 size_t hsize = offset - *already_hashed_to;
 193                                 if (rsize < hsize)
 194                                         hsize = rsize;
 195                                 if (hsize)
 196                                         the_hash_algo->update_fn(ctx, ibuf, hsize);
 197                                 *already_hashed_to = offset;
 198                         }
 199                         s.next_in = ibuf;
 200                         s.avail_in = rsize;
 201                         size -= rsize;
 202                 }
 203
 204                 status = git_deflate(&s, size ? 0 : Z_FINISH);
 205
 206                 if (!s.avail_out || status == Z_STREAM_END) {
 207                         if (write_object) {
 208                                 size_t written = s.next_out - obuf;
 209
 210                                 /* would we bust the size limit? */
 211                                 if (state->nr_written &&
 212                                     pack_size_limit_cfg &&
 213                                     pack_size_limit_cfg < state->offset + written) {
 214                                         git_deflate_abort(&s);
 215                                         return -1;
 216                                 }
 217
 218                                 hashwrite(state->f, obuf, written);
 219                                 state->offset += written;
 220                         }
 221                         s.next_out = obuf;
 222                         s.avail_out = sizeof(obuf);
 223                 }
 224
 225                 switch (status) {
 226                 case Z_OK:
 227                 case Z_BUF_ERROR:
 228                 case Z_STREAM_END:
 229                         continue;
 230                 default:
 231                         die("unexpected deflate failure: %d", status);
 232                 }
 233         }
 234         git_deflate_end(&s);
 235         return 0;
 236 }
 237
 238 /* Lazily create backing packfile for the state */
 239 static void prepare_to_stream(struct bulk_checkin_packfile *state,
 240                               unsigned flags)
 241 {
 242         if (!(flags & HASH_WRITE_OBJECT) || state->f)
 243                 return;
 244
 245         state->f = create_tmp_packfile(&state->pack_tmp_name);
 246         reset_pack_idx_option(&state->pack_idx_opts);
 247
 248         /* Pretend we are going to write only one object */
 249         state->offset = write_pack_header(state->f, 1);
 250         if (!state->offset)
 251                 die_errno("unable to write pack header");
 252 }
 253
 254 static int deflate_blob_to_pack(struct bulk_checkin_packfile *state,
 255                                 struct object_id *result_oid,
 256                                 int fd, size_t size,
 257                                 const char *path, unsigned flags)
 258 {
 259         off_t seekback, already_hashed_to;
 260         git_hash_ctx ctx;
 261         unsigned char obuf[16384];
 262         unsigned header_len;
 263         struct hashfile_checkpoint checkpoint = {0};
 264         struct pack_idx_entry *idx = NULL;
 265
 266         seekback = lseek(fd, 0, SEEK_CUR);
 267         if (seekback == (off_t) -1)
 268                 return error("cannot find the current offset");
 269
 270         header_len = format_object_header((char *)obuf, sizeof(obuf),
 271                                           OBJ_BLOB, size);
 272         the_hash_algo->init_fn(&ctx);
 273         the_hash_algo->update_fn(&ctx, obuf, header_len);
 274         the_hash_algo->init_fn(&checkpoint.ctx);
 275
 276         /* Note: idx is non-NULL when we are writing */
 277         if ((flags & HASH_WRITE_OBJECT) != 0)
 278                 CALLOC_ARRAY(idx, 1);
 279
 280         already_hashed_to = 0;
 281
 282         while (1) {
 283                 prepare_to_stream(state, flags);
 284                 if (idx) {
 285                         hashfile_checkpoint(state->f, &checkpoint);
 286                         idx->offset = state->offset;
 287                         crc32_begin(state->f);
 288                 }
 289                 if (!stream_blob_to_pack(state, &ctx, &already_hashed_to,
 290                                          fd, size, path, flags))
 291                         break;
 292                 /*
 293                  * Writing this object to the current pack will make
 294                  * it too big; we need to truncate it, start a new
 295                  * pack, and write into it.
 296                  */
 297                 if (!idx)
 298                         BUG("should not happen");
 299                 hashfile_truncate(state->f, &checkpoint);
 300                 state->offset = checkpoint.offset;
 301                 flush_bulk_checkin_packfile(state);
 302                 if (lseek(fd, seekback, SEEK_SET) == (off_t) -1)
 303                         return error("cannot seek back");
 304         }
 305         the_hash_algo->final_oid_fn(result_oid, &ctx);
 306         if (!idx)
 307                 return 0;
 308
 309         idx->crc32 = crc32_end(state->f);
 310         if (already_written(state, result_oid)) {
 311                 hashfile_truncate(state->f, &checkpoint);
 312                 state->offset = checkpoint.offset;
 313                 free(idx);
 314         } else {
 315                 oidcpy(&idx->oid, result_oid);
 316                 ALLOC_GROW(state->written,
 317                            state->nr_written + 1,
 318                            state->alloc_written);
 319                 state->written[state->nr_written++] = idx;
 320         }
 321         return 0;
 322 }
 323
 324 void prepare_loose_object_bulk_checkin(void)
 325 {
 326         /*
 327          * We lazily create the temporary object directory
 328          * the first time an object might be added, since
 329          * callers may not know whether any objects will be
 330          * added at the time they call begin_odb_transaction.
 331          */
 332         if (!odb_transaction_nesting || bulk_fsync_objdir)
 333                 return;
 334
 335         bulk_fsync_objdir = tmp_objdir_create("bulk-fsync");
 336         if (bulk_fsync_objdir)
 337                 tmp_objdir_replace_primary_odb(bulk_fsync_objdir, 0);
 338 }
 339
 340 void fsync_loose_object_bulk_checkin(int fd, const char *filename)
 341 {
 342         /*
 343          * If we have an active ODB transaction, we issue a call that
 344          * cleans the filesystem page cache but avoids a hardware flush
 345          * command. Later on we will issue a single hardware flush
 346          * before renaming the objects to their final names as part of
 347          * flush_batch_fsync.
 348          */
 349         if (!bulk_fsync_objdir ||
 350             git_fsync(fd, FSYNC_WRITEOUT_ONLY) < 0) {
 351                 if (errno == ENOSYS)
 352                         warning(_("core.fsyncMethod = batch is unsupported on this platform"));
 353                 fsync_or_die(fd, filename);
 354         }
 355 }
 356
 357 int index_blob_bulk_checkin(struct object_id *oid,
 358                             int fd, size_t size,
 359                             const char *path, unsigned flags)
 360 {
 361         int status = deflate_blob_to_pack(&bulk_checkin_packfile, oid, fd, size,
 362                                           path, flags);
 363         if (!odb_transaction_nesting)
 364                 flush_bulk_checkin_packfile(&bulk_checkin_packfile);
 365         return status;
 366 }
 367
 368 void begin_odb_transaction(void)
 369 {
 370         odb_transaction_nesting += 1;
 371 }
 372
 373 void flush_odb_transaction(void)
 374 {
 375         flush_batch_fsync();
 376         flush_bulk_checkin_packfile(&bulk_checkin_packfile);
 377 }
 378
 379 void end_odb_transaction(void)
 380 {
 381         odb_transaction_nesting -= 1;
 382         if (odb_transaction_nesting < 0)
 383                 BUG("Unbalanced ODB transaction nesting");
 384
 385         if (odb_transaction_nesting)
 386                 return;
 387
 388         flush_odb_transaction();
 389 }