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