int scr_swap_file_names(
const char* file_send, int rank_send,
char* file_recv, size_t size_recv, int rank_recv,
const char* dir_recv, MPI_Comm comm)
{
int rc = SCR_SUCCESS;
/* determine whether we have a file to send */
int have_outgoing = 0;
if (rank_send != MPI_PROC_NULL &&
file_send != NULL &&
strcmp(file_send, "") != 0)
{
have_outgoing = 1;
} else {
/* nothing to send, make sure to use PROC_NULL in sendrecv call */
rank_send = MPI_PROC_NULL;
}
/* determine whether we are expecting to receive a file */
int have_incoming = 0;
if (rank_recv != MPI_PROC_NULL &&
dir_recv != NULL &&
strcmp(dir_recv, "") != 0)
{
have_incoming = 1;
} else {
/* nothing to recv, make sure to use PROC_NULL in sendrecv call */
rank_recv = MPI_PROC_NULL;
}
/* exchange file names with partners, note that we initialize
* file_recv_orig to NULL in case we recv from MPI_PROC_NULL */
char* file_recv_orig = NULL;
scr_str_sendrecv(file_send, rank_send, &file_recv_orig, rank_recv, comm);
/* define the path to store our partner's file */
if (have_incoming) {
/* set path to file name */
scr_path* path_recv = scr_path_from_str(file_recv_orig);
scr_path_basename(path_recv);
scr_path_prepend_str(path_recv, dir_recv);
scr_path_strcpy(file_recv, size_recv, path_recv);
scr_path_delete(&path_recv);
/* free the file name we received */
scr_free(&file_recv_orig);
}
return rc;
}
/* for file name listed in meta, fetch that file from src_dir and store
* a copy in dst_dir, record full path to copy in newfile, and
* return whether operation succeeded */
static int scr_fetch_file(
const char* dst_file,
const char* src_dir,
const scr_meta* meta)
{
int rc = SCR_SUCCESS;
/* build full path to source file */
scr_path* path_src_file = scr_path_from_str(dst_file);
scr_path_basename(path_src_file);
scr_path_prepend_str(path_src_file, src_dir);
char* src_file = scr_path_strdup(path_src_file);
/* fetch the file */
uLong crc;
uLong* crc_p = NULL;
if (scr_crc_on_flush) {
crc_p = &crc;
}
rc = scr_file_copy(src_file, dst_file, scr_file_buf_size, crc_p);
/* check that crc matches crc stored in meta */
uLong meta_crc;
if (scr_meta_get_crc32(meta, &meta_crc) == SCR_SUCCESS) {
if (rc == SCR_SUCCESS && scr_crc_on_flush && crc != meta_crc) {
rc = SCR_FAILURE;
scr_err("CRC32 mismatch detected when fetching file from %s to %s @ %s:%d",
src_file, dst_file, __FILE__, __LINE__
);
/* TODO: would be good to log this, but right now only rank 0
* can write log entries */
/*
if (scr_log_enable) {
time_t now = scr_log_seconds();
scr_log_event("CRC32 MISMATCH", filename, NULL, &now, NULL);
}
*/
}
}
/* free path and string for source file */
scr_free(&src_file);
scr_path_delete(&path_src_file);
return rc;
}
/* allocates a new string (to be freed with scr_free)
* that is path to user config file */
static char* user_config_path()
{
char* file = NULL;
/* first, use SCR_CONF_FILE if it's set */
char* value = getenv("SCR_CONF_FILE");
if (value != NULL) {
file = strdup(value);
return file;
}
/* otherwise, look in the prefix directory */
char* prefix = NULL;
value = getenv("SCR_PREFIX");
if (value != NULL) {
/* user set SCR_PREFIX, strdup that value */
prefix = strdup(value);
} else {
/* if user didn't set with SCR_PREFIX,
* pick up the current working directory as a default */
char current_dir[SCR_MAX_FILENAME];
if (scr_getcwd(current_dir, sizeof(current_dir)) != SCR_SUCCESS) {
scr_abort(-1, "Problem reading current working directory @ %s:%d",
__FILE__, __LINE__
);
}
prefix = strdup(current_dir);
}
/* couldn't find a prefix directory, so bail */
if (prefix == NULL) {
return file;
}
/* tack file name on to directory */
scr_path* prefix_path = scr_path_from_str(prefix);
scr_path_append_str(prefix_path, SCR_CONFIG_FILE_USER);
file = scr_path_strdup(prefix_path);
scr_path_delete(&prefix_path);
/* free the prefix dir which we strdup'd */
scr_free(&prefix);
return file;
}
/* this moves all files of the specified dataset in the cache to
* make them accessible to new rank mapping */
static int scr_distribute_files(scr_filemap* map, const scr_reddesc* red, int id)
{
int i, round;
int rc = SCR_SUCCESS;
/* TODO: mark dataset as being distributed in filemap,
* because if we fail in the middle of a distribute,
* we can't trust the contents of the files anymore,
* at which point it should be deleted */
/* clean out any incomplete files before we start */
scr_cache_clean(map);
/* for this dataset, get list of ranks we have data for */
int nranks = 0;
int* ranks = NULL;
scr_filemap_list_ranks_by_dataset(map, id, &nranks, &ranks);
/* walk backwards through the list of ranks, and set our start index
* to the rank which is the first rank that is equal to or higher
* than our own rank -- when we assign round ids below, this offsetting
* helps distribute the load */
int start_index = 0;
int invalid_rank_found = 0;
for (i = nranks-1; i >= 0; i--) {
int rank = ranks[i];
/* pick the first rank whose rank id is equal to or higher than our own */
if (rank >= scr_my_rank_world) {
start_index = i;
}
/* while we're at it, check that the rank is within range */
if (rank < 0 || rank >= scr_ranks_world) {
scr_err("Invalid rank id %d in world of %d @ %s:%d",
rank, scr_ranks_world, __FILE__, __LINE__
);
invalid_rank_found = 1;
}
}
/* check that we didn't find an invalid rank on any process */
if (! scr_alltrue(invalid_rank_found == 0)) {
scr_free(&ranks);
return SCR_FAILURE;
}
/* allocate array to record the rank we can send to in each round */
int* have_rank_by_round = (int*) SCR_MALLOC(sizeof(int) * nranks);
int* send_flag_by_round = (int*) SCR_MALLOC(sizeof(int) * nranks);
/* check that we have all of the files for each rank,
* and determine the round we can send them */
scr_hash* send_hash = scr_hash_new();
scr_hash* recv_hash = scr_hash_new();
for (round = 0; round < nranks; round++) {
/* get the rank id */
int index = (start_index + round) % nranks;
int rank = ranks[index];
/* record the rank indexed by the round number */
have_rank_by_round[round] = rank;
/* assume we won't be sending to this rank in this round */
send_flag_by_round[round] = 0;
/* if we have files for this rank, specify the round we can
* send those files in */
if (scr_bool_have_files(map, id, rank)) {
scr_hash_setf(send_hash, NULL, "%d %d", rank, round);
}
}
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* search for the minimum round we can get our files */
int retrieve_rank = -1;
int retrieve_round = -1;
scr_hash_elem* elem = NULL;
for (elem = scr_hash_elem_first(recv_hash);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the rank id */
int rank = scr_hash_elem_key_int(elem);
/* get the round id */
scr_hash* round_hash = scr_hash_elem_hash(elem);
scr_hash_elem* round_elem = scr_hash_elem_first(round_hash);
char* round_str = scr_hash_elem_key(round_elem);
int round = atoi(round_str);
/* record this round and rank number if it's less than the current round */
if (round < retrieve_round || retrieve_round == -1) {
retrieve_round = round;
retrieve_rank = rank;
}
}
/* done with the round hashes, free them off */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
/* free off our list of ranks */
scr_free(&ranks);
/* for some redundancy schemes, we know at this point whether we
* can recover all files */
int can_get_files = (retrieve_rank != -1);
if (red->copy_type != SCR_COPY_XOR && !scr_alltrue(can_get_files)) {
/* print a debug message indicating which rank is missing files */
if (! can_get_files) {
scr_dbg(2, "Cannot find process that has my checkpoint files @ %s:%d",
__FILE__, __LINE__
);
}
return SCR_FAILURE;
}
/* get the maximum retrieve round */
int max_rounds = 0;
MPI_Allreduce(
&retrieve_round, &max_rounds, 1, MPI_INT, MPI_MAX, scr_comm_world
);
/* tell destination which round we'll take our files in */
send_hash = scr_hash_new();
recv_hash = scr_hash_new();
if (retrieve_rank != -1) {
scr_hash_setf(send_hash, NULL, "%d %d", retrieve_rank, retrieve_round);
}
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* determine which ranks want to fetch their files from us */
for(elem = scr_hash_elem_first(recv_hash);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the round id */
scr_hash* round_hash = scr_hash_elem_hash(elem);
scr_hash_elem* round_elem = scr_hash_elem_first(round_hash);
char* round_str = scr_hash_elem_key(round_elem);
int round = atoi(round_str);
/* record whether this rank wants its files from us */
if (round >= 0 && round < nranks) {
send_flag_by_round[round] = 1;
}
}
/* done with the round hashes, free them off */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
int tmp_rc = 0;
/* run through rounds and exchange files */
for (round = 0; round <= max_rounds; round++) {
/* assume we don't need to send or receive any files this round */
int send_rank = MPI_PROC_NULL;
int recv_rank = MPI_PROC_NULL;
int send_num = 0;
int recv_num = 0;
/* check whether I can potentially send to anyone in this round */
if (round < nranks) {
/* have someone's files, check whether they are asking
* for them this round */
if (send_flag_by_round[round]) {
/* need to send files this round, remember to whom and how many */
int dst_rank = have_rank_by_round[round];
send_rank = dst_rank;
send_num = scr_filemap_num_files(map, id, dst_rank);
}
}
/* if I'm supposed to get my files this round, set the recv_rank */
if (retrieve_round == round) {
recv_rank = retrieve_rank;
}
/* TODO: another special case is to just move files if the
* processes are on the same node */
/* if i'm sending to myself, just move (rename) each file */
if (send_rank == scr_my_rank_world) {
/* get our file list */
int numfiles = 0;
char** files = NULL;
scr_filemap_list_files(map, id, send_rank, &numfiles, &files);
/* TODO: sort files in reverse order by size */
/* iterate over and rename each file */
for (i=0; i < numfiles; i++) {
/* get the current file name */
char* file = files[i];
/* lookup meta data for this file */
scr_meta* meta = scr_meta_new();
scr_filemap_get_meta(map, id, send_rank, file, meta);
/* get the path for this file based on its type
* and dataset id */
char* dir = NULL;
if (scr_meta_check_filetype(meta, SCR_META_FILE_USER) == SCR_SUCCESS) {
dir = scr_cache_dir_get(red, id);
} else {
dir = scr_cache_dir_hidden_get(red, id);
}
/* build the new file name */
scr_path* path_newfile = scr_path_from_str(file);
scr_path_basename(path_newfile);
scr_path_prepend_str(path_newfile, dir);
char* newfile = scr_path_strdup(path_newfile);
/* if the new file name is different from the old name, rename it */
if (strcmp(file, newfile) != 0) {
/* record the new filename to our map and write it to disk */
scr_filemap_add_file(map, id, send_rank, newfile);
scr_filemap_set_meta(map, id, send_rank, newfile, meta);
scr_filemap_write(scr_map_file, map);
/* rename the file */
scr_dbg(2, "Round %d: rename(%s, %s)", round, file, newfile);
tmp_rc = rename(file, newfile);
if (tmp_rc != 0) {
/* TODO: to cross mount points, if tmp_rc == EXDEV,
* open new file, copy, and delete orig */
scr_err("Moving checkpoint file: rename(%s, %s) %s errno=%d @ %s:%d",
file, newfile, strerror(errno), errno, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* remove the old name from the filemap and write it to disk */
scr_filemap_remove_file(map, id, send_rank, file);
scr_filemap_write(scr_map_file, map);
}
/* free the path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
/* free directory string */
scr_free(&dir);
/* free meta data */
scr_meta_delete(&meta);
}
/* free the list of filename pointers */
scr_free(&files);
} else {
/* if we have files for this round, but the correspdonding
* rank doesn't need them, delete the files */
if (round < nranks && send_rank == MPI_PROC_NULL) {
int dst_rank = have_rank_by_round[round];
scr_unlink_rank(map, id, dst_rank);
}
/* sending to and/or recieving from another node */
if (send_rank != MPI_PROC_NULL || recv_rank != MPI_PROC_NULL) {
/* have someone to send to or receive from */
int have_outgoing = 0;
int have_incoming = 0;
if (send_rank != MPI_PROC_NULL) {
have_outgoing = 1;
}
if (recv_rank != MPI_PROC_NULL) {
have_incoming = 1;
}
/* first, determine how many files I will be receiving and
* tell how many I will be sending */
MPI_Request request[2];
MPI_Status status[2];
int num_req = 0;
if (have_incoming) {
MPI_Irecv(
&recv_num, 1, MPI_INT, recv_rank, 0,
scr_comm_world, &request[num_req]
);
num_req++;
}
if (have_outgoing) {
MPI_Isend(
&send_num, 1, MPI_INT, send_rank, 0,
scr_comm_world, &request[num_req]
);
num_req++;
}
if (num_req > 0) {
MPI_Waitall(num_req, request, status);
}
/* record how many files I will receive (need to distinguish
* between 0 files and not knowing) */
if (have_incoming) {
scr_filemap_set_expected_files(map, id, scr_my_rank_world, recv_num);
}
/* turn off send or receive flags if the file count is 0,
* nothing else to do */
if (send_num == 0) {
have_outgoing = 0;
send_rank = MPI_PROC_NULL;
}
if (recv_num == 0) {
have_incoming = 0;
recv_rank = MPI_PROC_NULL;
}
/* TODO: since we overwrite files in place in order to avoid
* running out of storage space, we should sort files in order
* of descending size for the next step */
/* get our file list for the destination */
int numfiles = 0;
char** files = NULL;
if (have_outgoing) {
scr_filemap_list_files(map, id, send_rank, &numfiles, &files);
}
/* while we have a file to send or receive ... */
while (have_incoming || have_outgoing) {
/* get the filename */
char* file = NULL;
scr_meta* send_meta = NULL;
if (have_outgoing) {
file = files[numfiles - send_num];
send_meta = scr_meta_new();
scr_filemap_get_meta(map, id, send_rank, file, send_meta);
}
/* exchange meta data so we can determine type of incoming file */
scr_meta* recv_meta = scr_meta_new();
scr_hash_sendrecv(send_meta, send_rank, recv_meta, recv_rank, scr_comm_world);
/* get the path for this file based on its type and dataset id */
char* dir = NULL;
if (have_incoming) {
if (scr_meta_check_filetype(recv_meta, SCR_META_FILE_USER) == SCR_SUCCESS) {
dir = scr_cache_dir_get(red, id);
} else {
dir = scr_cache_dir_hidden_get(red, id);
}
}
/* exhange file names with partners,
* building full path of incoming file */
char file_partner[SCR_MAX_FILENAME];
scr_swap_file_names(
file, send_rank, file_partner, sizeof(file_partner), recv_rank,
dir, scr_comm_world
);
/* free directory string */
scr_free(&dir);
/* free incoming meta data (we'll get this again later) */
scr_meta_delete(&recv_meta);
/* if we'll receive a file, record the name of our file
* in the filemap and write it to disk */
recv_meta = NULL;
if (recv_rank != MPI_PROC_NULL) {
recv_meta = scr_meta_new();
scr_filemap_add_file(map, id, scr_my_rank_world, file_partner);
scr_filemap_write(scr_map_file, map);
}
/* either sending or receiving a file this round, since we move files,
* it will be deleted or overwritten */
if (scr_swap_files(MOVE_FILES, file, send_meta, send_rank,
file_partner, recv_meta, recv_rank, scr_comm_world) != SCR_SUCCESS)
{
scr_err("Swapping files: %s to %d, %s from %d @ %s:%d",
file, send_rank, file_partner, recv_rank, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* if we received a file, record its meta data and decrement
* our receive count */
if (have_incoming) {
/* record meta data for the file we received */
scr_filemap_set_meta(map, id, scr_my_rank_world, file_partner, recv_meta);
scr_meta_delete(&recv_meta);
/* decrement receive count */
recv_num--;
if (recv_num == 0) {
have_incoming = 0;
recv_rank = MPI_PROC_NULL;
}
}
/* if we sent a file, remove it from the filemap and decrement
* our send count */
if (have_outgoing) {
/* remove file from the filemap */
scr_filemap_remove_file(map, id, send_rank, file);
scr_meta_delete(&send_meta);
/* decrement our send count */
send_num--;
if (send_num == 0) {
have_outgoing = 0;
send_rank = MPI_PROC_NULL;
}
}
/* update filemap on disk */
scr_filemap_write(scr_map_file, map);
}
/* free our file list */
scr_free(&files);
}
}
}
/* if we have more rounds than max rounds, delete the remainder of our files */
for (round = max_rounds+1; round < nranks; round++) {
/* have someone's files for this round, so delete them */
int dst_rank = have_rank_by_round[round];
scr_unlink_rank(map, id, dst_rank);
}
scr_free(&send_flag_by_round);
scr_free(&have_rank_by_round);
/* write out new filemap and free the memory resources */
scr_filemap_write(scr_map_file, map);
/* clean out any incomplete files */
scr_cache_clean(map);
/* TODO: if the exchange or redundancy rebuild failed,
* we should also delete any *good* files we received */
/* return whether distribute succeeded, it does not ensure we have
* all of our files, only that the transfer completed without failure */
return rc;
}
/* since on a restart we may end up with more or fewer ranks on a node than the
* previous run, rely on the master to read in and distribute the filemap to
* other ranks on the node */
int scr_scatter_filemaps(scr_filemap* my_map)
{
/* TODO: if the control directory is on a device shared by lots of procs,
* we should read and distribute this data in a more scalable way */
/* allocate empty send hash */
scr_hash* send_hash = scr_hash_new();
/* if i'm the master on this node, read in all filemaps */
if (scr_storedesc_cntl->rank == 0) {
/* create an empty filemap */
scr_filemap* all_map = scr_filemap_new();
/* read in the master map */
scr_hash* hash = scr_hash_new();
scr_hash_read_path(scr_master_map_file, hash);
/* for each filemap listed in the master map */
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(scr_hash_get(hash, "Filemap"));
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the filename of this filemap */
char* file = scr_hash_elem_key(elem);
/* TODO MEMFS: mount storage for each filemap */
/* read in the filemap */
scr_filemap* tmp_map = scr_filemap_new();
scr_path* path_file = scr_path_from_str(file);
scr_filemap_read(path_file, tmp_map);
scr_path_delete(&path_file);
/* merge it with the all_map */
scr_filemap_merge(all_map, tmp_map);
/* delete filemap */
scr_filemap_delete(&tmp_map);
/* TODO: note that if we fail after unlinking this file but before
* writing out the new file, we'll lose information */
/* delete the file */
scr_file_unlink(file);
}
/* free the hash object */
scr_hash_delete(&hash);
/* write out new local 0 filemap */
if (scr_filemap_num_ranks(all_map) > 0) {
scr_filemap_write(scr_map_file, all_map);
}
/* get global rank of each rank */
int* ranks = (int*) SCR_MALLOC(scr_storedesc_cntl->ranks * sizeof(int));
MPI_Gather(
&scr_my_rank_world, 1, MPI_INT, ranks, 1, MPI_INT,
0, scr_storedesc_cntl->comm
);
/* for each rank, send them their own file data if we have it */
int i;
for (i=0; i < scr_storedesc_cntl->ranks; i++) {
int rank = ranks[i];
if (scr_filemap_have_rank(all_map, rank)) {
/* extract the filemap for this rank */
scr_filemap* tmp_map = scr_filemap_extract_rank(all_map, rank);
/* get a reference to the hash object that we'll send to this rank,
* and merge this filemap into it */
scr_hash* tmp_hash = scr_hash_getf(send_hash, "%d", i);
if (tmp_hash == NULL) {
/* if we don't find an existing entry in the send_hash,
* create an empty hash and insert it */
scr_hash* empty_hash = scr_hash_new();
scr_hash_setf(send_hash, empty_hash, "%d", i);
tmp_hash = empty_hash;
}
scr_hash_merge(tmp_hash, tmp_map);
/* delete the filemap for this rank */
scr_filemap_delete(&tmp_map);
}
}
/* free our rank list */
scr_free(&ranks);
/* now just round robin the remainder across the set (load balancing) */
int num;
int* remaining_ranks = NULL;
scr_filemap_list_ranks(all_map, &num, &remaining_ranks);
int j = 0;
while (j < num) {
/* pick a rank in to send to */
i = j % scr_storedesc_cntl->ranks;
/* extract the filemap for this rank */
scr_filemap* tmp_map = scr_filemap_extract_rank(all_map, remaining_ranks[j]);
/* get a reference to the hash object that we'll send to this rank,
* and merge this filemap into it */
scr_hash* tmp_hash = scr_hash_getf(send_hash, "%d", i);
if (tmp_hash == NULL) {
/* if we don't find an existing entry in the send_hash,
* create an empty hash and insert it */
scr_hash* empty_hash = scr_hash_new();
scr_hash_setf(send_hash, empty_hash, "%d", i);
tmp_hash = empty_hash;
}
scr_hash_merge(tmp_hash, tmp_map);
/* delete the filemap for this rank */
scr_filemap_delete(&tmp_map);
j++;
}
scr_free(&remaining_ranks);
/* delete the filemap */
scr_filemap_delete(&all_map);
/* write out the new master filemap */
hash = scr_hash_new();
char file[SCR_MAX_FILENAME];
for (i=0; i < scr_storedesc_cntl->ranks; i++) {
sprintf(file, "%s/filemap_%d.scrinfo", scr_cntl_prefix, i);
scr_hash_set_kv(hash, "Filemap", file);
}
scr_hash_write_path(scr_master_map_file, hash);
scr_hash_delete(&hash);
} else {
/* send our global rank to the master */
MPI_Gather(
&scr_my_rank_world, 1, MPI_INT, NULL, 1, MPI_INT,
0, scr_storedesc_cntl->comm
);
}
/* receive our filemap from master */
scr_hash* recv_hash = scr_hash_new();
scr_hash_exchange(send_hash, recv_hash, scr_storedesc_cntl->comm);
/* merge map sent from master into our map */
scr_hash* map_from_master = scr_hash_getf(recv_hash, "%d", 0);
if (map_from_master != NULL) {
scr_hash_merge(my_map, map_from_master);
}
/* write out our local filemap */
if (scr_filemap_num_ranks(my_map) > 0) {
scr_filemap_write(scr_map_file, my_map);
}
/* free off our send and receive hashes */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
return SCR_SUCCESS;
}
/* given a filename, its meta data, its list of segments, and list of destination containers,
* copy file to container files */
static int scr_flush_file_to_containers(
const char* file,
scr_meta* meta,
scr_hash* segments,
const char* dst_dir)
{
/* check that we got something for a source file */
if (file == NULL || strcmp(file, "") == 0) {
scr_err("Invalid source file @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* check that our other arguments are valid */
if (meta == NULL || segments == NULL) {
scr_err("Invalid metadata or segments @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* open the file for reading */
int fd_src = scr_open(file, O_RDONLY);
if (fd_src < 0) {
scr_err("Opening file to copy: scr_open(%s) errno=%d %s @ %s:%d",
file, errno, strerror(errno), __FILE__, __LINE__
);
return SCR_FAILURE;
}
#if !defined(__APPLE__)
/* TODO:
posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL)
that tells the kernel that you don't ever need the pages
from the file again, and it won't bother keeping them in the page cache.
*/
posix_fadvise(fd_src, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL);
#endif
/* get the buffer size we'll use to write to the file */
unsigned long buf_size = scr_file_buf_size;
/* allocate buffer to read in file chunks */
char* buf = (char*) SCR_MALLOC(buf_size);
/* initialize crc value */
uLong crc;
if (scr_crc_on_flush) {
crc = crc32(0L, Z_NULL, 0);
}
int rc = SCR_SUCCESS;
/* write out each segment */
scr_hash_sort_int(segments, SCR_HASH_SORT_ASCENDING);
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(segments);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the container info for this segment */
scr_hash* hash = scr_hash_elem_hash(elem);
/* get the offset into the container and the length of the segment (both in bytes) */
char* container_name;
unsigned long container_offset, segment_length;
if (scr_container_get_name_offset_length(hash,
&container_name, &container_offset, &segment_length) != SCR_SUCCESS)
{
scr_err("Failed to get segment offset and length @ %s:%d",
__FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
/* build full name to destination file */
scr_path* dst_path = scr_path_from_str(dst_dir);
scr_path_append_str(dst_path, container_name);
scr_path_reduce(dst_path);
char* dst_file = scr_path_strdup(dst_path);
/* open container file for writing -- we don't truncate here because more than one
* process may be writing to the same file */
int fd_container = scr_open(dst_file, O_WRONLY);
if (fd_container < 0) {
scr_err("Opening file for writing: scr_open(%s) errno=%d %s @ %s:%d",
dst_file, errno, strerror(errno), __FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
#if !defined(__APPLE__)
/* TODO:
posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL)
that tells the kernel that you don't ever need the pages
from the file again, and it won't bother keeping them in the page cache.
*/
posix_fadvise(fd_container, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL);
#endif
/* seek to offset within container */
off_t pos = (off_t) container_offset;
if (lseek(fd_container, pos, SEEK_SET) == (off_t)-1) {
/* our seek failed, return an error */
scr_err("Failed to seek to byte %lu in %s @ %s:%d",
pos, dst_file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
/* copy data from file into container in chunks */
unsigned long remaining = segment_length;
while (remaining > 0) {
/* read / write up to buf_size bytes at a time from file */
unsigned long count = remaining;
if (count > buf_size) {
count = buf_size;
}
/* attempt to read buf_size bytes from file */
int nread = scr_read_attempt(file, fd_src, buf, count);
/* if we read some bytes, write them out */
if (nread > 0) {
/* optionally compute crc value as we go */
if (scr_crc_on_flush) {
crc = crc32(crc, (const Bytef*) buf, (uInt) nread);
}
/* write our nread bytes out */
int nwrite = scr_write_attempt(dst_file, fd_container, buf, nread);
/* check for a write error or a short write */
if (nwrite != nread) {
/* write had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
/* subtract the bytes we've processed from the number remaining */
remaining -= (unsigned long) nread;
}
/* assume a short read is an error */
if (nread < count) {
/* read had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
/* check for a read error, stop copying and return an error */
if (nread < 0) {
/* read had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
}
/* close container */
if (scr_close(dst_file, fd_container) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* free the container file name and path */
scr_free(&dst_file);
scr_path_delete(&dst_path);
}
/* close the source file */
if (scr_close(file, fd_src) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* free buffer */
scr_free(&buf);
/* verify / set crc value */
if (rc == SCR_SUCCESS) {
uLong crc2;
if (scr_crc_on_flush) {
if (scr_meta_get_crc32(meta, &crc2) == SCR_SUCCESS) {
/* if a crc is already set in the meta data, check that we computed the same value */
if (crc != crc2) {
scr_err("CRC32 mismatch detected when flushing file %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
} else {
/* if there is no crc set, let's set it now */
scr_meta_set_crc32(meta, crc);
}
}
}
return rc;
}
/* flush files specified in list, and record corresponding entries for summary file */
static int scr_flush_files_list(scr_hash* file_list, scr_hash* summary)
{
/* assume we will succeed in this flush */
int rc = SCR_SUCCESS;
/* flush each of my files and fill in summary data structure */
scr_hash_elem* elem = NULL;
scr_hash* files = scr_hash_get(file_list, SCR_KEY_FILE);
for (elem = scr_hash_elem_first(files);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the filename */
char* file = scr_hash_elem_key(elem);
/* convert file to path and extract name of file */
scr_path* path_name = scr_path_from_str(file);
scr_path_basename(path_name);
/* get the hash for this element */
scr_hash* hash = scr_hash_elem_hash(elem);
/* get meta data for this file */
scr_meta* meta = scr_hash_get(hash, SCR_KEY_META);
/* if segments are defined, we flush the file to its containers,
* otherwise we copy the file out as is */
scr_hash* segments = scr_hash_get(hash, SCR_SUMMARY_6_KEY_SEGMENT);
if (segments != NULL) {
/* TODO: PRESERVE get original filename here */
/* add this file to the summary file */
char* name = scr_path_strdup(path_name);
scr_hash* file_hash = scr_hash_set_kv(summary, SCR_SUMMARY_6_KEY_FILE, name);
scr_free(&name);
// USERDEF fixme!
/* flush the file to the containers listed in its segmenets */
if (scr_flush_file_to_containers(file, meta, segments, scr_prefix) == SCR_SUCCESS) {
/* successfully flushed this file, record the filesize */
unsigned long filesize = 0;
if (scr_meta_get_filesize(meta, &filesize) == SCR_SUCCESS) {
scr_hash_util_set_bytecount(file_hash, SCR_SUMMARY_6_KEY_SIZE, filesize);
}
/* record the crc32 if one was computed */
uLong crc = 0;
if (scr_meta_get_crc32(meta, &crc) == SCR_SUCCESS) {
scr_hash_util_set_crc32(file_hash, SCR_SUMMARY_6_KEY_CRC, crc);
}
/* record segment information in summary file */
scr_hash* segments_copy = scr_hash_new();
scr_hash_merge(segments_copy, segments);
scr_hash_set(file_hash, SCR_SUMMARY_6_KEY_SEGMENT, segments_copy);
} else {
/* the flush failed */
rc = SCR_FAILURE;
/* explicitly mark file as incomplete */
scr_hash_set_kv_int(file_hash, SCR_SUMMARY_6_KEY_COMPLETE, 0);
}
} else {
/* get directory to flush file to */
char* dir;
if (scr_hash_util_get_str(hash, SCR_KEY_PATH, &dir) == SCR_SUCCESS) {
/* create full path of destination file */
scr_path* path_full = scr_path_from_str(dir);
scr_path_append(path_full, path_name);
/* get relative path to flushed file from SCR_PREFIX directory */
scr_path* path_relative = scr_path_relative(scr_prefix_path, path_full);
if (! scr_path_is_null(path_relative)) {
/* record the name of the file in the summary hash, and get reference to a hash for this file */
char* name = scr_path_strdup(path_relative);
scr_hash* file_hash = scr_hash_set_kv(summary, SCR_SUMMARY_6_KEY_FILE, name);
scr_free(&name);
/* flush the file and fill in the meta data for this file */
if (scr_flush_a_file(file, dir, meta) == SCR_SUCCESS) {
/* successfully flushed this file, record the filesize */
unsigned long filesize = 0;
if (scr_meta_get_filesize(meta, &filesize) == SCR_SUCCESS) {
scr_hash_util_set_bytecount(file_hash, SCR_SUMMARY_6_KEY_SIZE, filesize);
}
/* record the crc32 if one was computed */
uLong crc = 0;
if (scr_meta_get_crc32(meta, &crc) == SCR_SUCCESS) {
scr_hash_util_set_crc32(file_hash, SCR_SUMMARY_6_KEY_CRC, crc);
}
} else {
/* the flush failed */
rc = SCR_FAILURE;
/* explicitly mark incomplete files */
scr_hash_set_kv_int(file_hash, SCR_SUMMARY_6_KEY_COMPLETE, 0);
}
} else {
scr_abort(-1, "Failed to get relative path to directory %s from %s @ %s:%d",
dir, scr_prefix, __FILE__, __LINE__
);
}
/* free relative and full paths */
scr_path_delete(&path_relative);
scr_path_delete(&path_full);
} else {
scr_abort(-1, "Failed to read directory to flush file to @ %s:%d",
__FILE__, __LINE__
);
}
}
/* free the file name path */
scr_path_delete(&path_name);
}
return rc;
}
/* flushes file named in src_file to dst_dir and fills in meta based on flush,
* returns success of flush */
static int scr_flush_a_file(const char* src_file, const char* dst_dir, scr_meta* meta)
{
int flushed = SCR_SUCCESS;
int tmp_rc;
/* build full name to destination file */
scr_path* dst_path = scr_path_from_str(src_file);
scr_path_basename(dst_path);
scr_path_prepend_str(dst_path, dst_dir);
scr_path_reduce(dst_path);
char* dst_file = scr_path_strdup(dst_path);
/* copy file */
int crc_valid = 0;
uLong crc;
uLong* crc_p = NULL;
if (scr_crc_on_flush) {
crc_valid = 1;
crc_p = &crc;
}
tmp_rc = scr_file_copy(src_file, dst_file, scr_file_buf_size, crc_p);
if (tmp_rc != SCR_SUCCESS) {
crc_valid = 0;
flushed = SCR_FAILURE;
}
scr_dbg(2, "scr_flush_a_file: Read and copied %s to %s with success code %d @ %s:%d",
src_file, dst_file, tmp_rc, __FILE__, __LINE__
);
/* if file has crc32, check it against the one computed during the copy,
* otherwise if scr_crc_on_flush is set, record crc32 */
if (crc_valid) {
uLong crc_meta;
if (scr_meta_get_crc32(meta, &crc_meta) == SCR_SUCCESS) {
if (crc != crc_meta) {
/* detected a crc mismatch during the copy */
/* TODO: unlink the copied file */
/* scr_file_unlink(dst_file); */
/* mark the file as invalid */
scr_meta_set_complete(meta, 0);
flushed = SCR_FAILURE;
scr_err("scr_flush_a_file: CRC32 mismatch detected when flushing file %s to %s @ %s:%d",
src_file, dst_file, __FILE__, __LINE__
);
/* TODO: would be good to log this, but right now only rank 0 can write log entries */
/*
if (scr_log_enable) {
time_t now = scr_log_seconds();
scr_log_event("CRC32 MISMATCH", dst_file, NULL, &now, NULL);
}
*/
}
} else {
/* the crc was not already in the metafile, but we just computed it, so set it */
scr_meta_set_crc32(meta, crc);
}
}
/* TODO: check that written filesize matches expected filesize */
/* fill out meta data, set complete field based on flush success */
/* (we don't update the meta file here, since perhaps the file in cache is ok and only the flush failed) */
int complete = (flushed == SCR_SUCCESS);
scr_meta_set_complete(meta, complete);
/* free destination file string and path */
scr_free(&dst_file);
scr_path_delete(&dst_path);
return flushed;
}
int main(int argc, char* argv[])
{
/* print usage if not enough arguments were given */
if (argc < 2) {
printf("Usage: scr_inspect_cache <cntldir>\n");
return 1;
}
scr_path* scr_master_map_file = scr_path_from_str(strdup(argv[1]));
/* get my hostname */
if (gethostname(scr_my_hostname, sizeof(scr_my_hostname)) != 0) {
scr_err("scr_inspect_cache: Call to gethostname failed @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
/* read in the master map */
scr_hash* hash = scr_hash_new();
scr_hash_read_path(scr_master_map_file, hash);
/* create an empty filemap */
scr_filemap* map = scr_filemap_new();
/* for each filemap listed in the master map */
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(scr_hash_get(hash, "Filemap"));
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the filename of this filemap */
char* file = scr_hash_elem_key(elem);
/* read in the filemap */
scr_filemap* tmp_map = scr_filemap_new();
scr_path* path_file = scr_path_from_str(file);
scr_filemap_read(path_file, tmp_map);
scr_path_delete(&path_file);
/* merge it with local 0 filemap */
scr_filemap_merge(map, tmp_map);
/* delete filemap */
scr_filemap_delete(&tmp_map);
}
/* scan each file for each rank of each dataset */
scr_hash_elem* dset_elem;
for (dset_elem = scr_filemap_first_dataset(map);
dset_elem != NULL;
dset_elem = scr_hash_elem_next(dset_elem))
{
/* get dataset id */
int dset = scr_hash_elem_key_int(dset_elem);
scr_hash_elem* rank_elem;
for (rank_elem = scr_filemap_first_rank_by_dataset(map, dset);
rank_elem != NULL;
rank_elem = scr_hash_elem_next(rank_elem))
{
/* get rank id */
int rank = scr_hash_elem_key_int(rank_elem);
int missing_file = 0;
int expected = scr_filemap_get_expected_files(map, dset, rank);
int num = scr_filemap_num_files(map, dset, rank);
if (expected == num) {
/* first time through the file list, check that we have each file */
scr_hash_elem* file_elem = NULL;
for (file_elem = scr_filemap_first_file(map, dset, rank);
file_elem != NULL;
file_elem = scr_hash_elem_next(file_elem))
{
/* get filename */
char* file = scr_hash_elem_key(file_elem);
/* check that we can read the file */
if (! scr_bool_have_file(map, dset, rank, file)) {
missing_file = 1;
scr_dbg(1, "File is unreadable or incomplete: Dataset %d, Rank %d, File: %s",
dset, rank, file
);
}
}
} else {
missing_file = 1;
}
/* TODO: print partner names */
/* if we're not missing a file for rank, print this info out */
if (! missing_file) {
scr_hash* desc = scr_hash_new();
scr_filemap_get_desc(map, dset, rank, desc);
char* type = scr_hash_elem_get_first_val(desc, SCR_CONFIG_KEY_TYPE);
char* groups_str = scr_hash_elem_get_first_val(desc, SCR_CONFIG_KEY_GROUPS);
char* group_id_str = scr_hash_elem_get_first_val(desc, SCR_CONFIG_KEY_GROUP_ID);
char* group_size_str = scr_hash_elem_get_first_val(desc, SCR_CONFIG_KEY_GROUP_SIZE);
char* group_rank_str = scr_hash_elem_get_first_val(desc, SCR_CONFIG_KEY_GROUP_RANK);
if (type != NULL && groups_str != NULL && group_id_str != NULL && group_size_str != NULL && group_rank_str != NULL) {
/* we already have a group id and rank, use that to rebuild the communicator */
int groups = atoi(groups_str);
int group_id = atoi(group_id_str);
int group_size = atoi(group_size_str);
int group_rank = atoi(group_rank_str);
printf("DSET=%d RANK=%d TYPE=%s GROUPS=%d GROUP_ID=%d GROUP_SIZE=%d GROUP_RANK=%d FILES=1\n",
dset, rank, type, groups, group_id, group_size, group_rank
);
}
}
}
}
scr_path_delete(&scr_master_map_file);
return 0;
}
int main(int argc, char* argv[])
{
int i, j;
int index = 1;
/* print usage if not enough arguments were given */
if (argc < 2) {
printf("Usage: scr_rebuild_xor <size> <root> <missing_xor_filename> <ordered_remaining_xor_filenames>\n");
return 1;
}
/* TODO: want to pass this on command line? */
/* get current working directory */
char dsetdir[SCR_MAX_FILENAME];
scr_getcwd(dsetdir, sizeof(dsetdir));
/* create and reduce path for dataset */
scr_path* path_dset = scr_path_from_str(dsetdir);
scr_path_reduce(path_dset);
/* allocate buffers */
char* buffer_A = malloc(buffer_size * sizeof(char));
char* buffer_B = malloc(buffer_size * sizeof(char));
if (buffer_A == NULL || buffer_B == NULL) {
scr_err("Failed to allocate buffer memory @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
/* read in the size of the XOR set */
int xor_set_size = (int) strtol(argv[index++], (char **)NULL, 10);
if (xor_set_size <= 0) {
scr_err("Invalid XOR set size argument %s @ %s:%d",
argv[index-1], __FILE__, __LINE__
);
return 1;
}
/* allocate memory for data structures based on the XOR set size */
int* num_files = malloc(xor_set_size * sizeof(int));
int* offsets = malloc(xor_set_size * sizeof(int));
char** xor_files = malloc(xor_set_size * sizeof(char*));
int* xor_fds = malloc(xor_set_size * sizeof(int));
scr_hash** xor_headers = malloc(xor_set_size * sizeof(scr_hash*));
if (num_files == NULL || offsets == NULL || xor_files == NULL || xor_fds == NULL || xor_headers == NULL) {
scr_err("Failed to allocate buffer memory @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
/* read in the rank of the missing process (the root) */
int root = (int) strtol(argv[index++], (char **)NULL, 10);
if (root < 0 || root >= xor_set_size) {
scr_err("Invalid root argument %s @ %s:%d",
argv[index-1], __FILE__, __LINE__
);
return 1;
}
/* read in the missing xor filename */
xor_files[0] = strdup(argv[index++]);
if (xor_files[0] == NULL) {
scr_err("Failed to dup XOR filename @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
/* read in the xor filenames (expected to be in order of XOR segment number) */
/* we order ranks so that root is index 0, the rank to the right of root is index 1, and so on */
for (i=0; i < xor_set_size; i++) {
xor_headers[i] = scr_hash_new();
/* we'll get the XOR file name for root from the header stored in the XOR file of the partner */
if (i == root) {
continue;
}
/* adjust the index relative to root */
j = i - root;
if (j < 0) {
j += xor_set_size;
}
/* copy the XOR file name */
xor_files[j] = strdup(argv[index++]);
if (xor_files[j] == NULL) {
scr_err("Failed to dup XOR filename @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
}
/* open each of the xor files and read in the headers */
for (i=1; i < xor_set_size; i++) {
/* open each xor file for reading */
xor_fds[i] = scr_open(xor_files[i], O_RDONLY);
if (xor_fds[i] < 0) {
scr_err("Opening xor segment file: scr_open(%s) errno=%d %s @ %s:%d",
xor_files[i], errno, strerror(errno), __FILE__, __LINE__
);
return 1;
}
/* read the header from this xor file */
if (scr_hash_read_fd(xor_files[i], xor_fds[i], xor_headers[i]) < 0) {
scr_err("Failed to read XOR header from %s @ %s:%d",
xor_files[i], __FILE__, __LINE__
);
return 1;
}
}
/* build header for missing XOR file */
int partner_rank = -1;
if (xor_set_size >= 2) {
scr_hash_merge(xor_headers[0], xor_headers[1]);
/* fetch our own file list from rank to our right */
scr_hash* rhs_hash = scr_hash_get(xor_headers[1], SCR_KEY_COPY_XOR_PARTNER);
scr_hash* current_hash = scr_hash_new();
scr_hash_merge(current_hash, rhs_hash);
scr_hash_set(xor_headers[0], SCR_KEY_COPY_XOR_CURRENT, current_hash);
/* we are the partner to the rank to our left */
scr_hash* lhs_hash = scr_hash_get(xor_headers[xor_set_size-1], SCR_KEY_COPY_XOR_CURRENT);
scr_hash* partner_hash = scr_hash_new();
scr_hash_merge(partner_hash, lhs_hash);
scr_hash_set(xor_headers[0], SCR_KEY_COPY_XOR_PARTNER, partner_hash);
/* get global rank of partner */
if (scr_hash_util_get_int(lhs_hash, SCR_KEY_COPY_XOR_RANK, &partner_rank) != SCR_SUCCESS) {
scr_err("Failed to read partner rank from XOR file header in %s @ %s:%d",
xor_files[xor_set_size-1], __FILE__, __LINE__
);
return 1;
}
}
/* get a pointer to the current hash for the missing rank */
scr_hash* missing_current_hash = scr_hash_get(xor_headers[0], SCR_KEY_COPY_XOR_CURRENT);
/* read the rank */
int my_rank = -1;
if (scr_hash_util_get_int(missing_current_hash, SCR_KEY_COPY_XOR_RANK, &my_rank) != SCR_SUCCESS) {
scr_err("Failed to read rank from XOR file header in %s @ %s:%d",
xor_files[0], __FILE__, __LINE__
);
return 1;
}
/* get the dataset */
scr_dataset* dataset = scr_hash_get(xor_headers[0], SCR_KEY_COPY_XOR_DATASET);
/* read the dataset id */
int dset_id = -1;
if (scr_dataset_get_id(dataset, &dset_id) != SCR_SUCCESS) {
scr_err("Failed to read dataset id from XOR file header in %s @ %s:%d",
xor_files[0], __FILE__, __LINE__
);
return 1;
}
/* read the ranks */
int num_ranks = -1;
if (scr_hash_util_get_int(xor_headers[0], SCR_KEY_COPY_XOR_RANKS, &num_ranks) != SCR_SUCCESS) {
scr_err("Failed to read ranks from XOR file header in %s @ %s:%d",
xor_files[0], __FILE__, __LINE__
);
return 1;
}
/* get name of partner's fmap */
scr_path* path_partner_map = scr_path_from_str(".scr");
scr_path_append_strf(path_partner_map, "fmap.%d.scr", partner_rank);
/* extract partner's flush descriptor */
scr_hash* flushdesc = scr_hash_new();
scr_filemap* partner_map = scr_filemap_new();
scr_filemap_read(path_partner_map, partner_map);
scr_filemap_get_flushdesc(partner_map, dset_id, partner_rank, flushdesc);
scr_filemap_delete(&partner_map);
/* delete partner map path */
scr_path_delete(&path_partner_map);
/* determine whether we should preserve user directories */
int preserve_dirs = 0;
scr_hash_util_get_int(flushdesc, SCR_SCAVENGE_KEY_PRESERVE, &preserve_dirs);
/* read the chunk size */
unsigned long chunk_size = 0;
if (scr_hash_util_get_unsigned_long(xor_headers[0], SCR_KEY_COPY_XOR_CHUNK, &chunk_size) != SCR_SUCCESS) {
scr_err("Failed to read chunk size from XOR file header in %s @ %s:%d",
xor_files[0], __FILE__, __LINE__
);
return 1;
}
/* determine number of files each member wrote in XOR set */
for (i=0; i < xor_set_size; i++) {
/* record the number of files for this rank */
scr_hash* current_hash = scr_hash_get(xor_headers[i], SCR_KEY_COPY_XOR_CURRENT);
if (scr_hash_util_get_int(current_hash, SCR_KEY_COPY_XOR_FILES, &num_files[i]) != SCR_SUCCESS) {
scr_err("Failed to read number of files from %s @ %s:%d",
xor_files[i], __FILE__, __LINE__
);
return 1;
}
}
/* count the total number of files and set the offsets array */
int total_num_files = 0;
for (i=0; i < xor_set_size; i++) {
offsets[i] = total_num_files;
total_num_files += num_files[i];
}
/* allocate space for a file descriptor, file name pointer, and filesize for each user file */
int* user_fds = (int*) malloc(total_num_files * sizeof(int));
char** user_files = (char**) malloc(total_num_files * sizeof(char*));
char** user_rel_files = (char**) malloc(total_num_files * sizeof(char*));
unsigned long* user_filesizes = (unsigned long*) malloc(total_num_files * sizeof(unsigned long));
if (user_fds == NULL || user_files == NULL || user_rel_files == NULL || user_filesizes == NULL) {
scr_err("Failed to allocate buffer memory @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
/* get file name, file size, and open each of the user files that we have */
for (i=0; i < xor_set_size; i++) {
scr_hash* current_hash = scr_hash_get(xor_headers[i], SCR_KEY_COPY_XOR_CURRENT);
/* for each file belonging to this rank, get filename, filesize, and open file */
for (j=0; j < num_files[i]; j++) {
int offset = offsets[i] + j;
/* get the meta data for this file */
scr_meta* meta = scr_hash_get_kv_int(current_hash, SCR_KEY_COPY_XOR_FILE, j);
if (meta == NULL) {
scr_err("Failed to read meta data for file %d in %s @ %s:%d",
j, xor_files[i], __FILE__, __LINE__
);
return 1;
}
/* record the filesize of this file */
if (scr_meta_get_filesize(meta, &user_filesizes[offset]) != SCR_SUCCESS) {
scr_err("Failed to read filesize field for file %d in %s @ %s:%d",
j, xor_files[i], __FILE__, __LINE__
);
return 1;
}
/* get filename */
char* origname;
if (scr_meta_get_origname(meta, &origname) != SCR_SUCCESS) {
scr_err("Failed to read original name for file %d in %s @ %s:%d",
j, xor_files[i], __FILE__, __LINE__
);
return 1;
}
/* construct full path to user file */
scr_path* path_user_full = scr_path_from_str(origname);
if (preserve_dirs) {
/* get original path of file */
char* origpath;
if (scr_meta_get_origpath(meta, &origpath) != SCR_SUCCESS) {
scr_err("Failed to read original path for file %d in %s @ %s:%d",
j, xor_files[i], __FILE__, __LINE__
);
return 1;
}
/* construct full path to file */
scr_path_prepend_str(path_user_full, origpath);
} else {
/* construct full path to file */
scr_path_prepend(path_user_full, path_dset);
}
/* reduce path to user file */
scr_path_reduce(path_user_full);
/* make a copy of the full path */
user_files[offset] = scr_path_strdup(path_user_full);
/* make a copy of relative path */
scr_path* path_user_rel = scr_path_relative(path_dset, path_user_full);
user_rel_files[offset] = scr_path_strdup(path_user_rel);
scr_path_delete(&path_user_rel);
/* free the full path */
scr_path_delete(&path_user_full);
/* open the file */
if (i == 0) {
/* create directory for file */
scr_path* user_dir_path = scr_path_from_str(user_files[offset]);
scr_path_reduce(user_dir_path);
scr_path_dirname(user_dir_path);
if (! scr_path_is_null(user_dir_path)) {
char* user_dir = scr_path_strdup(user_dir_path);
mode_t mode_dir = scr_getmode(1, 1, 1);
if (scr_mkdir(user_dir, mode_dir) != SCR_SUCCESS) {
scr_err("Failed to create directory for user file %s @ %s:%d",
user_dir, __FILE__, __LINE__
);
return 1;
}
scr_free(&user_dir);
}
scr_path_delete(&user_dir_path);
/* open missing file for writing */
mode_t mode_file = scr_getmode(1, 1, 0);
user_fds[offset] = scr_open(user_files[offset], O_WRONLY | O_CREAT | O_TRUNC, mode_file);
if (user_fds[offset] < 0) {
scr_err("Opening user file for writing: scr_open(%s) errno=%d %s @ %s:%d",
user_files[offset], errno, strerror(errno), __FILE__, __LINE__
);
return 1;
}
} else {
/* open existing file for reading */
user_fds[offset] = scr_open(user_files[offset], O_RDONLY);
if (user_fds[offset] < 0) {
scr_err("Opening user file for reading: scr_open(%s) errno=%d %s @ %s:%d",
user_files[offset], errno, strerror(errno), __FILE__, __LINE__
);
return 1;
}
}
}
}
/* finally, open the xor file for the missing rank */
mode_t mode_file = scr_getmode(1, 1, 0);
xor_fds[0] = scr_open(xor_files[0], O_WRONLY | O_CREAT | O_TRUNC, mode_file);
if (xor_fds[0] < 0) {
scr_err("Opening xor file to be reconstructed: scr_open(%s) errno=%d %s @ %s:%d",
xor_files[0], errno, strerror(errno), __FILE__, __LINE__
);
return 1;
}
int rc = 0;
/* write the header to the XOR file of the missing rank */
if (scr_hash_write_fd(xor_files[0], xor_fds[0], xor_headers[0]) < 0) {
rc = 1;
}
/* this offset array records the current position we are in the logical file for each rank */
unsigned long* offset = malloc(xor_set_size * sizeof(unsigned long));
if (offset == NULL) {
scr_err("Failed to allocate buffer memory @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
for (i=0; i < xor_set_size; i++) {
offset[i] = 0;
}
unsigned long write_pos = 0;
int chunk_id;
for (chunk_id = 0; chunk_id < xor_set_size && rc == 0; chunk_id++) {
size_t nread = 0;
while (nread < chunk_size && rc == 0) {
/* read upto buffer_size bytes at a time */
size_t count = chunk_size - nread;
if (count > buffer_size) {
count = buffer_size;
}
/* clear our buffer */
memset(buffer_A, 0, count);
/* read a segment from each rank and XOR it into our buffer */
for (i=1; i < xor_set_size; i++) {
/* read the next set of bytes for this chunk from my file into send_buf */
if (chunk_id != ((i + root) % xor_set_size)) {
/* read chunk from the logical file for this rank */
if (scr_read_pad_n(num_files[i], &user_files[offsets[i]], &user_fds[offsets[i]],
buffer_B, count, offset[i], &user_filesizes[offsets[i]]) != SCR_SUCCESS)
{
/* our read failed, set the return code to an error */
rc = 1;
count = 0;
}
offset[i] += count;
} else {
/* read chunk from the XOR file for this rank */
if (scr_read_attempt(xor_files[i], xor_fds[i], buffer_B, count) != count) {
/* our read failed, set the return code to an error */
rc = 1;
count = 0;
}
}
/* TODO: XORing with unsigned long would be faster here (if chunk size is multiple of this size) */
/* merge the blocks via xor operation */
for (j = 0; j < count; j++) {
buffer_A[j] ^= buffer_B[j];
}
}
/* at this point, we have the data from the missing rank, write it out */
if (chunk_id != root) {
/* write chunk to logical file for the missing rank */
if (scr_write_pad_n(num_files[0], &user_files[0], &user_fds[0],
buffer_A, count, write_pos, &user_filesizes[0]) != SCR_SUCCESS)
{
/* our write failed, set the return code to an error */
rc = 1;
}
write_pos += count;
} else {
/* write chunk to xor file for the missing rank */
if (scr_write_attempt(xor_files[0], xor_fds[0], buffer_A, count) != count) {
/* our write failed, set the return code to an error */
rc = 1;
}
}
nread += count;
}
}
/* close each of the user files */
for (i=0; i < total_num_files; i++) {
if (scr_close(user_files[i], user_fds[i]) != SCR_SUCCESS) {
rc = 1;
}
}
/* close each of the XOR files */
for (i=0; i < xor_set_size; i++) {
if (scr_close(xor_files[i], xor_fds[i]) != SCR_SUCCESS) {
rc = 1;
}
}
/* if the write failed, delete the files we just wrote, and return an error */
if (rc != 0) {
for (j=0; j < num_files[0]; j++) {
scr_file_unlink(user_files[j]);
}
scr_file_unlink(xor_files[0]);
return 1;
}
/* check that filesizes are correct */
unsigned long filesize;
for (j=0; j < num_files[0]; j++) {
filesize = scr_file_size(user_files[j]);
if (filesize != user_filesizes[j]) {
/* the filesize check failed, so delete the file */
scr_file_unlink(user_files[j]);
/* mark the file as incomplete */
scr_meta* meta = scr_hash_get_kv_int(missing_current_hash, SCR_KEY_COPY_XOR_FILE, j);
scr_meta_set_complete(meta, 0);
rc = 1;
}
}
/* TODO: we didn't record the filesize of the XOR file for the missing rank anywhere */
/* create a filemap for this rank */
scr_filemap* map = scr_filemap_new();
if (map == NULL) {
scr_err("Failed to allocate filemap @ %s:%d",
__FILE__, __LINE__
);
return 1;
}
/* record the dataset information in the filemap */
scr_filemap_set_dataset(map, dset_id, my_rank, dataset);
/* write meta data for each of the user files and add each one to the filemap */
for (j=0; j < num_files[0]; j++) {
/* add user file to filemap and record meta data */
char* user_file_relative = user_rel_files[j];
scr_filemap_add_file(map, dset_id, my_rank, user_file_relative);
scr_meta* meta = scr_hash_get_kv_int(missing_current_hash, SCR_KEY_COPY_XOR_FILE, j);
scr_filemap_set_meta(map, dset_id, my_rank, user_file_relative, meta);
}
/* write meta data for xor file and add it to the filemap */
scr_filemap_add_file(map, dset_id, my_rank, xor_files[0]);
unsigned long full_chunk_filesize = scr_file_size(xor_files[0]);
int missing_complete = 1;
scr_meta* meta_chunk = scr_meta_new();
scr_meta_set_filename(meta_chunk, xor_files[0]);
scr_meta_set_filetype(meta_chunk, SCR_META_FILE_XOR);
scr_meta_set_filesize(meta_chunk, full_chunk_filesize);
/* TODO: remove this from meta file, for now it's needed in scr_index.c */
scr_meta_set_ranks(meta_chunk, num_ranks);
scr_meta_set_complete(meta_chunk, missing_complete);
scr_filemap_set_meta(map, dset_id, my_rank, xor_files[0], meta_chunk);
/* set expected number of files for the missing rank */
int expected_num_files = scr_filemap_num_files(map, dset_id, my_rank);
scr_filemap_set_expected_files(map, dset_id, my_rank, expected_num_files);
/* compute, check, and store crc values with files */
for (j=0; j < num_files[0]; j++) {
/* compute crc on user file */
char* user_file_relative = user_rel_files[j];
if (scr_compute_crc(map, dset_id, my_rank, user_file_relative) != SCR_SUCCESS) {
/* the crc check failed, so delete the file */
scr_file_unlink(user_files[j]);
rc = 1;
}
}
if (scr_compute_crc(map, dset_id, my_rank, xor_files[0]) != SCR_SUCCESS) {
/* the crc check failed, so delete the file */
scr_file_unlink(xor_files[0]);
rc = 1;
}
/* store flush descriptor */
scr_filemap_set_flushdesc(map, dset_id, my_rank, flushdesc);
/* write filemap for this rank */
scr_path* path_map = scr_path_from_str(".scr");
scr_path_append_strf(path_map, "fmap.%d.scr", my_rank);
if (scr_filemap_write(path_map, map) != SCR_SUCCESS) {
rc = 1;
}
scr_path_delete(&path_map);
/* delete the map */
scr_filemap_delete(&map);
scr_meta_delete(&meta_chunk);
/* delete the flush/scavenge descriptor */
scr_hash_delete(&flushdesc);
scr_free(&offset);
for (i=0; i < total_num_files; i++) {
scr_free(&user_rel_files[i]);
scr_free(&user_files[i]);
}
scr_free(&user_filesizes);
scr_free(&user_rel_files);
scr_free(&user_files);
scr_free(&user_fds);
for (i=0; i < xor_set_size; i++) {
scr_hash_delete(&xor_headers[i]);
}
for (i=0; i < xor_set_size; i++) {
scr_free(&xor_files[i]);
}
scr_free(&xor_headers);
scr_free(&xor_fds);
scr_free(&xor_files);
scr_free(&offsets);
scr_free(&num_files);
scr_free(&buffer_B);
scr_free(&buffer_A);
scr_path_delete(&path_dset);
return rc;
}
/* read contents of summary file */
static int scr_fetch_summary(
const char* summary_dir,
scr_hash* file_list)
{
/* assume that we won't succeed in our fetch attempt */
int rc = SCR_SUCCESS;
/* check whether summary file exists and is readable */
if (scr_my_rank_world == 0) {
/* check that we can access the directory */
if (scr_file_is_readable(summary_dir) != SCR_SUCCESS) {
scr_err("Failed to access summary directory %s @ %s:%d",
summary_dir, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
}
/* broadcast success code from rank 0 */
MPI_Bcast(&rc, 1, MPI_INT, 0, scr_comm_world);
if (rc != SCR_SUCCESS) {
return rc;
}
/* add path to file list */
scr_hash_util_set_str(file_list, SCR_KEY_PATH, summary_dir);
/* build path to summary file */
scr_path* dataset_path = scr_path_from_str(summary_dir);
scr_path* meta_path = scr_path_dup(dataset_path);
scr_path_append_str(meta_path, ".scr");
scr_path_reduce(meta_path);
/* rank 0 reads the summary file */
scr_hash* header = scr_hash_new();
if (scr_my_rank_world == 0) {
/* build path to summary file */
scr_path* summary_path = scr_path_dup(meta_path);
scr_path_append_str(summary_path, "summary.scr");
const char* summary_file = scr_path_strdup(summary_path);
/* open file for reading */
int fd = scr_open(summary_file, O_RDONLY);
if (fd >= 0) {
/* read summary hash */
ssize_t header_size = scr_hash_read_fd(summary_file, fd, header);
if (header_size < 0) {
rc = SCR_FAILURE;
}
/* TODO: check that the version is correct */
/* close the file */
scr_close(summary_file, fd);
} else {
scr_err("Failed to open summary file %s @ %s:%d",
summary_file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* free summary path and string */
scr_free(&summary_file);
scr_path_delete(&summary_path);
}
/* broadcast success code from rank 0 */
MPI_Bcast(&rc, 1, MPI_INT, 0, scr_comm_world);
if (rc != SCR_SUCCESS) {
goto cleanup;
}
/* broadcast the summary hash */
scr_hash_bcast(header, 0, scr_comm_world);
/* extract and record the datast in file list */
scr_hash* dataset_hash = scr_hash_new();
scr_dataset* dataset = scr_hash_get(header, SCR_SUMMARY_6_KEY_DATASET);
scr_hash_merge(dataset_hash, dataset);
scr_hash_set(file_list, SCR_SUMMARY_6_KEY_DATASET, dataset_hash);
/* build path to rank2file map */
scr_path* rank2file_path = scr_path_dup(meta_path);
scr_path_append_str(rank2file_path, "rank2file.scr");
/* fetch file names and offsets containing file hash data */
int valid = 0;
char* file = NULL;
unsigned long offset = 0;
if (scr_my_rank_world == 0) {
/* rank 0 is only valid reader to start with */
valid = 1;
file = scr_path_strdup(rank2file_path);
offset = 0;
}
if (scr_fetch_rank2file_map(dataset_path, 1, &valid, &file, &offset)
!= SCR_SUCCESS)
{
rc = SCR_FAILURE;
}
/* create hashes to exchange data */
scr_hash* send = scr_hash_new();
scr_hash* recv = scr_hash_new();
/* read data from file */
if (valid) {
/* open file if necessary */
int fd = scr_open(file, O_RDONLY);
if (fd >= 0) {
/* create hash to hold file contents */
scr_hash* save = scr_hash_new();
/* read hash from file */
scr_lseek(file, fd, offset, SEEK_SET);
ssize_t readsize = scr_hash_read_fd(file, fd, save);
if (readsize < 0) {
scr_err("Failed to read rank2file map file %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* check that the number of ranks match */
int ranks = 0;
scr_hash_util_get_int(save, SCR_SUMMARY_6_KEY_RANKS, &ranks);
if (ranks != scr_ranks_world) {
scr_err("Invalid number of ranks in %s, got %d expected %d @ %s:%d",
file, ranks, scr_ranks_world, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* delete current send hash, set it to values from file,
* delete file hash */
scr_hash_delete(&send);
send = scr_hash_extract(save, SCR_SUMMARY_6_KEY_RANK);
scr_hash_delete(&save);
/* close the file */
scr_close(file, fd);
} else {
scr_err("Failed to open rank2file map %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* delete file name string */
scr_free(&file);
}
/* check that everyone read the data ok */
if (! scr_alltrue(rc == SCR_SUCCESS)) {
rc = SCR_FAILURE;
goto cleanup_hashes;
}
/* scatter to groups */
scr_hash_exchange_direction(send, recv, scr_comm_world, SCR_HASH_EXCHANGE_RIGHT);
/* iterate over the ranks that sent data to us, and set up our
* list of files */
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(recv);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* the key is the source rank, which we don't care about,
* the info we need is in the element hash */
scr_hash* elem_hash = scr_hash_elem_hash(elem);
/* get pointer to file hash */
scr_hash* file_hash = scr_hash_get(elem_hash, SCR_SUMMARY_6_KEY_FILE);
if (file_hash != NULL) {
/* TODO: parse summary file format */
scr_hash_merge(file_list, elem_hash);
} else {
rc = SCR_FAILURE;
}
}
/* fill in file list parameters */
if (rc == SCR_SUCCESS) {
/* if we're not using containers, add PATH entry for each of our
* files */
scr_hash* files = scr_hash_get(file_list, SCR_KEY_FILE);
for (elem = scr_hash_elem_first(files);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the file name */
char* file = scr_hash_elem_key(elem);
/* combine the file name with the summary directory to build a
* full path to the file */
scr_path* path_full = scr_path_dup(dataset_path);
scr_path_append_str(path_full, file);
/* subtract off last component to get just the path */
scr_path_dirname(path_full);
char* path = scr_path_strdup(path_full);
/* record path in file list */
scr_hash* hash = scr_hash_elem_hash(elem);
scr_hash_util_set_str(hash, SCR_KEY_PATH, path);
/* free the path and string */
scr_free(&path);
scr_path_delete(&path_full);
}
}
/* check that everyone read the data ok */
if (! scr_alltrue(rc == SCR_SUCCESS)) {
rc = SCR_FAILURE;
goto cleanup_hashes;
}
cleanup_hashes:
/* delete send and receive hashes */
scr_hash_delete(&recv);
scr_hash_delete(&send);
/* free string and path for rank2file map */
scr_path_delete(&rank2file_path);
cleanup:
/* free the header hash */
scr_hash_delete(&header);
/* free path for dataset directory */
scr_path_delete(&meta_path);
scr_path_delete(&dataset_path);
return rc;
}
/* fetch files listed in hash into specified cache directory,
* update filemap and fill in total number of bytes fetched,
* returns SCR_SUCCESS if successful */
static int scr_fetch_files_list(
const scr_hash* file_list,
const char* dir,
scr_filemap* map)
{
/* assume we'll succeed in fetching our files */
int rc = SCR_SUCCESS;
/* assume we don't have any files to fetch */
int my_num_files = 0;
/* get dataset id */
int id;
scr_dataset* dataset = scr_hash_get(file_list, SCR_KEY_DATASET);
scr_dataset_get_id(dataset, &id);
/* now iterate through the file list and fetch each file */
scr_hash_elem* file_elem = NULL;
scr_hash* files = scr_hash_get(file_list, SCR_KEY_FILE);
for (file_elem = scr_hash_elem_first(files);
file_elem != NULL;
file_elem = scr_hash_elem_next(file_elem))
{
/* get the filename */
char* file = scr_hash_elem_key(file_elem);
/* get a pointer to the hash for this file */
scr_hash* hash = scr_hash_elem_hash(file_elem);
/* check whether we are supposed to fetch this file */
/* TODO: this is a hacky way to avoid reading a redundancy file
* back in under the assumption that it's an original file, which
* breaks our redundancy computation due to a name conflict on
* the file names */
scr_hash_elem* no_fetch_hash = scr_hash_elem_get(hash, SCR_SUMMARY_6_KEY_NOFETCH);
if (no_fetch_hash != NULL) {
continue;
}
/* increment our file count */
my_num_files++;
/* build the destination file name */
scr_path* path_newfile = scr_path_from_str(file);
scr_path_basename(path_newfile);
scr_path_prepend_str(path_newfile, dir);
char* newfile = scr_path_strdup(path_newfile);
/* add the file to our filemap and write it to disk before creating
* the file, this way we have a record that it may exist before we
* actually start to fetch it */
scr_filemap_add_file(map, id, scr_my_rank_world, newfile);
scr_filemap_write(scr_map_file, map);
/* get the file size */
unsigned long filesize = 0;
if (scr_hash_util_get_unsigned_long(hash, SCR_KEY_SIZE, &filesize) != SCR_SUCCESS) {
scr_err("Failed to read file size from summary data @ %s:%d",
__FILE__, __LINE__
);
rc = SCR_FAILURE;
/* free path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
break;
}
/* check for a complete flag */
int complete = 1;
if (scr_hash_util_get_int(hash, SCR_KEY_COMPLETE, &complete) != SCR_SUCCESS) {
/* in summary file, the absence of a complete flag on a file
* implies the file is complete */
complete = 1;
}
/* create a new meta data object for this file */
scr_meta* meta = scr_meta_new();
/* set the meta data */
scr_meta_set_filename(meta, newfile);
scr_meta_set_filetype(meta, SCR_META_FILE_USER);
scr_meta_set_filesize(meta, filesize);
scr_meta_set_complete(meta, 1);
/* TODODSET: move the ranks field elsewhere, for now it's needed
* by scr_index.c */
scr_meta_set_ranks(meta, scr_ranks_world);
/* get the crc, if set, and add it to the meta data */
uLong crc;
if (scr_hash_util_get_crc32(hash, SCR_KEY_CRC, &crc) == SCR_SUCCESS) {
scr_meta_set_crc32(meta, crc);
}
/* fetch file from containers if they are defined, otherwise fetch
* the native file */
scr_hash* segments = scr_hash_get(hash, SCR_SUMMARY_6_KEY_SEGMENT);
if (segments != NULL) {
/* get source path */
char* from_dir;
if (scr_hash_util_get_str(file_list, SCR_KEY_PATH, &from_dir) == SCR_SUCCESS) {
/* fetch file from containers */
if (scr_fetch_file_from_containers(newfile, meta, segments, from_dir) != SCR_SUCCESS) {
/* failed to fetch file, mark it as incomplete */
scr_meta_set_complete(meta, 0);
rc = SCR_FAILURE;
}
} else {
/* failed to find base dataset directory in file list */
rc = SCR_FAILURE;
}
} else {
/* fetch native file, lookup directory for this file */
char* from_dir;
if (scr_hash_util_get_str(hash, SCR_KEY_PATH, &from_dir) == SCR_SUCCESS) {
if (scr_fetch_file(newfile, from_dir, meta) != SCR_SUCCESS) {
/* failed to fetch file, mark it as incomplete */
scr_meta_set_complete(meta, 0);
rc = SCR_FAILURE;
}
} else {
/* failed to read source directory, mark file as incomplete */
scr_meta_set_complete(meta, 0);
rc = SCR_FAILURE;
}
}
/* TODODSET: want to write out filemap before we start to fetch
* each file? */
/* mark the file as complete */
scr_filemap_set_meta(map, id, scr_my_rank_world, newfile, meta);
/* free the meta data object */
scr_meta_delete(&meta);
/* free path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
}
/* set the expected number of files for this dataset */
scr_filemap_set_expected_files(map, id, scr_my_rank_world, my_num_files);
scr_filemap_write(scr_map_file, map);
return rc;
}