static
void *
S3_Create_Or_Open_internal(char* testFileName,
IOR_param_t* param,
unsigned char createFile,
int multi_part_upload_p ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Create_Or_Open('%s', ,%d, %d)\n",
testFileName, createFile, multi_part_upload_p);
}
/* initialize curl, if needed */
s3_connect( param );
/* Check for unsupported flags */
if ( param->openFlags & IOR_EXCL ) {
fprintf( stdout, "Opening in Exclusive mode is not implemented in S3\n" );
}
if ( param->useO_DIRECT == TRUE ) {
fprintf( stdout, "Direct I/O mode is not implemented in S3\n" );
}
// easier to think
int n_to_n = param->filePerProc;
int n_to_1 = ! n_to_n;
/* check whether object needs reset to zero-length */
int needs_reset = 0;
if (! multi_part_upload_p)
needs_reset = 1; /* so "append" can work */
else if ( param->openFlags & IOR_TRUNC )
needs_reset = 1; /* so "append" can work */
else if (createFile) {
// AWS4C_CHECK( s3_head(param->io_buf, testFileName) );
// if ( ! AWS4C_OK(param->io_buf) )
needs_reset = 1;
}
if ( param->open == WRITE ) {
/* initializations for N:1 or N:N writes using multi-part upload */
if (multi_part_upload_p) {
// For N:N, all ranks do their own MPU open/close. For N:1, only
// rank0 does that. Either way, the response from the server
// includes an "uploadId", which must be used to upload parts to
// the same object.
if ( n_to_n || (rank == 0) ) {
// rank0 handles truncate
if ( needs_reset) {
aws_iobuf_reset(param->io_buf);
AWS4C_CHECK( s3_put(param->io_buf, testFileName) ); /* 0-length write */
AWS4C_CHECK_OK( param->io_buf );
}
// POST request with URL+"?uploads" initiates multi-part upload
snprintf(buff, BUFF_SIZE, "%s?uploads", testFileName);
IOBuf* response = aws_iobuf_new();
AWS4C_CHECK( s3_post2(param->io_buf, buff, NULL, response) );
AWS4C_CHECK_OK( param->io_buf );
// parse XML returned from server, into a tree structure
aws_iobuf_realloc(response);
xmlDocPtr doc = xmlReadMemory(response->first->buf,
response->first->len,
NULL, NULL, 0);
if (doc == NULL)
ERR_SIMPLE("Rank0 Failed to find POST response\n");
// navigate parsed XML-tree to find UploadId
xmlNode* root_element = xmlDocGetRootElement(doc);
const char* upload_id = find_element_named(root_element, (char*)"UploadId");
if (! upload_id)
ERR_SIMPLE("couldn't find 'UploadId' in returned XML\n");
if (param->verbose >= VERBOSE_3)
printf("got UploadId = '%s'\n", upload_id);
const size_t upload_id_len = strlen(upload_id);
if (upload_id_len > MAX_UPLOAD_ID_SIZE) {
snprintf(buff, BUFF_SIZE,
"UploadId length %d exceeds expected max (%d)",
upload_id_len, MAX_UPLOAD_ID_SIZE);
ERR_SIMPLE(buff);
}
// save the UploadId we found
memcpy(param->UploadId, upload_id, upload_id_len);
param->UploadId[upload_id_len] = 0;
// free storage for parsed XML tree
xmlFreeDoc(doc);
aws_iobuf_free(response);
// For N:1, share UploadId across all ranks
if (n_to_1)
MPI_Bcast(param->UploadId, MAX_UPLOAD_ID_SIZE, MPI_BYTE, 0, param->testComm);
}
else
// N:1, and we're not rank0. recv UploadID from Rank 0
MPI_Bcast(param->UploadId, MAX_UPLOAD_ID_SIZE, MPI_BYTE, 0, param->testComm);
}
/* initializations for N:N or N:1 writes using EMC byte-range extensions */
else {
/* maybe reset to zero-length, so "append" can work */
if (needs_reset) {
if (verbose >= VERBOSE_3) {
fprintf( stdout, "rank %d resetting\n",
rank);
}
aws_iobuf_reset(param->io_buf);
AWS4C_CHECK( s3_put(param->io_buf, testFileName) );
AWS4C_CHECK_OK( param->io_buf );
}
}
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Create_Or_Open\n");
}
return ((void *) testFileName );
}
int
main (int argc, char *argv[]) {
aws_init();
if(argv[3] != NULL) {
aws_set_debug(atoi(argv[3]));
}
IOBuf * aws_buf = aws_iobuf_new();
// Read credential file
int rv = aws_read_config("myteksi");
if ( rv )
{
fprintf(stderr, "Could not find a credential in the config file \n" );
fprintf(stderr, "Make sure your ~/.awsAuth file is correct \n" );
exit (1);
}
// Read config file
FILE *fp = NULL;
char getline[ LINE_MAX * sizeof(char) ];
if( (fp = fopen("s3config", "r")) == NULL) {
//File does not exist. Initialize it
if( (fp = fopen("s3config", "w+")) == NULL) {
fprintf(stderr, "ERROR: Unable to create config file.\n");
exit(0);
}
// Ask for bucket_name
fprintf(stdout, "Config file doesn't exist yet! Creating one now. \n");
fprintf(stdout, "Please specify the AWS S3 base address "
"[default s3.amazonaws.com] :");
char getInput[ LINE_MAX * sizeof(char) ];
if( fgets( getInput, sizeof(getInput) , stdin ) != NULL ) {
if( strcmp(getInput, "\n") != 0 ) {
S3_host = strndup(getInput, strlen(getInput) -1); // Remove trailing NL
}
else {
S3_host = strdup("s3.amazonaws.com");
}
}
int validbucketname = 0;
while( !validbucketname ) {
fprintf(stdout, "Please specify the bucket name: ");
if( fgets( getInput, sizeof(getInput) , stdin ) != NULL ) {
bucketname = strndup(getInput, strlen(getInput) -1);
validbucketname = 1;
}
}
char * buf = malloc( snprintf(NULL, 0, "S3_Base_Address=\"%s\"\n"
"bucket_name=\"%s\"\n", S3_host, bucketname));
sprintf(buf, "S3_Base_Address=\"%s\"\n"
"bucket_name=\"%s\"\n", S3_host, bucketname );
if( fputs( buf, fp ) == EOF ) {
fprintf(stderr, "ERROR: Unable to create config file.\n");
}
}
// Config file exist, parse it
else {
char delim[4] = {'=', '\"', '\n', '\0'};
char* left;
char* right;
while( fgets( getline, sizeof(getline) , fp ) != NULL ) {
if( (left = strtok(getline, delim)) != NULL ) {
right = strtok(NULL, delim);
}
else {
//Empty Line
}
// Match the strings
char* comparison = "S3_Base_Address";
if( strcmp(left, comparison) == 0) {
if(right != NULL) {
S3_host = strdup(right);
}
else {
S3_host = strdup("s3.amazonaws.com");
}
}
comparison = "bucket_name";
if( strcmp(left, comparison) == 0 && right != NULL) {
bucketname = strdup(right);
}
} // End while
if( S3_host == NULL || bucketname == NULL ) {
fprintf(stderr, "ERROR: Invalid entry in config file.\n");
}
}
// Set parameters in S3 library
s3_set_host(S3_host);
s3_set_bucket(bucketname);
s3_set_acl(S3_acl);
// Check for valid arguments
if ( argc != 3 && argc != 4 ) {
fprintf(stderr, "Usage: s3util <operation> <filename>\n");
fprintf(stderr, "Operation can be one of {PUT, GET, DELETE}\n");
exit(1);
}
// Check if operation is valid
operation = strdup(argv[1]);
filename = strdup(argv[2]);
// PUT file
if( strcmp(operation, "PUT") == 0 ) {
int rc;
char s3replyMD5[33];
rv = put_file( aws_buf, filename );
rc = -1;
if( aws_buf->eTag != NULL && strlen(aws_buf->eTag) > 2 ) {
memset(s3replyMD5, 0, 33);
memcpy(s3replyMD5, aws_buf->eTag + 1, 32);
rc = verifyMD5(filename, s3replyMD5);
}
if(rv != 0 || rc != 0) {
printf ( "PUT operation was unsuccessful \n" );
return rc;
}
printf ( "MD5SUM matches, file uploaded successfully \n" );
}
// GET file
else if( strcmp(operation, "GET") == 0 ) {
rv = get_file( aws_buf, filename );
if(rv == 0 && aws_buf->code == 200) {
printf ( "File was successfully downloaded \n" );
}
else {
printf ( "GET operation was unsuccessful \n" );
return(-1);
}
}
// DELETE FILE
else if( strcmp(operation, "DELETE") == 0 ) {
rv = delete_file( aws_buf, filename );
if(rv == 0 && aws_buf->code == 204) {
printf ( "File was successfully deleted \n" );
}
else {
printf ( "DELETE operation was unsuccessful \n" );
return(-1);
}
}
else {
fprintf(stderr, "Invalid operation, operation must be one of "
"{PUT, GET, DELETE}\n");
return(1);
}
/*
printf ( "RV %d\n", rv );
printf ( "CODE [%d] \n", aws_buf->code );
printf ( "RESULT [%s] \n", aws_buf->result );
printf ( "LEN [%d] \n", aws_buf->len );
printf ( "LASTMOD [%s] \n", aws_buf->lastMod );
printf ( "ETAG [%s] \n", aws_buf->eTag );
*/
aws_iobuf_free(aws_buf);
global_free();
return 0;
}
static
void
S3_Close_internal( void* fd,
IOR_param_t* param,
int multi_part_upload_p ) {
char* fname = (char*)fd; /* see NOTE above S3_Create_Or_Open() */
// easier to think
int n_to_n = param->filePerProc;
int n_to_1 = (! n_to_n);
int segmented = (param->segmentCount == 1);
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Close('%s', ,%d) %s\n",
fname,
multi_part_upload_p,
((n_to_n) ? "N:N" : ((segmented) ? "N:1(seg)" : "N:1(str)")));
}
if (param->open == WRITE) {
// finalizing Multi-Part Upload (for N:1 or N:N)
if (multi_part_upload_p) {
size_t etag_data_size = param->etags->write_count; /* local ETag data (bytes) */
size_t etags_per_rank = etag_data_size / ETAG_SIZE; /* number of local etags */
// --- create XML containing ETags in an IOBuf for "close" request
IOBuf* xml = NULL;
if (n_to_1) {
// for N:1, gather all Etags at Rank0
MPI_Datatype mpi_size_t;
if (sizeof(size_t) == sizeof(int))
mpi_size_t = MPI_INT;
else if (sizeof(size_t) == sizeof(long))
mpi_size_t = MPI_LONG;
else
mpi_size_t = MPI_LONG_LONG;
// Everybody should have the same number of ETags (?)
size_t etag_count_max = 0; /* highest number on any proc */
MPI_Allreduce(&etags_per_rank, &etag_count_max,
1, mpi_size_t, MPI_MAX, param->testComm);
if (etags_per_rank != etag_count_max) {
printf("Rank %d: etag count mismatch: max:%d, mine:%d\n",
rank, etag_count_max, etags_per_rank);
MPI_Abort(param->testComm, 1);
}
// collect ETag data at Rank0
aws_iobuf_realloc(param->etags); /* force single contiguous buffer */
char* etag_data = param->etags->first->buf; /* per-rank data, contiguous */
if (rank == 0) {
char* etag_ptr;
int i;
int j;
int rnk;
char* etag_vec = (char*)malloc((param->numTasks * etag_data_size) +1);
if (! etag_vec) {
fprintf(stderr, "rank 0 failed to malloc %d bytes\n",
param->numTasks * etag_data_size);
MPI_Abort(param->testComm, 1);
}
MPI_Gather(etag_data, etag_data_size, MPI_BYTE,
etag_vec, etag_data_size, MPI_BYTE, 0, MPI_COMM_WORLD);
// --- debugging: show the gathered etag data
// (This shows the raw concatenated etag-data from each node.)
if (param->verbose >= VERBOSE_4) {
printf("rank 0: gathered %d etags from all ranks:\n", etags_per_rank);
etag_ptr=etag_vec;
for (rnk=0; rnk<param->numTasks; ++rnk) {
printf("\t[%d]: '", rnk);
int ii;
for (ii=0; ii<etag_data_size; ++ii) /* NOT null-terminated! */
printf("%c", etag_ptr[ii]);
printf("'\n");
etag_ptr += etag_data_size;
}
}
// add XML for *all* the parts. The XML must be ordered by
// part-number. Each rank wrote <etags_per_rank> parts,
// locally. At rank0, the etags for each rank are now
// stored as a continguous block of text, with the blocks
// stored in rank order in etag_vec. In other words, our
// internal rep at rank 0 matches the "segmented" format.
// From this, we must select etags in an order matching how
// they appear in the actual object, and give sequential
// part-numbers to the resulting sequence.
//
// That ordering of parts in the actual written object
// varies according to whether we wrote in the "segmented"
// or "strided" format.
//
// supposing N ranks, and P parts per rank:
//
// segmented:
//
// all parts for a given rank are consecutive.
// rank r writes these parts:
//
// rP, rP+1, ... (r+1)P -1
//
// i.e. rank0 writes parts 0,1,2,3 ... P-1
//
//
// strided:
//
// rank r writes every P-th part, starting with r.
//
// r, P+r, ... (P-1)P + r
//
// i.e. rank0 writes parts 0,P,2P,3P ... (P-1)P
//
//
// NOTE: If we knew ahead of time how many parts each rank was
// going to write, we could assign part-number ranges, per
// rank, and then have nice locality here.
//
// Alternatively, we could have everyone format their own
// XML text and send that, instead of just the tags. This
// would increase the amount of data being sent, but would
// reduce the work for rank0 to format everything.
size_t i_max; // outer-loop
size_t j_max; // inner loop
size_t start_multiplier; // initial offset in collected data
size_t stride; // in etag_vec
if (segmented) { // segmented
i_max = param->numTasks;
j_max = etags_per_rank;
start_multiplier = etag_data_size; /* one rank's-worth of Etag data */
stride = ETAG_SIZE; /* one ETag */
}
else { // strided
i_max = etags_per_rank;
j_max = param->numTasks;
start_multiplier = ETAG_SIZE; /* one ETag */
stride = etag_data_size; /* one rank's-worth of Etag data */
}
xml = aws_iobuf_new();
aws_iobuf_growth_size(xml, 1024 * 8);
// write XML header ...
aws_iobuf_append_str(xml, "<CompleteMultipartUpload>\n");
int part = 0;
for (i=0; i<i_max; ++i) {
etag_ptr=etag_vec + (i * start_multiplier);
for (j=0; j<j_max; ++j) {
// etags were saved as contiguous text. Extract the next one.
char etag[ETAG_SIZE +1];
memcpy(etag, etag_ptr, ETAG_SIZE);
etag[ETAG_SIZE] = 0;
// write XML for next part, with Etag ...
snprintf(buff, BUFF_SIZE,
" <Part>\n"
" <PartNumber>%d</PartNumber>\n"
" <ETag>%s</ETag>\n"
" </Part>\n",
part, etag);
aws_iobuf_append_str(xml, buff);
etag_ptr += stride;
++ part;
}
}
// write XML tail ...
aws_iobuf_append_str(xml, "</CompleteMultipartUpload>\n");
}
else {
MPI_Gather(etag_data, etag_data_size, MPI_BYTE,
NULL, etag_data_size, MPI_BYTE, 0, MPI_COMM_WORLD);
}
}
else { /* N:N */
xml = aws_iobuf_new();
aws_iobuf_growth_size(xml, 1024 * 8);
// write XML header ...
aws_iobuf_append_str(xml, "<CompleteMultipartUpload>\n");
// all parts of our object were written from this rank.
char etag[ETAG_SIZE +1];
int part = 0;
int i;
for (i=0; i<etags_per_rank; ++i) {
// TBD: Instead of reading into etag, then sprintf'ing, then
// copying into xml, we could just read directly into xml
int sz = aws_iobuf_get_raw(param->etags, etag, ETAG_SIZE);
if (sz != ETAG_SIZE) {
snprintf(buff, BUFF_SIZE,
"Read of ETag %d had length %d (not %d)\n",
rank, i, sz, ETAG_SIZE);
ERR_SIMPLE(buff);
}
etag[ETAG_SIZE] = 0;
// write XML for next part, with Etag ...
snprintf(buff, BUFF_SIZE,
" <Part>\n"
" <PartNumber>%d</PartNumber>\n"
" <ETag>%s</ETag>\n"
" </Part>\n",
part, etag);
aws_iobuf_append_str(xml, buff);
++ part;
}
// write XML tail ...
aws_iobuf_append_str(xml, "</CompleteMultipartUpload>\n");
}
// send request to finalize MPU
if (n_to_n || (rank == 0)) {
// DEBUGGING: show the XML we constructed
if (param->verbose >= VERBOSE_3)
debug_iobuf(xml, 1, 1);
// --- POST our XML to the server.
snprintf(buff, BUFF_SIZE,
"%s?uploadId=%s",
fname, param->UploadId);
AWS4C_CHECK ( s3_post(xml, buff) );
AWS4C_CHECK_OK( xml );
aws_iobuf_free(xml);
}
// everybody reset MPU info. Allows another MPU, and frees memory.
s3_MPU_reset(param);
// Everybody meetup, so non-zero ranks won't go trying to stat the
// N:1 file until rank0 has finished the S3 multi-part finalize.
// The object will not appear to exist, until then.
if (n_to_1)
MPI_CHECK(MPI_Barrier(param->testComm), "barrier error");
}
else {
// No finalization is needed, when using EMC's byte-range writing
// support. However, we do need to make sure everyone has
// finished writing, before anyone starts reading.
if (n_to_1) {
MPI_CHECK(MPI_Barrier(param->testComm), "barrier error");
if (param->verbose >= VERBOSE_2)
printf("rank %d: passed barrier\n", rank);
}
}
// After writing, reset the CURL connection, so that caches won't be
// used for reads.
aws_reset_connection();
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Close\n");
}
}
