refs #5663 Merge #5663 onto main

git-svn-id: file:///svn/toku/tokudb@51238 c7de825b-a66e-492c-adef-691d508d4ae1

ft/dbufio.cc

@@ -11,6 +11,9 @@
 #include <unistd.h>
 #include "memory.h"
 #include <string.h>
+#include "ftloader-internal.h"
+#include "ft-internal.h"
+#include "ft.h"
 
 struct dbufio_file {
     // i/o thread owns these
@@ -18,7 +21,7 @@ struct dbufio_file {
 
     // consumers own these
     size_t offset_in_buf;
-    toku_off_t  offset_in_file;
+    toku_off_t  offset_in_uncompressed_file;
 
     // need the mutex to modify these
     struct dbufio_file *next;
@@ -49,6 +52,7 @@ struct dbufio_fileset {
     size_t bufsize; // the bufsize is the constant (the same for all buffers).
 
     bool panic;
+    bool compressed;
     int  panic_errno;
     toku_pthread_t iothread;
 };
@@ -75,6 +79,162 @@ static bool paniced (DBUFIO_FILESET bfs) {
     return bfs->panic;
 }
 
+static ssize_t dbf_read_some_compressed(struct dbufio_file *dbf, char *buf, size_t bufsize) {
+    ssize_t ret;
+    invariant(bufsize >= MAX_UNCOMPRESSED_BUF);
+    unsigned char *raw_block = NULL;
+
+    // deserialize the sub block header
+
+    // total_size
+    // num_sub_blocks
+    // compressed_size,uncompressed_size,xsum (repeated num_sub_blocks times)
+    ssize_t readcode;
+    const uint32_t header_size = sizeof(uint32_t);
+    char header[header_size];
+
+    readcode = toku_os_read(dbf->fd, &header, header_size);
+    if (readcode < 0) {
+        ret = -1;
+        goto exit;
+    }
+    if (readcode == 0) {
+        ret = 0;
+        goto exit;
+    }
+    if (readcode < header_size) {
+        errno = TOKUDB_NO_DATA;
+        ret = -1;
+        goto exit;
+    }
+    uint32_t total_size;
+    {
+        uint32_t *p = (uint32_t *) &header[0];
+        total_size = toku_dtoh32(p[0]);
+    }
+    if (total_size == 0 || total_size > (1<<30)) {
+        errno = toku_db_badformat(); 
+        ret = -1;
+        goto exit;
+    }
+
+    //Cannot use XMALLOC
+    MALLOC_N(total_size, raw_block);
+    if (raw_block == nullptr) {
+        errno = ENOMEM;
+        ret = -1;
+        goto exit;
+    }
+    readcode = toku_os_read(dbf->fd, raw_block, total_size);
+    if (readcode < 0) {
+        ret = -1;
+        goto exit;
+    }
+    if (readcode < total_size) {
+        errno = TOKUDB_NO_DATA;
+        ret = -1;
+        goto exit;
+    }
+
+    struct sub_block sub_block[max_sub_blocks];
+    uint32_t *sub_block_header;
+    sub_block_header = (uint32_t *) &raw_block[0];
+    int32_t n_sub_blocks;
+    n_sub_blocks = toku_dtoh32(sub_block_header[0]);
+    sub_block_header++;
+    size_t size_subblock_header;
+    size_subblock_header = sub_block_header_size(n_sub_blocks);
+    if (n_sub_blocks == 0 || n_sub_blocks > max_sub_blocks || size_subblock_header > total_size) {
+        errno = toku_db_badformat(); 
+        ret = -1;
+        goto exit;
+    }
+    for (int i = 0; i < n_sub_blocks; i++) {
+        sub_block_init(&sub_block[i]);
+        sub_block[i].compressed_size = toku_dtoh32(sub_block_header[0]);
+        sub_block[i].uncompressed_size = toku_dtoh32(sub_block_header[1]);
+        sub_block[i].xsum = toku_dtoh32(sub_block_header[2]);
+        sub_block_header += 3;
+    }
+
+    // verify sub block sizes
+    size_t total_compressed_size;
+    total_compressed_size = 0;
+    for (int i = 0; i < n_sub_blocks; i++) {
+        uint32_t compressed_size = sub_block[i].compressed_size;
+        if (compressed_size<=0   || compressed_size>(1<<30)) { 
+            errno = toku_db_badformat(); 
+            ret = -1;
+            goto exit;
+        }
+
+        uint32_t uncompressed_size = sub_block[i].uncompressed_size;
+        if (uncompressed_size<=0 || uncompressed_size>(1<<30)) { 
+            errno = toku_db_badformat(); 
+            ret = -1;
+            goto exit;
+        }
+        total_compressed_size += compressed_size;
+    }
+    if (total_size != total_compressed_size + size_subblock_header) {
+        errno = toku_db_badformat(); 
+        ret = -1;
+        goto exit;
+    }
+
+    // sum up the uncompressed size of the sub blocks
+    size_t uncompressed_size;
+    uncompressed_size = get_sum_uncompressed_size(n_sub_blocks, sub_block);
+    if (uncompressed_size > bufsize || uncompressed_size > MAX_UNCOMPRESSED_BUF) {
+        errno = toku_db_badformat(); 
+        ret = -1;
+        goto exit;
+    }
+
+    unsigned char *uncompressed_data;
+    uncompressed_data = (unsigned char *)buf;
+
+    // point at the start of the compressed data (past the node header, the sub block header, and the header checksum)
+    unsigned char *compressed_data;
+    compressed_data = raw_block + size_subblock_header;
+
+    // decompress all the compressed sub blocks into the uncompressed buffer
+    {
+        int r;
+        r = decompress_all_sub_blocks(n_sub_blocks, sub_block, compressed_data, uncompressed_data, get_num_cores(), get_ft_pool());
+        if (r != 0) {
+            fprintf(stderr, "%s:%d loader failed %d at %p size %" PRIu32"\n", __FUNCTION__, __LINE__, r, raw_block, total_size);
+            dump_bad_block(raw_block, total_size);
+            errno = r;
+            ret = -1;
+            goto exit;
+        }
+    }
+    ret = uncompressed_size;
+exit:
+    if (raw_block) {
+        toku_free(raw_block);
+    }
+    return ret;
+}
+
+static ssize_t dbf_read_compressed(struct dbufio_file *dbf, char *buf, size_t bufsize) {
+    invariant(bufsize >= MAX_UNCOMPRESSED_BUF);
+    size_t count = 0;
+
+    while (count + MAX_UNCOMPRESSED_BUF <= bufsize) {
+        ssize_t readcode = dbf_read_some_compressed(dbf, buf + count, bufsize - count);
+        if (readcode < 0) {
+            return readcode;
+        }
+        count += readcode;
+        if (readcode == 0) {
+            break;
+        }
+    }
+    return count;
+}
+
 static void* io_thread (void *v)
 // The dbuf_thread does all the asynchronous I/O.
 {
@@ -118,7 +278,13 @@ static void* io_thread (void *v)
             toku_mutex_unlock(&bfs->mutex);
 	    //printf("%s:%d Doing read fd=%d\n", __FILE__, __LINE__, dbf->fd);
 	    {
-		ssize_t readcode = toku_os_read(dbf->fd, dbf->buf[1], bfs->bufsize);
+		ssize_t readcode;
+                if (bfs->compressed) {
+                    readcode = dbf_read_compressed(dbf, dbf->buf[1], bfs->bufsize);
+                }
+                else {
+                    readcode = toku_os_read(dbf->fd, dbf->buf[1], bfs->bufsize);
+                }
 		//printf("%s:%d readcode=%ld\n", __FILE__, __LINE__, readcode);
 		if (readcode==-1) {
 		    // a real error.  Save the real error.
@@ -159,11 +325,14 @@ static void* io_thread (void *v)
     }
 }
 
-int create_dbufio_fileset (DBUFIO_FILESET *bfsp, int N, int fds[/*N*/], size_t bufsize) {
+int create_dbufio_fileset (DBUFIO_FILESET *bfsp, int N, int fds[/*N*/], size_t bufsize, bool compressed) {
     //printf("%s:%d here\n", __FILE__, __LINE__);
     int result = 0;
     DBUFIO_FILESET CALLOC(bfs);
     if (bfs==0) { result = get_error_errno(); }
+
+    bfs->compressed = compressed;
+
     bool mutex_inited = false, cond_inited = false;
     if (result==0) {
 	CALLOC_N(N, bfs->files);
@@ -190,7 +359,7 @@ int create_dbufio_fileset (DBUFIO_FILESET *bfsp, int N, int fds[/*N*/], size_t b
 	for (int i=0; i<N; i++) {
 	    bfs->files[i].fd = fds[i];
 	    bfs->files[i].offset_in_buf = 0;
-	    bfs->files[i].offset_in_file = 0;
+	    bfs->files[i].offset_in_uncompressed_file = 0;
 	    bfs->files[i].next = NULL;
 	    bfs->files[i].second_buf_ready = false;
 	    for (int j=0; j<2; j++) {
@@ -202,8 +371,14 @@ int create_dbufio_fileset (DBUFIO_FILESET *bfsp, int N, int fds[/*N*/], size_t b
 		bfs->files[i].error_code[j] = 0;
 	    }
 	    bfs->files[i].io_done = false;
+            ssize_t r;
+            if (bfs->compressed) {
+                r = dbf_read_compressed(&bfs->files[i], bfs->files[i].buf[0], bufsize);
+            } else {
+                //TODO: 5663 If compressed need to read differently
+		r = toku_os_read(bfs->files[i].fd, bfs->files[i].buf[0], bufsize);
+            }
             {
-		ssize_t r = toku_os_read(bfs->files[i].fd, bfs->files[i].buf[0], bufsize);
 		if (r<0) {
 		    result=get_error_errno();
 		    break;
@@ -297,7 +472,7 @@ int dbufio_fileset_read (DBUFIO_FILESET bfs, int filenum, void *buf_v, size_t co
 	// Enough data is present to do it all now
 	memcpy(buf, dbf->buf[0]+dbf->offset_in_buf, count);
 	dbf->offset_in_buf += count;
-	dbf->offset_in_file += count;
+	dbf->offset_in_uncompressed_file += count;
 	*n_read = count;
 	return 0;
     } else if (dbf->n_in_buf[0] > dbf->offset_in_buf) {
@@ -306,7 +481,7 @@ int dbufio_fileset_read (DBUFIO_FILESET bfs, int filenum, void *buf_v, size_t co
 	assert(dbf->offset_in_buf + this_count <= bfs->bufsize);
 	memcpy(buf, dbf->buf[0]+dbf->offset_in_buf, this_count);
 	dbf->offset_in_buf += this_count;
-	dbf->offset_in_file += this_count;
+	dbf->offset_in_uncompressed_file += this_count;
 	size_t sub_n_read;
 	int r = dbufio_fileset_read(bfs, filenum, buf+this_count, count-this_count, &sub_n_read);
 	if (r==0) {

ft/dbufio.h

@@ -14,7 +14,7 @@
 /* An implementation would typically use a separate thread or asynchronous I/O to fetch ahead data for each file.  The system will typically fill two buffers of size M for each file.  One buffer is being read out of using dbuf_read(), and the other buffer is either empty (waiting on the asynchronous I/O to start), being filled in by the asynchronous I/O mechanism, or is waiting for the caller to read data from it. */
 typedef struct dbufio_fileset *DBUFIO_FILESET; 
 
-int create_dbufio_fileset (DBUFIO_FILESET *bfsp, int N, int fds[/*N*/], size_t bufsize);
+int create_dbufio_fileset (DBUFIO_FILESET *bfsp, int N, int fds[/*N*/], size_t bufsize, bool compressed);
 
 int destroy_dbufio_fileset(DBUFIO_FILESET);
 

ft/ft.h

@@ -110,4 +110,7 @@ void toku_ft_set_blackhole(FT_HANDLE ft_handle);
 //         The difference between the two is MVCC garbage.
 void toku_ft_get_garbage(FT ft, uint64_t *total_space, uint64_t *used_space);
 
+int get_num_cores(void);
+struct toku_thread_pool *get_ft_pool(void);
+void dump_bad_block(unsigned char *vp, uint64_t size);
 #endif

ft/ft_node-serialize.cc

@@ -10,6 +10,7 @@
 #include <portability/toku_atomic.h>
 #include <util/sort.h>
 #include <util/threadpool.h>
+#include "ft.h"
 
 static FT_UPGRADE_STATUS_S ft_upgrade_status;
 
@@ -57,6 +58,14 @@ static inline void do_toku_trace(const char *cp, int len) {
 static int num_cores = 0; // cache the number of cores for the parallelization
 static struct toku_thread_pool *ft_pool = NULL;
 
+int get_num_cores(void) {
+    return num_cores;
+}
+
+struct toku_thread_pool *get_ft_pool(void) {
+    return ft_pool;
+}
+
 void 
 toku_ft_serialize_layer_init(void) {
     num_cores = toku_os_get_number_active_processors();
@@ -980,7 +989,7 @@ deserialize_child_buffer(NONLEAF_CHILDINFO bnc, struct rbuf *rbuf,
 
 // dump a buffer to stderr
 // no locking around this for now
-static void
+void
 dump_bad_block(unsigned char *vp, uint64_t size) {
     const uint64_t linesize = 64;
     uint64_t n = size / linesize;
@@ -1181,14 +1190,7 @@ read_and_decompress_sub_block(struct rbuf *rb, struct sub_block *sb)
         goto exit;
     }
 
-    sb->uncompressed_ptr = toku_xmalloc(sb->uncompressed_size);
-
-    toku_decompress(
-        (Bytef *) sb->uncompressed_ptr,
-        sb->uncompressed_size,
-        (Bytef *) sb->compressed_ptr,
-        sb->compressed_size
-        );
+    just_decompress_sub_block(sb);
 exit:
     return r;
 }

ft/ftloader-internal.h

@@ -12,6 +12,19 @@
 #include <toku_pthread.h>
 #include "dbufio.h"
 
+enum { EXTRACTOR_QUEUE_DEPTH = 2,
+       FILE_BUFFER_SIZE  = 1<<24,
+       MIN_ROWSET_MEMORY = 1<<23,
+       MIN_MERGE_FANIN   = 2,
+       FRACTAL_WRITER_QUEUE_DEPTH = 3,
+       FRACTAL_WRITER_ROWSETS = FRACTAL_WRITER_QUEUE_DEPTH + 2,
+       DBUFIO_DEPTH = 2,
+       TARGET_MERGE_BUF_SIZE = 1<<24, // we'd like the merge buffer to be this big.
+       MIN_MERGE_BUF_SIZE = 1<<20, // always use at least this much
+       MAX_UNCOMPRESSED_BUF = MIN_MERGE_BUF_SIZE
+};
+
+
 /* These functions are exported to allow the tests to compile. */
 
 /* These structures maintain a collection of all the open temporary files used by the loader. */
@@ -56,7 +69,7 @@ int init_rowset (struct rowset *rows, uint64_t memory_budget);
 void destroy_rowset (struct rowset *rows);
 int add_row (struct rowset *rows, DBT *key, DBT *val);
 
-int loader_write_row(DBT *key, DBT *val, FIDX data, FILE*, uint64_t *dataoff, FTLOADER bl);
+int loader_write_row(DBT *key, DBT *val, FIDX data, FILE*, uint64_t *dataoff, struct wbuf *wb, FTLOADER bl);
 int loader_read_row (FILE *f, DBT *key, DBT *val);
 
 struct merge_fileset {
@@ -146,6 +159,7 @@ struct ft_loader_s {
 
     CACHETABLE cachetable;
     bool did_reserve_memory;
+    bool compress_intermediates;
     uint64_t   reserved_memory; // how much memory are we allowed to use?
 
     /* To make it easier to recover from errors, we don't use FILE*, instead we use an index into the file_infos. */
@@ -243,7 +257,8 @@ int toku_ft_loader_internal_init (/* out */ FTLOADER *blp,
 				   const char *temp_file_template,
 				   LSN load_lsn,
                                    TOKUTXN txn,
-                                   bool reserve_memory);
+                                   bool reserve_memory,
+                                   bool compress_intermediates);
 
 void toku_ft_loader_internal_destroy (FTLOADER bl, bool is_error);
 

ft/ftloader.h

@@ -26,7 +26,8 @@ int toku_ft_loader_open (FTLOADER *bl,
 			  const char *temp_file_template,
                           LSN load_lsn,
                           TOKUTXN txn,
-                          bool reserve_memory);
+                          bool reserve_memory,
+                          bool compress_intermediates);
 
 int toku_ft_loader_put (FTLOADER bl, DBT *key, DBT *val);
 

ft/sub_block.h

@@ -44,7 +44,7 @@ sub_block_header_size(int n_sub_blocks);
 void
 set_compressed_size_bound(struct sub_block *se, enum toku_compression_method method);
 
-// get the sum of the sub block compressed sizes 
+// get the sum of the sub block compressed bound sizes 
 size_t 
 get_sum_compressed_size_bound(int n_sub_blocks, struct sub_block sub_block[], enum toku_compression_method method);
 

ft/tests/dbufio-test-destroy.cc

@@ -42,7 +42,7 @@ static void test1 (size_t chars_per_file, size_t UU(bytes_per_read)) {
     }
     DBUFIO_FILESET bfs;
     {
-	int r = create_dbufio_fileset(&bfs, N, fds, M);
+	int r = create_dbufio_fileset(&bfs, N, fds, M, false);
 	assert(r==0);
     }
 

ft/tests/dbufio-test.cc

@@ -41,7 +41,7 @@ static void test1 (size_t chars_per_file, size_t bytes_per_read) {
     }
     DBUFIO_FILESET bfs;
     {
-	int r = create_dbufio_fileset(&bfs, N, fds, M);
+	int r = create_dbufio_fileset(&bfs, N, fds, M, false);
 	assert(r==0);
     }
     while (n_live>0) {

ft/tests/ftloader-test-bad-generate.cc

@@ -87,7 +87,7 @@ static void test_extractor(int nrows, int nrowsets, bool expect_fail) {
     }
 
     FTLOADER loader;
-    r = toku_ft_loader_open(&loader, NULL, generate, NULL, N, brts, dbs, fnames, compares, "tempXXXXXX", ZERO_LSN, TXNID_NONE, true);
+    r = toku_ft_loader_open(&loader, NULL, generate, NULL, N, brts, dbs, fnames, compares, "tempXXXXXX", ZERO_LSN, TXNID_NONE, true, false);
     assert(r == 0);
 
     struct rowset *rowset[nrowsets];

ft/tests/ftloader-test-extractor-errors.cc

@@ -99,7 +99,7 @@ static void test_extractor(int nrows, int nrowsets, bool expect_fail, const char
     sprintf(temp, "%s/%s", testdir, "tempXXXXXX");
 
     FTLOADER loader;
-    r = toku_ft_loader_open(&loader, NULL, generate, NULL, N, brts, dbs, fnames, compares, "tempXXXXXX", ZERO_LSN, TXNID_NONE, true);
+    r = toku_ft_loader_open(&loader, NULL, generate, NULL, N, brts, dbs, fnames, compares, "tempXXXXXX", ZERO_LSN, TXNID_NONE, true, false);
     assert(r == 0);
 
     struct rowset *rowset[nrowsets];

ft/tests/ftloader-test-extractor.cc

@@ -319,7 +319,7 @@ static void test_extractor(int nrows, int nrowsets, const char *testdir) {
     sprintf(temp, "%s/%s", testdir, "tempXXXXXX");
 
     FTLOADER loader;
-    r = toku_ft_loader_open(&loader, NULL, generate, NULL, N, brts, dbs, fnames, compares, temp, ZERO_LSN, TXNID_NONE, true);
+    r = toku_ft_loader_open(&loader, NULL, generate, NULL, N, brts, dbs, fnames, compares, temp, ZERO_LSN, TXNID_NONE, true, false);
     assert(r == 0);
 
     struct rowset *rowset[nrowsets];

ft/tests/ftloader-test-merge-files-dbufio.cc

@@ -326,7 +326,7 @@ static void test (const char *directory, bool is_error) {
 					       bt_compare_functions,
 					       "tempxxxxxx",
 					       *lsnp,
-                                               TXNID_NONE, true);
+                                               TXNID_NONE, true, false);
 	assert(r==0);
     }
 
@@ -340,7 +340,7 @@ static void test (const char *directory, bool is_error) {
     { int r = queue_create(&q, 1000); assert(r==0); }
     DBUFIO_FILESET bfs;
     const int MERGE_BUF_SIZE = 100000; // bigger than 64K so that we will trigger malloc issues.
-    { int r = create_dbufio_fileset(&bfs, N_SOURCES, fds, MERGE_BUF_SIZE);  assert(r==0); }
+    { int r = create_dbufio_fileset(&bfs, N_SOURCES, fds, MERGE_BUF_SIZE, false);  assert(r==0); }
     FIDX *XMALLOC_N(N_SOURCES, src_fidxs);
     assert(bl->file_infos.n_files==0);
     bl->file_infos.n_files = N_SOURCES;

ft/tests/ftloader-test-open.cc

@@ -57,7 +57,7 @@ static void test_loader_open(int ndbs) {
     for (i = 0; ; i++) {
         set_my_malloc_trigger(i+1);
 
-        r = toku_ft_loader_open(&loader, NULL, NULL, NULL, ndbs, brts, dbs, fnames, compares, "", ZERO_LSN, TXNID_NONE, true);
+        r = toku_ft_loader_open(&loader, NULL, NULL, NULL, ndbs, brts, dbs, fnames, compares, "", ZERO_LSN, TXNID_NONE, true, false);
         if (r == 0)
             break;
     }

ft/tests/ftloader-test.cc

@@ -187,7 +187,7 @@ static void test_read_write_rows (char *tf_template) {
         toku_fill_dbt(&key, keystrings[i], strlen(keystrings[i]));
 	DBT val;
         toku_fill_dbt(&val, valstrings[i], strlen(valstrings[i]));
-	r = loader_write_row(&key, &val, file, toku_bl_fidx2file(&bl, file), &dataoff, &bl);
+	r = loader_write_row(&key, &val, file, toku_bl_fidx2file(&bl, file), &dataoff, nullptr, &bl);
 	CKERR(r);
 	actual_size+=key.size + val.size + 8;
     }

portability/file.cc

@@ -430,3 +430,8 @@ int toku_fsync_directory(const char *fname) {
     toku_free(dirname);
     return result;
 }
+
+FILE *toku_os_fmemopen(void *buf, size_t size, const char *mode) {
+    return fmemopen(buf, size, mode);
+}
+

src/loader.cc

@@ -174,6 +174,7 @@ toku_loader_create_loader(DB_ENV *env,
 
     DB_LOADER *loader = NULL;
     bool puts_allowed = !(loader_flags & LOADER_DISALLOW_PUTS);
+    bool compress_intermediates = (loader_flags & LOADER_COMPRESS_INTERMEDIATES) != 0;
     XCALLOC(loader);       // init to all zeroes (thus initializing the error_callback and poll_func)
     XCALLOC(loader->i);    // init to all zeroes (thus initializing all pointers to NULL)
 
@@ -252,7 +253,8 @@ toku_loader_create_loader(DB_ENV *env,
                                  loader->i->temp_file_template,
                                  load_lsn,
                                  ttxn,
-                                 puts_allowed);
+                                 puts_allowed,
+                                 compress_intermediates);
         if ( rval!=0 ) {
             toku_free(new_inames_in_env);
             toku_free(brts);

src/tests/CMakeLists.txt

@@ -546,14 +546,22 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
   add_test(ydb/loader-stress-test2.tdb loader-stress-test.tdb -r 5000 -s -e dir.loader-stress-test2.tdb)
   add_test(ydb/loader-stress-test3.tdb loader-stress-test.tdb -u -c -e dir.loader-stress-test3.tdb)
   add_test(ydb/loader-stress-test4.tdb loader-stress-test.tdb -r 10000000 -c -e dir.loader-stress-test4.tdb)
-  add_test(ydb/loader-stress-test5.tdb loader-stress-test.tdb -c -z -e dir.loader-stress-test5.tdb)
+  add_test(ydb/loader-stress-test0z.tdb loader-stress-test.tdb -c -e dir.loader-stress-test0z.tdb -z)
+  add_test(ydb/loader-stress-test1z.tdb loader-stress-test.tdb -c -p -e dir.loader-stress-test1z.tdb -z)
+  add_test(ydb/loader-stress-test2z.tdb loader-stress-test.tdb -r 5000 -s -e dir.loader-stress-test2z.tdb -z)
+  add_test(ydb/loader-stress-test3z.tdb loader-stress-test.tdb -u -c -e dir.loader-stress-test3z.tdb -z)
+  add_test(ydb/loader-stress-test4z.tdb loader-stress-test.tdb -r 10000000 -c -e dir.loader-stress-test4z.tdb -z)
   set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES
     dir.loader-stress-test0.tdb
     dir.loader-stress-test1.tdb
     dir.loader-stress-test2.tdb
     dir.loader-stress-test3.tdb
     dir.loader-stress-test4.tdb
-    dir.loader-stress-test5.tdb
+    dir.loader-stress-test0z.tdb
+    dir.loader-stress-test1z.tdb
+    dir.loader-stress-test2z.tdb
+    dir.loader-stress-test3z.tdb
+    dir.loader-stress-test4z.tdb
     )
 
   list(REMOVE_ITEM loader_tests loader-dup-test.loader)
@@ -563,6 +571,12 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
   add_test(ydb/loader-dup-test3.tdb loader-dup-test.tdb -d 1 -s -r 100 -e dir.loader-dup-test3.tdb)
   add_test(ydb/loader-dup-test4.tdb loader-dup-test.tdb -d 1 -s -r 1000 -e dir.loader-dup-test4.tdb)
   add_test(ydb/loader-dup-test5.tdb loader-dup-test.tdb -d 1 -s -r 1000 -E -e dir.loader-dup-test5.tdb)
+  add_test(ydb/loader-dup-test0z.tdb loader-dup-test.tdb -e dir.loader-dup-test0z.tdb -z)
+  add_test(ydb/loader-dup-test1z.tdb loader-dup-test.tdb -d 1 -r 500000 -e dir.loader-dup-test1z.tdb -z)
+  add_test(ydb/loader-dup-test2z.tdb loader-dup-test.tdb -d 1 -r 1000000 -e dir.loader-dup-test2z.tdb -z)
+  add_test(ydb/loader-dup-test3z.tdb loader-dup-test.tdb -d 1 -s -r 100 -e dir.loader-dup-test3z.tdb -z)
+  add_test(ydb/loader-dup-test4z.tdb loader-dup-test.tdb -d 1 -s -r 1000 -e dir.loader-dup-test4z.tdb -z)
+  add_test(ydb/loader-dup-test5z.tdb loader-dup-test.tdb -d 1 -s -r 1000 -E -e dir.loader-dup-test5z.tdb -z)
   set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES
     dir.loader-dup-test0.tdb
     dir.loader-dup-test1.tdb
@@ -570,6 +584,12 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
     dir.loader-dup-test3.tdb
     dir.loader-dup-test4.tdb
     dir.loader-dup-test5.tdb
+    dir.loader-dup-test0z.tdb
+    dir.loader-dup-test1z.tdb
+    dir.loader-dup-test2z.tdb
+    dir.loader-dup-test3z.tdb
+    dir.loader-dup-test4z.tdb
+    dir.loader-dup-test5z.tdb
     )
 
   ## as part of #4503, we took out test 1 and 3
@@ -578,6 +598,8 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
   #add_test(ydb/loader-cleanup-test1.tdb loader-cleanup-test.tdb -s -r 800 -p -e dir.loader-cleanup-test1.tdb)
   add_test(ydb/loader-cleanup-test2.tdb loader-cleanup-test.tdb -s -r 8000 -e dir.loader-cleanup-test2.tdb)
   #add_test(ydb/loader-cleanup-test3.tdb loader-cleanup-test.tdb -s -r 8000 -p -e dir.loader-cleanup-test3.tdb)
+  add_test(ydb/loader-cleanup-test0z.tdb loader-cleanup-test.tdb -s -r 800 -e dir.loader-cleanup-test0z.tdb -z)
+  add_test(ydb/loader-cleanup-test2z.tdb loader-cleanup-test.tdb -s -r 8000 -e dir.loader-cleanup-test2z.tdb -z)
   set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES
     dir.loader-cleanup-test0.tdb
     dir.loader-cleanup-test1.tdb
@@ -606,11 +628,15 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
     )
 
   declare_custom_tests(maxsize-for-loader.tdb)
-  add_test(ydb/maxsize-for-loader-A.tdb maxsize-for-loader.tdb -e dir.maxsize-for-loader-A.tdb -f)
-  add_test(ydb/maxsize-for-loader-B.tdb maxsize-for-loader.tdb -e dir.maxsize-for-loader-B.tdb)
+  add_test(ydb/maxsize-for-loader-A.tdb maxsize-for-loader.tdb -e dir.maxsize-for-loader-A.tdb -f -c)
+  add_test(ydb/maxsize-for-loader-B.tdb maxsize-for-loader.tdb -e dir.maxsize-for-loader-B.tdb -c)
+  add_test(ydb/maxsize-for-loader-Az.tdb maxsize-for-loader.tdb -e dir.maxsize-for-loader-Az.tdb -f -z -c)
+  add_test(ydb/maxsize-for-loader-Bz.tdb maxsize-for-loader.tdb -e dir.maxsize-for-loader-Bz.tdb -z -c)
   set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES
     dir.maxsize-for-loader-A.tdb
     dir.maxsize-for-loader-B.tdb
+    dir.maxsize-for-loader-Az.tdb
+    dir.maxsize-for-loader-Bz.tdb
     )
 
   declare_custom_tests(hotindexer-undo-do-test.tdb)
@@ -667,6 +693,8 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
     ydb/env-put-multiple.tdb
     ydb/filesize.tdb
     ydb/loader-cleanup-test0.tdb
+    ydb/loader-cleanup-test0z.tdb
+    ydb/loader-cleanup-test2z.tdb
     ydb/manyfiles.tdb
     ydb/recover-loader-test.abortrecover
     ydb/recovery_fileops_stress.tdb
@@ -712,6 +740,7 @@ if(BUILD_TESTING OR BUILD_SRC_TESTS)
     ydb/loader-stress-del.nop.loader
     ydb/loader-stress-del.p.loader
     ydb/loader-stress-del.comp.loader
+    ydb/loader-stress-test4z.tdb
     ydb/test3039.tdb
     ydb/test3529.tdb
     ydb/test_update_stress.tdb

toku_include/toku_portability.h

@@ -266,6 +266,7 @@ int toku_os_close(int fd);
 int toku_os_fclose(FILE * stream);
 ssize_t toku_os_read(int fd, void *buf, size_t count);
 ssize_t toku_os_pread(int fd, void *buf, size_t count, off_t offset);
+FILE *toku_os_fmemopen(void *buf, size_t size, const char *mode);
 
 // wrapper around fsync
 void toku_file_fsync_without_accounting(int fd);