*
* Copyright (C) 1998 Andrew Tridgell
* Copyright (C) 2002 Martin Pool
- * Copyright (C) 2004-2008 Wayne Davison
+ * Copyright (C) 2004-2020 Wayne Davison
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
*/
#include "rsync.h"
-#include "ifuncs.h"
+#include "inums.h"
#ifndef ENODATA
#define ENODATA EAGAIN
#endif
+/* We want all reads to be aligned on 1K boundaries. */
+#define ALIGN_BOUNDARY 1024
+/* How far past the boundary is an offset? */
+#define ALIGNED_OVERSHOOT(oft) ((oft) & (ALIGN_BOUNDARY-1))
+/* Round up a length to the next boundary */
+#define ALIGNED_LENGTH(len) ((((len) - 1) | (ALIGN_BOUNDARY-1)) + 1)
+
extern int sparse_files;
-static char last_byte;
-static size_t sparse_seek = 0;
+OFF_T preallocated_len = 0;
+
+static OFF_T sparse_seek = 0;
+static OFF_T sparse_past_write = 0;
-int sparse_end(int f)
+int sparse_end(int f, OFF_T size)
{
int ret;
+ sparse_past_write = 0;
+
if (!sparse_seek)
return 0;
- do_lseek(f, sparse_seek-1, SEEK_CUR);
- sparse_seek = 0;
+#ifdef HAVE_FTRUNCATE
+ ret = do_ftruncate(f, size);
+#else
+ if (do_lseek(f, sparse_seek-1, SEEK_CUR) != size-1)
+ ret = -1;
+ else {
+ do {
+ ret = write(f, "", 1);
+ } while (ret < 0 && errno == EINTR);
+
+ ret = ret <= 0 ? -1 : 0;
+ }
+#endif
- do {
- ret = write(f, "", 1);
- } while (ret < 0 && errno == EINTR);
+ sparse_seek = 0;
- return ret <= 0 ? -1 : 0;
+ return ret;
}
-
-static int write_sparse(int f, char *buf, size_t len)
+/* Note that the offset is just the caller letting us know where
+ * the current file position is in the file. The use_seek arg tells
+ * us that we should seek over matching data instead of writing it. */
+static int write_sparse(int f, int use_seek, OFF_T offset, const char *buf, int len)
{
- size_t l1 = 0, l2 = 0;
+ int l1 = 0, l2 = 0;
int ret;
for (l1 = 0; l1 < len && buf[l1] == 0; l1++) {}
for (l2 = 0; l2 < len-l1 && buf[len-(l2+1)] == 0; l2++) {}
- /* XXX Riddle me this: why does this function SLOW DOWN when I
- * remove the following (unneeded) line?? Core Duo weirdness? */
- last_byte = buf[len-1];
-
sparse_seek += l1;
if (l1 == len)
return len;
- if (sparse_seek)
- do_lseek(f, sparse_seek, SEEK_CUR);
+ if (sparse_seek) {
+ if (sparse_past_write >= preallocated_len) {
+ if (do_lseek(f, sparse_seek, SEEK_CUR) < 0)
+ return -1;
+ } else if (do_punch_hole(f, sparse_past_write, sparse_seek) < 0) {
+ sparse_seek = 0;
+ return -1;
+ }
+ }
sparse_seek = l2;
+ sparse_past_write = offset + len - l2;
+
+ if (use_seek) {
+ /* The in-place data already matches. */
+ if (do_lseek(f, len - (l1+l2), SEEK_CUR) < 0)
+ return -1;
+ return len;
+ }
while ((ret = write(f, buf + l1, len - (l1+l2))) <= 0) {
if (ret < 0 && errno == EINTR)
continue;
+ sparse_seek = 0;
return ret;
}
- if (ret != (int)(len - (l1+l2)))
+ if (ret != (int)(len - (l1+l2))) {
+ sparse_seek = 0;
return l1+ret;
+ }
return len;
}
-
static char *wf_writeBuf;
static size_t wf_writeBufSize;
static size_t wf_writeBufCnt;
return ret;
}
-
-/*
- * write_file does not allow incomplete writes. It loops internally
- * until len bytes are written or errno is set.
- */
-int write_file(int f,char *buf,size_t len)
+/* write_file does not allow incomplete writes. It loops internally
+ * until len bytes are written or errno is set. Note that use_seek and
+ * offset are only used in sparse processing (see write_sparse()). */
+int write_file(int f, int use_seek, OFF_T offset, const char *buf, int len)
{
int ret = 0;
int r1;
if (sparse_files > 0) {
int len1 = MIN(len, SPARSE_WRITE_SIZE);
- r1 = write_sparse(f, buf, len1);
+ r1 = write_sparse(f, use_seek, offset, buf, len1);
+ offset += r1;
} else {
if (!wf_writeBuf) {
wf_writeBufSize = WRITE_SIZE * 8;
wf_writeBufCnt = 0;
wf_writeBuf = new_array(char, wf_writeBufSize);
- if (!wf_writeBuf)
- out_of_memory("write_file");
}
- r1 = MIN(len, wf_writeBufSize - wf_writeBufCnt);
+ r1 = (int)MIN((size_t)len, wf_writeBufSize - wf_writeBufCnt);
if (r1) {
memcpy(wf_writeBuf + wf_writeBufCnt, buf, r1);
wf_writeBufCnt += r1;
return ret;
}
+/* An in-place update found identical data at an identical location. We either
+ * just seek past it, or (for an in-place sparse update), we give the data to
+ * the sparse processor with the use_seek flag set. */
+int skip_matched(int fd, OFF_T offset, const char *buf, int len)
+{
+ OFF_T pos;
+
+ if (sparse_files > 0) {
+ if (write_file(fd, 1, offset, buf, len) != len)
+ return -1;
+ return 0;
+ }
+
+ if (flush_write_file(fd) < 0)
+ return -1;
+
+ if ((pos = do_lseek(fd, len, SEEK_CUR)) != offset + len) {
+ rsyserr(FERROR_XFER, errno, "lseek returned %s, not %s",
+ big_num(pos), big_num(offset));
+ return -1;
+ }
+
+ return 0;
+}
/* This provides functionality somewhat similar to mmap() but using read().
* It gives sliding window access to a file. mmap() is not used because of
* the possibility of another program (such as a mailer) truncating the
* file thus giving us a SIGBUS. */
-struct map_struct *map_file(int fd, OFF_T len, int32 read_size,
- int32 blk_size)
+struct map_struct *map_file(int fd, OFF_T len, int32 read_size, int32 blk_size)
{
struct map_struct *map;
- if (!(map = new0(struct map_struct)))
- out_of_memory("map_file");
+ map = new0(struct map_struct);
if (blk_size && (read_size % blk_size))
read_size += blk_size - (read_size % blk_size);
map->fd = fd;
map->file_size = len;
- map->def_window_size = read_size;
+ map->def_window_size = ALIGNED_LENGTH(read_size);
return map;
}
/* slide the read window in the file */
char *map_ptr(struct map_struct *map, OFF_T offset, int32 len)
{
- int32 nread;
OFF_T window_start, read_start;
- int32 window_size, read_size, read_offset;
+ int32 window_size, read_size, read_offset, align_fudge;
if (len == 0)
return NULL;
return map->p + (offset - map->p_offset);
/* nope, we are going to have to do a read. Work out our desired window */
- window_start = offset;
+ align_fudge = (int32)ALIGNED_OVERSHOOT(offset);
+ window_start = offset - align_fudge;
window_size = map->def_window_size;
if (window_start + window_size > map->file_size)
window_size = (int32)(map->file_size - window_start);
- if (len > window_size)
- window_size = len;
+ if (window_size < len + align_fudge)
+ window_size = ALIGNED_LENGTH(len + align_fudge);
/* make sure we have allocated enough memory for the window */
if (window_size > map->p_size) {
map->p = realloc_array(map->p, char, window_size);
- if (!map->p)
- out_of_memory("map_ptr");
map->p_size = window_size;
}
- /* Now try to avoid re-reading any bytes by reusing any bytes
- * from the previous buffer. */
- if (window_start >= map->p_offset &&
- window_start < map->p_offset + map->p_len &&
- window_start + window_size >= map->p_offset + map->p_len) {
+ /* Now try to avoid re-reading any bytes by reusing any bytes from the previous buffer. */
+ if (window_start >= map->p_offset && window_start < map->p_offset + map->p_len
+ && window_start + window_size >= map->p_offset + map->p_len) {
read_start = map->p_offset + map->p_len;
read_offset = (int32)(read_start - window_start);
read_size = window_size - read_offset;
map->p_len = window_size;
while (read_size > 0) {
- nread = read(map->fd, map->p + read_offset, read_size);
+ int32 nread = read(map->fd, map->p + read_offset, read_size);
if (nread <= 0) {
if (!map->status)
map->status = nread ? errno : ENODATA;
read_size -= nread;
}
- return map->p;
+ return map->p + align_fudge;
}
-
int unmap_file(struct map_struct *map)
{
int ret;
map->p = NULL;
}
ret = map->status;
- memset(map, 0, sizeof map[0]);
+#if 0 /* I don't think we really need this. */
+ force_memzero(map, sizeof map[0]);
+#endif
free(map);
return ret;