1 From: Tobi Oetiker tobi{at}oetiker.ch
4 I am using rsync for hard-link backup. I found that there is a
5 major problem with frequent backup filling up the file system cache
6 with all the data from the files being backed up. The effect is
7 that all the other 'sensible' data in the cache gets thrown out in
8 the process. This is rather unfortunate as the performance of the
9 system becomes very bad after running rsync.
11 Some research showed, that
13 posix_fadvise64(fd, 0, 0,POSIX_FADV_DONTNEED);
15 would tell the OS that it should not keep the file in cache. I
16 have written a patch for rsync that adds the
20 option which activates posix_fadvise64.
22 There are some caveats though:
24 * When calling posix_fadvise64 while writing a file, only the
25 part of the cache will be release which has already been
26 written to disk. This means we have to call fdatasync before
27 calling posix_fadvise64 and this will unfortunately slow down
28 operations considerably. On my test system I get 240 KByte/s.
30 The patch has been optimized, so that the impact on large files
31 will be considerably lowered by calling posix_fadvise64 only
32 after a few megabytes have been written.
34 * When reading a file which has been cached *Before* rsync read
35 it, the content of the file will be released from cache never
36 the less, which may not be intended. I have unfortunately not
37 found a method for determining if a file is in cache or not
40 I found that running rsync of an lvm snapshot is a good way
41 around this problem, since the snapshot data is cached
42 separately from the original. It has the additional benefit of
43 making the backups more consistent.
45 * I don't really know the rsync code, so it may be that the patch
46 is calling fadvise for files where this would not be necessary.
48 * The patch is tested only on Linux 2.6.18
50 If you have any input on this, please let me know.
52 You can get the latest edition of the patch from
54 http://tobi.oetiker.ch/patches/
63 * pass --drop-cache on to the remote server
66 diff --git a/checksum.c b/checksum.c
69 @@ -26,6 +26,10 @@ extern int protocol_version;
71 int csum_length = SHORT_SUM_LENGTH; /* initial value */
73 +#ifdef HAVE_POSIX_FADVISE64
74 +#define close(fd) fadv_close(fd)
78 a simple 32 bit checksum that can be upadted from either end
79 (inspired by Mark Adler's Adler-32 checksum)
80 diff --git a/cleanup.c b/cleanup.c
83 @@ -46,7 +46,13 @@ void close_all(void)
89 +#ifdef HAVE_POSIX_FADVISE64
93 +#ifdef SHUTDOWN_ALL_SOCKETS
94 max_fd = sysconf(_SC_OPEN_MAX) - 1;
95 for (fd = max_fd; fd >= 0; fd--) {
96 if ((ret = do_fstat(fd, &st)) == 0) {
97 diff --git a/configure.in b/configure.in
100 @@ -552,7 +552,7 @@ AC_CHECK_FUNCS(waitpid wait4 getcwd strdup chown chmod lchmod mknod mkfifo \
101 strlcat strlcpy strtol mallinfo getgroups setgroups geteuid getegid \
102 setlocale setmode open64 lseek64 mkstemp64 mtrace va_copy __va_copy \
103 strerror putenv iconv_open locale_charset nl_langinfo getxattr \
104 - extattr_get_link sigaction sigprocmask setattrlist)
105 + extattr_get_link sigaction sigprocmask setattrlist posix_fadvise64)
107 AC_CHECK_FUNCS(getpgrp tcgetpgrp)
108 if test $ac_cv_func_getpgrp = yes; then
109 diff --git a/fileio.c b/fileio.c
112 @@ -30,6 +30,12 @@ extern int sparse_files;
113 static char last_byte;
114 static size_t sparse_seek = 0;
116 +#ifdef HAVE_POSIX_FADVISE64
117 +#define close(fd) fadv_close(fd)
118 +#define read(fd,buf,len) fadv_read(fd,buf,len)
119 +#define write(fd,buf,len) fadv_write(fd,buf,len)
122 int sparse_end(int f)
125 diff --git a/generator.c b/generator.c
128 @@ -112,6 +112,10 @@ static int need_retouch_dir_times;
129 static int need_retouch_dir_perms;
130 static const char *solo_file = NULL;
132 +#ifdef HAVE_POSIX_FADVISE64
133 +#define close(fd) fadv_close(fd)
136 /* For calling delete_item() and delete_dir_contents(). */
137 #define DEL_RECURSE (1<<1) /* recurse */
138 #define DEL_DIR_IS_EMPTY (1<<2) /* internal delete_FUNCTIONS use only */
139 diff --git a/options.c b/options.c
142 @@ -59,6 +59,7 @@ int preserve_uid = 0;
143 int preserve_gid = 0;
144 int preserve_times = 0;
150 @@ -321,6 +322,9 @@ void usage(enum logcode F)
151 rprintf(F," --backup-dir=DIR make backups into hierarchy based in DIR\n");
152 rprintf(F," --suffix=SUFFIX set backup suffix (default %s w/o --backup-dir)\n",BACKUP_SUFFIX);
153 rprintf(F," -u, --update skip files that are newer on the receiver\n");
154 +#ifdef HAVE_POSIX_FADVISE64
155 + rprintf(F," --drop-cache tell OS to drop caching of file data\n");
157 rprintf(F," --inplace update destination files in-place (SEE MAN PAGE)\n");
158 rprintf(F," --append append data onto shorter files\n");
159 rprintf(F," --append-verify like --append, but with old data in file checksum\n");
160 @@ -525,6 +529,9 @@ static struct poptOption long_options[] = {
161 {"size-only", 0, POPT_ARG_NONE, &size_only, 0, 0, 0 },
162 {"one-file-system", 'x', POPT_ARG_NONE, 0, 'x', 0, 0 },
163 {"update", 'u', POPT_ARG_NONE, &update_only, 0, 0, 0 },
164 +#ifdef HAVE_POSIX_FADVISE64
165 + {"drop-cache", 0, POPT_ARG_NONE, &drop_cache, 0, 0, 0 },
167 {"existing", 0, POPT_ARG_NONE, &ignore_non_existing, 0, 0, 0 },
168 {"ignore-non-existing",0,POPT_ARG_NONE, &ignore_non_existing, 0, 0, 0 },
169 {"ignore-existing", 0, POPT_ARG_NONE, &ignore_existing, 0, 0, 0 },
170 @@ -1670,6 +1677,11 @@ void server_options(char **args, int *argc_p)
172 args[ac++] = "--sender";
174 +#ifdef HAVE_POSIX_FADVISE64
176 + args[ac++] = "--drop-cache";
182 diff --git a/receiver.c b/receiver.c
185 @@ -62,6 +62,10 @@ static int phase = 0, redoing = 0;
186 /* We're either updating the basis file or an identical copy: */
187 static int updating_basis_or_equiv;
189 +#ifdef HAVE_POSIX_FADVISE64
190 +#define close(fd) fadv_close(fd)
194 * get_tmpname() - create a tmp filename for a given filename
196 diff --git a/rsync.yo b/rsync.yo
199 @@ -352,6 +352,7 @@ to the detailed description below for a complete description. verb(
200 --super receiver attempts super-user activities
201 --fake-super store/recover privileged attrs using xattrs
202 -S, --sparse handle sparse files efficiently
203 + --drop-cache tell OS to drop caching of file data
204 -n, --dry-run perform a trial run with no changes made
205 -W, --whole-file copy files whole (w/o delta-xfer algorithm)
206 -x, --one-file-system don't cross filesystem boundaries
207 @@ -1023,6 +1024,10 @@ NOTE: Don't use this option when the destination is a Solaris "tmpfs"
208 filesystem. It doesn't seem to handle seeks over null regions
209 correctly and ends up corrupting the files.
211 +dit(bf(--drop-cache)) Tell the OS to drop the caching of the file data. This
212 +prevents rsync from filling up the filesystem cache. This can sometimes help
213 +to make a system perform better by keeping non-rsync files in the disk cache.
215 dit(bf(-n, --dry-run)) This makes rsync perform a trial run that doesn't
216 make any changes (and produces mostly the same output as a real run). It
217 is most commonly used in combination with the bf(-v, --verbose) and/or
218 diff --git a/sender.c b/sender.c
221 @@ -46,6 +46,10 @@ extern int write_batch;
222 extern struct stats stats;
223 extern struct file_list *cur_flist, *first_flist, *dir_flist;
225 +#ifdef HAVE_POSIX_FADVISE64
226 +#define close(fd) fadv_close(fd)
232 diff --git a/t_unsafe.c b/t_unsafe.c
235 @@ -28,6 +28,7 @@ int am_root = 0;
240 int preserve_perms = 0;
243 diff --git a/util.c b/util.c
249 extern int module_id;
250 +extern int drop_cache;
251 extern int modify_window;
252 extern int relative_paths;
253 extern int human_readable;
254 @@ -41,6 +42,131 @@ char curr_dir[MAXPATHLEN];
255 unsigned int curr_dir_len;
256 int curr_dir_depth; /* This is only set for a sanitizing daemon. */
258 +#ifdef HAVE_POSIX_FADVISE64
259 +#define FADV_BUFFER_SIZE 1024*1024*16
261 +static struct stat fadv_fd_stat[1024];
262 +static off_t fadv_fd_pos[1024];
263 +static int fadv_fd_init = 0;
264 +static int fadv_max_fd = 0;
265 +static int fadv_close_ring_tail = 0;
266 +static int fadv_close_ring_head = 0;
267 +static int fadv_close_ring_size = 0;
268 +static int fadv_close_ring[1024];
269 +static int fadv_close_buffer_size = 0;
271 +static void fadv_fd_init_func(void)
273 + if (fadv_fd_init == 0) {
276 + if (fadv_max_fd == 0){
277 + fadv_max_fd = sysconf(_SC_OPEN_MAX) - 20;
278 + if (fadv_max_fd < 0)
280 + if (fadv_max_fd > 1000)
281 + fadv_max_fd = 1000;
283 + for (i = 0; i < fadv_max_fd; i++) {
284 + fadv_fd_pos[i] = 0;
285 + fadv_fd_stat[i].st_dev = 0;
286 + fadv_fd_stat[i].st_ino = 0;
291 +static void fadv_drop(int fd, int sync)
296 + /* Trail 1 MB behind in dropping. we do this to make
297 + * sure that the same block or stripe does not have
298 + * to be written twice. */
299 + if (fd > fadv_max_fd)
301 + pos = lseek(fd, 0, SEEK_CUR) - 1024*1024;
302 + fadv_fd_init_func();
304 + if (fadv_fd_stat[fd].st_dev == sb.st_dev
305 + && fadv_fd_stat[fd].st_ino == sb.st_ino) {
306 + if (fadv_fd_pos[fd] < pos - FADV_BUFFER_SIZE) {
308 + /* If the file is not flushed to disk before calling fadvise,
309 + * then the Cache will not be freed and the advise gets ignored
310 + * this does give a severe hit on performance. If only there
311 + * was a way to mark cache so that it gets release once the data
312 + * is written to disk. */
315 + posix_fadvise64(fd, 0, pos, POSIX_FADV_DONTNEED);
316 + fadv_fd_pos[fd] = pos;
319 + fadv_fd_stat[fd].st_dev = sb.st_dev;
320 + fadv_fd_stat[fd].st_ino = sb.st_ino;
321 + fadv_fd_pos[fd] = 0;
325 +ssize_t fadv_write(int fd, const void *buf, size_t count)
327 + int ret = write(fd, buf, count);
333 +ssize_t fadv_read(int fd, void *buf, size_t count)
335 + int ret = read(fd, buf, count);
341 +void fadv_close_all(void)
343 + while (fadv_close_ring_size > 0){
344 + fdatasync(fadv_close_ring[fadv_close_ring_tail]);
345 + posix_fadvise64(fadv_close_ring[fadv_close_ring_tail], 0, 0,POSIX_FADV_DONTNEED);
346 + fadv_close_ring_size--;
347 + close(fadv_close_ring[fadv_close_ring_tail]);
348 + fadv_close_ring_tail = (fadv_close_ring_tail + 1) % fadv_max_fd;
349 + fadv_close_buffer_size = 0;
353 +int fadv_close(int fd)
356 + /* If the file is not flushed to disk before calling fadvise,
357 + * then the Cache will not be freed and the advise gets ignored
358 + * this does give a severe hit on performance. So instead of doing
359 + * it right away, we save us a copy of the filehandle and do it
360 + * some time before we are out of filehandles. This speeds
361 + * up operation for small files massively. It is directly
362 + * related to the number of spare file handles you have. */
363 + int newfd = dup(fd);
364 + int pos = lseek(fd, 0, SEEK_CUR);
365 + fadv_fd_init_func();
366 + fadv_close_buffer_size += pos - fadv_fd_pos[fd];
367 + fadv_close_ring[fadv_close_ring_head] = newfd;
368 + fadv_close_ring_head = (fadv_close_ring_head + 1) % fadv_max_fd;
369 + fadv_close_ring_size ++;
370 + if (fadv_close_ring_size == fadv_max_fd || fadv_close_buffer_size > 1024*1024 ){
371 + /* it seems fastest to drop things 'in groups' */
378 +#define close(fd) fadv_close(fd)
379 +#define read(fd,buf,len) fadv_read(fd,buf,len)
380 +#define write(fd,buf,len) fadv_write(fd,buf,len)
383 /* Set a fd into nonblocking mode. */
384 void set_nonblocking(int fd)