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 To use this patch, run these commands for a successful build:
68 patch -p1 <patches/drop-cache.diff
69 ./configure (optional if already run)
72 diff --git a/checksum.c b/checksum.c
73 index 811b5b6..f1f51cb 100644
77 extern int checksum_seed;
78 extern int protocol_version;
80 +#ifdef HAVE_POSIX_FADVISE64
81 +#define close(fd) fadv_close(fd)
85 a simple 32 bit checksum that can be upadted from either end
86 (inspired by Mark Adler's Adler-32 checksum)
87 diff --git a/cleanup.c b/cleanup.c
88 index 19ef072..84a6cf3 100644
91 @@ -47,7 +47,13 @@ void close_all(void)
97 +#ifdef HAVE_POSIX_FADVISE64
101 +#ifdef SHUTDOWN_ALL_SOCKETS
102 max_fd = sysconf(_SC_OPEN_MAX) - 1;
103 for (fd = max_fd; fd >= 0; fd--) {
104 if ((ret = do_fstat(fd, &st)) == 0) {
105 diff --git a/configure.in b/configure.in
106 index bc7d4a7..e9ff45a 100644
109 @@ -553,7 +553,7 @@ AC_CHECK_FUNCS(waitpid wait4 getcwd strdup chown chmod lchmod mknod mkfifo \
110 setlocale setmode open64 lseek64 mkstemp64 mtrace va_copy __va_copy \
111 strerror putenv iconv_open locale_charset nl_langinfo getxattr \
112 extattr_get_link sigaction sigprocmask setattrlist getgrouplist \
114 + initgroups posix_fadvise64)
116 dnl cygwin iconv.h defines iconv_open as libiconv_open
117 if test x"$ac_cv_func_iconv_open" != x"yes"; then
118 diff --git a/fileio.c b/fileio.c
119 index 0faa619..0326fe8 100644
122 @@ -31,6 +31,12 @@ extern int sparse_files;
123 static char last_byte;
124 static OFF_T sparse_seek = 0;
126 +#ifdef HAVE_POSIX_FADVISE64
127 +#define close(fd) fadv_close(fd)
128 +#define read(fd,buf,len) fadv_read(fd,buf,len)
129 +#define write(fd,buf,len) fadv_write(fd,buf,len)
132 int sparse_end(int f)
135 diff --git a/generator.c b/generator.c
136 index 12007a1..f0c7a91 100644
139 @@ -112,6 +112,10 @@ static int need_retouch_dir_times;
140 static int need_retouch_dir_perms;
141 static const char *solo_file = NULL;
143 +#ifdef HAVE_POSIX_FADVISE64
144 +#define close(fd) fadv_close(fd)
148 TYPE_DIR, TYPE_SPECIAL, TYPE_DEVICE, TYPE_SYMLINK
150 diff --git a/options.c b/options.c
151 index e7c6c61..ce55d8e 100644
154 @@ -60,6 +60,7 @@ int preserve_uid = 0;
155 int preserve_gid = 0;
156 int preserve_times = 0;
162 @@ -670,6 +671,9 @@ void usage(enum logcode F)
163 rprintf(F," --backup-dir=DIR make backups into hierarchy based in DIR\n");
164 rprintf(F," --suffix=SUFFIX set backup suffix (default %s w/o --backup-dir)\n",BACKUP_SUFFIX);
165 rprintf(F," -u, --update skip files that are newer on the receiver\n");
166 +#ifdef HAVE_POSIX_FADVISE64
167 + rprintf(F," --drop-cache tell OS to drop caching of file data\n");
169 rprintf(F," --inplace update destination files in-place (SEE MAN PAGE)\n");
170 rprintf(F," --append append data onto shorter files\n");
171 rprintf(F," --append-verify like --append, but with old data in file checksum\n");
172 @@ -891,6 +895,9 @@ static struct poptOption long_options[] = {
173 {"no-one-file-system",'x',POPT_ARG_VAL, &one_file_system, 0, 0, 0 },
174 {"no-x", 'x', POPT_ARG_VAL, &one_file_system, 0, 0, 0 },
175 {"update", 'u', POPT_ARG_NONE, &update_only, 0, 0, 0 },
176 +#ifdef HAVE_POSIX_FADVISE64
177 + {"drop-cache", 0, POPT_ARG_NONE, &drop_cache, 0, 0, 0 },
179 {"existing", 0, POPT_ARG_NONE, &ignore_non_existing, 0, 0, 0 },
180 {"ignore-non-existing",0,POPT_ARG_NONE, &ignore_non_existing, 0, 0, 0 },
181 {"ignore-existing", 0, POPT_ARG_NONE, &ignore_existing, 0, 0, 0 },
182 @@ -2235,6 +2242,11 @@ void server_options(char **args, int *argc_p)
184 args[ac++] = "--sender";
186 +#ifdef HAVE_POSIX_FADVISE64
188 + args[ac++] = "--drop-cache";
194 diff --git a/receiver.c b/receiver.c
195 index 4325e30..a3da64e 100644
198 @@ -64,6 +64,10 @@ static flist_ndx_list batch_redo_list;
199 /* We're either updating the basis file or an identical copy: */
200 static int updating_basis_or_equiv;
202 +#ifdef HAVE_POSIX_FADVISE64
203 +#define close(fd) fadv_close(fd)
207 * get_tmpname() - create a tmp filename for a given filename
209 diff --git a/rsync.yo b/rsync.yo
210 index 941f7a5..512aa6b 100644
213 @@ -359,6 +359,7 @@ to the detailed description below for a complete description. verb(
214 --super receiver attempts super-user activities
215 --fake-super store/recover privileged attrs using xattrs
216 -S, --sparse handle sparse files efficiently
217 + --drop-cache tell OS to drop caching of file data
218 -n, --dry-run perform a trial run with no changes made
219 -W, --whole-file copy files whole (w/o delta-xfer algorithm)
220 -x, --one-file-system don't cross filesystem boundaries
221 @@ -1120,6 +1121,10 @@ NOTE: Don't use this option when the destination is a Solaris "tmpfs"
222 filesystem. It doesn't seem to handle seeks over null regions
223 correctly and ends up corrupting the files.
225 +dit(bf(--drop-cache)) Tell the OS to drop the caching of the file data. This
226 +prevents rsync from filling up the filesystem cache. This can sometimes help
227 +to make a system perform better by keeping non-rsync files in the disk cache.
229 dit(bf(-n, --dry-run)) This makes rsync perform a trial run that doesn't
230 make any changes (and produces mostly the same output as a real run). It
231 is most commonly used in combination with the bf(-v, --verbose) and/or
232 diff --git a/sender.c b/sender.c
233 index bf8221d..9e23dbb 100644
236 @@ -45,6 +45,10 @@ extern int write_batch;
237 extern struct stats stats;
238 extern struct file_list *cur_flist, *first_flist, *dir_flist;
240 +#ifdef HAVE_POSIX_FADVISE64
241 +#define close(fd) fadv_close(fd)
247 diff --git a/t_unsafe.c b/t_unsafe.c
248 index 9ba0aaa..3cb55e9 100644
251 @@ -27,6 +27,7 @@ int dry_run = 0;
256 int human_readable = 0;
257 int preserve_perms = 0;
258 int preserve_executability = 0;
259 diff --git a/util.c b/util.c
260 index 0cafed6..06d8770 100644
266 extern int module_id;
267 +extern int drop_cache;
268 extern int modify_window;
269 extern int relative_paths;
270 extern int preserve_xattrs;
271 @@ -42,6 +43,131 @@ char curr_dir[MAXPATHLEN];
272 unsigned int curr_dir_len;
273 int curr_dir_depth; /* This is only set for a sanitizing daemon. */
275 +#ifdef HAVE_POSIX_FADVISE64
276 +#define FADV_BUFFER_SIZE 1024*1024*16
278 +static struct stat fadv_fd_stat[1024];
279 +static off_t fadv_fd_pos[1024];
280 +static int fadv_fd_init = 0;
281 +static int fadv_max_fd = 0;
282 +static int fadv_close_ring_tail = 0;
283 +static int fadv_close_ring_head = 0;
284 +static int fadv_close_ring_size = 0;
285 +static int fadv_close_ring[1024];
286 +static int fadv_close_buffer_size = 0;
288 +static void fadv_fd_init_func(void)
290 + if (fadv_fd_init == 0) {
293 + if (fadv_max_fd == 0){
294 + fadv_max_fd = sysconf(_SC_OPEN_MAX) - 20;
295 + if (fadv_max_fd < 0)
297 + if (fadv_max_fd > 1000)
298 + fadv_max_fd = 1000;
300 + for (i = 0; i < fadv_max_fd; i++) {
301 + fadv_fd_pos[i] = 0;
302 + fadv_fd_stat[i].st_dev = 0;
303 + fadv_fd_stat[i].st_ino = 0;
308 +static void fadv_drop(int fd, int sync)
313 + /* Trail 1 MB behind in dropping. we do this to make
314 + * sure that the same block or stripe does not have
315 + * to be written twice. */
316 + if (fd > fadv_max_fd)
318 + pos = lseek(fd, 0, SEEK_CUR) - 1024*1024;
319 + fadv_fd_init_func();
321 + if (fadv_fd_stat[fd].st_dev == sb.st_dev
322 + && fadv_fd_stat[fd].st_ino == sb.st_ino) {
323 + if (fadv_fd_pos[fd] < pos - FADV_BUFFER_SIZE) {
325 + /* If the file is not flushed to disk before calling fadvise,
326 + * then the Cache will not be freed and the advise gets ignored
327 + * this does give a severe hit on performance. If only there
328 + * was a way to mark cache so that it gets release once the data
329 + * is written to disk. */
332 + posix_fadvise64(fd, 0, pos, POSIX_FADV_DONTNEED);
333 + fadv_fd_pos[fd] = pos;
336 + fadv_fd_stat[fd].st_dev = sb.st_dev;
337 + fadv_fd_stat[fd].st_ino = sb.st_ino;
338 + fadv_fd_pos[fd] = 0;
342 +ssize_t fadv_write(int fd, const void *buf, size_t count)
344 + int ret = write(fd, buf, count);
350 +ssize_t fadv_read(int fd, void *buf, size_t count)
352 + int ret = read(fd, buf, count);
358 +void fadv_close_all(void)
360 + while (fadv_close_ring_size > 0){
361 + fdatasync(fadv_close_ring[fadv_close_ring_tail]);
362 + posix_fadvise64(fadv_close_ring[fadv_close_ring_tail], 0, 0,POSIX_FADV_DONTNEED);
363 + fadv_close_ring_size--;
364 + close(fadv_close_ring[fadv_close_ring_tail]);
365 + fadv_close_ring_tail = (fadv_close_ring_tail + 1) % fadv_max_fd;
366 + fadv_close_buffer_size = 0;
370 +int fadv_close(int fd)
373 + /* If the file is not flushed to disk before calling fadvise,
374 + * then the Cache will not be freed and the advise gets ignored
375 + * this does give a severe hit on performance. So instead of doing
376 + * it right away, we save us a copy of the filehandle and do it
377 + * some time before we are out of filehandles. This speeds
378 + * up operation for small files massively. It is directly
379 + * related to the number of spare file handles you have. */
380 + int newfd = dup(fd);
381 + int pos = lseek(fd, 0, SEEK_CUR);
382 + fadv_fd_init_func();
383 + fadv_close_buffer_size += pos - fadv_fd_pos[fd];
384 + fadv_close_ring[fadv_close_ring_head] = newfd;
385 + fadv_close_ring_head = (fadv_close_ring_head + 1) % fadv_max_fd;
386 + fadv_close_ring_size ++;
387 + if (fadv_close_ring_size == fadv_max_fd || fadv_close_buffer_size > 1024*1024 ){
388 + /* it seems fastest to drop things 'in groups' */
395 +#define close(fd) fadv_close(fd)
396 +#define read(fd,buf,len) fadv_read(fd,buf,len)
397 +#define write(fd,buf,len) fadv_write(fd,buf,len)
400 /* Set a fd into nonblocking mode. */
401 void set_nonblocking(int fd)