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
72 based-on: 40afd365cc8ca968fd16e161d24df5b8a8a520cc
73 diff --git a/checksum.c b/checksum.c
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
90 @@ -47,7 +47,13 @@ void close_all(void)
96 +#ifdef HAVE_POSIX_FADVISE64
100 +#ifdef SHUTDOWN_ALL_SOCKETS
101 max_fd = sysconf(_SC_OPEN_MAX) - 1;
102 for (fd = max_fd; fd >= 0; fd--) {
103 if ((ret = do_fstat(fd, &st)) == 0) {
104 diff --git a/configure.ac b/configure.ac
107 @@ -570,6 +570,7 @@ AC_FUNC_ALLOCA
108 AC_CHECK_FUNCS(waitpid wait4 getcwd strdup chown chmod lchmod mknod mkfifo \
109 fchmod fstat ftruncate strchr readlink link utime utimes lutimes strftime \
110 memmove lchown vsnprintf snprintf vasprintf asprintf setsid strpbrk \
112 strlcat strlcpy strtol mallinfo getgroups setgroups geteuid getegid \
113 setlocale setmode open64 lseek64 mkstemp64 mtrace va_copy __va_copy \
114 seteuid strerror putenv iconv_open locale_charset nl_langinfo getxattr \
115 diff --git a/fileio.c b/fileio.c
118 @@ -29,6 +29,12 @@ extern int sparse_files;
120 static OFF_T sparse_seek = 0;
122 +#ifdef HAVE_POSIX_FADVISE64
123 +#define close(fd) fadv_close(fd)
124 +#define read(fd,buf,len) fadv_read(fd,buf,len)
125 +#define write(fd,buf,len) fadv_write(fd,buf,len)
128 int sparse_end(int f, OFF_T size)
131 diff --git a/generator.c b/generator.c
134 @@ -113,6 +113,10 @@ static int need_retouch_dir_times;
135 static int need_retouch_dir_perms;
136 static const char *solo_file = NULL;
138 +#ifdef HAVE_POSIX_FADVISE64
139 +#define close(fd) fadv_close(fd)
142 /* For calling delete_item() and delete_dir_contents(). */
143 #define DEL_NO_UID_WRITE (1<<0) /* file/dir has our uid w/o write perm */
144 #define DEL_RECURSE (1<<1) /* if dir, delete all contents */
145 diff --git a/options.c b/options.c
148 @@ -60,6 +60,7 @@ int preserve_uid = 0;
149 int preserve_gid = 0;
150 int preserve_times = 0;
156 @@ -325,6 +326,9 @@ void usage(enum logcode F)
157 rprintf(F," --backup-dir=DIR make backups into hierarchy based in DIR\n");
158 rprintf(F," --suffix=SUFFIX set backup suffix (default %s w/o --backup-dir)\n",BACKUP_SUFFIX);
159 rprintf(F," -u, --update skip files that are newer on the receiver\n");
160 +#ifdef HAVE_POSIX_FADVISE64
161 + rprintf(F," --drop-cache tell OS to drop caching of file data\n");
163 rprintf(F," --inplace update destination files in-place (SEE MAN PAGE)\n");
164 rprintf(F," --append append data onto shorter files\n");
165 rprintf(F," --append-verify like --append, but with old data in file checksum\n");
166 @@ -533,6 +537,9 @@ static struct poptOption long_options[] = {
167 {"no-one-file-system",0, POPT_ARG_VAL, &one_file_system, 0, 0, 0 },
168 {"no-x", 0, POPT_ARG_VAL, &one_file_system, 0, 0, 0 },
169 {"update", 'u', POPT_ARG_NONE, &update_only, 0, 0, 0 },
170 +#ifdef HAVE_POSIX_FADVISE64
171 + {"drop-cache", 0, POPT_ARG_NONE, &drop_cache, 0, 0, 0 },
173 {"existing", 0, POPT_ARG_NONE, &ignore_non_existing, 0, 0, 0 },
174 {"ignore-non-existing",0,POPT_ARG_NONE, &ignore_non_existing, 0, 0, 0 },
175 {"ignore-existing", 0, POPT_ARG_NONE, &ignore_existing, 0, 0, 0 },
176 @@ -1730,6 +1737,11 @@ void server_options(char **args, int *argc_p)
178 args[ac++] = "--sender";
180 +#ifdef HAVE_POSIX_FADVISE64
182 + args[ac++] = "--drop-cache";
188 diff --git a/receiver.c b/receiver.c
191 @@ -64,6 +64,10 @@ static flist_ndx_list batch_redo_list;
192 /* We're either updating the basis file or an identical copy: */
193 static int updating_basis_or_equiv;
195 +#ifdef HAVE_POSIX_FADVISE64
196 +#define close(fd) fadv_close(fd)
199 #define TMPNAME_SUFFIX ".XXXXXX"
200 #define TMPNAME_SUFFIX_LEN ((int)sizeof TMPNAME_SUFFIX - 1)
202 diff --git a/rsync.yo b/rsync.yo
205 @@ -369,6 +369,7 @@ to the detailed description below for a complete description. verb(
206 --super receiver attempts super-user activities
207 --fake-super store/recover privileged attrs using xattrs
208 -S, --sparse handle sparse files efficiently
209 + --drop-cache tell OS to drop caching of file data
210 -n, --dry-run perform a trial run with no changes made
211 -W, --whole-file copy files whole (w/o delta-xfer algorithm)
212 -x, --one-file-system don't cross filesystem boundaries
213 @@ -1122,6 +1123,10 @@ dit(bf(-S, --sparse)) Try to handle sparse files efficiently so they take
214 up less space on the destination. Conflicts with bf(--inplace) because it's
215 not possible to overwrite data in a sparse fashion.
217 +dit(bf(--drop-cache)) Tell the OS to drop the caching of the file data. This
218 +prevents rsync from filling up the filesystem cache. This can sometimes help
219 +to make a system perform better by keeping non-rsync files in the disk cache.
221 dit(bf(-n, --dry-run)) This makes rsync perform a trial run that doesn't
222 make any changes (and produces mostly the same output as a real run). It
223 is most commonly used in combination with the bf(-v, --verbose) and/or
224 diff --git a/sender.c b/sender.c
227 @@ -45,6 +45,10 @@ extern int write_batch;
228 extern struct stats stats;
229 extern struct file_list *cur_flist, *first_flist, *dir_flist;
231 +#ifdef HAVE_POSIX_FADVISE64
232 +#define close(fd) fadv_close(fd)
238 diff --git a/t_unsafe.c b/t_unsafe.c
241 @@ -28,6 +28,7 @@ int am_root = 0;
246 int preserve_perms = 0;
247 int preserve_executability = 0;
249 diff --git a/util.c b/util.c
255 extern int module_id;
256 +extern int drop_cache;
257 extern int modify_window;
258 extern int relative_paths;
259 extern int preserve_times;
260 @@ -42,6 +43,131 @@ char curr_dir[MAXPATHLEN];
261 unsigned int curr_dir_len;
262 int curr_dir_depth; /* This is only set for a sanitizing daemon. */
264 +#ifdef HAVE_POSIX_FADVISE64
265 +#define FADV_BUFFER_SIZE 1024*1024*16
267 +static struct stat fadv_fd_stat[1024];
268 +static off_t fadv_fd_pos[1024];
269 +static int fadv_fd_init = 0;
270 +static int fadv_max_fd = 0;
271 +static int fadv_close_ring_tail = 0;
272 +static int fadv_close_ring_head = 0;
273 +static int fadv_close_ring_size = 0;
274 +static int fadv_close_ring[1024];
275 +static int fadv_close_buffer_size = 0;
277 +static void fadv_fd_init_func(void)
279 + if (fadv_fd_init == 0) {
282 + if (fadv_max_fd == 0){
283 + fadv_max_fd = sysconf(_SC_OPEN_MAX) - 20;
284 + if (fadv_max_fd < 0)
286 + if (fadv_max_fd > 1000)
287 + fadv_max_fd = 1000;
289 + for (i = 0; i < fadv_max_fd; i++) {
290 + fadv_fd_pos[i] = 0;
291 + fadv_fd_stat[i].st_dev = 0;
292 + fadv_fd_stat[i].st_ino = 0;
297 +static void fadv_drop(int fd, int sync)
302 + /* Trail 1 MB behind in dropping. we do this to make
303 + * sure that the same block or stripe does not have
304 + * to be written twice. */
305 + if (fd > fadv_max_fd)
307 + pos = lseek(fd, 0, SEEK_CUR) - 1024*1024;
308 + fadv_fd_init_func();
310 + if (fadv_fd_stat[fd].st_dev == sb.st_dev
311 + && fadv_fd_stat[fd].st_ino == sb.st_ino) {
312 + if (fadv_fd_pos[fd] < pos - FADV_BUFFER_SIZE) {
314 + /* If the file is not flushed to disk before calling fadvise,
315 + * then the Cache will not be freed and the advise gets ignored
316 + * this does give a severe hit on performance. If only there
317 + * was a way to mark cache so that it gets release once the data
318 + * is written to disk. */
321 + posix_fadvise64(fd, 0, pos, POSIX_FADV_DONTNEED);
322 + fadv_fd_pos[fd] = pos;
325 + fadv_fd_stat[fd].st_dev = sb.st_dev;
326 + fadv_fd_stat[fd].st_ino = sb.st_ino;
327 + fadv_fd_pos[fd] = 0;
331 +ssize_t fadv_write(int fd, const void *buf, size_t count)
333 + int ret = write(fd, buf, count);
339 +ssize_t fadv_read(int fd, void *buf, size_t count)
341 + int ret = read(fd, buf, count);
347 +void fadv_close_all(void)
349 + while (fadv_close_ring_size > 0){
350 + fdatasync(fadv_close_ring[fadv_close_ring_tail]);
351 + posix_fadvise64(fadv_close_ring[fadv_close_ring_tail], 0, 0,POSIX_FADV_DONTNEED);
352 + fadv_close_ring_size--;
353 + close(fadv_close_ring[fadv_close_ring_tail]);
354 + fadv_close_ring_tail = (fadv_close_ring_tail + 1) % fadv_max_fd;
355 + fadv_close_buffer_size = 0;
359 +int fadv_close(int fd)
362 + /* If the file is not flushed to disk before calling fadvise,
363 + * then the Cache will not be freed and the advise gets ignored
364 + * this does give a severe hit on performance. So instead of doing
365 + * it right away, we save us a copy of the filehandle and do it
366 + * some time before we are out of filehandles. This speeds
367 + * up operation for small files massively. It is directly
368 + * related to the number of spare file handles you have. */
369 + int newfd = dup(fd);
370 + int pos = lseek(fd, 0, SEEK_CUR);
371 + fadv_fd_init_func();
372 + fadv_close_buffer_size += pos - fadv_fd_pos[fd];
373 + fadv_close_ring[fadv_close_ring_head] = newfd;
374 + fadv_close_ring_head = (fadv_close_ring_head + 1) % fadv_max_fd;
375 + fadv_close_ring_size ++;
376 + if (fadv_close_ring_size == fadv_max_fd || fadv_close_buffer_size > 1024*1024 ){
377 + /* it seems fastest to drop things 'in groups' */
384 +#define close(fd) fadv_close(fd)
385 +#define read(fd,buf,len) fadv_read(fd,buf,len)
386 +#define write(fd,buf,len) fadv_write(fd,buf,len)
389 /* Set a fd into nonblocking mode. */
390 void set_nonblocking(int fd)