I'm trying to reduce the load a bit on our NFS ccache server. I noted the
following pattern of accesses on a cache hit:
# stat("$CACHE_DIR", {...}) = 0
# stat("$CACHE_DIR/tmp", {...}) = 0
# stat("$CACHE_DIR/CACHEDIR.TAG", {...}) = 0
* stat("$CACHE_DIR/3", {...}) = 0
* stat("$CACHE_DIR/3/d", {...}) = 0
# stat("$CACHE_DIR/3/d/3", {...}) = 0
open("$CACHE_DIR/3/d/3/df7ee0d1f41405e97cc3c5a639dba-14689.manifest",
O_RDONLY) = 4
* stat("$CACHE_DIR/f", {...}) = 0
* stat("$CACHE_DIR/f/6", {...}) = 0
# stat("$CACHE_DIR/f/6/c", {...}) = 0
* stat("$CACHE_DIR/f", {...}) = 0
* stat("$CACHE_DIR/f/6", {...}) = 0
# stat("$CACHE_DIR/f/6/c", {...}) = 0
* stat("$CACHE_DIR/f", {...}) = 0
* stat("$CACHE_DIR/f/6", {...}) = 0
# stat("$CACHE_DIR/f/6/c", {...}) = 0
stat("$CACHE_DIR/f/6/c/51577239e589de02d30b94152958b-2123817.o", {...}) = 0
open("$CACHE_DIR/f/6/c/51577239e589de02d30b94152958b-2123817.o", O_RDONLY) = 4
utimes("$CACHE_DIR/f/6/c/51577239e589de02d30b94152958b-2123817.o", NULL) = 0
utimes("$CACHE_DIR/f/6/c/51577239e589de02d30b94152958b-2123817.stderr", NULL)
= -1 ENOENT (No such file or directory)
open("$CACHE_DIR/f/6/c/51577239e589de02d30b94152958b-2123817.stderr",
O_RDONLY) = -1 ENOENT (No such file or directory)
symlink("lion02:19843:1286222896", "$CACHE_DIR/f/stats.lock") = 0
# stat("$CACHE_DIR/f/stats", {...}) = 0
open("$CACHE_DIR/f/stats", O_RDONLY) = 4
open("$CACHE_DIR/f/stats.tmp.lion02.19843", O_WRONLY|O_CREAT|O_TRUNC, 0666) =
4
rename("$CACHE_DIR/f/stats.tmp.lion02.19843", "$CACHE_DIR/f/stats") = 0
unlink("$CACHE_DIR/f/stats.lock") = 0
This patch removes those with *s.
There are still extra calls (the #s); it would be best if
the code assumed directories existed and just tried to read
the manifest or .o. Then the directory created only if a
new file is added. But, that does more violence then I'm
willing to undertake at the moment. Improvements to do that
are welcome.
Thanks
diff --git a/ccache.c b/ccache.c
index b84c84e..eed076f 100644
--- a/ccache.c
+++ b/ccache.c
@@ -253,6 +253,10 @@ clean_up_tmp_files()
/*
* Transform a name to a full path into the cache directory, creating needed
* sublevels if needed. Caller frees.
+ *
+ * stat() calls can be expensive on the cache, so we both avoid checking a
+ * single path multiple times, and also avoid stat()ing each level in a
+ * many level cache.
*/
static char *
get_path_in_cache(const char *name, const char *suffix)
@@ -260,17 +264,32 @@ get_path_in_cache(const char *name, const char *suffix)
int i;
char *path;
char *result;
+ struct stat st;
+ /* Form the entire path */
path = x_strdup(cache_dir);
for (i = 0; i < nlevels; ++i) {
char *p = format("%s/%c", path, name[i]);
free(path);
path = p;
- if (create_dir(path) != 0) {
- cc_log("Failed to create %s", path);
- failed();
+ }
+
+ /* First see if whole path exists, so can avoid directory one-by-one
check */
+ if (!(stat(path, &st) == 0
+ && S_ISDIR(st.st_mode))) {
+ free(path);
+ path = x_strdup(cache_dir);
+ for (i = 0; i < nlevels; ++i) {
+ char *p = format("%s/%c", path, name[i]);
+ free(path);
+ path = p;
+ if (create_dir(path) != 0) {
+ cc_log("Failed to create %s", path);
+ failed();
+ }
}
}
+
result = format("%s/%s%s", path, name + nlevels, suffix);
free(path);
return result;
___
ccache mailing list
ccache@lists.samba.org
https://lists.samba.org/mailman/listinfo/ccache
