readahead: chunk on spinning media

Readahead has all sorts of bad side effects depending on your
storage media. On rotating disks, it may be degrading startup
performance if enough requests are queued spanning linearly
over all blocks early at boot, and mount, blkid and friends
want to insert reads to the start of these block devices after.

The end result is that on spinning disks with ext3/4 that udev
and mounts take a very long time, and nothing really happens until
readahead is completely finished.

This has the net effect that the CPU is almost entirely idle
for the entire period that readahead is working. We could have
finished starting up quite a lot of services in this time if
we were smarter at how we do readahead.

This patch sorts all requests into 2 second "chunks" and sub-sorts
each chunk by block. This adds a single cross-drive seek per "chunk"
but has the benefit that we will have a lot of the blocks we need
early on in the boot sequence loaded into memory faster.

For a comparison of how before/after bootcharts look (ext4 on a
mobile 5400rpm 250GB drive) please look at:

    http://foo-projects.org/~sofar/blocked-tests/

There are bootcharts in the "before" and "after" folders where you
should be able to see that many low-level services finish 5-7
seconds earlier with the patch applied (after).

2 files changed, 26 insertions, 4 deletions

diff --git a/Makefile.am b/Makefile.am
index 37c1cc2692..58619760f5 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -2956,7 +2956,7 @@ systemd_readahead_SOURCES = \
 systemd_readahead_LDADD = \
 	libsystemd-shared.la \
 	libsystemd-daemon.la \
-	libudev.la
+	libudev.la -lm
 
 dist_doc_DATA += \
 	src/readahead/sd-readahead.c \
diff --git a/src/readahead/readahead-collect.c b/src/readahead/readahead-collect.c
index 5d07f4704a..5d22949a12 100644
--- a/src/readahead/readahead-collect.c
+++ b/src/readahead/readahead-collect.c
@@ -42,6 +42,7 @@
 #include <sys/vfs.h>
 #include <getopt.h>
 #include <sys/inotify.h>
+#include <math.h>
 
 #ifdef HAVE_FANOTIFY_INIT
 #include <sys/fanotify.h>
@@ -67,6 +68,7 @@
  */
 
 static ReadaheadShared *shared = NULL;
+static struct timespec starttime;
 
 /* Avoid collisions with the NULL pointer */
 #define SECTOR_TO_PTR(s) ULONG_TO_PTR((s)+1)
@@ -205,6 +207,7 @@ static unsigned long fd_first_block(int fd) {
 struct item {
         const char *path;
         unsigned long block;
+        unsigned long bin;
 };
 
 static int qsort_compare(const void *a, const void *b) {
@@ -213,6 +216,13 @@ static int qsort_compare(const void *a, const void *b) {
         i = a;
         j = b;
 
+        /* sort by bin first */
+        if (i->bin < j->bin)
+                return -1;
+        if (i->bin > j->bin)
+                return 1;
+
+        /* then sort by sector */
         if (i->block < j->block)
                 return -1;
         if (i->block > j->block)
@@ -250,6 +260,8 @@ static int collect(const char *root) {
                 goto finish;
         }
 
+        clock_gettime(CLOCK_MONOTONIC, &starttime);
+
         /* If there's no pack file yet we lower the kernel readahead
          * so that mincore() is accurate. If there is a pack file
          * already we assume it is accurate enough so that kernel
@@ -447,10 +459,21 @@ static int collect(const char *root) {
                                         free(p);
                                 else {
                                         unsigned long ul;
+                                        struct timespec ts;
+                                        struct item *entry;
+
+                                        entry = new0(struct item, 1);
 
                                         ul = fd_first_block(m->fd);
 
-                                        if ((k = hashmap_put(files, p, SECTOR_TO_PTR(ul))) < 0) {
+                                        clock_gettime(CLOCK_MONOTONIC, &ts);
+
+                                        entry->block = ul;
+                                        entry->path = strdup(p);
+                                        entry->bin = round((ts.tv_sec - starttime.tv_sec +
+                                                     ((ts.tv_nsec - starttime.tv_nsec) / 1000000000.0)) / 2.0);
+
+                                        if ((k = hashmap_put(files, p, entry)) < 0) {
                                                 log_warning("set_put() failed: %s", strerror(-k));
                                                 free(p);
                                         }
@@ -518,8 +541,7 @@ done:
 
                 j = ordered;
                 HASHMAP_FOREACH_KEY(q, p, files, i) {
-                        j->path = p;
-                        j->block = PTR_TO_SECTOR(q);
+                        memcpy(j, q, sizeof(struct item));
                         j++;
                 }