gfs2: low-memory forced flush fixes
authorAndreas Gruenbacher <agruenba@redhat.com>
Thu, 10 Aug 2023 15:15:46 +0000 (17:15 +0200)
committerAndreas Gruenbacher <agruenba@redhat.com>
Tue, 5 Sep 2023 13:58:16 +0000 (15:58 +0200)
First, function gfs2_ail_flush_reqd checks the SDF_FORCE_AIL_FLUSH flag
to determine if an AIL flush should be forced in low-memory situations.
However, it also immediately clears the flag, and when called repeatedly
as in function gfs2_logd, the flag will be lost.  Fix that by pulling
the SDF_FORCE_AIL_FLUSH flag check out of gfs2_ail_flush_reqd.

Second, function gfs2_writepages sets the SDF_FORCE_AIL_FLUSH flag
whether or not enough pages were written.  If enough pages could be
written, flushing the AIL is unnecessary, though.

Third, gfs2_writepages doesn't wake up logd after setting the
SDF_FORCE_AIL_FLUSH flag, so it can take a long time for logd to react.
It would be preferable to wake up logd, but that hurts the performance
of some workloads and we don't quite understand why so far, so don't
wake up logd so far.

Fixes: b066a4eebd4f ("gfs2: forcibly flush ail to relieve memory pressure")
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
fs/gfs2/aops.c
fs/gfs2/log.c

index 5f02542..add62aa 100644 (file)
@@ -183,13 +183,13 @@ static int gfs2_writepages(struct address_space *mapping,
        int ret;
 
        /*
-        * Even if we didn't write any pages here, we might still be holding
+        * Even if we didn't write enough pages here, we might still be holding
         * dirty pages in the ail. We forcibly flush the ail because we don't
         * want balance_dirty_pages() to loop indefinitely trying to write out
         * pages held in the ail that it can't find.
         */
        ret = iomap_writepages(mapping, wbc, &wpc, &gfs2_writeback_ops);
-       if (ret == 0)
+       if (ret == 0 && wbc->nr_to_write > 0)
                set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
        return ret;
 }
index d3da259..aaca22f 100644 (file)
@@ -1282,9 +1282,6 @@ static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
 {
        unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
 
-       if (test_and_clear_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags))
-               return 1;
-
        return used_blocks + atomic_read(&sdp->sd_log_blks_needed) >=
                atomic_read(&sdp->sd_log_thresh2);
 }
@@ -1325,7 +1322,9 @@ int gfs2_logd(void *data)
                                                  GFS2_LFC_LOGD_JFLUSH_REQD);
                }
 
-               if (gfs2_ail_flush_reqd(sdp)) {
+               if (test_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags) ||
+                   gfs2_ail_flush_reqd(sdp)) {
+                       clear_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
                        gfs2_ail1_start(sdp);
                        gfs2_ail1_wait(sdp);
                        gfs2_ail1_empty(sdp, 0);
@@ -1338,6 +1337,7 @@ int gfs2_logd(void *data)
                try_to_freeze();
 
                t = wait_event_interruptible_timeout(sdp->sd_logd_waitq,
+                               test_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags) ||
                                gfs2_ail_flush_reqd(sdp) ||
                                gfs2_jrnl_flush_reqd(sdp) ||
                                kthread_should_stop(),