fix: mm/writeback: discard NR_UNSTABLE_NFS, use NR_WRITEBACK (v5.8)
authorMichael Jeanson <mjeanson@efficios.com>
Mon, 15 Jun 2020 15:12:24 +0000 (11:12 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Mon, 15 Jun 2020 19:57:32 +0000 (15:57 -0400)
See upstream commit:

  commit 8d92890bd6b8502d6aee4b37430ae6444ade7a8c
  Author: NeilBrown <neilb@suse.de>
  Date:   Mon Jun 1 21:48:21 2020 -0700

    mm/writeback: discard NR_UNSTABLE_NFS, use NR_WRITEBACK instead

    After an NFS page has been written it is considered "unstable" until a
    COMMIT request succeeds.  If the COMMIT fails, the page will be
    re-written.

    These "unstable" pages are currently accounted as "reclaimable", either
    in WB_RECLAIMABLE, or in NR_UNSTABLE_NFS which is included in a
    'reclaimable' count.  This might have made sense when sending the COMMIT
    required a separate action by the VFS/MM (e.g.  releasepage() used to
    send a COMMIT).  However now that all writes generated by ->writepages()
    will automatically be followed by a COMMIT (since commit 919e3bd9a875
    ("NFS: Ensure we commit after writeback is complete")) it makes more
    sense to treat them as writeback pages.

    So this patch removes NR_UNSTABLE_NFS and accounts unstable pages in
    NR_WRITEBACK and WB_WRITEBACK.

    A particular effect of this change is that when
    wb_check_background_flush() calls wb_over_bg_threshold(), the latter
    will report 'true' a lot less often as the 'unstable' pages are no
    longer considered 'dirty' (as there is nothing that writeback can do
    about them anyway).

    Currently wb_check_background_flush() will trigger writeback to NFS even
    when there are relatively few dirty pages (if there are lots of unstable
    pages), this can result in small writes going to the server (10s of
    Kilobytes rather than a Megabyte) which hurts throughput.  With this
    patch, there are fewer writes which are each larger on average.

    Where the NR_UNSTABLE_NFS count was included in statistics
    virtual-files, the entry is retained, but the value is hard-coded as
    zero.  static trace points and warning printks which mentioned this
    counter no longer report it.

Change-Id: I18080ca62bc6c1cd7d6da4cb27cc1521fbdca5e1
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
instrumentation/events/lttng-module/writeback.h

index feef94503d8376f7d08aa364c46843887d91a9a3..affb4eb55f06b3f0057a8ced2e1c3e97465bc75c 100644 (file)
@@ -399,7 +399,30 @@ LTTNG_TRACEPOINT_EVENT(writeback_queue_io,
        )
 )
 
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,14,0))
+#if (LINUX_VERSION_CODE >= KERNEL_VERSION(5,8,0))
+LTTNG_TRACEPOINT_EVENT_MAP(global_dirty_state,
+
+       writeback_global_dirty_state,
+
+       TP_PROTO(unsigned long background_thresh,
+                unsigned long dirty_thresh
+       ),
+
+       TP_ARGS(background_thresh,
+               dirty_thresh
+       ),
+
+       TP_FIELDS(
+               ctf_integer(unsigned long, nr_dirty, global_node_page_state(NR_FILE_DIRTY))
+               ctf_integer(unsigned long, nr_writeback, global_node_page_state(NR_WRITEBACK))
+               ctf_integer(unsigned long, nr_dirtied, global_node_page_state(NR_DIRTIED))
+               ctf_integer(unsigned long, nr_written, global_node_page_state(NR_WRITTEN))
+               ctf_integer(unsigned long, background_thresh, background_thresh)
+               ctf_integer(unsigned long, dirty_thresh, dirty_thresh)
+               ctf_integer(unsigned long, dirty_limit, global_dirty_limit)
+       )
+)
+#elif (LINUX_VERSION_CODE >= KERNEL_VERSION(4,14,0))
 LTTNG_TRACEPOINT_EVENT_MAP(global_dirty_state,
 
        writeback_global_dirty_state,
This page took 0.027203 seconds and 4 git commands to generate.