]> git.baikalelectronics.ru Git - kernel.git/commitdiff
ring-buffer: Allow splice to read previous partially read pages
authorSteven Rostedt (Google) <rostedt@goodmis.org>
Tue, 27 Sep 2022 18:43:17 +0000 (14:43 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 26 Oct 2022 11:22:19 +0000 (13:22 +0200)
commit 2a46fcd01ae9de4c3b5b89906d936b612a0c2565 upstream.

If a page is partially read, and then the splice system call is run
against the ring buffer, it will always fail to read, no matter how much
is in the ring buffer. That's because the code path for a partial read of
the page does will fail if the "full" flag is set.

The splice system call wants full pages, so if the read of the ring buffer
is not yet full, it should return zero, and the splice will block. But if
a previous read was done, where the beginning has been consumed, it should
still be given to the splice caller if the rest of the page has been
written to.

This caused the splice command to never consume data in this scenario, and
let the ring buffer just fill up and lose events.

Link: https://lkml.kernel.org/r/20220927144317.46be6b80@gandalf.local.home
Cc: stable@vger.kernel.org
Fixes: 29711b6a37032 ("ring-buffer: read page interface")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/trace/ring_buffer.c

index 5e1b9f6e77f31f1436e60c75bd3b1338c18f8e90..708fec4d5df9351a3fbb98e80000c97eac9ec1ff 100644 (file)
@@ -4825,7 +4825,15 @@ int ring_buffer_read_page(struct ring_buffer *buffer,
                unsigned int pos = 0;
                unsigned int size;
 
-               if (full)
+               /*
+                * If a full page is expected, this can still be returned
+                * if there's been a previous partial read and the
+                * rest of the page can be read and the commit page is off
+                * the reader page.
+                */
+               if (full &&
+                   (!read || (len < (commit - read)) ||
+                    cpu_buffer->reader_page == cpu_buffer->commit_page))
                        goto out_unlock;
 
                if (len > (commit - read))