ring-buffer: Include dropped pages in counting dirty patches
authorSteven Rostedt (Google) <rostedt@goodmis.org>
Fri, 21 Oct 2022 16:30:13 +0000 (12:30 -0400)
committerSteven Rostedt (Google) <rostedt@goodmis.org>
Wed, 16 Nov 2022 23:08:29 +0000 (18:08 -0500)
The function ring_buffer_nr_dirty_pages() was created to find out how many
pages are filled in the ring buffer. There's two running counters. One is
incremented whenever a new page is touched (pages_touched) and the other
is whenever a page is read (pages_read). The dirty count is the number
touched minus the number read. This is used to determine if a blocked task
should be woken up if the percentage of the ring buffer it is waiting for
is hit.

The problem is that it does not take into account dropped pages (when the
new writes overwrite pages that were not read). And then the dirty pages
will always be greater than the percentage.

This makes the "buffer_percent" file inaccurate, as the number of dirty
pages end up always being larger than the percentage, event when it's not
and this causes user space to be woken up more than it wants to be.

Add a new counter to keep track of lost pages, and include that in the
accounting of dirty pages so that it is actually accurate.

Link: https://lkml.kernel.org/r/20221021123013.55fb6055@gandalf.local.home
Fixes: 2c2b0a78b3739 ("ring-buffer: Add percentage of ring buffer full to wake up reader")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
kernel/trace/ring_buffer.c

index 089b1ec..a19369c 100644 (file)
@@ -519,6 +519,7 @@ struct ring_buffer_per_cpu {
        local_t                         committing;
        local_t                         commits;
        local_t                         pages_touched;
+       local_t                         pages_lost;
        local_t                         pages_read;
        long                            last_pages_touch;
        size_t                          shortest_full;
@@ -894,10 +895,18 @@ size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
 {
        size_t read;
+       size_t lost;
        size_t cnt;
 
        read = local_read(&buffer->buffers[cpu]->pages_read);
+       lost = local_read(&buffer->buffers[cpu]->pages_lost);
        cnt = local_read(&buffer->buffers[cpu]->pages_touched);
+
+       if (WARN_ON_ONCE(cnt < lost))
+               return 0;
+
+       cnt -= lost;
+
        /* The reader can read an empty page, but not more than that */
        if (cnt < read) {
                WARN_ON_ONCE(read > cnt + 1);
@@ -2031,6 +2040,7 @@ rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
                         */
                        local_add(page_entries, &cpu_buffer->overrun);
                        local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
+                       local_inc(&cpu_buffer->pages_lost);
                }
 
                /*
@@ -2515,6 +2525,7 @@ rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
                 */
                local_add(entries, &cpu_buffer->overrun);
                local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
+               local_inc(&cpu_buffer->pages_lost);
 
                /*
                 * The entries will be zeroed out when we move the
@@ -5265,6 +5276,7 @@ rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
        local_set(&cpu_buffer->committing, 0);
        local_set(&cpu_buffer->commits, 0);
        local_set(&cpu_buffer->pages_touched, 0);
+       local_set(&cpu_buffer->pages_lost, 0);
        local_set(&cpu_buffer->pages_read, 0);
        cpu_buffer->last_pages_touch = 0;
        cpu_buffer->shortest_full = 0;