ring-buffer: read page interface
Steven Rostedt [Tue, 2 Dec 2008 20:34:07 +0000 (15:34 -0500)]
Impact: new API to ring buffer

This patch adds a new interface into the ring buffer that allows a
page to be read from the ring buffer on a given CPU. For every page
read, one must also be given to allow for a "swap" of the pages.

 rpage = ring_buffer_alloc_read_page(buffer);
 if (!rpage)
goto err;
 ret = ring_buffer_read_page(buffer, &rpage, cpu, full);
 if (!ret)
goto empty;
 process_page(rpage);
 ring_buffer_free_read_page(rpage);

The caller of these functions must handle any waits that are
needed to wait for new data. The ring_buffer_read_page will simply
return 0 if there is no data, or if "full" is set and the writer
is still on the current page.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

include/linux/ring_buffer.h
kernel/trace/ring_buffer.c

index 3bb87a7..1a350a8 100644 (file)
@@ -124,6 +124,11 @@ void tracing_on(void);
 void tracing_off(void);
 void tracing_off_permanent(void);
 
+void *ring_buffer_alloc_read_page(struct ring_buffer *buffer);
+void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data);
+int ring_buffer_read_page(struct ring_buffer *buffer,
+                         void **data_page, int cpu, int full);
+
 enum ring_buffer_flags {
        RB_FL_OVERWRITE         = 1 << 0,
 };
index 8619c53..50b74d3 100644 (file)
@@ -687,6 +687,12 @@ static inline int rb_null_event(struct ring_buffer_event *event)
        return event->type == RINGBUF_TYPE_PADDING;
 }
 
+static inline void *
+__rb_data_page_index(struct buffer_data_page *page, unsigned index)
+{
+       return page->data + index;
+}
+
 static inline void *__rb_page_index(struct buffer_page *page, unsigned index)
 {
        return page->page->data + index;
@@ -2232,6 +2238,166 @@ int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
        return 0;
 }
 
+static void rb_remove_entries(struct ring_buffer_per_cpu *cpu_buffer,
+                             struct buffer_data_page *page)
+{
+       struct ring_buffer_event *event;
+       unsigned long head;
+
+       __raw_spin_lock(&cpu_buffer->lock);
+       for (head = 0; head < local_read(&page->commit);
+            head += rb_event_length(event)) {
+
+               event = __rb_data_page_index(page, head);
+               if (RB_WARN_ON(cpu_buffer, rb_null_event(event)))
+                       return;
+               /* Only count data entries */
+               if (event->type != RINGBUF_TYPE_DATA)
+                       continue;
+               cpu_buffer->entries--;
+       }
+       __raw_spin_unlock(&cpu_buffer->lock);
+}
+
+/**
+ * ring_buffer_alloc_read_page - allocate a page to read from buffer
+ * @buffer: the buffer to allocate for.
+ *
+ * This function is used in conjunction with ring_buffer_read_page.
+ * When reading a full page from the ring buffer, these functions
+ * can be used to speed up the process. The calling function should
+ * allocate a few pages first with this function. Then when it
+ * needs to get pages from the ring buffer, it passes the result
+ * of this function into ring_buffer_read_page, which will swap
+ * the page that was allocated, with the read page of the buffer.
+ *
+ * Returns:
+ *  The page allocated, or NULL on error.
+ */
+void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
+{
+       unsigned long addr;
+       struct buffer_data_page *page;
+
+       addr = __get_free_page(GFP_KERNEL);
+       if (!addr)
+               return NULL;
+
+       page = (void *)addr;
+
+       return page;
+}
+
+/**
+ * ring_buffer_free_read_page - free an allocated read page
+ * @buffer: the buffer the page was allocate for
+ * @data: the page to free
+ *
+ * Free a page allocated from ring_buffer_alloc_read_page.
+ */
+void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
+{
+       free_page((unsigned long)data);
+}
+
+/**
+ * ring_buffer_read_page - extract a page from the ring buffer
+ * @buffer: buffer to extract from
+ * @data_page: the page to use allocated from ring_buffer_alloc_read_page
+ * @cpu: the cpu of the buffer to extract
+ * @full: should the extraction only happen when the page is full.
+ *
+ * This function will pull out a page from the ring buffer and consume it.
+ * @data_page must be the address of the variable that was returned
+ * from ring_buffer_alloc_read_page. This is because the page might be used
+ * to swap with a page in the ring buffer.
+ *
+ * for example:
+ *     rpage = ring_buffer_alloc_page(buffer);
+ *     if (!rpage)
+ *             return error;
+ *     ret = ring_buffer_read_page(buffer, &rpage, cpu, 0);
+ *     if (ret)
+ *             process_page(rpage);
+ *
+ * When @full is set, the function will not return true unless
+ * the writer is off the reader page.
+ *
+ * Note: it is up to the calling functions to handle sleeps and wakeups.
+ *  The ring buffer can be used anywhere in the kernel and can not
+ *  blindly call wake_up. The layer that uses the ring buffer must be
+ *  responsible for that.
+ *
+ * Returns:
+ *  1 if data has been transferred
+ *  0 if no data has been transferred.
+ */
+int ring_buffer_read_page(struct ring_buffer *buffer,
+                           void **data_page, int cpu, int full)
+{
+       struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
+       struct ring_buffer_event *event;
+       struct buffer_data_page *page;
+       unsigned long flags;
+       int ret = 0;
+
+       if (!data_page)
+               return 0;
+
+       page = *data_page;
+       if (!page)
+               return 0;
+
+       spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
+
+       /*
+        * rb_buffer_peek will get the next ring buffer if
+        * the current reader page is empty.
+        */
+       event = rb_buffer_peek(buffer, cpu, NULL);
+       if (!event)
+               goto out;
+
+       /* check for data */
+       if (!local_read(&cpu_buffer->reader_page->page->commit))
+               goto out;
+       /*
+        * If the writer is already off of the read page, then simply
+        * switch the read page with the given page. Otherwise
+        * we need to copy the data from the reader to the writer.
+        */
+       if (cpu_buffer->reader_page == cpu_buffer->commit_page) {
+               unsigned int read = cpu_buffer->reader_page->read;
+
+               if (full)
+                       goto out;
+               /* The writer is still on the reader page, we must copy */
+               page = cpu_buffer->reader_page->page;
+               memcpy(page->data,
+                      cpu_buffer->reader_page->page->data + read,
+                      local_read(&page->commit) - read);
+
+               /* consume what was read */
+               cpu_buffer->reader_page += read;
+
+       } else {
+               /* swap the pages */
+               rb_init_page(page);
+               page = cpu_buffer->reader_page->page;
+               cpu_buffer->reader_page->page = *data_page;
+               cpu_buffer->reader_page->read = 0;
+               *data_page = page;
+       }
+       ret = 1;
+
+       /* update the entry counter */
+       rb_remove_entries(cpu_buffer, page);
+ out:
+       spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
+
+       return ret;
+}
+
 static ssize_t
 rb_simple_read(struct file *filp, char __user *ubuf,
               size_t cnt, loff_t *ppos)