Fix: mmap: caches aliased on virtual addresses

[lttng-modules.git] / lib / ringbuffer / ring_buffer_frontend.c

diff --git a/lib/ringbuffer/ring_buffer_frontend.c b/lib/ringbuffer/ring_buffer_frontend.c
index a477832ef45a9d30914d1dff1473d1e119e4e1d4..3d60c1435c329baf32a27ed9f3148dc8ed8ad88f 100644 (file)
--- a/lib/ringbuffer/ring_buffer_frontend.c
+++ b/lib/ringbuffer/ring_buffer_frontend.c
@@ -54,6 +54,7 @@
 #include <linux/delay.h>
 #include <linux/module.h>
 #include <linux/percpu.h>
+#include <asm/cacheflush.h>
 
 #include <wrapper/ringbuffer/config.h>
 #include <wrapper/ringbuffer/backend.h>
@@ -1180,6 +1181,47 @@ void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf,
 }
 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer);
 
+#if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
+static void lib_ring_buffer_flush_read_subbuf_dcache(
+               const struct lib_ring_buffer_config *config,
+               struct channel *chan,
+               struct lib_ring_buffer *buf)
+{
+       struct lib_ring_buffer_backend_pages *pages;
+       unsigned long sb_bindex, id, i, nr_pages;
+
+       if (config->output != RING_BUFFER_MMAP)
+               return;
+
+       /*
+        * Architectures with caches aliased on virtual addresses may
+        * use different cache lines for the linear mapping vs
+        * user-space memory mapping. Given that the ring buffer is
+        * based on the kernel linear mapping, aligning it with the
+        * user-space mapping is not straightforward, and would require
+        * extra TLB entries. Therefore, simply flush the dcache for the
+        * entire sub-buffer before reading it.
+        */
+       id = buf->backend.buf_rsb.id;
+       sb_bindex = subbuffer_id_get_index(config, id);
+       pages = buf->backend.array[sb_bindex];
+       nr_pages = buf->backend.num_pages_per_subbuf;
+       for (i = 0; i < nr_pages; i++) {
+               struct lib_ring_buffer_backend_page *backend_page;
+
+               backend_page = &pages->p[i];
+               flush_dcache_page(pfn_to_page(backend_page->pfn));
+       }
+}
+#else
+static void lib_ring_buffer_flush_read_subbuf_dcache(
+               const struct lib_ring_buffer_config *config,
+               struct channel *chan,
+               struct lib_ring_buffer *buf)
+{
+}
+#endif
+
 /**
  * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
  * @buf: ring buffer
@@ -1322,6 +1364,8 @@ retry:
        buf->get_subbuf_consumed = consumed;
        buf->get_subbuf = 1;
 
+       lib_ring_buffer_flush_read_subbuf_dcache(config, chan, buf);
+
        return 0;
 
 nodata:
@@ -1495,8 +1539,7 @@ void lib_ring_buffer_print_errors(struct channel *chan,
 /*
  * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
  *
- * Only executed by SWITCH_FLUSH, which can be issued while tracing is active
- * or at buffer finalization (destroy).
+ * Only executed when the buffer is finalized, in SWITCH_FLUSH.
  */
 static
 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf,
@@ -1692,14 +1735,12 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode,
                unsigned long sb_index, commit_count;
 
                /*
-                * We are performing a SWITCH_FLUSH. There may be concurrent
-                * writes into the buffer if e.g. invoked while performing a
-                * snapshot on an active trace.
+                * We are performing a SWITCH_FLUSH. At this stage, there are no
+                * concurrent writes into the buffer.
                 *
-                * If the client does not save any header information (sub-buffer
-                * header size == 0), don't switch empty subbuffer on finalize,
-                * because it is invalid to deliver a completely empty
-                * subbuffer.
+                * The client does not save any header information.  Don't
+                * switch empty subbuffer on finalize, because it is invalid to
+                * deliver a completely empty subbuffer.
                 */
                if (!config->cb.subbuffer_header_size())
                        return -1;
@@ -1864,6 +1905,7 @@ static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf,
        put_online_cpus();
 }
 
+/* Switch sub-buffer if current sub-buffer is non-empty. */
 void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf)
 {
        _lib_ring_buffer_switch_remote(buf, SWITCH_ACTIVE);
@@ -1888,7 +1930,8 @@ static
 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf,
                                     struct channel *chan,
                                     struct switch_offsets *offsets,
-                                    struct lib_ring_buffer_ctx *ctx)
+                                    struct lib_ring_buffer_ctx *ctx,
+                                    void *client_ctx)
 {
        const struct lib_ring_buffer_config *config = &chan->backend.config;
        unsigned long reserve_commit_diff, offset_cmp;
@@ -1914,7 +1957,7 @@ retry:
                offsets->size = config->cb.record_header_size(config, chan,
                                                offsets->begin,
                                                &offsets->pre_header_padding,
-                                               ctx);
+                                               ctx, client_ctx);
                offsets->size +=
                        lib_ring_buffer_align(offsets->begin + offsets->size,
                                              ctx->largest_align)
@@ -1998,7 +2041,7 @@ retry:
                        config->cb.record_header_size(config, chan,
                                                offsets->begin,
                                                &offsets->pre_header_padding,
-                                               ctx);
+                                               ctx, client_ctx);
                offsets->size +=
                        lib_ring_buffer_align(offsets->begin + offsets->size,
                                              ctx->largest_align)
@@ -2062,7 +2105,8 @@ EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big);
  * -EIO for other errors, else returns 0.
  * It will take care of sub-buffer switching.
  */
-int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx)
+int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx,
+               void *client_ctx)
 {
        struct channel *chan = ctx->chan;
        const struct lib_ring_buffer_config *config = &chan->backend.config;
@@ -2075,7 +2119,7 @@ int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx)
 
        do {
                ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets,
-                                                      ctx);
+                                                      ctx, client_ctx);
                if (unlikely(ret))
                        return ret;
        } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old,