gallium/u_threaded: flush batch when hitting mapping limit

tc_transfer_map maps buffers directly, but the unmap operation is executed
in the driver thread.
When an application does a lot of map/unmap operations, without flushing,
this increase the RAM used (and eventually get the app killed by the oom-killer).

This commit allows tc to keep track of how many bytes were mapped during
the current batch. When this estimation becomes higher than a threshold,
we flush the batch.

See: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2735
Reviewed-by: Marek Olšák <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4508>

2 files changed, 24 insertions, 0 deletions

diff --git a/src/gallium/auxiliary/util/u_threaded_context.c b/src/gallium/auxiliary/util/u_threaded_context.c
index 66e05fb1772..cfe88d310e8 100644
--- a/src/gallium/auxiliary/util/u_threaded_context.c
+++ b/src/gallium/auxiliary/util/u_threaded_context.c
@@ -108,6 +108,7 @@ tc_batch_flush(struct threaded_context *tc)
    tc_assert(next->num_total_call_slots != 0);
    tc_batch_check(next);
    tc_debug_check(tc);
+   tc->bytes_mapped_estimate = 0;
    p_atomic_add(&tc->num_offloaded_slots, next->num_total_call_slots);
 
    if (next->token) {
@@ -204,6 +205,7 @@ _tc_sync(struct threaded_context *tc, UNUSED const char *info, UNUSED const char
    /* .. and execute unflushed calls directly. */
    if (next->num_total_call_slots) {
       p_atomic_add(&tc->num_direct_slots, next->num_total_call_slots);
+      tc->bytes_mapped_estimate = 0;
       tc_batch_execute(next, 0);
       synced = true;
    }
@@ -1489,6 +1491,8 @@ tc_transfer_map(struct pipe_context *_pipe,
                       usage & PIPE_TRANSFER_DISCARD_RANGE ? "  discard_range" :
                       usage & PIPE_TRANSFER_READ ? "  read" : "  ??");
 
+   tc->bytes_mapped_estimate += box->width;
+
    return pipe->transfer_map(pipe, tres->latest ? tres->latest : resource,
                              level, usage, box, transfer);
 }
@@ -1585,6 +1589,10 @@ tc_call_transfer_unmap(struct pipe_context *pipe, union tc_payload *payload)
 }
 
 static void
+tc_flush(struct pipe_context *_pipe, struct pipe_fence_handle **fence,
+         unsigned flags);
+
+static void
 tc_transfer_unmap(struct pipe_context *_pipe, struct pipe_transfer *transfer)
 {
    struct threaded_context *tc = threaded_context(_pipe);
@@ -1606,6 +1614,16 @@ tc_transfer_unmap(struct pipe_context *_pipe, struct pipe_transfer *transfer)
    }
 
    tc_add_small_call(tc, TC_CALL_transfer_unmap)->transfer = transfer;
+
+   /* tc_transfer_map directly maps the buffers, but tc_transfer_unmap
+    * defers the unmap operation to the batch execution.
+    * bytes_mapped_estimate is an estimation of the map/unmap bytes delta
+    * and if it goes over an optional limit the current batch is flushed,
+    * to reclaim some RAM. */
+   if (!ttrans->staging && tc->bytes_mapped_limit &&
+       tc->bytes_mapped_estimate > tc->bytes_mapped_limit) {
+      tc_flush(_pipe, NULL, PIPE_FLUSH_ASYNC);
+   }
 }
 
 struct tc_buffer_subdata {
diff --git a/src/gallium/auxiliary/util/u_threaded_context.h b/src/gallium/auxiliary/util/u_threaded_context.h
index a32f893592a..c54dec0b1d7 100644
--- a/src/gallium/auxiliary/util/u_threaded_context.h
+++ b/src/gallium/auxiliary/util/u_threaded_context.h
@@ -361,6 +361,12 @@ struct threaded_context {
    unsigned num_direct_slots;
    unsigned num_syncs;
 
+   /* Estimation of how much vram/gtt bytes are mmap'd in
+    * the current tc_batch.
+    */
+   uint64_t bytes_mapped_estimate;
+   uint64_t bytes_mapped_limit;
+
    struct util_queue queue;
    struct util_queue_fence *fence;