virgl: Add perfetto performance calls conditionally

Author: Tomeu Vizoso <tomeu.vizoso@collabora.com> (v1)
v2: - Include perfetto_min.h only when perf tracing is enabled
    - Make stubs inline code so it can be optimized away.
    - Add compile with tracing info to configuration status
    - Use static for stubs (Chia-I Wu)
    - pass trace option as a string and add defines so that
      other tracing tools cold be accomodated
      (inspired by Chia-I Wu & Tomeu)
    - Use macros to wrap the trace calls

v3: - Use __attribute__((cleanup)) to handle trace_begin/trace_end
      instead of using macros to wrap the calls (Chia-I Wu)
    - remove cpp from language support

v4: - Fix __attribute__ check in meson.build
    - replace TRACE_FUNC macro to not take a paramter
    - move trace initialization to a seperate function
      (all suggestions by Chia-I Wu)

v5: use vsnprintf to avoid buffer overflow (John Bates)

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com> (v4)

config.h.meson

@@ -7,3 +7,4 @@
 #mesondefine HAVE_FUNC_ATTRIBUTE_VISIBILITY
 #mesondefine HAVE_EVENTFD_H
 #mesondefine HAVE_DLFCN_H
+#mesondefine ENABLE_TRACING

meson.build

@@ -82,6 +82,20 @@ m_dep = cc.find_library('m')
 conf_data = configuration_data()
 conf_data.set('VERSION', '0.8.1')
 
+with_tracing = get_option('tracing')
+
+if with_tracing != 'none'
+  if not cc.compiles('void f(void* v){}; int main () { void *dummy __attribute__((cleanup (f))) = 0;}')
+     error('Tracing requires compiler support for __attribute__((cleanup))')
+endif
+
+endif
+
+if with_tracing == 'perfetto'
+   vperfetto_min_dep = dependency('vperfetto_min')
+   conf_data.set('ENABLE_TRACING', 'TRACE_WITH_PERFETTO')
+endif
+
 if cc.has_header('sys/uio.h')
    conf_data.set('HAVE_SYS_UIO_H', 1)
 endif
@@ -199,6 +213,7 @@ lines += 'minigbm_alloc: ' + (with_minigbm_allocation ? 'yes' : 'no' )
 lines += ''
 lines += 'tests:     ' + (with_tests ? 'yes' : 'no' )
 lines += 'fuzzer:    ' + (with_fuzzer ? 'yes' : 'no' )
+lines += 'tracing:   ' + with_tracing
 
 indent = '   '
 summary = indent + ('\n' + indent).join(lines)

meson_options.txt

@@ -58,3 +58,11 @@ option(
   value : 'false',
   description : 'enable running unit tests with valgrind'
 )
+
+option(
+  'tracing',
+  type : 'combo',
+  value : 'none',
+  choices : [ 'perfetto', 'none' ],
+  description : 'enable emitting traces for Perfetto'
+)

src/meson.build

@@ -79,9 +79,13 @@ virgl_depends = [
    epoxy_dep,
    libdrm_dep,
    thread_dep,
-   m_dep
+   m_dep,
 ]
 
+if with_tracing == 'perfetto'
+   virgl_depends += [vperfetto_min_dep]
+endif
+
 virgl_sources += vrend_sources
 
 if have_egl

src/virgl_util.c

@@ -36,6 +36,17 @@
 
 #include "util/u_pointer.h"
 
+#include <stdarg.h>
+#include <stdio.h>
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#if ENABLE_TRACING == TRACE_WITH_PERFETTO
+#include <vperfetto-min.h>
+#endif
+
 unsigned hash_func_u32(void *key)
 {
    intptr_t ip = pointer_to_intptr(key);
@@ -98,3 +109,33 @@ void flush_eventfd(int fd)
        len = read(fd, &value, sizeof(value));
     } while ((len == -1 && errno == EINTR) || len == sizeof(value));
 }
+
+#if ENABLE_TRACING == TRACE_WITH_PERFETTO
+void trace_init(void)
+{
+   struct vperfetto_min_config config = {
+      .init_flags = VPERFETTO_INIT_FLAG_USE_SYSTEM_BACKEND,
+            .filename = NULL,
+            .shmem_size_hint_kb = 32 * 1024,
+   };
+
+   vperfetto_min_startTracing(&config);
+}
+
+char *trace_begin(const char* format, ...)
+{
+   char buffer[1024];
+   va_list args;
+   va_start (args, format);
+   vsnprintf (buffer, sizeof(buffer), format, args);
+   va_end (args);
+   vperfetto_min_beginTrackEvent_VMM(buffer);
+   return (void *)1;
+}
+
+void trace_end(char **dummy)
+{
+   (void)dummy;
+   vperfetto_min_endTrackEvent_VMM();
+}
+#endif

src/virgl_util.h

@@ -28,6 +28,13 @@
 #include <stdint.h>
 #include <stdbool.h>
 
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#define TRACE_WITH_PERFETTO 1
+
+
 #define BIT(n)                   (UINT32_C(1) << (n))
 
 static inline bool has_bit(uint32_t mask, uint32_t bit)
@@ -54,4 +61,24 @@ int create_eventfd(unsigned int initval);
 int write_eventfd(int fd, uint64_t val);
 void flush_eventfd(int fd);
 
+#ifdef ENABLE_TRACING
+void trace_init(void);
+char *trace_begin(const char* format, ...);
+void trace_end(char **dummy);
+
+#define TRACE_INIT() trace_init()
+#define TRACE_FUNC() \
+   char *trace_dummy __attribute__((cleanup (trace_end), unused)) = \
+   trace_begin("%s", __func__)
+
+#define TRACE_SCOPE(FORMAT, ...) \
+   char *trace_dummy __attribute__((cleanup (trace_end), unused)) = \
+   trace_begin(FORMAT, __VA_ARGS__)
+
+#else
+#define TRACE_INIT()
+#define TRACE_FUNC()
+#define TRACE_SCOPE(FORMAT, ...)
+#endif
+
 #endif /* VIRGL_UTIL_H */

src/virglrenderer.c

@@ -31,6 +31,7 @@
 #include <fcntl.h>
 #include <errno.h>
 #include <unistd.h>
+
 #include "pipe/p_state.h"
 #include "util/u_format.h"
 #include "util/u_math.h"
@@ -99,11 +100,13 @@ static int virgl_renderer_resource_create_internal(struct virgl_renderer_resourc
 int virgl_renderer_resource_create(struct virgl_renderer_resource_create_args *args,
                                    struct iovec *iov, uint32_t num_iovs)
 {
+   TRACE_FUNC();
    return virgl_renderer_resource_create_internal(args, iov, num_iovs, NULL);
 }
 
 int virgl_renderer_resource_import_eglimage(struct virgl_renderer_resource_create_args *args, void *image)
 {
+   TRACE_FUNC();
    return virgl_renderer_resource_create_internal(args, NULL, 0, image);
 }
 
@@ -165,6 +168,8 @@ int virgl_renderer_context_create(uint32_t handle, uint32_t nlen, const char *na
    struct virgl_context *ctx;
    int ret;
 
+   TRACE_FUNC();
+
    /* user context id must be greater than 0 */
    if (handle == 0)
       return EINVAL;
@@ -187,6 +192,7 @@ int virgl_renderer_context_create(uint32_t handle, uint32_t nlen, const char *na
 
 void virgl_renderer_context_destroy(uint32_t handle)
 {
+   TRACE_FUNC();
    virgl_context_remove(handle);
 }
 
@@ -194,6 +200,7 @@ int virgl_renderer_submit_cmd(void *buffer,
                               int ctx_id,
                               int ndw)
 {
+   TRACE_FUNC();
    struct virgl_context *ctx = virgl_context_lookup(ctx_id);
    if (!ctx)
       return EINVAL;
@@ -210,6 +217,8 @@ int virgl_renderer_transfer_write_iov(uint32_t handle,
                                       struct iovec *iovec,
                                       unsigned int iovec_cnt)
 {
+   TRACE_FUNC();
+
    struct virgl_resource *res = virgl_resource_lookup(handle);
    struct vrend_transfer_info transfer_info;
 
@@ -248,6 +257,7 @@ int virgl_renderer_transfer_read_iov(uint32_t handle, uint32_t ctx_id,
                                      uint64_t offset, struct iovec *iovec,
                                      int iovec_cnt)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(handle);
    struct vrend_transfer_info transfer_info;
 
@@ -282,6 +292,7 @@ int virgl_renderer_transfer_read_iov(uint32_t handle, uint32_t ctx_id,
 int virgl_renderer_resource_attach_iov(int res_handle, struct iovec *iov,
                                        int num_iovs)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!res)
       return EINVAL;
@@ -291,6 +302,7 @@ int virgl_renderer_resource_attach_iov(int res_handle, struct iovec *iov,
 
 void virgl_renderer_resource_detach_iov(int res_handle, struct iovec **iov_p, int *num_iovs_p)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!res)
       return;
@@ -305,16 +317,19 @@ void virgl_renderer_resource_detach_iov(int res_handle, struct iovec **iov_p, in
 
 int virgl_renderer_create_fence(int client_fence_id, uint32_t ctx_id)
 {
+   TRACE_FUNC();
    return vrend_renderer_create_fence(client_fence_id, ctx_id);
 }
 
 void virgl_renderer_force_ctx_0(void)
 {
+   TRACE_FUNC();
    vrend_renderer_force_ctx_0();
 }
 
 void virgl_renderer_ctx_attach_resource(int ctx_id, int res_handle)
 {
+   TRACE_FUNC();
    struct virgl_context *ctx = virgl_context_lookup(ctx_id);
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!ctx || !res)
@@ -324,6 +339,7 @@ void virgl_renderer_ctx_attach_resource(int ctx_id, int res_handle)
 
 void virgl_renderer_ctx_detach_resource(int ctx_id, int res_handle)
 {
+   TRACE_FUNC();
    struct virgl_context *ctx = virgl_context_lookup(ctx_id);
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!ctx || !res)
@@ -334,6 +350,7 @@ void virgl_renderer_ctx_detach_resource(int ctx_id, int res_handle)
 int virgl_renderer_resource_get_info(int res_handle,
                                      struct virgl_renderer_resource_info *info)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
 
    if (!res || !res->pipe_resource)
@@ -357,6 +374,7 @@ int virgl_renderer_resource_get_info(int res_handle,
 void virgl_renderer_get_cap_set(uint32_t cap_set, uint32_t *max_ver,
                                 uint32_t *max_size)
 {
+   TRACE_FUNC();
    switch (cap_set) {
    case VIRGL_RENDERER_CAPSET_VIRGL:
    case VIRGL_RENDERER_CAPSET_VIRGL2:
@@ -372,6 +390,7 @@ void virgl_renderer_get_cap_set(uint32_t cap_set, uint32_t *max_ver,
 void virgl_renderer_get_rect(int resource_id, struct iovec *iov, unsigned int num_iovs,
                              uint32_t offset, int x, int y, int width, int height)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(resource_id);
    if (!res || !res->pipe_resource)
       return;
@@ -439,12 +458,14 @@ void *virgl_renderer_get_cursor_data(uint32_t resource_id, uint32_t *width, uint
 
 void virgl_renderer_poll(void)
 {
+   TRACE_FUNC();
    if (state.vrend_initialized)
       vrend_renderer_check_fences();
 }
 
 void virgl_renderer_cleanup(UNUSED void *cookie)
 {
+   TRACE_FUNC();
    if (state.vrend_initialized)
       vrend_renderer_prepare_reset();
 
@@ -465,6 +486,9 @@ void virgl_renderer_cleanup(UNUSED void *cookie)
 
 int virgl_renderer_init(void *cookie, int flags, struct virgl_renderer_callbacks *cbs)
 {
+   TRACE_INIT();
+   TRACE_FUNC();
+
    int ret;
 
    /* VIRGL_RENDERER_THREAD_SYNC is a hint and can be silently ignored */
@@ -543,6 +567,7 @@ fail:
 
 int virgl_renderer_get_fd_for_texture(uint32_t tex_id, int *fd)
 {
+   TRACE_FUNC();
    if (state.winsys_initialized)
       return vrend_winsys_get_fd_for_texture(tex_id, fd);
    return -1;
@@ -550,6 +575,7 @@ int virgl_renderer_get_fd_for_texture(uint32_t tex_id, int *fd)
 
 int virgl_renderer_get_fd_for_texture2(uint32_t tex_id, int *fd, int *stride, int *offset)
 {
+   TRACE_FUNC();
    if (state.winsys_initialized)
       return vrend_winsys_get_fd_for_texture2(tex_id, fd, stride, offset);
    return -1;
@@ -557,6 +583,7 @@ int virgl_renderer_get_fd_for_texture2(uint32_t tex_id, int *fd, int *stride, in
 
 void virgl_renderer_reset(void)
 {
+   TRACE_FUNC();
    if (state.vrend_initialized)
       vrend_renderer_prepare_reset();
 
@@ -572,6 +599,7 @@ void virgl_renderer_reset(void)
 
 int virgl_renderer_get_poll_fd(void)
 {
+   TRACE_FUNC();
    if (state.vrend_initialized)
       return vrend_renderer_get_poll_fd();
 
@@ -622,6 +650,7 @@ static int virgl_renderer_supported_structures(void *execute_args, uint32_t exec
 
 int virgl_renderer_execute(void *execute_args, uint32_t execute_size)
 {
+   TRACE_FUNC();
    struct virgl_renderer_hdr *hdr = execute_args;
    if (hdr->stype_version != 0)
       return -EINVAL;
@@ -638,6 +667,7 @@ int virgl_renderer_execute(void *execute_args, uint32_t execute_size)
 
 int virgl_renderer_resource_create_blob(const struct virgl_renderer_resource_create_blob_args *args)
 {
+   TRACE_FUNC();
    struct virgl_context *ctx;
    struct virgl_context_blob blob;
    bool has_host_storage;
@@ -719,6 +749,7 @@ int virgl_renderer_resource_create_blob(const struct virgl_renderer_resource_cre
 
 int virgl_renderer_resource_map(uint32_t res_handle, void **map, uint64_t *out_size)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!res || !res->pipe_resource)
       return -EINVAL;
@@ -728,6 +759,7 @@ int virgl_renderer_resource_map(uint32_t res_handle, void **map, uint64_t *out_s
 
 int virgl_renderer_resource_unmap(uint32_t res_handle)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!res || !res->pipe_resource)
       return -EINVAL;
@@ -737,6 +769,7 @@ int virgl_renderer_resource_unmap(uint32_t res_handle)
 
 int virgl_renderer_resource_get_map_info(uint32_t res_handle, uint32_t *map_info)
 {
+   TRACE_FUNC();
    struct virgl_resource *res = virgl_resource_lookup(res_handle);
    if (!res || !res->pipe_resource)
       return -EINVAL;
@@ -768,5 +801,6 @@ virgl_renderer_resource_export_blob(uint32_t res_id, uint32_t *fd_type, int *fd)
 int
 virgl_renderer_export_fence(uint32_t client_fence_id, int *fd)
 {
+   TRACE_FUNC();
    return vrend_renderer_export_fence(client_fence_id, fd);
 }

src/vrend_decode.c

@@ -26,6 +26,7 @@
 #include <stdio.h>
 #include <errno.h>
 #include <epoxy/gl.h>
+#include <fcntl.h>
 
 #include "util/u_memory.h"
 #include "pipe/p_defines.h"
@@ -38,6 +39,7 @@
 #include "tgsi/tgsi_text.h"
 #include "vrend_debug.h"
 #include "vrend_tweaks.h"
+#include "virgl_util.h"
 
 /* decode side */
 #define DECODE_MAX_TOKENS 8000
@@ -764,6 +766,8 @@ static int vrend_decode_create_object(struct vrend_decode_ctx *ctx, int length)
    VREND_DEBUG(dbg_object, ctx->grctx,"  CREATE %-18s handle:0x%x len:%d\n",
                vrend_get_object_type_name(obj_type), handle, length);
 
+   TRACE_SCOPE("CREATE %-18s", vrend_get_object_type_name(obj_type));
+
    switch (obj_type){
    case VIRGL_OBJECT_BLEND:
       ret = vrend_decode_create_blend(ctx, handle, length);
@@ -1466,6 +1470,7 @@ struct virgl_context *vrend_renderer_context_create(uint32_t handle,
 
 static void vrend_decode_ctx_destroy(struct virgl_context *ctx)
 {
+   TRACE_FUNC();
    struct vrend_decode_ctx *dctx = (struct vrend_decode_ctx *)ctx;
 
    vrend_destroy_context(dctx->grctx);
@@ -1475,8 +1480,8 @@ static void vrend_decode_ctx_destroy(struct virgl_context *ctx)
 static void vrend_decode_ctx_attach_resource(struct virgl_context *ctx,
                                              struct virgl_resource *res)
 {
+   TRACE_FUNC();
    struct vrend_decode_ctx *dctx = (struct vrend_decode_ctx *)ctx;
-
    /* in the future, we should import to create the pipe resource */
    if (!res->pipe_resource)
       return;
@@ -1488,6 +1493,7 @@ static void vrend_decode_ctx_attach_resource(struct virgl_context *ctx,
 static void vrend_decode_ctx_detach_resource(struct virgl_context *ctx,
                                              struct virgl_resource *res)
 {
+   TRACE_FUNC();
    struct vrend_decode_ctx *dctx = (struct vrend_decode_ctx *)ctx;
    vrend_renderer_detach_res_ctx(dctx->grctx, res->res_id);
 }
@@ -1497,6 +1503,7 @@ static int vrend_decode_ctx_transfer_3d(struct virgl_context *ctx,
                                         const struct vrend_transfer_info *info,
                                         int transfer_mode)
 {
+   TRACE_FUNC();
    struct vrend_decode_ctx *dctx = (struct vrend_decode_ctx *)ctx;
    return vrend_renderer_transfer_iov(dctx->grctx, res->res_id, info,
                                       transfer_mode);
@@ -1507,6 +1514,7 @@ static int vrend_decode_ctx_get_blob(struct virgl_context *ctx,
                                      UNUSED uint32_t blob_flags,
                                      struct virgl_context_blob *blob)
 {
+   TRACE_FUNC();
    struct vrend_decode_ctx *dctx = (struct vrend_decode_ctx *)ctx;
 
    blob->type = VIRGL_RESOURCE_FD_INVALID;
@@ -1520,6 +1528,7 @@ static int vrend_decode_ctx_submit_cmd(struct virgl_context *ctx,
                                        const void *buffer,
                                        size_t size)
 {
+   TRACE_FUNC();
    struct vrend_decode_ctx *gdctx = (struct vrend_decode_ctx *)ctx;
    bool bret;
    int ret;
@@ -1546,6 +1555,8 @@ static int vrend_decode_ctx_submit_cmd(struct virgl_context *ctx,
       VREND_DEBUG(dbg_cmd, gdctx->grctx,"%-4d %-20s len:%d\n",
                   gdctx->ds->buf_offset, vrend_get_comand_name(header & 0xff), len);
 
+      TRACE_SCOPE("%s", vrend_get_comand_name(header & 0xff));
+
       switch (header & 0xff) {
       case VIRGL_CCMD_CREATE_OBJECT:
          ret = vrend_decode_create_object(gdctx, len);