freedreno/drm: Add rd dumper support

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30083>

7 files changed, 108 insertions, 5 deletions

diff --git a/src/freedreno/common/freedreno_rd_output.c b/src/freedreno/common/freedreno_rd_output.c
index db42ea5956b..231b6f33c9c 100644
--- a/src/freedreno/common/freedreno_rd_output.c
+++ b/src/freedreno/common/freedreno_rd_output.c
@@ -70,7 +70,7 @@ fd_rd_output_sanitize_name(char *name)
 }
 
 void
-fd_rd_output_init(struct fd_rd_output *output, char* output_name)
+fd_rd_output_init(struct fd_rd_output *output, const char* output_name)
 {
    const char *test_name = os_get_option("FD_RD_DUMP_TESTNAME");
    ASSERTED int name_len;
@@ -90,7 +90,7 @@ fd_rd_output_init(struct fd_rd_output *output, char* output_name)
       output->combine = true;
 
       char file_path[PATH_MAX];
-      snprintf(file_path, sizeof(file_path), "%s/%s_combined.rd",
+      snprintf(file_path, sizeof(file_path), "%s/%s_combined.rd.gz",
                fd_rd_output_base_path, output->name);
       output->file = gzopen(file_path, "w");
    }
diff --git a/src/freedreno/common/freedreno_rd_output.h b/src/freedreno/common/freedreno_rd_output.h
index 7d597b79ebe..b3d2aebe33e 100644
--- a/src/freedreno/common/freedreno_rd_output.h
+++ b/src/freedreno/common/freedreno_rd_output.h
@@ -44,7 +44,7 @@ struct fd_rd_output {
 };
 
 void
-fd_rd_output_init(struct fd_rd_output *output, char* output_name);
+fd_rd_output_init(struct fd_rd_output *output, const char* output_name);
 
 void
 fd_rd_output_fini(struct fd_rd_output *output);
diff --git a/src/freedreno/drm/freedreno_bo.c b/src/freedreno/drm/freedreno_bo.c
index 83d49c37bba..80b5a453d9c 100644
--- a/src/freedreno/drm/freedreno_bo.c
+++ b/src/freedreno/drm/freedreno_bo.c
@@ -640,7 +640,8 @@ fd_bo_map_os_mmap(struct fd_bo *bo)
                   bo->dev->fd, offset);
 }
 
-static void *
+/* For internal use only, does not check FD_BO_NOMAP: */
+void *
 __fd_bo_map(struct fd_bo *bo)
 {
    if (!bo->map) {
diff --git a/src/freedreno/drm/freedreno_device.c b/src/freedreno/drm/freedreno_device.c
index 7c00b8610a0..f4adf6a543e 100644
--- a/src/freedreno/drm/freedreno_device.c
+++ b/src/freedreno/drm/freedreno_device.c
@@ -29,6 +29,9 @@
 #include <sys/types.h>
 
 #include "util/os_file.h"
+#include "util/u_process.h"
+
+#include "freedreno_rd_output.h"
 
 #include "freedreno_drmif.h"
 #include "freedreno_drm_perfetto.h"
@@ -91,6 +94,9 @@ out:
 
    fd_drm_perfetto_init();
 
+   fd_rd_dump_env_init();
+   fd_rd_output_init(&dev->rd, util_get_process_name());
+
    p_atomic_set(&dev->refcnt, 1);
    dev->fd = fd;
    dev->handle_table =
@@ -184,6 +190,8 @@ fd_device_del(struct fd_device *dev)
    if (!unref(&dev->refcnt))
       return;
 
+   fd_rd_output_fini(&dev->rd);
+
    assert(list_is_empty(&dev->deferred_submits));
    assert(!dev->deferred_submits_fence);
 
diff --git a/src/freedreno/drm/freedreno_priv.h b/src/freedreno/drm/freedreno_priv.h
index b59629a7ee5..b7686cfc4ab 100644
--- a/src/freedreno/drm/freedreno_priv.h
+++ b/src/freedreno/drm/freedreno_priv.h
@@ -52,6 +52,7 @@
 #include "freedreno_common.h"
 #include "freedreno_dev_info.h"
 #include "freedreno_drmif.h"
+#include "freedreno_rd_output.h"
 #include "freedreno_ringbuffer.h"
 
 extern simple_mtx_t table_lock;
@@ -272,6 +273,8 @@ struct fd_device {
    simple_mtx_t suballoc_lock;
 
    struct util_queue submit_queue;
+
+   struct fd_rd_output rd;
 };
 
 static inline bool
@@ -464,6 +467,7 @@ struct fd_bo_funcs {
 
 void fd_bo_add_fence(struct fd_bo *bo, struct fd_fence *fence);
 void *fd_bo_map_os_mmap(struct fd_bo *bo);
+void *__fd_bo_map(struct fd_bo *bo);
 
 enum fd_bo_state {
    FD_BO_STATE_IDLE,
diff --git a/src/freedreno/drm/msm/msm_priv.h b/src/freedreno/drm/msm/msm_priv.h
index ed9c785420a..c17228f4bb8 100644
--- a/src/freedreno/drm/msm/msm_priv.h
+++ b/src/freedreno/drm/msm/msm_priv.h
@@ -27,9 +27,12 @@
 #ifndef MSM_PRIV_H_
 #define MSM_PRIV_H_
 
+#include "freedreno_drmif.h"
 #include "freedreno_priv.h"
+#include "freedreno_rd_output.h"
 
 #include "util/timespec.h"
+#include "util/u_process.h"
 
 #ifndef __user
 #define __user
@@ -99,6 +102,90 @@ msm_dump_submit(struct drm_msm_gem_submit *req)
    }
 }
 
+static inline bool
+__should_dump(struct fd_bo *bo)
+{
+   return (bo->reloc_flags & FD_RELOC_DUMP) || FD_RD_DUMP(FULL);
+}
+
+static inline void
+__snapshot_buf(struct fd_rd_output *rd, struct fd_bo *bo, uint64_t iova,
+               uint32_t size, bool full)
+{
+   uint64_t offset = 0;
+
+   if (iova) {
+      offset = iova - fd_bo_get_iova(bo);
+   } else {
+      iova = fd_bo_get_iova(bo);
+      size = bo->size;
+   }
+
+   fd_rd_output_write_section(rd, RD_GPUADDR, (uint32_t[]){
+      iova, size, iova >> 32
+   }, 12);
+
+   if (!full)
+      return;
+
+   const char *buf = __fd_bo_map(bo);
+   buf += offset;
+   fd_rd_output_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
+}
+
+static inline void
+msm_dump_rd(struct fd_pipe *pipe, struct drm_msm_gem_submit *req)
+{
+   struct fd_rd_output *rd = &pipe->dev->rd;
+
+   if (!fd_rd_dump_env.flags || !req->nr_cmds ||
+       !fd_rd_output_begin(rd, req->fence))
+      return;
+
+   if (FD_RD_DUMP(FULL)) {
+      fd_pipe_wait(pipe, &(struct fd_fence) {
+         /* this is cheating a bit, but msm_pipe_wait only needs kfence */
+         .kfence = req->fence,
+      });
+   }
+
+   const char *procname = util_get_process_name();
+   fd_rd_output_write_section(rd, RD_CHIP_ID, &to_msm_pipe(pipe)->chip_id, 8);
+   fd_rd_output_write_section(rd, RD_CMD, procname, strlen(procname));
+
+   struct drm_msm_gem_submit_bo *bos = U642VOID(req->bos);
+   struct drm_msm_gem_submit_cmd *cmds = U642VOID(req->cmds);
+
+   for (unsigned i = 0; i < req->nr_bos; i++) {
+      /* This size param to fd_bo_from_handle() only matters if the bo isn't already in
+       * the handle table.  Which it should be.
+       */
+      struct fd_bo *bo = fd_bo_from_handle(pipe->dev, bos[i].handle, 0);
+
+      __snapshot_buf(rd, bo, 0, 0, __should_dump(bo));
+
+      fd_bo_del(bo);
+   }
+
+   for (unsigned i = 0; i < req->nr_cmds; i++) {
+      struct drm_msm_gem_submit_cmd *cmd = &cmds[i];
+      struct fd_bo *bo = fd_bo_from_handle(pipe->dev, bos[cmd->submit_idx].handle, 0);
+      uint64_t iova = fd_bo_get_iova(bo) + cmd->submit_offset;
+
+      /* snapshot cmdstream bo's (if we haven't already): */
+      if (!__should_dump(bo))
+         __snapshot_buf(rd, bo, iova, cmd->size, true);
+
+      fd_rd_output_write_section(rd, RD_CMDSTREAM_ADDR, (uint32_t[]){
+         iova, cmd->size >> 2, iova >> 32
+      }, 12);
+
+      fd_bo_del(bo);
+   }
+
+   fd_rd_output_end(rd);
+}
+
 static inline void
 get_abs_timeout(struct drm_msm_timespec *tv, uint64_t ns)
 {
diff --git a/src/freedreno/drm/msm/msm_ringbuffer_sp.c b/src/freedreno/drm/msm/msm_ringbuffer_sp.c
index 36876d4e8ca..b555f8407dd 100644
--- a/src/freedreno/drm/msm/msm_ringbuffer_sp.c
+++ b/src/freedreno/drm/msm/msm_ringbuffer_sp.c
@@ -31,6 +31,7 @@
 #include "util/os_file.h"
 
 #include "drm/freedreno_ringbuffer_sp.h"
+#include "freedreno_rd_output.h"
 #include "msm_priv.h"
 
 static int
@@ -146,7 +147,7 @@ flush_submit_list(struct list_head *submit_list)
 
    DEBUG_MSG("nr_cmds=%u, nr_bos=%u", req.nr_cmds, req.nr_bos);
 
-   ret = drmCommandWriteRead(msm_pipe->base.dev->fd, DRM_MSM_GEM_SUBMIT, &req,
+   ret = drmCommandWriteRead(pipe->dev->fd, DRM_MSM_GEM_SUBMIT, &req,
                              sizeof(req));
    if (ret) {
       ERROR_MSG("submit failed: %d (%s)", ret, strerror(errno));
@@ -156,6 +157,8 @@ flush_submit_list(struct list_head *submit_list)
       fd_submit->out_fence->fence_fd = req.fence_fd;
    }
 
+   msm_dump_rd(pipe, &req);
+
    if (!bos_on_stack)
       free(submit_bos);