intel: tools: import intel_aubdump
authorLionel Landwerlin <lionel.g.landwerlin@intel.com>
Sat, 16 Jun 2018 16:42:13 +0000 (17:42 +0100)
committerLionel Landwerlin <lionel.g.landwerlin@intel.com>
Thu, 5 Jul 2018 10:57:45 +0000 (11:57 +0100)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Rafael Antognolli <rafael.antognolli@intel.com>
src/intel/Makefile.am
src/intel/tools/intel_dump_gpu.c [new file with mode: 0644]
src/intel/tools/intel_dump_gpu.in [new file with mode: 0755]
src/intel/tools/meson.build

index 3e098a7ac9b21a64512595dc4c6754a5cfe2526a..8448640983f6262ba99a994ccd44e6c5c32fad46 100644 (file)
@@ -71,6 +71,8 @@ EXTRA_DIST = \
        isl/meson.build \
        tools/intel_sanitize_gpu.c \
        tools/intel_sanitize_gpu.in \
+       tools/intel_dump_gpu.c \
+       tools/intel_dump_gpu.in \
        tools/meson.build \
        vulkan/meson.build \
        meson.build
diff --git a/src/intel/tools/intel_dump_gpu.c b/src/intel/tools/intel_dump_gpu.c
new file mode 100644 (file)
index 0000000..86c133d
--- /dev/null
@@ -0,0 +1,1313 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdint.h>
+#include <stdbool.h>
+#include <signal.h>
+#include <stdarg.h>
+#include <fcntl.h>
+#include <sys/types.h>
+#include <sys/sysmacros.h>
+#include <sys/stat.h>
+#include <sys/ioctl.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sys/mman.h>
+#include <dlfcn.h>
+#include <i915_drm.h>
+
+#include "intel_aub.h"
+
+#include "dev/gen_device_info.h"
+#include "util/macros.h"
+
+#ifndef ALIGN
+#define ALIGN(x, y) (((x) + (y)-1) & ~((y)-1))
+#endif
+
+#define MI_LOAD_REGISTER_IMM_n(n) ((0x22 << 23) | (2 * (n) - 1))
+#define MI_LRI_FORCE_POSTED       (1<<12)
+
+#define MI_BATCH_BUFFER_END (0xA << 23)
+
+#define min(a, b) ({                            \
+         __typeof(a) _a = (a);                  \
+         __typeof(b) _b = (b);                  \
+         _a < _b ? _a : _b;                     \
+      })
+
+#define HWS_PGA_RCSUNIT      0x02080
+#define HWS_PGA_VCSUNIT0   0x12080
+#define HWS_PGA_BCSUNIT      0x22080
+
+#define GFX_MODE_RCSUNIT   0x0229c
+#define GFX_MODE_VCSUNIT0   0x1229c
+#define GFX_MODE_BCSUNIT   0x2229c
+
+#define EXECLIST_SUBMITPORT_RCSUNIT   0x02230
+#define EXECLIST_SUBMITPORT_VCSUNIT0   0x12230
+#define EXECLIST_SUBMITPORT_BCSUNIT   0x22230
+
+#define EXECLIST_STATUS_RCSUNIT      0x02234
+#define EXECLIST_STATUS_VCSUNIT0   0x12234
+#define EXECLIST_STATUS_BCSUNIT      0x22234
+
+#define EXECLIST_SQ_CONTENTS0_RCSUNIT   0x02510
+#define EXECLIST_SQ_CONTENTS0_VCSUNIT0   0x12510
+#define EXECLIST_SQ_CONTENTS0_BCSUNIT   0x22510
+
+#define EXECLIST_CONTROL_RCSUNIT   0x02550
+#define EXECLIST_CONTROL_VCSUNIT0   0x12550
+#define EXECLIST_CONTROL_BCSUNIT   0x22550
+
+#define MEMORY_MAP_SIZE (64 /* MiB */ * 1024 * 1024)
+
+#define PTE_SIZE 4
+#define GEN8_PTE_SIZE 8
+
+#define NUM_PT_ENTRIES (ALIGN(MEMORY_MAP_SIZE, 4096) / 4096)
+#define PT_SIZE ALIGN(NUM_PT_ENTRIES * GEN8_PTE_SIZE, 4096)
+
+#define RING_SIZE         (1 * 4096)
+#define PPHWSP_SIZE         (1 * 4096)
+#define GEN10_LR_CONTEXT_RENDER_SIZE   (19 * 4096)
+#define GEN8_LR_CONTEXT_OTHER_SIZE   (2 * 4096)
+
+#define STATIC_GGTT_MAP_START 0
+
+#define RENDER_RING_ADDR STATIC_GGTT_MAP_START
+#define RENDER_CONTEXT_ADDR (RENDER_RING_ADDR + RING_SIZE)
+
+#define BLITTER_RING_ADDR (RENDER_CONTEXT_ADDR + PPHWSP_SIZE + GEN10_LR_CONTEXT_RENDER_SIZE)
+#define BLITTER_CONTEXT_ADDR (BLITTER_RING_ADDR + RING_SIZE)
+
+#define VIDEO_RING_ADDR (BLITTER_CONTEXT_ADDR + PPHWSP_SIZE + GEN8_LR_CONTEXT_OTHER_SIZE)
+#define VIDEO_CONTEXT_ADDR (VIDEO_RING_ADDR + RING_SIZE)
+
+#define STATIC_GGTT_MAP_END (VIDEO_CONTEXT_ADDR + PPHWSP_SIZE + GEN8_LR_CONTEXT_OTHER_SIZE)
+#define STATIC_GGTT_MAP_SIZE (STATIC_GGTT_MAP_END - STATIC_GGTT_MAP_START)
+
+#define CONTEXT_FLAGS (0x229)   /* Normal Priority | L3-LLC Coherency |
+                                   Legacy Context with no 64 bit VA support | Valid */
+
+#define RENDER_CONTEXT_DESCRIPTOR  ((uint64_t)1 << 32 | RENDER_CONTEXT_ADDR  | CONTEXT_FLAGS)
+#define BLITTER_CONTEXT_DESCRIPTOR ((uint64_t)2 << 32 | BLITTER_CONTEXT_ADDR | CONTEXT_FLAGS)
+#define VIDEO_CONTEXT_DESCRIPTOR   ((uint64_t)3 << 32 | VIDEO_CONTEXT_ADDR   | CONTEXT_FLAGS)
+
+static const uint32_t render_context_init[GEN10_LR_CONTEXT_RENDER_SIZE /
+                                          sizeof(uint32_t)] = {
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(14) | MI_LRI_FORCE_POSTED,
+   0x2244 /* CONTEXT_CONTROL */,      0x90009 /* Inhibit Synchronous Context Switch | Engine Context Restore Inhibit */,
+   0x2034 /* RING_HEAD */,         0,
+   0x2030 /* RING_TAIL */,         0,
+   0x2038 /* RING_BUFFER_START */,      RENDER_RING_ADDR,
+   0x203C /* RING_BUFFER_CONTROL */,   (RING_SIZE - 4096) | 1 /* Buffer Length | Ring Buffer Enable */,
+   0x2168 /* BB_HEAD_U */,         0,
+   0x2140 /* BB_HEAD_L */,         0,
+   0x2110 /* BB_STATE */,         0,
+   0x211C /* SECOND_BB_HEAD_U */,      0,
+   0x2114 /* SECOND_BB_HEAD_L */,      0,
+   0x2118 /* SECOND_BB_STATE */,      0,
+   0x21C0 /* BB_PER_CTX_PTR */,      0,
+   0x21C4 /* RCS_INDIRECT_CTX */,      0,
+   0x21C8 /* RCS_INDIRECT_CTX_OFFSET */,   0,
+   /* MI_NOOP */
+   0, 0,
+
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(9) | MI_LRI_FORCE_POSTED,
+   0x23A8 /* CTX_TIMESTAMP */,   0,
+   0x228C /* PDP3_UDW */,      0,
+   0x2288 /* PDP3_LDW */,      0,
+   0x2284 /* PDP2_UDW */,      0,
+   0x2280 /* PDP2_LDW */,      0,
+   0x227C /* PDP1_UDW */,      0,
+   0x2278 /* PDP1_LDW */,      0,
+   0x2274 /* PDP0_UDW */,      0,
+   0x2270 /* PDP0_LDW */,      0,
+   /* MI_NOOP */
+   0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(1),
+   0x20C8 /* R_PWR_CLK_STATE */, 0x7FFFFFFF,
+   MI_BATCH_BUFFER_END
+};
+
+static const uint32_t blitter_context_init[GEN8_LR_CONTEXT_OTHER_SIZE /
+                                           sizeof(uint32_t)] = {
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(11) | MI_LRI_FORCE_POSTED,
+   0x22244 /* CONTEXT_CONTROL */,      0x90009 /* Inhibit Synchronous Context Switch | Engine Context Restore Inhibit */,
+   0x22034 /* RING_HEAD */,      0,
+   0x22030 /* RING_TAIL */,      0,
+   0x22038 /* RING_BUFFER_START */,   BLITTER_RING_ADDR,
+   0x2203C /* RING_BUFFER_CONTROL */,   (RING_SIZE - 4096) | 1 /* Buffer Length | Ring Buffer Enable */,
+   0x22168 /* BB_HEAD_U */,      0,
+   0x22140 /* BB_HEAD_L */,      0,
+   0x22110 /* BB_STATE */,         0,
+   0x2211C /* SECOND_BB_HEAD_U */,      0,
+   0x22114 /* SECOND_BB_HEAD_L */,      0,
+   0x22118 /* SECOND_BB_STATE */,      0,
+   /* MI_NOOP */
+   0, 0, 0, 0, 0, 0, 0, 0,
+
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(9) | MI_LRI_FORCE_POSTED,
+   0x223A8 /* CTX_TIMESTAMP */,   0,
+   0x2228C /* PDP3_UDW */,      0,
+   0x22288 /* PDP3_LDW */,      0,
+   0x22284 /* PDP2_UDW */,      0,
+   0x22280 /* PDP2_LDW */,      0,
+   0x2227C /* PDP1_UDW */,      0,
+   0x22278 /* PDP1_LDW */,      0,
+   0x22274 /* PDP0_UDW */,      0,
+   0x22270 /* PDP0_LDW */,      0,
+   /* MI_NOOP */
+   0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+
+   MI_BATCH_BUFFER_END
+};
+
+static const uint32_t video_context_init[GEN8_LR_CONTEXT_OTHER_SIZE /
+                                         sizeof(uint32_t)] = {
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(11) | MI_LRI_FORCE_POSTED,
+   0x1C244 /* CONTEXT_CONTROL */,      0x90009 /* Inhibit Synchronous Context Switch | Engine Context Restore Inhibit */,
+   0x1C034 /* RING_HEAD */,      0,
+   0x1C030 /* RING_TAIL */,      0,
+   0x1C038 /* RING_BUFFER_START */,   VIDEO_RING_ADDR,
+   0x1C03C /* RING_BUFFER_CONTROL */,   (RING_SIZE - 4096) | 1 /* Buffer Length | Ring Buffer Enable */,
+   0x1C168 /* BB_HEAD_U */,      0,
+   0x1C140 /* BB_HEAD_L */,      0,
+   0x1C110 /* BB_STATE */,         0,
+   0x1C11C /* SECOND_BB_HEAD_U */,      0,
+   0x1C114 /* SECOND_BB_HEAD_L */,      0,
+   0x1C118 /* SECOND_BB_STATE */,      0,
+   /* MI_NOOP */
+   0, 0, 0, 0, 0, 0, 0, 0,
+
+   0 /* MI_NOOP */,
+   MI_LOAD_REGISTER_IMM_n(9) | MI_LRI_FORCE_POSTED,
+   0x1C3A8 /* CTX_TIMESTAMP */,   0,
+   0x1C28C /* PDP3_UDW */,      0,
+   0x1C288 /* PDP3_LDW */,      0,
+   0x1C284 /* PDP2_UDW */,      0,
+   0x1C280 /* PDP2_LDW */,      0,
+   0x1C27C /* PDP1_UDW */,      0,
+   0x1C278 /* PDP1_LDW */,      0,
+   0x1C274 /* PDP0_UDW */,      0,
+   0x1C270 /* PDP0_LDW */,      0,
+   /* MI_NOOP */
+   0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+
+   MI_BATCH_BUFFER_END
+};
+
+static int close_init_helper(int fd);
+static int ioctl_init_helper(int fd, unsigned long request, ...);
+
+static int (*libc_close)(int fd) = close_init_helper;
+static int (*libc_ioctl)(int fd, unsigned long request, ...) = ioctl_init_helper;
+
+static int drm_fd = -1;
+static char *filename = NULL;
+static FILE *files[2] = { NULL, NULL };
+static struct gen_device_info devinfo = {0};
+static int verbose = 0;
+static bool device_override;
+static uint32_t device;
+static int addr_bits = 0;
+
+#define MAX_BO_COUNT 64 * 1024
+
+struct bo {
+   uint32_t size;
+   uint64_t offset;
+   void *map;
+};
+
+static struct bo *bos;
+
+#define DRM_MAJOR 226
+
+#ifndef DRM_I915_GEM_USERPTR
+
+#define DRM_I915_GEM_USERPTR      0x33
+#define DRM_IOCTL_I915_GEM_USERPTR   DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_USERPTR, struct drm_i915_gem_userptr)
+
+struct drm_i915_gem_userptr {
+   __u64 user_ptr;
+   __u64 user_size;
+   __u32 flags;
+#define I915_USERPTR_READ_ONLY 0x1
+#define I915_USERPTR_UNSYNCHRONIZED 0x80000000
+   /**
+    * Returned handle for the object.
+    *
+    * Object handles are nonzero.
+    */
+   __u32 handle;
+};
+
+#endif
+
+/* We set bit 0 in the map pointer for userptr BOs so we know not to
+ * munmap them on DRM_IOCTL_GEM_CLOSE.
+ */
+#define USERPTR_FLAG 1
+#define IS_USERPTR(p) ((uintptr_t) (p) & USERPTR_FLAG)
+#define GET_PTR(p) ( (void *) ((uintptr_t) p & ~(uintptr_t) 1) )
+
+#ifndef I915_EXEC_BATCH_FIRST
+#define I915_EXEC_BATCH_FIRST (1 << 18)
+#endif
+
+static void __attribute__ ((format(__printf__, 2, 3)))
+fail_if(int cond, const char *format, ...)
+{
+   va_list args;
+
+   if (!cond)
+      return;
+
+   va_start(args, format);
+   vfprintf(stderr, format, args);
+   va_end(args);
+
+   raise(SIGTRAP);
+}
+
+static struct bo *
+get_bo(uint32_t handle)
+{
+   struct bo *bo;
+
+   fail_if(handle >= MAX_BO_COUNT, "bo handle too large\n");
+   bo = &bos[handle];
+
+   return bo;
+}
+
+static inline uint32_t
+align_u32(uint32_t v, uint32_t a)
+{
+   return (v + a - 1) & ~(a - 1);
+}
+
+static inline uint64_t
+align_u64(uint64_t v, uint64_t a)
+{
+   return (v + a - 1) & ~(a - 1);
+}
+
+static void
+dword_out(uint32_t data)
+{
+   for (int i = 0; i < ARRAY_SIZE (files); i++) {
+      if (files[i] == NULL)
+         continue;
+
+      fail_if(fwrite(&data, 1, 4, files[i]) == 0,
+              "Writing to output failed\n");
+   }
+}
+
+static void
+data_out(const void *data, size_t size)
+{
+   if (size == 0)
+      return;
+
+   for (int i = 0; i < ARRAY_SIZE (files); i++) {
+      if (files[i] == NULL)
+         continue;
+
+      fail_if(fwrite(data, 1, size, files[i]) == 0,
+              "Writing to output failed\n");
+   }
+}
+
+static uint32_t
+gtt_size(void)
+{
+   return NUM_PT_ENTRIES * (addr_bits > 32 ? GEN8_PTE_SIZE : PTE_SIZE);
+}
+
+static void
+mem_trace_memory_write_header_out(uint64_t addr, uint32_t len,
+                                  uint32_t addr_space)
+{
+   uint32_t dwords = ALIGN(len, sizeof(uint32_t)) / sizeof(uint32_t);
+
+   dword_out(CMD_MEM_TRACE_MEMORY_WRITE | (5 + dwords - 1));
+   dword_out(addr & 0xFFFFFFFF);   /* addr lo */
+   dword_out(addr >> 32);   /* addr hi */
+   dword_out(addr_space);   /* gtt */
+   dword_out(len);
+}
+
+static void
+register_write_out(uint32_t addr, uint32_t value)
+{
+   uint32_t dwords = 1;
+
+   dword_out(CMD_MEM_TRACE_REGISTER_WRITE | (5 + dwords - 1));
+   dword_out(addr);
+   dword_out(AUB_MEM_TRACE_REGISTER_SIZE_DWORD |
+             AUB_MEM_TRACE_REGISTER_SPACE_MMIO);
+   dword_out(0xFFFFFFFF);   /* mask lo */
+   dword_out(0x00000000);   /* mask hi */
+   dword_out(value);
+}
+
+static void
+gen8_emit_ggtt_pte_for_range(uint64_t start, uint64_t end)
+{
+   uint64_t entry_addr;
+   uint64_t page_num;
+   uint64_t end_aligned = align_u64(end, 4096);
+
+   if (start >= end || end > (1ull << 32))
+      return;
+
+   entry_addr = start & ~(4096 - 1);
+   do {
+      uint64_t last_page_entry, num_entries;
+
+      page_num = entry_addr >> 21;
+      last_page_entry = min((page_num + 1) << 21, end_aligned);
+      num_entries = (last_page_entry - entry_addr) >> 12;
+      mem_trace_memory_write_header_out(
+         entry_addr >> 9, num_entries * GEN8_PTE_SIZE,
+         AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT_ENTRY);
+      while (num_entries-- > 0) {
+         dword_out((entry_addr & ~(4096 - 1)) |
+                   3 /* read/write | present */);
+         dword_out(entry_addr >> 32);
+         entry_addr += 4096;
+      }
+   } while (entry_addr < end);
+}
+
+/**
+ * Sets bits `start` through `end` - 1 in the bitmap array.
+ */
+static void
+set_bitmap_range(uint32_t *bitmap, uint32_t start, uint32_t end)
+{
+   uint32_t pos = start;
+   while (pos < end) {
+      const uint32_t bit = 1 << (pos & 0x1f);
+      if (bit == 1 && (end - pos) > 32) {
+         bitmap[pos >> 5] = 0xffffffff;
+         pos += 32;
+      } else {
+         bitmap[pos >> 5] |= bit;
+         pos++;
+      }
+   }
+}
+
+/**
+ * Finds the next `set` (or clear) bit in the bitmap array.
+ *
+ * The search starts at `*start` and only checks until `end` - 1.
+ *
+ * If found, returns true, and the found bit index in `*start`.
+ */
+static bool
+find_bitmap_bit(uint32_t *bitmap, bool set, uint32_t *start, uint32_t end)
+{
+   uint32_t pos = *start;
+   const uint32_t neg_dw = set ? 0 : -1;
+   while (pos < end) {
+      const uint32_t dw = bitmap[pos >> 5];
+      const uint32_t bit = 1 << (pos & 0x1f);
+      if (!!(dw & bit) == set) {
+         *start = pos;
+         return true;
+      } else if (bit == 1 && dw == neg_dw)
+         pos += 32;
+      else
+         pos++;
+   }
+   return false;
+}
+
+/**
+ * Finds a range of clear bits within the bitmap array.
+ *
+ * The search starts at `*start` and only checks until `*end` - 1.
+ *
+ * If found, returns true, and `*start` and `*end` are set for the
+ * range of clear bits.
+ */
+static bool
+find_bitmap_clear_bit_range(uint32_t *bitmap, uint32_t *start, uint32_t *end)
+{
+   if (find_bitmap_bit(bitmap, false, start, *end)) {
+      uint32_t found_end = *start;
+      if (find_bitmap_bit(bitmap, true, &found_end, *end))
+         *end = found_end;
+      return true;
+   }
+   return false;
+}
+
+static void
+gen8_map_ggtt_range(uint64_t start, uint64_t end)
+{
+   uint32_t pos1, pos2, end_pos;
+   static uint32_t *bitmap = NULL;
+   if (bitmap == NULL) {
+      /* 4GiB (32-bits) of 4KiB pages (12-bits) in dwords (5-bits) */
+      bitmap = calloc(1 << (32 - 12 - 5), sizeof(*bitmap));
+      if (bitmap == NULL)
+         return;
+   }
+
+   pos1 = start >> 12;
+   end_pos = (end + 4096 - 1) >> 12;
+   while (pos1 < end_pos) {
+      pos2 = end_pos;
+      if (!find_bitmap_clear_bit_range(bitmap, &pos1, &pos2))
+         break;
+
+      if (verbose)
+         printf("MAPPING 0x%08lx-0x%08lx\n",
+                (uint64_t)pos1 << 12, (uint64_t)pos2 << 12);
+      gen8_emit_ggtt_pte_for_range((uint64_t)pos1 << 12,
+                                   (uint64_t)pos2 << 12);
+      set_bitmap_range(bitmap, (uint64_t)pos1, (uint64_t)pos2);
+      pos1 = pos2;
+   }
+}
+
+static void
+gen8_map_base_size(uint64_t base, uint64_t size)
+{
+   gen8_map_ggtt_range(base, base + size);
+}
+
+static void
+gen10_write_header(void)
+{
+   char app_name[8 * 4];
+   int app_name_len, dwords;
+
+   app_name_len =
+      snprintf(app_name, sizeof(app_name), "PCI-ID=0x%X %s", device,
+               program_invocation_short_name);
+   app_name_len = ALIGN(app_name_len, sizeof(uint32_t));
+
+   dwords = 5 + app_name_len / sizeof(uint32_t);
+   dword_out(CMD_MEM_TRACE_VERSION | (dwords - 1));
+   dword_out(AUB_MEM_TRACE_VERSION_FILE_VERSION);
+   dword_out(AUB_MEM_TRACE_VERSION_DEVICE_CNL |
+             AUB_MEM_TRACE_VERSION_METHOD_PHY);
+   dword_out(0);      /* version */
+   dword_out(0);      /* version */
+   data_out(app_name, app_name_len);
+
+   /* RENDER_RING */
+   gen8_map_base_size(RENDER_RING_ADDR, RING_SIZE);
+   mem_trace_memory_write_header_out(RENDER_RING_ADDR, RING_SIZE,
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   for (uint32_t i = 0; i < RING_SIZE; i += sizeof(uint32_t))
+      dword_out(0);
+
+   /* RENDER_PPHWSP */
+   gen8_map_base_size(RENDER_CONTEXT_ADDR,
+                      PPHWSP_SIZE + sizeof(render_context_init));
+   mem_trace_memory_write_header_out(RENDER_CONTEXT_ADDR,
+                                     PPHWSP_SIZE +
+                                     sizeof(render_context_init),
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   for (uint32_t i = 0; i < PPHWSP_SIZE; i += sizeof(uint32_t))
+      dword_out(0);
+
+   /* RENDER_CONTEXT */
+   data_out(render_context_init, sizeof(render_context_init));
+
+   /* BLITTER_RING */
+   gen8_map_base_size(BLITTER_RING_ADDR, RING_SIZE);
+   mem_trace_memory_write_header_out(BLITTER_RING_ADDR, RING_SIZE,
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   for (uint32_t i = 0; i < RING_SIZE; i += sizeof(uint32_t))
+      dword_out(0);
+
+   /* BLITTER_PPHWSP */
+   gen8_map_base_size(BLITTER_CONTEXT_ADDR,
+                      PPHWSP_SIZE + sizeof(blitter_context_init));
+   mem_trace_memory_write_header_out(BLITTER_CONTEXT_ADDR,
+                                     PPHWSP_SIZE +
+                                     sizeof(blitter_context_init),
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   for (uint32_t i = 0; i < PPHWSP_SIZE; i += sizeof(uint32_t))
+      dword_out(0);
+
+   /* BLITTER_CONTEXT */
+   data_out(blitter_context_init, sizeof(blitter_context_init));
+
+   /* VIDEO_RING */
+   gen8_map_base_size(VIDEO_RING_ADDR, RING_SIZE);
+   mem_trace_memory_write_header_out(VIDEO_RING_ADDR, RING_SIZE,
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   for (uint32_t i = 0; i < RING_SIZE; i += sizeof(uint32_t))
+      dword_out(0);
+
+   /* VIDEO_PPHWSP */
+   gen8_map_base_size(VIDEO_CONTEXT_ADDR,
+                      PPHWSP_SIZE + sizeof(video_context_init));
+   mem_trace_memory_write_header_out(VIDEO_CONTEXT_ADDR,
+                                     PPHWSP_SIZE +
+                                     sizeof(video_context_init),
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   for (uint32_t i = 0; i < PPHWSP_SIZE; i += sizeof(uint32_t))
+      dword_out(0);
+
+   /* VIDEO_CONTEXT */
+   data_out(video_context_init, sizeof(video_context_init));
+
+   register_write_out(HWS_PGA_RCSUNIT, RENDER_CONTEXT_ADDR);
+   register_write_out(HWS_PGA_VCSUNIT0, VIDEO_CONTEXT_ADDR);
+   register_write_out(HWS_PGA_BCSUNIT, BLITTER_CONTEXT_ADDR);
+
+   register_write_out(GFX_MODE_RCSUNIT, 0x80008000 /* execlist enable */);
+   register_write_out(GFX_MODE_VCSUNIT0, 0x80008000 /* execlist enable */);
+   register_write_out(GFX_MODE_BCSUNIT, 0x80008000 /* execlist enable */);
+}
+
+static void write_header(void)
+{
+   char app_name[8 * 4];
+   char comment[16];
+   int comment_len, comment_dwords, dwords;
+   uint32_t entry = 0x200003;
+
+   comment_len = snprintf(comment, sizeof(comment), "PCI-ID=0x%x", device);
+   comment_dwords = ((comment_len + 3) / 4);
+
+   /* Start with a (required) version packet. */
+   dwords = 13 + comment_dwords;
+   dword_out(CMD_AUB_HEADER | (dwords - 2));
+   dword_out((4 << AUB_HEADER_MAJOR_SHIFT) |
+             (0 << AUB_HEADER_MINOR_SHIFT));
+
+   /* Next comes a 32-byte application name. */
+   strncpy(app_name, program_invocation_short_name, sizeof(app_name));
+   app_name[sizeof(app_name) - 1] = 0;
+   data_out(app_name, sizeof(app_name));
+
+   dword_out(0); /* timestamp */
+   dword_out(0); /* timestamp */
+   dword_out(comment_len);
+   data_out(comment, comment_dwords * 4);
+
+   /* Set up the GTT. The max we can handle is 64M */
+   dword_out(CMD_AUB_TRACE_HEADER_BLOCK | ((addr_bits > 32 ? 6 : 5) - 2));
+   dword_out(AUB_TRACE_MEMTYPE_GTT_ENTRY |
+             AUB_TRACE_TYPE_NOTYPE | AUB_TRACE_OP_DATA_WRITE);
+   dword_out(0); /* subtype */
+   dword_out(0); /* offset */
+   dword_out(gtt_size()); /* size */
+   if (addr_bits > 32)
+      dword_out(0);
+   for (uint32_t i = 0; i < NUM_PT_ENTRIES; i++) {
+      dword_out(entry + 0x1000 * i);
+      if (addr_bits > 32)
+         dword_out(0);
+   }
+}
+
+/**
+ * Break up large objects into multiple writes.  Otherwise a 128kb VBO
+ * would overflow the 16 bits of size field in the packet header and
+ * everything goes badly after that.
+ */
+static void
+aub_write_trace_block(uint32_t type, void *virtual, uint32_t size, uint64_t gtt_offset)
+{
+   uint32_t block_size;
+   uint32_t subtype = 0;
+   static const char null_block[8 * 4096];
+
+   for (uint32_t offset = 0; offset < size; offset += block_size) {
+      block_size = size - offset;
+
+      if (block_size > 8 * 4096)
+         block_size = 8 * 4096;
+
+      if (devinfo.gen >= 10) {
+         mem_trace_memory_write_header_out(gtt_offset + offset,
+                                           block_size,
+                                           AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+      } else {
+         dword_out(CMD_AUB_TRACE_HEADER_BLOCK |
+                   ((addr_bits > 32 ? 6 : 5) - 2));
+         dword_out(AUB_TRACE_MEMTYPE_GTT |
+                   type | AUB_TRACE_OP_DATA_WRITE);
+         dword_out(subtype);
+         dword_out(gtt_offset + offset);
+         dword_out(align_u32(block_size, 4));
+         if (addr_bits > 32)
+            dword_out((gtt_offset + offset) >> 32);
+      }
+
+      if (virtual)
+         data_out(((char *) GET_PTR(virtual)) + offset, block_size);
+      else
+         data_out(null_block, block_size);
+
+      /* Pad to a multiple of 4 bytes. */
+      data_out(null_block, -block_size & 3);
+   }
+}
+
+static void
+write_reloc(void *p, uint64_t v)
+{
+   if (addr_bits > 32) {
+      /* From the Broadwell PRM Vol. 2a,
+       * MI_LOAD_REGISTER_MEM::MemoryAddress:
+       *
+       *   "This field specifies the address of the memory
+       *   location where the register value specified in the
+       *   DWord above will read from.  The address specifies
+       *   the DWord location of the data. Range =
+       *   GraphicsVirtualAddress[63:2] for a DWord register
+       *   GraphicsAddress [63:48] are ignored by the HW and
+       *   assumed to be in correct canonical form [63:48] ==
+       *   [47]."
+       *
+       * In practice, this will always mean the top bits are zero
+       * because of the GTT size limitation of the aubdump tool.
+       */
+      const int shift = 63 - 47;
+      *(uint64_t *)p = (((int64_t)v) << shift) >> shift;
+   } else {
+      *(uint32_t *)p = v;
+   }
+}
+
+static void
+aub_dump_execlist(uint64_t batch_offset, int ring_flag)
+{
+   uint32_t ring_addr;
+   uint64_t descriptor;
+   uint32_t elsp_reg;
+   uint32_t elsq_reg;
+   uint32_t status_reg;
+   uint32_t control_reg;
+
+   switch (ring_flag) {
+   case I915_EXEC_DEFAULT:
+   case I915_EXEC_RENDER:
+      ring_addr = RENDER_RING_ADDR;
+      descriptor = RENDER_CONTEXT_DESCRIPTOR;
+      elsp_reg = EXECLIST_SUBMITPORT_RCSUNIT;
+      elsq_reg = EXECLIST_SQ_CONTENTS0_RCSUNIT;
+      status_reg = EXECLIST_STATUS_RCSUNIT;
+      control_reg = EXECLIST_CONTROL_RCSUNIT;
+      break;
+   case I915_EXEC_BSD:
+      ring_addr = VIDEO_RING_ADDR;
+      descriptor = VIDEO_CONTEXT_DESCRIPTOR;
+      elsp_reg = EXECLIST_SUBMITPORT_VCSUNIT0;
+      elsq_reg = EXECLIST_SQ_CONTENTS0_VCSUNIT0;
+      status_reg = EXECLIST_STATUS_VCSUNIT0;
+      control_reg = EXECLIST_CONTROL_VCSUNIT0;
+      break;
+   case I915_EXEC_BLT:
+      ring_addr = BLITTER_RING_ADDR;
+      descriptor = BLITTER_CONTEXT_DESCRIPTOR;
+      elsp_reg = EXECLIST_SUBMITPORT_BCSUNIT;
+      elsq_reg = EXECLIST_SQ_CONTENTS0_BCSUNIT;
+      status_reg = EXECLIST_STATUS_BCSUNIT;
+      control_reg = EXECLIST_CONTROL_BCSUNIT;
+      break;
+   }
+
+   mem_trace_memory_write_header_out(ring_addr, 16,
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   dword_out(AUB_MI_BATCH_BUFFER_START | (3 - 2));
+   dword_out(batch_offset & 0xFFFFFFFF);
+   dword_out(batch_offset >> 32);
+   dword_out(0 /* MI_NOOP */);
+
+   mem_trace_memory_write_header_out(ring_addr + 8192 + 20, 4,
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   dword_out(0); /* RING_BUFFER_HEAD */
+   mem_trace_memory_write_header_out(ring_addr + 8192 + 28, 4,
+                                     AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_LOCAL);
+   dword_out(16); /* RING_BUFFER_TAIL */
+
+   if (devinfo.gen >= 11) {
+      register_write_out(elsq_reg, descriptor & 0xFFFFFFFF);
+      register_write_out(elsq_reg + sizeof(uint32_t), descriptor >> 32);
+      register_write_out(control_reg, 1);
+   } else {
+      register_write_out(elsp_reg, 0);
+      register_write_out(elsp_reg, 0);
+      register_write_out(elsp_reg, descriptor >> 32);
+      register_write_out(elsp_reg, descriptor & 0xFFFFFFFF);
+   }
+
+   dword_out(CMD_MEM_TRACE_REGISTER_POLL | (5 + 1 - 1));
+   dword_out(status_reg);
+   dword_out(AUB_MEM_TRACE_REGISTER_SIZE_DWORD |
+             AUB_MEM_TRACE_REGISTER_SPACE_MMIO);
+   if (devinfo.gen >= 11) {
+      dword_out(0x00000001);   /* mask lo */
+      dword_out(0x00000000);   /* mask hi */
+      dword_out(0x00000001);
+   } else {
+      dword_out(0x00000010);   /* mask lo */
+      dword_out(0x00000000);   /* mask hi */
+      dword_out(0x00000000);
+   }
+}
+
+static void
+aub_dump_ringbuffer(uint64_t batch_offset, uint64_t offset, int ring_flag)
+{
+   uint32_t ringbuffer[4096];
+   unsigned aub_mi_bbs_len;
+   int ring = AUB_TRACE_TYPE_RING_PRB0; /* The default ring */
+   int ring_count = 0;
+
+   if (ring_flag == I915_EXEC_BSD)
+      ring = AUB_TRACE_TYPE_RING_PRB1;
+   else if (ring_flag == I915_EXEC_BLT)
+      ring = AUB_TRACE_TYPE_RING_PRB2;
+
+   /* Make a ring buffer to execute our batchbuffer. */
+   memset(ringbuffer, 0, sizeof(ringbuffer));
+
+   aub_mi_bbs_len = addr_bits > 32 ? 3 : 2;
+   ringbuffer[ring_count] = AUB_MI_BATCH_BUFFER_START | (aub_mi_bbs_len - 2);
+   write_reloc(&ringbuffer[ring_count + 1], batch_offset);
+   ring_count += aub_mi_bbs_len;
+
+   /* Write out the ring.  This appears to trigger execution of
+    * the ring in the simulator.
+    */
+   dword_out(CMD_AUB_TRACE_HEADER_BLOCK |
+             ((addr_bits > 32 ? 6 : 5) - 2));
+   dword_out(AUB_TRACE_MEMTYPE_GTT | ring | AUB_TRACE_OP_COMMAND_WRITE);
+   dword_out(0); /* general/surface subtype */
+   dword_out(offset);
+   dword_out(ring_count * 4);
+   if (addr_bits > 32)
+      dword_out(offset >> 32);
+
+   data_out(ringbuffer, ring_count * 4);
+}
+
+static void *
+relocate_bo(struct bo *bo, const struct drm_i915_gem_execbuffer2 *execbuffer2,
+            const struct drm_i915_gem_exec_object2 *obj)
+{
+   const struct drm_i915_gem_exec_object2 *exec_objects =
+      (struct drm_i915_gem_exec_object2 *) (uintptr_t) execbuffer2->buffers_ptr;
+   const struct drm_i915_gem_relocation_entry *relocs =
+      (const struct drm_i915_gem_relocation_entry *) (uintptr_t) obj->relocs_ptr;
+   void *relocated;
+   int handle;
+
+   relocated = malloc(bo->size);
+   fail_if(relocated == NULL, "intel_aubdump: out of memory\n");
+   memcpy(relocated, GET_PTR(bo->map), bo->size);
+   for (size_t i = 0; i < obj->relocation_count; i++) {
+      fail_if(relocs[i].offset >= bo->size, "intel_aubdump: reloc outside bo\n");
+
+      if (execbuffer2->flags & I915_EXEC_HANDLE_LUT)
+         handle = exec_objects[relocs[i].target_handle].handle;
+      else
+         handle = relocs[i].target_handle;
+
+      write_reloc(((char *)relocated) + relocs[i].offset,
+                  get_bo(handle)->offset + relocs[i].delta);
+   }
+
+   return relocated;
+}
+
+static int
+gem_ioctl(int fd, unsigned long request, void *argp)
+{
+   int ret;
+
+   do {
+      ret = libc_ioctl(fd, request, argp);
+   } while (ret == -1 && (errno == EINTR || errno == EAGAIN));
+
+   return ret;
+}
+
+static void *
+gem_mmap(int fd, uint32_t handle, uint64_t offset, uint64_t size)
+{
+   struct drm_i915_gem_mmap mmap = {
+      .handle = handle,
+      .offset = offset,
+      .size = size
+   };
+
+   if (gem_ioctl(fd, DRM_IOCTL_I915_GEM_MMAP, &mmap) == -1)
+      return MAP_FAILED;
+
+   return (void *)(uintptr_t) mmap.addr_ptr;
+}
+
+static int
+gem_get_param(int fd, uint32_t param)
+{
+   int value;
+   drm_i915_getparam_t gp = {
+      .param = param,
+      .value = &value
+   };
+
+   if (gem_ioctl(fd, DRM_IOCTL_I915_GETPARAM, &gp) == -1)
+      return 0;
+
+   return value;
+}
+
+static void
+dump_execbuffer2(int fd, struct drm_i915_gem_execbuffer2 *execbuffer2)
+{
+   struct drm_i915_gem_exec_object2 *exec_objects =
+      (struct drm_i915_gem_exec_object2 *) (uintptr_t) execbuffer2->buffers_ptr;
+   uint32_t ring_flag = execbuffer2->flags & I915_EXEC_RING_MASK;
+   uint32_t offset;
+   struct drm_i915_gem_exec_object2 *obj;
+   struct bo *bo, *batch_bo;
+   int batch_index;
+   void *data;
+
+   /* We can't do this at open time as we're not yet authenticated. */
+   if (device == 0) {
+      device = gem_get_param(fd, I915_PARAM_CHIPSET_ID);
+      fail_if(device == 0 || devinfo.gen == 0, "failed to identify chipset\n");
+   }
+   if (devinfo.gen == 0) {
+      fail_if(!gen_get_device_info(device, &devinfo),
+              "failed to identify chipset=0x%x\n", device);
+
+      addr_bits = devinfo.gen >= 8 ? 48 : 32;
+
+      if (devinfo.gen >= 10)
+         gen10_write_header();
+      else
+         write_header();
+
+      if (verbose)
+         printf("[intel_aubdump: running, "
+                "output file %s, chipset id 0x%04x, gen %d]\n",
+                filename, device, devinfo.gen);
+   }
+
+   if (devinfo.gen >= 10)
+      offset = STATIC_GGTT_MAP_END;
+   else
+      offset = gtt_size();
+
+   if (verbose)
+      printf("Dumping execbuffer2:\n");
+
+   for (uint32_t i = 0; i < execbuffer2->buffer_count; i++) {
+      obj = &exec_objects[i];
+      bo = get_bo(obj->handle);
+
+      /* If bo->size == 0, this means they passed us an invalid
+       * buffer.  The kernel will reject it and so should we.
+       */
+      if (bo->size == 0) {
+         if (verbose)
+            printf("BO #%d is invalid!\n", obj->handle);
+         return;
+      }
+
+      if (obj->flags & EXEC_OBJECT_PINNED) {
+         bo->offset = obj->offset;
+         if (verbose)
+            printf("BO #%d (%dB) pinned @ 0x%lx\n",
+                   obj->handle, bo->size, bo->offset);
+      } else {
+         if (obj->alignment != 0)
+            offset = align_u32(offset, obj->alignment);
+         bo->offset = offset;
+         if (verbose)
+            printf("BO #%d (%dB) @ 0x%lx\n", obj->handle,
+                   bo->size, bo->offset);
+         offset = align_u32(offset + bo->size + 4095, 4096);
+      }
+
+      if (bo->map == NULL && bo->size > 0)
+         bo->map = gem_mmap(fd, obj->handle, 0, bo->size);
+      fail_if(bo->map == MAP_FAILED, "intel_aubdump: bo mmap failed\n");
+
+      if (devinfo.gen >= 10)
+         gen8_map_ggtt_range(bo->offset, bo->offset + bo->size);
+   }
+
+   batch_index = (execbuffer2->flags & I915_EXEC_BATCH_FIRST) ? 0 :
+      execbuffer2->buffer_count - 1;
+   batch_bo = get_bo(exec_objects[batch_index].handle);
+   for (uint32_t i = 0; i < execbuffer2->buffer_count; i++) {
+      obj = &exec_objects[i];
+      bo = get_bo(obj->handle);
+
+      if (obj->relocation_count > 0)
+         data = relocate_bo(bo, execbuffer2, obj);
+      else
+         data = bo->map;
+
+      if (bo == batch_bo) {
+         aub_write_trace_block(AUB_TRACE_TYPE_BATCH,
+                               data, bo->size, bo->offset);
+      } else {
+         aub_write_trace_block(AUB_TRACE_TYPE_NOTYPE,
+                               data, bo->size, bo->offset);
+      }
+      if (data != bo->map)
+         free(data);
+   }
+
+   if (devinfo.gen >= 10) {
+      aub_dump_execlist(batch_bo->offset +
+                        execbuffer2->batch_start_offset, ring_flag);
+   } else {
+      /* Dump ring buffer */
+      aub_dump_ringbuffer(batch_bo->offset +
+                          execbuffer2->batch_start_offset, offset,
+                          ring_flag);
+   }
+
+   for (int i = 0; i < ARRAY_SIZE(files); i++) {
+      if (files[i] != NULL)
+         fflush(files[i]);
+   }
+
+   if (device_override &&
+       (execbuffer2->flags & I915_EXEC_FENCE_ARRAY) != 0) {
+      struct drm_i915_gem_exec_fence *fences =
+         (void*)(uintptr_t)execbuffer2->cliprects_ptr;
+      for (uint32_t i = 0; i < execbuffer2->num_cliprects; i++) {
+         if ((fences[i].flags & I915_EXEC_FENCE_SIGNAL) != 0) {
+            struct drm_syncobj_array arg = {
+               .handles = (uintptr_t)&fences[i].handle,
+               .count_handles = 1,
+               .pad = 0,
+            };
+            libc_ioctl(fd, DRM_IOCTL_SYNCOBJ_SIGNAL, &arg);
+         }
+      }
+   }
+}
+
+static void
+add_new_bo(int handle, uint64_t size, void *map)
+{
+   struct bo *bo = &bos[handle];
+
+   fail_if(handle >= MAX_BO_COUNT, "intel_aubdump: bo handle out of range\n");
+   fail_if(size == 0, "intel_aubdump: bo size is invalid\n");
+
+   bo->size = size;
+   bo->map = map;
+}
+
+static void
+remove_bo(int handle)
+{
+   struct bo *bo = get_bo(handle);
+
+   if (bo->map && !IS_USERPTR(bo->map))
+      munmap(bo->map, bo->size);
+   bo->size = 0;
+   bo->map = NULL;
+}
+
+__attribute__ ((visibility ("default"))) int
+close(int fd)
+{
+   if (fd == drm_fd)
+      drm_fd = -1;
+
+   return libc_close(fd);
+}
+
+static FILE *
+launch_command(char *command)
+{
+   int i = 0, fds[2];
+   char **args = calloc(strlen(command), sizeof(char *));
+   char *iter = command;
+
+   args[i++] = iter = command;
+
+   while ((iter = strstr(iter, ",")) != NULL) {
+      *iter = '\0';
+      iter += 1;
+      args[i++] = iter;
+   }
+
+   if (pipe(fds) == -1)
+      return NULL;
+
+   switch (fork()) {
+   case 0:
+      dup2(fds[0], 0);
+      fail_if(execvp(args[0], args) == -1,
+              "intel_aubdump: failed to launch child command\n");
+      return NULL;
+
+   default:
+      free(args);
+      return fdopen(fds[1], "w");
+
+   case -1:
+      return NULL;
+   }
+}
+
+static void
+maybe_init(void)
+{
+   static bool initialized = false;
+   FILE *config;
+   char *key, *value;
+
+   if (initialized)
+      return;
+
+   initialized = true;
+
+   config = fdopen(3, "r");
+   while (fscanf(config, "%m[^=]=%m[^\n]\n", &key, &value) != EOF) {
+      if (!strcmp(key, "verbose")) {
+         verbose = 1;
+      } else if (!strcmp(key, "device")) {
+         fail_if(sscanf(value, "%i", &device) != 1,
+                 "intel_aubdump: failed to parse device id '%s'",
+                 value);
+         device_override = true;
+      } else if (!strcmp(key, "file")) {
+         filename = strdup(value);
+         files[0] = fopen(filename, "w+");
+         fail_if(files[0] == NULL,
+                 "intel_aubdump: failed to open file '%s'\n",
+                 filename);
+      } else if (!strcmp(key,  "command")) {
+         files[1] = launch_command(value);
+         fail_if(files[1] == NULL,
+                 "intel_aubdump: failed to launch command '%s'\n",
+                 value);
+      } else {
+         fprintf(stderr, "intel_aubdump: unknown option '%s'\n", key);
+      }
+
+      free(key);
+      free(value);
+   }
+   fclose(config);
+
+   bos = calloc(MAX_BO_COUNT, sizeof(bos[0]));
+   fail_if(bos == NULL, "intel_aubdump: out of memory\n");
+}
+
+#define LOCAL_IOCTL_I915_GEM_EXECBUFFER2_WR                             \
+   DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_EXECBUFFER2, struct drm_i915_gem_execbuffer2)
+
+__attribute__ ((visibility ("default"))) int
+ioctl(int fd, unsigned long request, ...)
+{
+   va_list args;
+   void *argp;
+   int ret;
+   struct stat buf;
+
+   va_start(args, request);
+   argp = va_arg(args, void *);
+   va_end(args);
+
+   if (_IOC_TYPE(request) == DRM_IOCTL_BASE &&
+       drm_fd != fd && fstat(fd, &buf) == 0 &&
+       (buf.st_mode & S_IFMT) == S_IFCHR && major(buf.st_rdev) == DRM_MAJOR) {
+      drm_fd = fd;
+      if (verbose)
+         printf("[intel_aubdump: intercept drm ioctl on fd %d]\n", fd);
+   }
+
+   if (fd == drm_fd) {
+      maybe_init();
+
+      switch (request) {
+      case DRM_IOCTL_I915_GETPARAM: {
+         struct drm_i915_getparam *getparam = argp;
+
+         if (device_override && getparam->param == I915_PARAM_CHIPSET_ID) {
+            *getparam->value = device;
+            return 0;
+         }
+
+         ret = libc_ioctl(fd, request, argp);
+
+         /* If the application looks up chipset_id
+          * (they typically do), we'll piggy-back on
+          * their ioctl and store the id for later
+          * use. */
+         if (getparam->param == I915_PARAM_CHIPSET_ID)
+            device = *getparam->value;
+
+         return ret;
+      }
+
+      case DRM_IOCTL_I915_GEM_EXECBUFFER: {
+         static bool once;
+         if (!once) {
+            fprintf(stderr, "intel_aubdump: "
+                    "application uses DRM_IOCTL_I915_GEM_EXECBUFFER, not handled\n");
+            once = true;
+         }
+         return libc_ioctl(fd, request, argp);
+      }
+
+      case DRM_IOCTL_I915_GEM_EXECBUFFER2:
+      case LOCAL_IOCTL_I915_GEM_EXECBUFFER2_WR: {
+         dump_execbuffer2(fd, argp);
+         if (device_override)
+            return 0;
+
+         return libc_ioctl(fd, request, argp);
+      }
+
+      case DRM_IOCTL_I915_GEM_CREATE: {
+         struct drm_i915_gem_create *create = argp;
+
+         ret = libc_ioctl(fd, request, argp);
+         if (ret == 0)
+            add_new_bo(create->handle, create->size, NULL);
+
+         return ret;
+      }
+
+      case DRM_IOCTL_I915_GEM_USERPTR: {
+         struct drm_i915_gem_userptr *userptr = argp;
+
+         ret = libc_ioctl(fd, request, argp);
+         if (ret == 0)
+            add_new_bo(userptr->handle, userptr->user_size,
+                       (void *) (uintptr_t) (userptr->user_ptr | USERPTR_FLAG));
+         return ret;
+      }
+
+      case DRM_IOCTL_GEM_CLOSE: {
+         struct drm_gem_close *close = argp;
+
+         remove_bo(close->handle);
+
+         return libc_ioctl(fd, request, argp);
+      }
+
+      case DRM_IOCTL_GEM_OPEN: {
+         struct drm_gem_open *open = argp;
+
+         ret = libc_ioctl(fd, request, argp);
+         if (ret == 0)
+            add_new_bo(open->handle, open->size, NULL);
+
+         return ret;
+      }
+
+      case DRM_IOCTL_PRIME_FD_TO_HANDLE: {
+         struct drm_prime_handle *prime = argp;
+
+         ret = libc_ioctl(fd, request, argp);
+         if (ret == 0) {
+            off_t size;
+
+            size = lseek(prime->fd, 0, SEEK_END);
+            fail_if(size == -1, "intel_aubdump: failed to get prime bo size\n");
+            add_new_bo(prime->handle, size, NULL);
+         }
+
+         return ret;
+      }
+
+      default:
+         return libc_ioctl(fd, request, argp);
+      }
+   } else {
+      return libc_ioctl(fd, request, argp);
+   }
+}
+
+static void
+init(void)
+{
+   libc_close = dlsym(RTLD_NEXT, "close");
+   libc_ioctl = dlsym(RTLD_NEXT, "ioctl");
+   fail_if(libc_close == NULL || libc_ioctl == NULL,
+           "intel_aubdump: failed to get libc ioctl or close\n");
+}
+
+static int
+close_init_helper(int fd)
+{
+   init();
+   return libc_close(fd);
+}
+
+static int
+ioctl_init_helper(int fd, unsigned long request, ...)
+{
+   va_list args;
+   void *argp;
+
+   va_start(args, request);
+   argp = va_arg(args, void *);
+   va_end(args);
+
+   init();
+   return libc_ioctl(fd, request, argp);
+}
+
+static void __attribute__ ((destructor))
+fini(void)
+{
+   free(filename);
+   for (int i = 0; i < ARRAY_SIZE(files); i++) {
+      if (files[i] != NULL)
+         fclose(files[i]);
+   }
+   free(bos);
+}
diff --git a/src/intel/tools/intel_dump_gpu.in b/src/intel/tools/intel_dump_gpu.in
new file mode 100755 (executable)
index 0000000..875a67e
--- /dev/null
@@ -0,0 +1,107 @@
+#!/bin/bash
+# -*- mode: sh -*-
+
+function show_help() {
+    cat <<EOF
+Usage: intel_dump_gpu [OPTION]... [--] COMMAND ARGUMENTS
+
+Run COMMAND with ARGUMENTS and dump an AUB file that captures buffer
+contents and execution of the GEM application.
+
+  -o, --output=FILE  Name of AUB file. Defaults to COMMAND.aub
+
+  -c, --command=CMD  Execute CMD and write the AUB file's content to its
+                     standard input
+
+      --device=ID    Override PCI ID of the reported device
+
+  -v                 Enable verbose output
+
+      --help         Display this help message and exit
+
+EOF
+
+    exit 0
+}
+
+args=""
+command=""
+file=""
+
+function add_arg() {
+    arg=$1
+    args="$args$arg\n"
+}
+
+function build_command () {
+    command=""
+    for i in $1; do
+        if [ -z $command ]; then
+            command=$i
+        else
+            command="$command,$i"
+        fi;
+    done
+}
+
+while true; do
+    case "$1" in
+        -o)
+            file=$2
+            add_arg "file=${file:-$(basename ${file}).aub}"
+            shift 2
+            ;;
+        -v)
+            add_arg "verbose=1"
+            shift 1
+            ;;
+        -o*)
+            file=${1##-o}
+            add_arg "file=${file:-$(basename ${file}).aub}"
+            shift
+            ;;
+        --output=*)
+            file=${1##--output=}
+            add_arg "file=${file:-$(basename ${file}).aub}"
+            shift
+            ;;
+        -c)
+            build_command "$2"
+            add_arg "command=$command"
+            shift 2
+            ;;
+        --command=*)
+            build_command "${1##--command=}"
+            add_arg "command=$command"
+            shift
+            ;;
+        --device=*)
+            add_arg "device=${1##--device=}"
+            shift
+            ;;
+        --help)
+            show_help
+            ;;
+        --)
+            shift
+            break
+            ;;
+        -*)
+            echo "intel_aubdump: invalid option: $1"
+            echo
+            show_help
+            ;;
+        *)
+            break
+            ;;
+    esac
+done
+
+[ -z $1 ] && show_help
+
+[ -z $file ] && [ -z $command ] && add_arg "file=intel.aub"
+
+LD_PRELOAD="@install_libexecdir@/libintel_dump_gpu.so${LD_PPRELOAD:+:$LD_PRELOAD}" \
+          exec -- "$@" 3<<EOF
+`echo -e $args`
+EOF
index 0be530546c3e00fd37da5b7c9e34c3ed3ba79cc7..705a353f26a8ef0f901fabe51216072d2e2a4c57 100644 (file)
@@ -63,4 +63,22 @@ if with_tools.contains('intel')
     install_dir: get_option('libexecdir'),
     install: true,
   )
+
+  configure_file(
+    input : 'intel_dump_gpu.in',
+    output : '@BASENAME@',
+    install_dir: get_option('bindir'),
+    configuration: sanitize_data,
+  )
+
+  libintel_dump_gpu = shared_library(
+    'intel_dump_gpu',
+    files('intel_dump_gpu.c'),
+    dependencies : dep_dl,
+    include_directories : [inc_common, inc_intel, inc_drm_uapi],
+    link_with : libintel_dev,
+    c_args : [c_vis_args, no_override_init_args],
+    install_dir: get_option('libexecdir'),
+    install : true
+  )
 endif