Re: [Mesa-dev] [PATCH v2 04/16] intel: aubinator: drop the 1Tb GTT mapping
Patch is
Reviewed-by: Rafael Antognolli
On Tue, Jun 19, 2018 at 02:45:19PM +0100, Lionel Landwerlin wrote:
> Now that we're softpinning the address of our BOs in anv & i965, the
> addresses selected start at the top of the addressing space. This is a
> problem for the current implementation of aubinator which uses only a
> 40bit mmapped address space.
>
> This change keeps track of all the memory writes from the aub file and
> fetch them on request by the batch decoder. As a result we can get rid
> of the 1<<40 mmapped address space and only rely on the mmap aub file
> \o/
>
> Signed-off-by: Lionel Landwerlin
> ---
> src/intel/tools/aubinator.c | 130
> 1 file changed, 72 insertions(+), 58 deletions(-)
>
> diff --git a/src/intel/tools/aubinator.c b/src/intel/tools/aubinator.c
> index 0438f96cd1b..f70038376be 100644
> --- a/src/intel/tools/aubinator.c
> +++ b/src/intel/tools/aubinator.c
> @@ -37,10 +37,12 @@
> #include
> #include
>
> +#include "util/list.h"
> #include "util/macros.h"
>
> #include "common/gen_decoder.h"
> #include "common/gen_disasm.h"
> +#include "common/gen_gem.h"
> #include "intel_aub.h"
>
> /* Below is the only command missing from intel_aub.h in libdrm
> @@ -68,8 +70,12 @@ char *input_file = NULL, *xml_path = NULL;
> struct gen_device_info devinfo;
> struct gen_batch_decode_ctx batch_ctx;
>
> -uint64_t gtt_size, gtt_end;
> -void *gtt;
> +struct bo_map {
> + struct list_head link;
> + struct gen_batch_decode_bo bo;
> +};
> +
> +static struct list_head maps;
>
> FILE *outfile;
>
> @@ -85,10 +91,32 @@ field(uint32_t value, int start, int end)
>
> struct brw_instruction;
>
> -static inline int
> -valid_offset(uint32_t offset)
> +static void
> +add_gtt_bo_map(struct gen_batch_decode_bo bo)
> {
> - return offset < gtt_end;
> + struct bo_map *m = calloc(1, sizeof(*m));
> +
> + m->bo = bo;
> + list_add(&m->link, &maps);
> +}
> +
> +static void
> +clear_bo_maps(void)
> +{
> + list_for_each_entry_safe(struct bo_map, i, &maps, link) {
> + list_del(&i->link);
> + free(i);
> + }
> +}
> +
> +static struct gen_batch_decode_bo
> +get_gen_batch_bo(void *user_data, uint64_t address)
> +{
> + list_for_each_entry(struct bo_map, i, &maps, link)
> + if (i->bo.addr <= address && i->bo.addr + i->bo.size > address)
> + return i->bo;
> +
> + return (struct gen_batch_decode_bo) { .map = NULL };
> }
>
> #define GEN_ENGINE_RENDER 1
> @@ -100,26 +128,23 @@ handle_trace_block(uint32_t *p)
> int operation = p[1] & AUB_TRACE_OPERATION_MASK;
> int type = p[1] & AUB_TRACE_TYPE_MASK;
> int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
> - uint64_t offset = p[3];
> - uint32_t size = p[4];
> int header_length = p[0] & 0x;
> - uint32_t *data = p + header_length + 2;
> int engine = GEN_ENGINE_RENDER;
> -
> - if (devinfo.gen >= 8)
> - offset += (uint64_t) p[5] << 32;
> + struct gen_batch_decode_bo bo = {
> + .map = p + header_length + 2,
> + /* Addresses written by aubdump here are in canonical form but the
> batch
> + * decoder always gives us addresses with the top 16bits zeroed, so do
> + * the same here.
> + */
> + .addr = gen_48b_address((devinfo.gen >= 8 ? ((uint64_t) p[5] << 32) :
> 0) |
> + ((uint64_t) p[3])),
> + .size = p[4],
> + };
>
> switch (operation) {
> case AUB_TRACE_OP_DATA_WRITE:
> - if (address_space != AUB_TRACE_MEMTYPE_GTT)
> - break;
> - if (gtt_size < offset + size) {
> - fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
> - exit(EXIT_FAILURE);
> - }
> - memcpy((char *) gtt + offset, data, size);
> - if (gtt_end < offset + size)
> - gtt_end = offset + size;
> + if (address_space == AUB_TRACE_MEMTYPE_GTT)
> + add_gtt_bo_map(bo);
>break;
> case AUB_TRACE_OP_COMMAND_WRITE:
>switch (type) {
> @@ -135,27 +160,13 @@ handle_trace_block(uint32_t *p)
>}
>
>(void)engine; /* TODO */
> - gen_print_batch(&batch_ctx, data, size, 0);
> + gen_print_batch(&batch_ctx, bo.map, bo.size, 0);
>
> - gtt_end = 0;
> + clear_bo_maps();
>break;
> }
> }
>
> -static struct gen_batch_decode_bo
> -get_gen_batch_bo(void *user_data, uint64_t address)
> -{
> - if (address > gtt_end)
> - return (struct gen_batch_decode_bo) { .map = NULL };
> -
> - /* We really only have one giant address range */
> - return (struct gen_batch_decode_bo) {
> - .addr = 0,
> - .map = gtt,
> - .size = gtt_size
> - };
> -}
> -
> static void
> aubinator_init(uint16_t aub_pci_id, const char *app_name)
> {
> @@ -305,34 +316,44 @@ handle_memtrace_reg_write(uint32_t *p)
> }
>
> const uint32_t pphwsp_size = 4096;
> - uint32_t *context = (uint32_t*)(gtt + (context_descriptor & 0xf000) +
> pphwsp_size);
> + uint32_t pphw
[Mesa-dev] [PATCH v2 04/16] intel: aubinator: drop the 1Tb GTT mapping
Now that we're softpinning the address of our BOs in anv & i965, the
addresses selected start at the top of the addressing space. This is a
problem for the current implementation of aubinator which uses only a
40bit mmapped address space.
This change keeps track of all the memory writes from the aub file and
fetch them on request by the batch decoder. As a result we can get rid
of the 1<<40 mmapped address space and only rely on the mmap aub file
\o/
Signed-off-by: Lionel Landwerlin
---
src/intel/tools/aubinator.c | 130
1 file changed, 72 insertions(+), 58 deletions(-)
diff --git a/src/intel/tools/aubinator.c b/src/intel/tools/aubinator.c
index 0438f96cd1b..f70038376be 100644
--- a/src/intel/tools/aubinator.c
+++ b/src/intel/tools/aubinator.c
@@ -37,10 +37,12 @@
#include
#include
+#include "util/list.h"
#include "util/macros.h"
#include "common/gen_decoder.h"
#include "common/gen_disasm.h"
+#include "common/gen_gem.h"
#include "intel_aub.h"
/* Below is the only command missing from intel_aub.h in libdrm
@@ -68,8 +70,12 @@ char *input_file = NULL, *xml_path = NULL;
struct gen_device_info devinfo;
struct gen_batch_decode_ctx batch_ctx;
-uint64_t gtt_size, gtt_end;
-void *gtt;
+struct bo_map {
+ struct list_head link;
+ struct gen_batch_decode_bo bo;
+};
+
+static struct list_head maps;
FILE *outfile;
@@ -85,10 +91,32 @@ field(uint32_t value, int start, int end)
struct brw_instruction;
-static inline int
-valid_offset(uint32_t offset)
+static void
+add_gtt_bo_map(struct gen_batch_decode_bo bo)
{
- return offset < gtt_end;
+ struct bo_map *m = calloc(1, sizeof(*m));
+
+ m->bo = bo;
+ list_add(&m->link, &maps);
+}
+
+static void
+clear_bo_maps(void)
+{
+ list_for_each_entry_safe(struct bo_map, i, &maps, link) {
+ list_del(&i->link);
+ free(i);
+ }
+}
+
+static struct gen_batch_decode_bo
+get_gen_batch_bo(void *user_data, uint64_t address)
+{
+ list_for_each_entry(struct bo_map, i, &maps, link)
+ if (i->bo.addr <= address && i->bo.addr + i->bo.size > address)
+ return i->bo;
+
+ return (struct gen_batch_decode_bo) { .map = NULL };
}
#define GEN_ENGINE_RENDER 1
@@ -100,26 +128,23 @@ handle_trace_block(uint32_t *p)
int operation = p[1] & AUB_TRACE_OPERATION_MASK;
int type = p[1] & AUB_TRACE_TYPE_MASK;
int address_space = p[1] & AUB_TRACE_ADDRESS_SPACE_MASK;
- uint64_t offset = p[3];
- uint32_t size = p[4];
int header_length = p[0] & 0x;
- uint32_t *data = p + header_length + 2;
int engine = GEN_ENGINE_RENDER;
-
- if (devinfo.gen >= 8)
- offset += (uint64_t) p[5] << 32;
+ struct gen_batch_decode_bo bo = {
+ .map = p + header_length + 2,
+ /* Addresses written by aubdump here are in canonical form but the batch
+ * decoder always gives us addresses with the top 16bits zeroed, so do
+ * the same here.
+ */
+ .addr = gen_48b_address((devinfo.gen >= 8 ? ((uint64_t) p[5] << 32) : 0)
|
+ ((uint64_t) p[3])),
+ .size = p[4],
+ };
switch (operation) {
case AUB_TRACE_OP_DATA_WRITE:
- if (address_space != AUB_TRACE_MEMTYPE_GTT)
- break;
- if (gtt_size < offset + size) {
- fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
- exit(EXIT_FAILURE);
- }
- memcpy((char *) gtt + offset, data, size);
- if (gtt_end < offset + size)
- gtt_end = offset + size;
+ if (address_space == AUB_TRACE_MEMTYPE_GTT)
+ add_gtt_bo_map(bo);
break;
case AUB_TRACE_OP_COMMAND_WRITE:
switch (type) {
@@ -135,27 +160,13 @@ handle_trace_block(uint32_t *p)
}
(void)engine; /* TODO */
- gen_print_batch(&batch_ctx, data, size, 0);
+ gen_print_batch(&batch_ctx, bo.map, bo.size, 0);
- gtt_end = 0;
+ clear_bo_maps();
break;
}
}
-static struct gen_batch_decode_bo
-get_gen_batch_bo(void *user_data, uint64_t address)
-{
- if (address > gtt_end)
- return (struct gen_batch_decode_bo) { .map = NULL };
-
- /* We really only have one giant address range */
- return (struct gen_batch_decode_bo) {
- .addr = 0,
- .map = gtt,
- .size = gtt_size
- };
-}
-
static void
aubinator_init(uint16_t aub_pci_id, const char *app_name)
{
@@ -305,34 +316,44 @@ handle_memtrace_reg_write(uint32_t *p)
}
const uint32_t pphwsp_size = 4096;
- uint32_t *context = (uint32_t*)(gtt + (context_descriptor & 0xf000) +
pphwsp_size);
+ uint32_t pphwsp_addr = context_descriptor & 0xf000;
+ struct gen_batch_decode_bo pphwsp_bo = get_gen_batch_bo(NULL, pphwsp_addr);
+ uint32_t *context = (uint32_t *)((uint8_t *)pphwsp_bo.map +
+(pphwsp_bo.addr - pphwsp_addr) +
+pphwsp_size);
+
uint32_t ring_buffer_head = context[5];
uint32_t ring_buffer_tail = context[7];
uint32_t ring_buffer_start = co
