#include <inttypes.h>
#include <signal.h>
+#include <stdarg.h>
#include <stdlib.h>
#include <string.h>
-#include "i915_drm.h"
+#include "drm-uapi/i915_drm.h"
#include "intel_aub.h"
+#include "gen_context.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_NON_SECURE_I965 (1 << 8)
-#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 GEN11_LR_CONTEXT_RENDER_SIZE (14 * 4096)
-#define GEN10_LR_CONTEXT_RENDER_SIZE (19 * 4096)
-#define GEN9_LR_CONTEXT_RENDER_SIZE (22 * 4096)
-#define GEN8_LR_CONTEXT_RENDER_SIZE (20 * 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 PML4_PHYS_ADDR ((uint64_t)(STATIC_GGTT_MAP_END))
-
-#define CONTEXT_FLAGS (0x339) /* Normal Priority | L3-LLC Coherency |
- * PPGTT Enabled |
- * Legacy Context with 64 bit VA support |
- * Valid
- */
-
-#define RENDER_CONTEXT_DESCRIPTOR ((uint64_t)1 << 62 | RENDER_CONTEXT_ADDR | CONTEXT_FLAGS)
-#define BLITTER_CONTEXT_DESCRIPTOR ((uint64_t)2 << 62 | BLITTER_CONTEXT_ADDR | CONTEXT_FLAGS)
-#define VIDEO_CONTEXT_DESCRIPTOR ((uint64_t)3 << 62 | VIDEO_CONTEXT_ADDR | CONTEXT_FLAGS)
-
-static const uint32_t render_context_init[GEN9_LR_CONTEXT_RENDER_SIZE / /* Choose the largest */
- 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 */, PML4_PHYS_ADDR >> 32,
- 0x2270 /* PDP0_LDW */, PML4_PHYS_ADDR,
- /* 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 */, PML4_PHYS_ADDR >> 32,
- 0x22270 /* PDP0_LDW */, PML4_PHYS_ADDR,
- /* 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 */, PML4_PHYS_ADDR >> 32,
- 0x1C270 /* PDP0_LDW */, PML4_PHYS_ADDR,
- /* MI_NOOP */
- 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-
- MI_BATCH_BUFFER_END
-};
+static struct aub_context *aub_context_new(struct aub_file *aub, uint32_t new_id);
+static void mem_trace_memory_write_header_out(struct aub_file *aub, uint64_t addr,
+ uint32_t len, uint32_t addr_space,
+ const char *desc);
static void __attribute__ ((format(__printf__, 2, 3)))
fail_if(int cond, const char *format, ...)
}
static void
-aub_ppgtt_table_finish(struct aub_ppgtt_table *table)
+aub_ppgtt_table_finish(struct aub_ppgtt_table *table, int level)
{
+ if (level == 1)
+ return;
+
for (unsigned i = 0; i < ARRAY_SIZE(table->subtables); i++) {
- aub_ppgtt_table_finish(table->subtables[i]);
- free(table->subtables[i]);
+ if (table->subtables[i]) {
+ aub_ppgtt_table_finish(table->subtables[i], level - 1);
+ free(table->subtables[i]);
+ }
}
}
+static void
+data_out(struct aub_file *aub, const void *data, size_t size)
+{
+ if (size == 0)
+ return;
+
+ fail_if(fwrite(data, 1, size, aub->file) == 0,
+ "Writing to output failed\n");
+}
+
+static void
+dword_out(struct aub_file *aub, uint32_t data)
+{
+ data_out(aub, &data, sizeof(data));
+}
+
+static void
+write_execlists_header(struct aub_file *aub, const char *name)
+{
+ char app_name[8 * 4];
+ int app_name_len, dwords;
+
+ app_name_len =
+ snprintf(app_name, sizeof(app_name), "PCI-ID=0x%X %s",
+ aub->pci_id, name);
+ app_name_len = ALIGN(app_name_len, sizeof(uint32_t));
+
+ dwords = 5 + app_name_len / sizeof(uint32_t);
+ dword_out(aub, CMD_MEM_TRACE_VERSION | (dwords - 1));
+ dword_out(aub, AUB_MEM_TRACE_VERSION_FILE_VERSION);
+ dword_out(aub, aub->devinfo.simulator_id << AUB_MEM_TRACE_VERSION_DEVICE_SHIFT);
+ dword_out(aub, 0); /* version */
+ dword_out(aub, 0); /* version */
+ data_out(aub, app_name, app_name_len);
+}
+
+static void
+write_legacy_header(struct aub_file *aub, const char *name)
+{
+ char app_name[8 * 4];
+ char comment[16];
+ int comment_len, comment_dwords, dwords;
+
+ comment_len = snprintf(comment, sizeof(comment), "PCI-ID=0x%x", aub->pci_id);
+ comment_dwords = ((comment_len + 3) / 4);
+
+ /* Start with a (required) version packet. */
+ dwords = 13 + comment_dwords;
+ dword_out(aub, CMD_AUB_HEADER | (dwords - 2));
+ dword_out(aub, (4 << AUB_HEADER_MAJOR_SHIFT) |
+ (0 << AUB_HEADER_MINOR_SHIFT));
+
+ /* Next comes a 32-byte application name. */
+ strncpy(app_name, name, sizeof(app_name));
+ app_name[sizeof(app_name) - 1] = 0;
+ data_out(aub, app_name, sizeof(app_name));
+
+ dword_out(aub, 0); /* timestamp */
+ dword_out(aub, 0); /* timestamp */
+ dword_out(aub, comment_len);
+ data_out(aub, comment, comment_dwords * 4);
+}
+
+
+static void
+aub_write_header(struct aub_file *aub, const char *app_name)
+{
+ if (aub_use_execlists(aub))
+ write_execlists_header(aub, app_name);
+ else
+ write_legacy_header(aub, app_name);
+}
+
void
-aub_file_init(struct aub_file *aub, FILE *file, uint16_t pci_id)
+aub_file_init(struct aub_file *aub, FILE *file, FILE *debug, uint16_t pci_id, const char *app_name)
{
memset(aub, 0, sizeof(*aub));
+ aub->verbose_log_file = debug;
aub->file = file;
aub->pci_id = pci_id;
- fail_if(!gen_get_device_info(pci_id, &aub->devinfo),
+ fail_if(!gen_get_device_info_from_pci_id(pci_id, &aub->devinfo),
"failed to identify chipset=0x%x\n", pci_id);
aub->addr_bits = aub->devinfo.gen >= 8 ? 48 : 32;
- aub->pml4.phys_addr = PML4_PHYS_ADDR;
+ aub_write_header(aub, app_name);
+
+ aub->phys_addrs_allocator = 0;
+ aub->ggtt_addrs_allocator = 0;
+ aub->pml4.phys_addr = aub->phys_addrs_allocator++ << 12;
+
+ mem_trace_memory_write_header_out(aub, aub->ggtt_addrs_allocator++,
+ GEN8_PTE_SIZE,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT_ENTRY,
+ "GGTT PT");
+ dword_out(aub, 1);
+ dword_out(aub, 0);
+
+ aub->next_context_handle = 1;
+ aub_context_new(aub, 0); /* Default context */
}
void
aub_file_finish(struct aub_file *aub)
{
- aub_ppgtt_table_finish(&aub->pml4);
+ aub_ppgtt_table_finish(&aub->pml4, 4);
fclose(aub->file);
}
return NUM_PT_ENTRIES * (aub->addr_bits > 32 ? GEN8_PTE_SIZE : PTE_SIZE);
}
-static void
-data_out(struct aub_file *aub, const void *data, size_t size)
-{
- if (size == 0)
- return;
-
- fail_if(fwrite(data, 1, size, aub->file) == 0,
- "Writing to output failed\n");
-}
-
-static void
-dword_out(struct aub_file *aub, uint32_t data)
-{
- data_out(aub, &data, sizeof(data));
-}
-
static void
mem_trace_memory_write_header_out(struct aub_file *aub, uint64_t addr,
- uint32_t len, uint32_t addr_space)
+ uint32_t len, uint32_t addr_space,
+ const char *desc)
{
uint32_t dwords = ALIGN(len, sizeof(uint32_t)) / sizeof(uint32_t);
+ if (aub->verbose_log_file) {
+ fprintf(aub->verbose_log_file,
+ " MEM WRITE (0x%016" PRIx64 "-0x%016" PRIx64 ") %s\n",
+ addr, addr + len, desc);
+ }
+
dword_out(aub, CMD_MEM_TRACE_MEMORY_WRITE | (5 + dwords - 1));
dword_out(aub, addr & 0xFFFFFFFF); /* addr lo */
dword_out(aub, addr >> 32); /* addr hi */
{
uint32_t dwords = 1;
+ if (aub->verbose_log_file) {
+ fprintf(aub->verbose_log_file,
+ " MMIO WRITE (0x%08x = 0x%08x)\n", addr, value);
+ }
+
dword_out(aub, CMD_MEM_TRACE_REGISTER_WRITE | (5 + dwords - 1));
dword_out(aub, addr);
dword_out(aub, AUB_MEM_TRACE_REGISTER_SIZE_DWORD |
populate_ppgtt_table(struct aub_file *aub, struct aub_ppgtt_table *table,
int start, int end, int level)
{
- static uint64_t phys_addrs_allocator = (PML4_PHYS_ADDR >> 12) + 1;
uint64_t entries[512] = {0};
int dirty_start = 512, dirty_end = 0;
dirty_end = max(dirty_end, i);
if (level == 1) {
table->subtables[i] =
- (void *)(phys_addrs_allocator++ << 12);
+ (void *)(aub->phys_addrs_allocator++ << 12);
if (aub->verbose_log_file) {
fprintf(aub->verbose_log_file,
" Adding entry: %x, phys_addr: 0x%016" PRIx64 "\n",
table->subtables[i] =
calloc(1, sizeof(struct aub_ppgtt_table));
table->subtables[i]->phys_addr =
- phys_addrs_allocator++ << 12;
+ aub->phys_addrs_allocator++ << 12;
if (aub->verbose_log_file) {
fprintf(aub->verbose_log_file,
" Adding entry: %x, phys_addr: 0x%016" PRIx64 "\n",
uint64_t write_size = (dirty_end - dirty_start + 1) *
sizeof(uint64_t);
mem_trace_memory_write_header_out(aub, write_addr, write_size,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_PHYSICAL);
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_PHYSICAL,
+ "PPGTT update");
data_out(aub, entries + dirty_start, write_size);
}
}
return (uint64_t)L1_table(ppgtt_addr)->subtables[L1_index(ppgtt_addr)];
}
+static const struct engine {
+ const char *name;
+ enum drm_i915_gem_engine_class engine_class;
+ uint32_t hw_class;
+ uint32_t elsp_reg;
+ uint32_t elsq_reg;
+ uint32_t status_reg;
+ uint32_t control_reg;
+} engines[] = {
+ [I915_ENGINE_CLASS_RENDER] = {
+ .name = "RENDER",
+ .engine_class = I915_ENGINE_CLASS_RENDER,
+ .hw_class = 1,
+ .elsp_reg = EXECLIST_SUBMITPORT_RCSUNIT,
+ .elsq_reg = EXECLIST_SQ_CONTENTS0_RCSUNIT,
+ .status_reg = EXECLIST_STATUS_RCSUNIT,
+ .control_reg = EXECLIST_CONTROL_RCSUNIT,
+ },
+ [I915_ENGINE_CLASS_VIDEO] = {
+ .name = "VIDEO",
+ .engine_class = I915_ENGINE_CLASS_VIDEO,
+ .hw_class = 3,
+ .elsp_reg = EXECLIST_SUBMITPORT_VCSUNIT0,
+ .elsq_reg = EXECLIST_SQ_CONTENTS0_VCSUNIT0,
+ .status_reg = EXECLIST_STATUS_VCSUNIT0,
+ .control_reg = EXECLIST_CONTROL_VCSUNIT0,
+ },
+ [I915_ENGINE_CLASS_COPY] = {
+ .name = "BLITTER",
+ .engine_class = I915_ENGINE_CLASS_COPY,
+ .hw_class = 2,
+ .elsp_reg = EXECLIST_SUBMITPORT_BCSUNIT,
+ .elsq_reg = EXECLIST_SQ_CONTENTS0_BCSUNIT,
+ .status_reg = EXECLIST_STATUS_BCSUNIT,
+ .control_reg = EXECLIST_CONTROL_BCSUNIT,
+ },
+};
+
static void
-write_execlists_header(struct aub_file *aub, const char *name)
+aub_map_ggtt(struct aub_file *aub, uint64_t virt_addr, uint64_t size)
{
- char app_name[8 * 4];
- int app_name_len, dwords;
-
- app_name_len =
- snprintf(app_name, sizeof(app_name), "PCI-ID=0x%X %s",
- aub->pci_id, name);
- app_name_len = ALIGN(app_name_len, sizeof(uint32_t));
-
- dwords = 5 + app_name_len / sizeof(uint32_t);
- dword_out(aub, CMD_MEM_TRACE_VERSION | (dwords - 1));
- dword_out(aub, AUB_MEM_TRACE_VERSION_FILE_VERSION);
- dword_out(aub, aub->devinfo.simulator_id << AUB_MEM_TRACE_VERSION_DEVICE_SHIFT);
- dword_out(aub, 0); /* version */
- dword_out(aub, 0); /* version */
- data_out(aub, app_name, app_name_len);
+ /* Makes the code below a bit simpler. In practice all of the write we
+ * receive from error2aub are page aligned.
+ */
+ assert(virt_addr % 4096 == 0);
+ assert((aub->phys_addrs_allocator + size) < (1UL << 32));
/* GGTT PT */
- uint32_t ggtt_ptes = STATIC_GGTT_MAP_SIZE >> 12;
+ uint32_t ggtt_ptes = DIV_ROUND_UP(size, 4096);
+ uint64_t phys_addr = aub->phys_addrs_allocator << 12;
+ aub->phys_addrs_allocator += ggtt_ptes;
+
+ if (aub->verbose_log_file) {
+ fprintf(aub->verbose_log_file,
+ " Mapping GGTT address: 0x%" PRIx64 ", size: %" PRIu64" phys_addr=0x%" PRIx64 " entries=%u\n",
+ virt_addr, size, phys_addr, ggtt_ptes);
+ }
- mem_trace_memory_write_header_out(aub, STATIC_GGTT_MAP_START >> 12,
+ mem_trace_memory_write_header_out(aub,
+ (virt_addr >> 12) * GEN8_PTE_SIZE,
ggtt_ptes * GEN8_PTE_SIZE,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT_ENTRY);
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT_ENTRY,
+ "GGTT PT");
for (uint32_t i = 0; i < ggtt_ptes; i++) {
- dword_out(aub, 1 + 0x1000 * i + STATIC_GGTT_MAP_START);
+ dword_out(aub, 1 + phys_addr + i * 4096);
dword_out(aub, 0);
}
+}
- /* RENDER_RING */
- mem_trace_memory_write_header_out(aub, RENDER_RING_ADDR, RING_SIZE,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
- for (uint32_t i = 0; i < RING_SIZE; i += sizeof(uint32_t))
- dword_out(aub, 0);
+void
+aub_write_ggtt(struct aub_file *aub, uint64_t virt_addr, uint64_t size, const void *data)
+{
+ /* Default setup assumes a 1 to 1 mapping between physical and virtual GGTT
+ * addresses. This is somewhat incompatible with the aub_write_ggtt()
+ * function. In practice it doesn't matter as the GGTT writes are used to
+ * replace the default setup and we've taken care to setup the PML4 as the
+ * top of the GGTT.
+ */
+ assert(!aub->has_default_setup);
- /* RENDER_PPHWSP */
- mem_trace_memory_write_header_out(aub, RENDER_CONTEXT_ADDR,
- PPHWSP_SIZE +
- sizeof(render_context_init),
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
- for (uint32_t i = 0; i < PPHWSP_SIZE; i += sizeof(uint32_t))
- dword_out(aub, 0);
+ aub_map_ggtt(aub, virt_addr, size);
- /* RENDER_CONTEXT */
- data_out(aub, render_context_init, sizeof(render_context_init));
+ /* We write the GGTT buffer through the GGTT aub command rather than the
+ * PHYSICAL aub command. This is because the Gen9 simulator seems to have 2
+ * different set of memory pools for GGTT and physical (probably someone
+ * didn't really understand the concept?).
+ */
+ static const char null_block[8 * 4096];
+ for (uint64_t offset = 0; offset < size; offset += 4096) {
+ uint32_t block_size = min(4096, size - offset);
- /* BLITTER_RING */
- mem_trace_memory_write_header_out(aub, BLITTER_RING_ADDR, RING_SIZE,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
- for (uint32_t i = 0; i < RING_SIZE; i += sizeof(uint32_t))
- dword_out(aub, 0);
+ mem_trace_memory_write_header_out(aub, virt_addr + offset, block_size,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT,
+ "GGTT buffer");
+ data_out(aub, (char *) data + offset, block_size);
- /* BLITTER_PPHWSP */
- mem_trace_memory_write_header_out(aub, BLITTER_CONTEXT_ADDR,
- PPHWSP_SIZE +
- sizeof(blitter_context_init),
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
- for (uint32_t i = 0; i < PPHWSP_SIZE; i += sizeof(uint32_t))
- dword_out(aub, 0);
+ /* Pad to a multiple of 4 bytes. */
+ data_out(aub, null_block, -block_size & 3);
+ }
+}
- /* BLITTER_CONTEXT */
- data_out(aub, blitter_context_init, sizeof(blitter_context_init));
+static const struct engine *
+engine_from_engine_class(enum drm_i915_gem_engine_class engine_class)
+{
+ switch (engine_class) {
+ case I915_ENGINE_CLASS_RENDER:
+ case I915_ENGINE_CLASS_COPY:
+ case I915_ENGINE_CLASS_VIDEO:
+ return &engines[engine_class];
+ default:
+ unreachable("unknown ring");
+ }
+}
+
+static void
+get_context_init(const struct gen_device_info *devinfo,
+ const struct gen_context_parameters *params,
+ enum drm_i915_gem_engine_class engine_class,
+ uint32_t *data,
+ uint32_t *size)
+{
+ static const gen_context_init_t gen8_contexts[] = {
+ [I915_ENGINE_CLASS_RENDER] = gen8_render_context_init,
+ [I915_ENGINE_CLASS_COPY] = gen8_blitter_context_init,
+ [I915_ENGINE_CLASS_VIDEO] = gen8_video_context_init,
+ };
+ static const gen_context_init_t gen10_contexts[] = {
+ [I915_ENGINE_CLASS_RENDER] = gen10_render_context_init,
+ [I915_ENGINE_CLASS_COPY] = gen10_blitter_context_init,
+ [I915_ENGINE_CLASS_VIDEO] = gen10_video_context_init,
+ };
+
+ assert(devinfo->gen >= 8);
+
+ if (devinfo->gen <= 10)
+ gen8_contexts[engine_class](params, data, size);
+ else
+ gen10_contexts[engine_class](params, data, size);
+}
+
+static uint64_t
+alloc_ggtt_address(struct aub_file *aub, uint64_t size)
+{
+ uint32_t ggtt_ptes = DIV_ROUND_UP(size, 4096);
+ uint64_t addr = aub->ggtt_addrs_allocator << 12;
- /* VIDEO_RING */
- mem_trace_memory_write_header_out(aub, VIDEO_RING_ADDR, RING_SIZE,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
+ aub->ggtt_addrs_allocator += ggtt_ptes;
+ aub_map_ggtt(aub, addr, size);
+
+ return addr;
+}
+
+static void
+write_hwsp(struct aub_file *aub,
+ enum drm_i915_gem_engine_class engine_class)
+{
+ uint32_t reg = 0;
+ switch (engine_class) {
+ case I915_ENGINE_CLASS_RENDER: reg = HWS_PGA_RCSUNIT; break;
+ case I915_ENGINE_CLASS_COPY: reg = HWS_PGA_BCSUNIT; break;
+ case I915_ENGINE_CLASS_VIDEO: reg = HWS_PGA_VCSUNIT0; break;
+ default:
+ unreachable("unknown ring");
+ }
+
+ register_write_out(aub, reg, aub->engine_setup[engine_class].hwsp_addr);
+}
+
+static uint32_t
+write_engine_execlist_setup(struct aub_file *aub,
+ uint32_t ctx_id,
+ struct aub_hw_context *hw_ctx,
+ enum drm_i915_gem_engine_class engine_class)
+{
+ const struct engine *cs = engine_from_engine_class(engine_class);
+ uint32_t context_size;
+
+ get_context_init(&aub->devinfo, NULL, engine_class, NULL, &context_size);
+
+ /* GGTT PT */
+ uint32_t total_size = RING_SIZE + PPHWSP_SIZE + context_size;
+ char name[80];
+ uint64_t ggtt_addr = alloc_ggtt_address(aub, total_size);
+
+ snprintf(name, sizeof(name), "%s (ctx id: %d) GGTT PT", cs->name, ctx_id);
+
+ /* RING */
+ hw_ctx->ring_addr = ggtt_addr;
+ snprintf(name, sizeof(name), "%s RING", cs->name);
+ mem_trace_memory_write_header_out(aub, ggtt_addr, RING_SIZE,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT,
+ name);
for (uint32_t i = 0; i < RING_SIZE; i += sizeof(uint32_t))
dword_out(aub, 0);
-
- /* VIDEO_PPHWSP */
- mem_trace_memory_write_header_out(aub, VIDEO_CONTEXT_ADDR,
- PPHWSP_SIZE +
- sizeof(video_context_init),
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
+ ggtt_addr += RING_SIZE;
+
+ /* PPHWSP */
+ hw_ctx->pphwsp_addr = ggtt_addr;
+ snprintf(name, sizeof(name), "%s PPHWSP", cs->name);
+ mem_trace_memory_write_header_out(aub, ggtt_addr,
+ PPHWSP_SIZE + context_size,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT,
+ name);
for (uint32_t i = 0; i < PPHWSP_SIZE; i += sizeof(uint32_t))
dword_out(aub, 0);
- /* VIDEO_CONTEXT */
- data_out(aub, video_context_init, sizeof(video_context_init));
+ /* CONTEXT */
+ struct gen_context_parameters params = {
+ .ring_addr = hw_ctx->ring_addr,
+ .ring_size = RING_SIZE,
+ .pml4_addr = aub->pml4.phys_addr,
+ };
+ uint32_t *context_data = calloc(1, context_size);
+ get_context_init(&aub->devinfo, ¶ms, engine_class, context_data, &context_size);
+ data_out(aub, context_data, context_size);
+ free(context_data);
- register_write_out(aub, HWS_PGA_RCSUNIT, RENDER_CONTEXT_ADDR);
- register_write_out(aub, HWS_PGA_VCSUNIT0, VIDEO_CONTEXT_ADDR);
- register_write_out(aub, HWS_PGA_BCSUNIT, BLITTER_CONTEXT_ADDR);
+ hw_ctx->initialized = true;
+ return total_size;
+}
+
+static void
+write_execlists_default_setup(struct aub_file *aub)
+{
register_write_out(aub, GFX_MODE_RCSUNIT, 0x80008000 /* execlist enable */);
register_write_out(aub, GFX_MODE_VCSUNIT0, 0x80008000 /* execlist enable */);
register_write_out(aub, GFX_MODE_BCSUNIT, 0x80008000 /* execlist enable */);
}
-static void write_legacy_header(struct aub_file *aub, const char *name)
+static void write_legacy_default_setup(struct aub_file *aub)
{
- 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", aub->pci_id);
- comment_dwords = ((comment_len + 3) / 4);
-
- /* Start with a (required) version packet. */
- dwords = 13 + comment_dwords;
- dword_out(aub, CMD_AUB_HEADER | (dwords - 2));
- dword_out(aub, (4 << AUB_HEADER_MAJOR_SHIFT) |
- (0 << AUB_HEADER_MINOR_SHIFT));
-
- /* Next comes a 32-byte application name. */
- strncpy(app_name, name, sizeof(app_name));
- app_name[sizeof(app_name) - 1] = 0;
- data_out(aub, app_name, sizeof(app_name));
-
- dword_out(aub, 0); /* timestamp */
- dword_out(aub, 0); /* timestamp */
- dword_out(aub, comment_len);
- data_out(aub, comment, comment_dwords * 4);
-
/* Set up the GTT. The max we can handle is 64M */
dword_out(aub, CMD_AUB_TRACE_HEADER_BLOCK |
((aub->addr_bits > 32 ? 6 : 5) - 2));
}
}
+/**
+ * Sets up a default GGTT/PPGTT address space and execlists context (when
+ * supported).
+ */
void
-aub_write_header(struct aub_file *aub, const char *app_name)
+aub_write_default_setup(struct aub_file *aub)
{
if (aub_use_execlists(aub))
- write_execlists_header(aub, app_name);
+ write_execlists_default_setup(aub);
else
- write_legacy_header(aub, app_name);
+ write_legacy_default_setup(aub);
+
+ aub->has_default_setup = true;
+}
+
+static struct aub_context *
+aub_context_new(struct aub_file *aub, uint32_t new_id)
+{
+ assert(aub->num_contexts < MAX_CONTEXT_COUNT);
+
+ struct aub_context *ctx = &aub->contexts[aub->num_contexts++];
+ memset(ctx, 0, sizeof(*ctx));
+ ctx->id = new_id;
+
+ return ctx;
+}
+
+uint32_t
+aub_write_context_create(struct aub_file *aub, uint32_t *ctx_id)
+{
+ uint32_t new_id = ctx_id ? *ctx_id : aub->next_context_handle;
+
+ aub_context_new(aub, new_id);
+
+ if (!ctx_id)
+ aub->next_context_handle++;
+
+ return new_id;
+}
+
+static struct aub_context *
+aub_context_find(struct aub_file *aub, uint32_t id)
+{
+ for (int i = 0; i < aub->num_contexts; i++) {
+ if (aub->contexts[i].id == id)
+ return &aub->contexts[i];
+ }
+
+ return NULL;
+}
+
+static struct aub_hw_context *
+aub_write_ensure_context(struct aub_file *aub, uint32_t ctx_id,
+ enum drm_i915_gem_engine_class engine_class)
+{
+ struct aub_context *ctx = aub_context_find(aub, ctx_id);
+ assert(ctx != NULL);
+
+ struct aub_hw_context *hw_ctx = &ctx->hw_contexts[engine_class];
+ if (!hw_ctx->initialized)
+ write_engine_execlist_setup(aub, ctx->id, hw_ctx, engine_class);
+
+ return hw_ctx;
+}
+
+static uint64_t
+get_context_descriptor(struct aub_file *aub,
+ const struct engine *cs,
+ struct aub_hw_context *hw_ctx)
+{
+ return cs->hw_class | hw_ctx->pphwsp_addr | CONTEXT_FLAGS;
}
/**
mem_trace_memory_write_header_out(aub,
ppgtt_lookup(aub, gtt_offset + offset),
block_size,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_PHYSICAL);
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_PHYSICAL,
+ "Trace Block");
} else {
dword_out(aub, CMD_AUB_TRACE_HEADER_BLOCK |
((aub->addr_bits > 32 ? 6 : 5) - 2));
}
static void
-aub_dump_execlist(struct aub_file *aub, uint64_t batch_offset, int ring_flag)
+aub_dump_ring_buffer_execlist(struct aub_file *aub,
+ struct aub_hw_context *hw_ctx,
+ const struct engine *cs,
+ uint64_t batch_offset)
{
- 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;
- default:
- unreachable("unknown ring");
- }
-
- mem_trace_memory_write_header_out(aub, ring_addr, 16,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
+ mem_trace_memory_write_header_out(aub, hw_ctx->ring_addr, 16,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT,
+ "RING MI_BATCH_BUFFER_START user");
dword_out(aub, AUB_MI_BATCH_BUFFER_START | MI_BATCH_NON_SECURE_I965 | (3 - 2));
dword_out(aub, batch_offset & 0xFFFFFFFF);
dword_out(aub, batch_offset >> 32);
dword_out(aub, 0 /* MI_NOOP */);
- mem_trace_memory_write_header_out(aub, ring_addr + 8192 + 20, 4,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
+ mem_trace_memory_write_header_out(aub, hw_ctx->ring_addr + 8192 + 20, 4,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT,
+ "RING BUFFER HEAD");
dword_out(aub, 0); /* RING_BUFFER_HEAD */
- mem_trace_memory_write_header_out(aub, ring_addr + 8192 + 28, 4,
- AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT);
+ mem_trace_memory_write_header_out(aub, hw_ctx->ring_addr + 8192 + 28, 4,
+ AUB_MEM_TRACE_MEMORY_ADDRESS_SPACE_GGTT,
+ "RING BUFFER TAIL");
dword_out(aub, 16); /* RING_BUFFER_TAIL */
+}
+static void
+aub_dump_execlist(struct aub_file *aub, const struct engine *cs, uint64_t descriptor)
+{
if (aub->devinfo.gen >= 11) {
- register_write_out(aub, elsq_reg, descriptor & 0xFFFFFFFF);
- register_write_out(aub, elsq_reg + sizeof(uint32_t), descriptor >> 32);
- register_write_out(aub, control_reg, 1);
+ register_write_out(aub, cs->elsq_reg, descriptor & 0xFFFFFFFF);
+ register_write_out(aub, cs->elsq_reg + sizeof(uint32_t), descriptor >> 32);
+ register_write_out(aub, cs->control_reg, 1);
} else {
- register_write_out(aub, elsp_reg, 0);
- register_write_out(aub, elsp_reg, 0);
- register_write_out(aub, elsp_reg, descriptor >> 32);
- register_write_out(aub, elsp_reg, descriptor & 0xFFFFFFFF);
+ register_write_out(aub, cs->elsp_reg, 0);
+ register_write_out(aub, cs->elsp_reg, 0);
+ register_write_out(aub, cs->elsp_reg, descriptor >> 32);
+ register_write_out(aub, cs->elsp_reg, descriptor & 0xFFFFFFFF);
}
dword_out(aub, CMD_MEM_TRACE_REGISTER_POLL | (5 + 1 - 1));
- dword_out(aub, status_reg);
+ dword_out(aub, cs->status_reg);
dword_out(aub, AUB_MEM_TRACE_REGISTER_SIZE_DWORD |
AUB_MEM_TRACE_REGISTER_SPACE_MMIO);
if (aub->devinfo.gen >= 11) {
}
static void
-aub_dump_ringbuffer(struct aub_file *aub, uint64_t batch_offset,
- uint64_t offset, int ring_flag)
+aub_dump_ring_buffer_legacy(struct aub_file *aub,
+ uint64_t batch_offset,
+ uint64_t offset,
+ enum drm_i915_gem_engine_class engine_class)
{
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;
+ static const int engine_class_to_ring[] = {
+ [I915_ENGINE_CLASS_RENDER] = AUB_TRACE_TYPE_RING_PRB0,
+ [I915_ENGINE_CLASS_VIDEO] = AUB_TRACE_TYPE_RING_PRB1,
+ [I915_ENGINE_CLASS_COPY] = AUB_TRACE_TYPE_RING_PRB2,
+ };
+ int ring = engine_class_to_ring[engine_class];
/* Make a ring buffer to execute our batchbuffer. */
memset(ringbuffer, 0, sizeof(ringbuffer));
data_out(aub, ringbuffer, ring_count * 4);
}
+static void
+aub_write_ensure_hwsp(struct aub_file *aub,
+ enum drm_i915_gem_engine_class engine_class)
+{
+ uint64_t *hwsp_addr = &aub->engine_setup[engine_class].hwsp_addr;
+
+ if (*hwsp_addr != 0)
+ return;
+
+ *hwsp_addr = alloc_ggtt_address(aub, 4096);
+ write_hwsp(aub, engine_class);
+}
+
void
-aub_write_exec(struct aub_file *aub, uint64_t batch_addr,
- uint64_t offset, int ring_flag)
+aub_write_exec(struct aub_file *aub, uint32_t ctx_id, uint64_t batch_addr,
+ uint64_t offset, enum drm_i915_gem_engine_class engine_class)
{
+ const struct engine *cs = engine_from_engine_class(engine_class);
+
if (aub_use_execlists(aub)) {
- aub_dump_execlist(aub, batch_addr, ring_flag);
+ struct aub_hw_context *hw_ctx =
+ aub_write_ensure_context(aub, ctx_id, engine_class);
+ uint64_t descriptor = get_context_descriptor(aub, cs, hw_ctx);
+ aub_write_ensure_hwsp(aub, engine_class);
+ aub_dump_ring_buffer_execlist(aub, hw_ctx, cs, batch_addr);
+ aub_dump_execlist(aub, cs, descriptor);
} else {
/* Dump ring buffer */
- aub_dump_ringbuffer(aub, batch_addr, offset, ring_flag);
+ aub_dump_ring_buffer_legacy(aub, batch_addr, offset, engine_class);
}
fflush(aub->file);
}
+
+void
+aub_write_context_execlists(struct aub_file *aub, uint64_t context_addr,
+ enum drm_i915_gem_engine_class engine_class)
+{
+ const struct engine *cs = engine_from_engine_class(engine_class);
+ uint64_t descriptor = ((uint64_t)1 << 62 | context_addr | CONTEXT_FLAGS);
+ aub_dump_execlist(aub, cs, descriptor);
+}