[Mesa-dev] [PATCH 2/9] intel/tools/aubinator: aubinate ppgtt aubs
Lionel Landwerlin
lionel.g.landwerlin at intel.com
Tue Jun 19 13:40:51 UTC 2018
Hey Rafael,
Thanks for taking the time to look into that.
I kept the commit as Scott wrote it to preserve ownership.
I've just broke this commit into 3 parts (register writes changes, ggtt
& ppgtt).
Sending a new series.
Cheers,
-
Lionel
On 18/06/18 23:25, Rafael Antognolli wrote:
> Hi Lionel,
>
> I've been going through the patch and I think the content so far is
> mostly fine. However, it has a lot of things going on for a single patch
> IMHO. I see changes for the execution list submission port, ppgtt, using
> the RB tree for ggtt too, etc...
>
> It definitely make it more painful to read (at least for me). Are these
> changes so related that they couldn't be split up?
>
> Rafael
>
> On Thu, Jun 14, 2018 at 06:11:38PM +0100, Lionel Landwerlin wrote:
>> From: Scott D Phillips <scott.d.phillips at intel.com>
>>
>> v2: by Lionel
>> Fix memfd_create compilation issue
>> Fix pml4 address stored on 32 instead of 64bits
>> Return no buffer if first ppgtt page is not mapped
>>
>> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
>> ---
>> src/intel/tools/aubinator.c | 460 ++++++++++++++++++++++++++++++++----
>> 1 file changed, 410 insertions(+), 50 deletions(-)
>>
>> diff --git a/src/intel/tools/aubinator.c b/src/intel/tools/aubinator.c
>> index 3120e82b22e..99cd010dd9d 100644
>> --- a/src/intel/tools/aubinator.c
>> +++ b/src/intel/tools/aubinator.c
>> @@ -37,12 +37,24 @@
>> #include <sys/wait.h>
>> #include <sys/mman.h>
>>
>> +#include "util/list.h"
>> #include "util/macros.h"
>> +#include "util/rb_tree.h"
>>
>> #include "common/gen_decoder.h"
>> #include "common/gen_disasm.h"
>> #include "intel_aub.h"
>>
>> +#ifndef HAVE_MEMFD_CREATE
>> +#include <sys/syscall.h>
>> +
>> +static inline int
>> +memfd_create(const char *name, unsigned int flags)
>> +{
>> + return syscall(SYS_memfd_create, name, flags);
>> +}
>> +#endif
>> +
>> /* Below is the only command missing from intel_aub.h in libdrm
>> * So, reuse intel_aub.h from libdrm and #define the
>> * AUB_MI_BATCH_BUFFER_END as below
>> @@ -70,6 +82,31 @@ 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;
>> +};
>> +
>> +struct ggtt_entry {
>> + struct rb_node node;
>> + uint64_t virt_addr;
>> + uint64_t phys_addr;
>> +};
>> +
>> +struct phys_mem {
>> + struct rb_node node;
>> + uint64_t fd_offset;
>> + uint64_t phys_addr;
>> + uint8_t *data;
>> +};
>> +
>> +static struct list_head maps;
>> +static struct rb_tree ggtt = {NULL};
>> +static struct rb_tree mem = {NULL};
>> +int mem_fd = -1;
>> +off_t mem_fd_len = 0;
>> +
>> uint64_t general_state_base;
>> uint64_t surface_state_base;
>> uint64_t dynamic_state_base;
>> @@ -99,6 +136,191 @@ valid_offset(uint32_t offset)
>> #define GEN_ENGINE_RENDER 1
>> #define GEN_ENGINE_BLITTER 2
>>
>> +static inline struct ggtt_entry *
>> +ggtt_entry_next(struct ggtt_entry *entry)
>> +{
>> + if (!entry)
>> + return NULL;
>> + struct rb_node *node = rb_node_next(&entry->node);
>> + if (!node)
>> + return NULL;
>> + return rb_node_data(struct ggtt_entry, node, node);
>> +}
>> +
>> +static inline int
>> +cmp_uint64(uint64_t a, uint64_t b)
>> +{
>> + if (a < b)
>> + return -1;
>> + if (a > b)
>> + return 1;
>> + return 0;
>> +}
>> +
>> +static inline int
>> +cmp_ggtt_entry(const struct rb_node *node, const void *addr)
>> +{
>> + struct ggtt_entry *entry = rb_node_data(struct ggtt_entry, node, node);
>> + return cmp_uint64(entry->virt_addr, *(uint64_t *)addr);
>> +}
>> +
>> +static struct ggtt_entry *
>> +ensure_ggtt_entry(struct rb_tree *tree, uint64_t virt_addr)
>> +{
>> + struct rb_node *node = rb_tree_search_sloppy(&ggtt, &virt_addr,
>> + cmp_ggtt_entry);
>> + int cmp = 0;
>> + if (!node || (cmp = cmp_ggtt_entry(node, &virt_addr))) {
>> + struct ggtt_entry *new_entry = calloc(1, sizeof(*new_entry));
>> + new_entry->virt_addr = virt_addr;
>> + rb_tree_insert_at(&ggtt, node, &new_entry->node, cmp > 0);
>> + node = &new_entry->node;
>> + }
>> +
>> + return rb_node_data(struct ggtt_entry, node, node);
>> +}
>> +
>> +static struct ggtt_entry *
>> +search_ggtt_entry(uint64_t virt_addr)
>> +{
>> + virt_addr &= ~0xfff;
>> +
>> + struct rb_node *node = rb_tree_search(&ggtt, &virt_addr, cmp_ggtt_entry);
>> +
>> + if (!node)
>> + return NULL;
>> +
>> + return rb_node_data(struct ggtt_entry, node, node);
>> +}
>> +
>> +static inline int
>> +cmp_phys_mem(const struct rb_node *node, const void *addr)
>> +{
>> + struct phys_mem *mem = rb_node_data(struct phys_mem, node, node);
>> + return cmp_uint64(mem->phys_addr, *(uint64_t *)addr);
>> +}
>> +
>> +static struct phys_mem *
>> +ensure_phys_mem(uint64_t phys_addr)
>> +{
>> + struct rb_node *node = rb_tree_search_sloppy(&mem, &phys_addr, cmp_phys_mem);
>> + int cmp = 0;
>> + if (!node || (cmp = cmp_phys_mem(node, &phys_addr))) {
>> + struct phys_mem *new_mem = calloc(1, sizeof(*new_mem));
>> + new_mem->phys_addr = phys_addr;
>> + new_mem->fd_offset = mem_fd_len;
>> +
>> + int ftruncate_res = ftruncate(mem_fd, mem_fd_len += 4096);
>> + assert(ftruncate_res == 0);
>> +
>> + new_mem->data = mmap(NULL, 4096, PROT_READ | PROT_WRITE, MAP_SHARED,
>> + mem_fd, new_mem->fd_offset);
>> + assert(new_mem->data != MAP_FAILED);
>> +
>> + rb_tree_insert_at(&mem, node, &new_mem->node, cmp > 0);
>> + node = &new_mem->node;
>> + }
>> +
>> + return rb_node_data(struct phys_mem, node, node);
>> +}
>> +
>> +static struct phys_mem *
>> +search_phys_mem(uint64_t phys_addr)
>> +{
>> + phys_addr &= ~0xfff;
>> +
>> + struct rb_node *node = rb_tree_search(&mem, &phys_addr, cmp_phys_mem);
>> +
>> + if (!node)
>> + return NULL;
>> +
>> + return rb_node_data(struct phys_mem, node, node);
>> +}
>> +
>> +static void
>> +handle_ggtt_entry_write(uint64_t address, void *_data, uint32_t _size)
>> +{
>> + uint64_t virt_addr = (address / sizeof(uint64_t)) << 12;
>> + uint64_t *data = _data;
>> + size_t size = _size / sizeof(*data);
>> + for (uint64_t *entry = data;
>> + entry < data + size;
>> + entry++, virt_addr += 4096) {
>> + struct ggtt_entry *pt = ensure_ggtt_entry(&ggtt, virt_addr);
>> + pt->phys_addr = *entry;
>> + }
>> +}
>> +
>> +static void
>> +handle_physical_write(uint64_t address, void *data, uint32_t size)
>> +{
>> + uint32_t to_write = size;
>> + for (uint64_t page = address & ~0xfff; page < address + size; page += 4096) {
>> + struct phys_mem *mem = ensure_phys_mem(page);
>> + uint64_t offset = MAX2(page, address) - page;
>> + uint32_t size_this_page = MIN2(to_write, 4096 - offset);
>> + to_write -= size_this_page;
>> + memcpy(mem->data + offset, data, size_this_page);
>> + data = (uint8_t *)data + size_this_page;
>> + }
>> +}
>> +
>> +static void
>> +handle_ggtt_write(uint64_t address, void *data, uint32_t size)
>> +{
>> + uint32_t to_write = size;
>> + for (uint64_t page = address & ~0xfff; page < address + size; page += 4096) {
>> + struct ggtt_entry *entry = search_ggtt_entry(page);
>> + assert(entry && entry->phys_addr & 0x1);
>> +
>> + uint64_t offset = MAX2(page, address) - page;
>> + uint32_t size_this_page = MIN2(to_write, 4096 - offset);
>> + to_write -= size_this_page;
>> +
>> + handle_physical_write(page + offset, data, size_this_page);
>> + data = (uint8_t *)data + size_this_page;
>> + }
>> +}
>> +
>> +static struct phys_mem *
>> +ppgtt_walk(uint64_t pml4, uint64_t address)
>> +{
>> + uint64_t shift = 39;
>> + uint64_t addr = pml4;
>> + for (int level = 4; level > 0; level--) {
>> + struct phys_mem *table = search_phys_mem(addr);
>> + if (!table)
>> + return NULL;
>> + int index = (address >> shift) & 0x1ff;
>> + uint64_t entry = ((uint64_t *)table->data)[index];
>> + if (!(entry & 1))
>> + return NULL;
>> + addr = entry & ~0xfff;
>> + shift -= 9;
>> + }
>> + return search_phys_mem(addr);
>> +}
>> +
>> +static bool
>> +ppgtt_mapped(uint64_t pml4, uint64_t address)
>> +{
>> + return ppgtt_walk(pml4, address) != NULL;
>> +}
>> +
>> +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
>> handle_trace_block(uint32_t *p)
>> {
>> @@ -140,6 +362,7 @@ handle_trace_block(uint32_t *p)
>> }
>>
>> (void)engine; /* TODO */
>> + batch_ctx.get_bo = get_gen_batch_bo;
>> gen_print_batch(&batch_ctx, data, size, 0);
>>
>> gtt_end = 0;
>> @@ -148,17 +371,103 @@ handle_trace_block(uint32_t *p)
>> }
>>
>> static struct gen_batch_decode_bo
>> -get_gen_batch_bo(void *user_data, uint64_t address)
>> +get_ggtt_batch_bo(void *user_data, uint64_t address)
>> {
>> - if (address > gtt_end)
>> - return (struct gen_batch_decode_bo) { .map = NULL };
>> + struct gen_batch_decode_bo bo = {0};
>> +
>> + address &= ~0xfff;
>> +
>> + 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;
>> +
>> + struct ggtt_entry *start =
>> + (struct ggtt_entry *)rb_tree_search_sloppy(&ggtt, &address,
>> + cmp_ggtt_entry);
>> + if (start && start->virt_addr < address)
>> + start = ggtt_entry_next(start);
>> + if (!start)
>> + return bo;
>> +
>> + struct ggtt_entry *last = start;
>> + for (struct ggtt_entry *i = ggtt_entry_next(last);
>> + i && last->virt_addr + 4096 == i->virt_addr;
>> + last = i, i = ggtt_entry_next(last))
>> + ;
>> +
>> + bo.addr = MIN2(address, start->virt_addr);
>> + bo.size = last->virt_addr - bo.addr + 4096;
>> + bo.map = mmap(NULL, bo.size, PROT_READ, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>> + assert(bo.map != MAP_FAILED);
>> +
>> + for (struct ggtt_entry *i = start;
>> + i;
>> + i = i == last ? NULL : ggtt_entry_next(i)) {
>> + uint64_t phys_addr = i->phys_addr & ~0xfff;
>> + struct phys_mem *phys_mem = search_phys_mem(phys_addr);
>> +
>> + if (!phys_mem)
>> + continue;
>> +
>> + uint32_t map_offset = i->virt_addr - address;
>> + void *res = mmap((uint8_t *)bo.map + map_offset, 4096, PROT_READ,
>> + MAP_SHARED | MAP_FIXED, mem_fd, phys_mem->fd_offset);
>> + assert(res != MAP_FAILED);
>> + }
>>
>> - /* We really only have one giant address range */
>> - return (struct gen_batch_decode_bo) {
>> - .addr = 0,
>> - .map = gtt,
>> - .size = gtt_size
>> - };
>> + struct bo_map *m = calloc(1, sizeof(*m));
>> + m->bo = bo;
>> + list_add(&m->link, &maps);
>> +
>> + return bo;
>> +}
>> +
>> +static struct gen_batch_decode_bo
>> +get_ppgtt_batch_bo(void *user_data, uint64_t address)
>> +{
>> + struct gen_batch_decode_bo bo = {0};
>> + uint64_t pml4 = *(uint64_t *)user_data;
>> +
>> + address &= ~0xfff;
>> +
>> + if (!ppgtt_mapped(pml4, address))
>> + return bo;
>> +
>> + /* Map everything until the first gap since we don't know how much the
>> + * decoder actually needs.
>> + */
>> + uint64_t end = address;
>> + while (ppgtt_mapped(pml4, end))
>> + end += 4096;
>> +
>> + bo.addr = address;
>> + bo.size = end - address;
>> + bo.map = mmap(NULL, bo.size, PROT_READ, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>> + assert(bo.map != MAP_FAILED);
>> +
>> + for (uint64_t page = address; page < end; page += 4096) {
>> + struct phys_mem *phys_mem = ppgtt_walk(pml4, page);
>> +
>> + void *res = mmap((uint8_t *)bo.map + (page - bo.addr), 4096, PROT_READ,
>> + MAP_SHARED | MAP_FIXED, mem_fd, phys_mem->fd_offset);
>> + assert(res != MAP_FAILED);
>> + }
>> +
>> + struct bo_map *m = calloc(1, sizeof(*m));
>> + m->bo = bo;
>> + list_add(&m->link, &maps);
>> +
>> + return bo;
>> +}
>> +
>> +static void
>> +clear_bo_maps(void)
>> +{
>> + list_for_each_entry_safe(struct bo_map, i, &maps, link) {
>> + munmap((void *)i->bo.map, i->bo.size);
>> + list_del(&i->link);
>> + free(i);
>> + }
>> }
>>
>> static void
>> @@ -179,7 +488,7 @@ aubinator_init(uint16_t aub_pci_id, const char *app_name)
>> batch_flags |= GEN_BATCH_DECODE_FLOATS;
>>
>> gen_batch_decode_ctx_init(&batch_ctx, &devinfo, outfile, batch_flags,
>> - xml_path, get_gen_batch_bo, NULL, NULL);
>> + xml_path, NULL, NULL, NULL);
>> batch_ctx.max_vbo_decoded_lines = max_vbo_lines;
>>
>> char *color = GREEN_HEADER, *reset_color = NORMAL;
>> @@ -245,52 +554,98 @@ handle_memtrace_version(uint32_t *p)
>> static void
>> handle_memtrace_reg_write(uint32_t *p)
>> {
>> + static struct execlist_regs {
>> + uint32_t render_elsp[4];
>> + int render_elsp_index;
>> + uint32_t blitter_elsp[4];
>> + int blitter_elsp_index;
>> + } state = {};
>> +
>> uint32_t offset = p[1];
>> uint32_t value = p[5];
>> +
>> int engine;
>> - static int render_elsp_writes = 0;
>> - static int blitter_elsp_writes = 0;
>> - static int render_elsq0 = 0;
>> - static int blitter_elsq0 = 0;
>> - uint8_t *pphwsp;
>> -
>> - if (offset == 0x2230) {
>> - render_elsp_writes++;
>> + uint64_t context_descriptor;
>> +
>> + switch (offset) {
>> + case 0x2230: /* render elsp */
>> + state.render_elsp[state.render_elsp_index++] = value;
>> + if (state.render_elsp_index < 4)
>> + return;
>> +
>> + state.render_elsp_index = 0;
>> engine = GEN_ENGINE_RENDER;
>> - } else if (offset == 0x22230) {
>> - blitter_elsp_writes++;
>> + context_descriptor = (uint64_t)state.render_elsp[2] << 32 |
>> + state.render_elsp[3];
>> + break;
>> + case 0x22230: /* blitter elsp */
>> + state.blitter_elsp[state.blitter_elsp_index++] = value;
>> + if (state.blitter_elsp_index < 4)
>> + return;
>> +
>> + state.blitter_elsp_index = 0;
>> engine = GEN_ENGINE_BLITTER;
>> - } else if (offset == 0x2510) {
>> - render_elsq0 = value;
>> - } else if (offset == 0x22510) {
>> - blitter_elsq0 = value;
>> - } else if (offset == 0x2550 || offset == 0x22550) {
>> - /* nothing */;
>> - } else {
>> + context_descriptor = (uint64_t)state.blitter_elsp[2] << 32 |
>> + state.blitter_elsp[3];
>> + break;
>> + case 0x2510: /* render elsq0 lo */
>> + state.render_elsp[3] = value;
>> return;
>> - }
>> -
>> - if (render_elsp_writes > 3 || blitter_elsp_writes > 3) {
>> - render_elsp_writes = blitter_elsp_writes = 0;
>> - pphwsp = (uint8_t*)gtt + (value & 0xfffff000);
>> - } else if (offset == 0x2550) {
>> + break;
>> + case 0x2514: /* render elsq0 hi */
>> + state.render_elsp[2] = value;
>> + return;
>> + break;
>> + case 0x22510: /* blitter elsq0 lo */
>> + state.blitter_elsp[3] = value;
>> + return;
>> + break;
>> + case 0x22514: /* blitter elsq0 hi */
>> + state.blitter_elsp[2] = value;
>> + return;
>> + break;
>> + case 0x2550: /* render elsc */
>> engine = GEN_ENGINE_RENDER;
>> - pphwsp = (uint8_t*)gtt + (render_elsq0 & 0xfffff000);
>> - } else if (offset == 0x22550) {
>> + context_descriptor = (uint64_t)state.render_elsp[2] << 32 |
>> + state.render_elsp[3];
>> + break;
>> + case 0x22550: /* blitter elsc */
>> engine = GEN_ENGINE_BLITTER;
>> - pphwsp = (uint8_t*)gtt + (blitter_elsq0 & 0xfffff000);
>> - } else {
>> + context_descriptor = (uint64_t)state.blitter_elsp[2] << 32 |
>> + state.blitter_elsp[3];
>> + break;
>> + default:
>> return;
>> }
>>
>> const uint32_t pphwsp_size = 4096;
>> - uint32_t *context = (uint32_t*)(pphwsp + pphwsp_size);
>> + uint32_t pphwsp_addr = context_descriptor & 0xfffff000;
>> + struct gen_batch_decode_bo pphwsp_bo = get_ggtt_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 = context[9];
>> - uint32_t *commands = (uint32_t*)((uint8_t*)gtt + ring_buffer_start + ring_buffer_head);
>> + uint64_t pml4 = (uint64_t)context[49] << 32 | context[51];
>> +
>> + struct gen_batch_decode_bo ring_bo = get_ggtt_batch_bo(NULL,
>> + ring_buffer_start);
>> + assert(ring_bo.size > 0);
>> + void *commands = (uint8_t *)ring_bo.map + (ring_bo.addr - ring_buffer_start);
>> +
>> + if (context_descriptor & 0x100 /* ppgtt */) {
>> + batch_ctx.get_bo = get_ppgtt_batch_bo;
>> + batch_ctx.user_data = &pml4;
>> + } else {
>> + batch_ctx.get_bo = get_ggtt_batch_bo;
>> + }
>> +
>> (void)engine; /* TODO */
>> - gen_print_batch(&batch_ctx, commands, ring_buffer_tail - ring_buffer_head, 0);
>> + gen_print_batch(&batch_ctx, commands, ring_buffer_tail - ring_buffer_head,
>> + 0);
>> + clear_bo_maps();
>> }
>>
>> static void
>> @@ -301,17 +656,18 @@ handle_memtrace_mem_write(uint32_t *p)
>> uint32_t size = p[4];
>> uint32_t *data = p + 5;
>>
>> - if (address_space != 1)
>> - return;
>> -
>> - if (gtt_size < address + size) {
>> - fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
>> - exit(EXIT_FAILURE);
>> + switch (address_space) {
>> + case 0: /* GGTT */
>> + handle_ggtt_write(address, data, size);
>> + break;
>> + case 1:
>> + case 2: /* Physical */
>> + handle_physical_write(address, data, size);
>> + break;
>> + case 4: /* GGTT Entry */
>> + handle_ggtt_entry_write(address, data, size);
>> + break;
>> }
>> -
>> - memcpy((char *) gtt + address, data, size);
>> - if (gtt_end < address + size)
>> - gtt_end = address + size;
>> }
>>
>> struct aub_file {
>> @@ -646,6 +1002,10 @@ int main(int argc, char *argv[])
>> exit(EXIT_FAILURE);
>> }
>>
>> + mem_fd = memfd_create("phys memory", 0);
>> +
>> + list_inithead(&maps);
>> +
>> while (aub_file_more_stuff(file)) {
>> switch (aub_file_decode_batch(file)) {
>> case AUB_ITEM_DECODE_OK:
>> --
>> 2.17.1
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list