[Mesa-dev] [PATCH v2 06/16] intel: aubinator: handle GGTT mappings
Rafael Antognolli
rafael.antognolli at intel.com
Wed Jun 20 18:19:40 UTC 2018
On Tue, Jun 19, 2018 at 02:45:21PM +0100, Lionel Landwerlin wrote:
> We use memfd to store physical pages as they get read/written to and
> the GGTT entries translating virtual address to physical pages.
>
> Based on a commit by Scott Phillips.
>
> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
> ---
> src/intel/tools/aubinator.c | 256 ++++++++++++++++++++++++++++++++++--
> 1 file changed, 243 insertions(+), 13 deletions(-)
>
> diff --git a/src/intel/tools/aubinator.c b/src/intel/tools/aubinator.c
> index f70038376be..962546d360c 100644
> --- a/src/intel/tools/aubinator.c
> +++ b/src/intel/tools/aubinator.c
> @@ -39,12 +39,23 @@
>
> #include "util/list.h"
> #include "util/macros.h"
> +#include "util/rb_tree.h"
>
> #include "common/gen_decoder.h"
> #include "common/gen_disasm.h"
> #include "common/gen_gem.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
> @@ -73,9 +84,27 @@ struct gen_batch_decode_ctx batch_ctx;
> struct bo_map {
> struct list_head link;
> struct gen_batch_decode_bo bo;
> + bool unmap_after_use;
> +};
> +
> +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;
>
> FILE *outfile;
>
> @@ -92,11 +121,12 @@ field(uint32_t value, int start, int end)
> struct brw_instruction;
>
> static void
> -add_gtt_bo_map(struct gen_batch_decode_bo bo)
> +add_gtt_bo_map(struct gen_batch_decode_bo bo, bool unmap_after_use)
> {
> struct bo_map *m = calloc(1, sizeof(*m));
>
> m->bo = bo;
> + m->unmap_after_use = unmap_after_use;
> list_add(&m->link, &maps);
> }
>
> @@ -104,21 +134,208 @@ static void
> clear_bo_maps(void)
> {
> list_for_each_entry_safe(struct bo_map, i, &maps, link) {
> + if (i->unmap_after_use)
> + munmap((void *)i->bo.map, i->bo.size);
> list_del(&i->link);
> free(i);
> }
> }
>
> +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, const void *_data, uint32_t _size)
> +{
> + uint64_t virt_addr = (address / sizeof(uint64_t)) << 12;
> + const uint64_t *data = _data;
> + size_t size = _size / sizeof(*data);
> + for (const 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, const 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 = (const uint8_t *)data + size_this_page;
> + }
> +}
> +
> +static void
> +handle_ggtt_write(uint64_t address, const 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);
I don't really get this function :(
Isn't "address" a ggtt mapped virtual address? And if so, shouldn't we
pass entry->phys_addr instead of page, which is based on the virtual
one?
> + data = (const uint8_t *)data + size_this_page;
> + }
> +}
> +
> 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)
> {
> + struct gen_batch_decode_bo bo = {0};
> +
> 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 };
> -}
> + address &= ~0xfff;
> +
> + 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);
> + }
>
> + add_gtt_bo_map(bo, true);
> +
> + return bo;
> +}
> #define GEN_ENGINE_RENDER 1
> #define GEN_ENGINE_BLITTER 2
>
> @@ -144,7 +361,7 @@ handle_trace_block(uint32_t *p)
> switch (operation) {
> case AUB_TRACE_OP_DATA_WRITE:
> if (address_space == AUB_TRACE_MEMTYPE_GTT)
> - add_gtt_bo_map(bo);
> + add_gtt_bo_map(bo, false);
> break;
> case AUB_TRACE_OP_COMMAND_WRITE:
> switch (type) {
> @@ -185,7 +402,7 @@ 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, get_ggtt_batch_bo, NULL, NULL);
> batch_ctx.max_vbo_decoded_lines = max_vbo_lines;
>
> char *color = GREEN_HEADER, *reset_color = NORMAL;
> @@ -317,7 +534,7 @@ handle_memtrace_reg_write(uint32_t *p)
>
> const uint32_t pphwsp_size = 4096;
> uint32_t pphwsp_addr = context_descriptor & 0xfffff000;
> - struct gen_batch_decode_bo pphwsp_bo = get_gen_batch_bo(NULL, pphwsp_addr);
> + 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);
> @@ -326,10 +543,11 @@ handle_memtrace_reg_write(uint32_t *p)
> uint32_t ring_buffer_tail = context[7];
> uint32_t ring_buffer_start = context[9];
>
> - struct gen_batch_decode_bo ring_bo = get_gen_batch_bo(NULL,
> - ring_buffer_start);
> + 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);
> +
> (void)engine; /* TODO */
> gen_print_batch(&batch_ctx, commands, ring_buffer_tail - ring_buffer_head,
> 0);
> @@ -350,10 +568,20 @@ handle_memtrace_mem_write(uint32_t *p)
> };
> uint32_t address_space = p[3] >> 28;
>
> - if (address_space != 1)
> - return;
> -
> - add_gtt_bo_map(bo);
> + switch (address_space) {
> + case 0: /* GGTT */
> + handle_ggtt_write(bo.addr, bo.map, bo.size);
> + break;
> + case 1: /* Local */
> + add_gtt_bo_map(bo, false);
> + break;
> + case 2: /* Physical */
> + handle_physical_write(bo.addr, bo.map, bo.size);
> + break;
> + case 4: /* GGTT Entry */
> + handle_ggtt_entry_write(bo.addr, bo.map, bo.size);
> + break;
> + }
> }
>
> struct aub_file {
> @@ -617,6 +845,8 @@ int main(int argc, char *argv[])
> if (isatty(1) && pager)
> setup_pager();
>
> + mem_fd = memfd_create("phys memory", 0);
> +
> list_inithead(&maps);
>
> file = aub_file_open(input_file);
> --
> 2.17.1
>
More information about the mesa-dev
mailing list