radv: a support for a secure compile fork at device creation
authorTimothy Arceri <tarceri@itsqueeze.com>
Wed, 31 Jul 2019 04:03:53 +0000 (14:03 +1000)
committerTimothy Arceri <tarceri@itsqueeze.com>
Sat, 26 Oct 2019 02:04:12 +0000 (13:04 +1100)
This added support for the fork, the installation of the seccomp
filter, and the main loop for the actual compilation to be called
from i.e. run_secure_compile_device().

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/vulkan/radv_device.c

index eeac9d3adca3e21d4306b6f9a29c818c7a148762..10211c8f9bcae95eb663c3326be940efa4bbd05b 100644 (file)
  * IN THE SOFTWARE.
  */
 
+#include <errno.h>
+#include <fcntl.h>
+#include <linux/audit.h>
+#include <linux/bpf.h>
+#include <linux/filter.h>
+#include <linux/seccomp.h>
+#include <linux/unistd.h>
 #include <stdbool.h>
+#include <stddef.h>
+#include <stdio.h>
 #include <string.h>
+#include <sys/prctl.h>
+#include <sys/wait.h>
 #include <unistd.h>
 #include <fcntl.h>
 #include <llvm/Config/llvm-config.h>
+
 #include "radv_debug.h"
 #include "radv_private.h"
 #include "radv_shader.h"
@@ -1920,6 +1932,270 @@ radv_get_int_debug_option(const char *name, int default_value)
        return result;
 }
 
+static int install_seccomp_filter() {
+
+       struct sock_filter filter[] = {
+               /* Check arch is 64bit x86 */
+               BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, arch))),
+               BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, AUDIT_ARCH_X86_64, 0, 10),
+
+               /* Allow system exit calls for the forked process */
+               BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+               BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_exit_group, 9, 0),
+
+               /* Allow system read calls */
+               BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+               BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_read, 7, 0),
+
+               /* Allow system write calls */
+               BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+               BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_write, 5, 0),
+
+               /* Allow system brk calls (we need this for malloc) */
+               BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+               BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_brk, 3, 0),
+
+               /* Futex is required for mutex locks */
+               BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+               BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_futex, 1, 0),
+
+               /* Return error if we hit a system call not on the whitelist */
+               BPF_STMT(BPF_RET + BPF_K, SECCOMP_RET_ERRNO | (EPERM & SECCOMP_RET_DATA)),
+
+               /* Allow whitelisted system calls */
+               BPF_STMT(BPF_RET + BPF_K, SECCOMP_RET_ALLOW),
+       };
+
+       struct sock_fprog prog = {
+               .len = (unsigned short)(sizeof(filter) / sizeof(filter[0])),
+               .filter = filter,
+       };
+
+       if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0))
+               return -1;
+
+       if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &prog))
+               return -1;
+
+       return 0;
+}
+
+static void run_secure_compile_device(struct radv_device *device, unsigned process,
+                                     int *fd_secure_input, int *fd_secure_output)
+{
+       enum radv_secure_compile_type sc_type;
+       if (install_seccomp_filter() == -1) {
+               sc_type = RADV_SC_TYPE_INIT_FAILURE;
+       } else {
+               sc_type = RADV_SC_TYPE_INIT_SUCCESS;
+               device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input[0];
+               device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output[1];
+       }
+
+       write(fd_secure_output[1], &sc_type, sizeof(sc_type));
+
+       if (sc_type == RADV_SC_TYPE_INIT_FAILURE)
+               goto secure_compile_exit;
+
+       while (true) {
+               read(fd_secure_input[0], &sc_type, sizeof(sc_type));
+
+               if (sc_type == RADV_SC_TYPE_COMPILE_PIPELINE) {
+                       struct radv_pipeline *pipeline;
+
+                       pipeline = vk_zalloc2(&device->alloc, NULL, sizeof(*pipeline), 8,
+                                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+                       pipeline->device = device;
+
+                       /* Read pipeline layout */
+                       struct radv_pipeline_layout layout;
+                       read(fd_secure_input[0], &layout, sizeof(struct radv_pipeline_layout));
+                       read(fd_secure_input[0], &layout.num_sets, sizeof(uint32_t));
+                       for (uint32_t set = 0; set < layout.num_sets; set++) {
+                               uint32_t layout_size;
+                               read(fd_secure_input[0], &layout_size, sizeof(uint32_t));
+                               layout.set[set].layout = malloc(layout_size);
+                               layout.set[set].layout->layout_size = layout_size;
+                               read(fd_secure_input[0], layout.set[set].layout, layout.set[set].layout->layout_size);
+                       }
+
+                       pipeline->layout = &layout;
+
+                       /* Read pipeline key */
+                       struct radv_pipeline_key key;
+                       read(fd_secure_input[0], &key, sizeof(struct radv_pipeline_key));
+
+                       /* Read pipeline create flags */
+                       VkPipelineCreateFlags flags;
+                       read(fd_secure_input[0], &flags, sizeof(VkPipelineCreateFlags));
+
+                       /* Read stage and shader information */
+                       uint32_t num_stages;
+                       const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
+                       read(fd_secure_input[0], &num_stages, sizeof(uint32_t));
+                       for (uint32_t i = 0; i < num_stages; i++) {
+
+                               /* Read stage */
+                               gl_shader_stage stage;
+                               read(fd_secure_input[0], &stage, sizeof(gl_shader_stage));
+
+                               VkPipelineShaderStageCreateInfo *pStage = calloc(1, sizeof(VkPipelineShaderStageCreateInfo));
+
+                               /* Read entry point name */
+                               size_t name_size;
+                               read(fd_secure_input[0], &name_size, sizeof(size_t));
+                               char *ep_name = malloc(name_size);
+                               read(fd_secure_input[0], ep_name, name_size);
+                               pStage->pName = ep_name;
+
+                               /* Read shader module */
+                               size_t module_size;
+                               read(fd_secure_input[0], &module_size, sizeof(size_t));
+                               struct radv_shader_module *module = malloc(module_size);
+                               read(fd_secure_input[0], module, module_size);
+                               pStage->module = radv_shader_module_to_handle(module);
+
+                               /* Read specialization info */
+                               bool has_spec_info;
+                               read(fd_secure_input[0], &has_spec_info, sizeof(bool));
+                               if (has_spec_info) {
+                                       VkSpecializationInfo *specInfo = malloc(sizeof(VkSpecializationInfo));
+                                       pStage->pSpecializationInfo = specInfo;
+
+                                       read(fd_secure_input[0], &specInfo->dataSize, sizeof(size_t));
+
+                                       void *si_data = malloc(specInfo->dataSize);
+                                       read(fd_secure_input[0], si_data, specInfo->dataSize);
+                                       specInfo->pData = si_data;
+
+                                       read(fd_secure_input[0], &specInfo->mapEntryCount, sizeof(uint32_t));
+                                       VkSpecializationMapEntry *mapEntries = malloc(sizeof(VkSpecializationMapEntry) * specInfo->mapEntryCount);
+                                       for (uint32_t j = 0; j < specInfo->mapEntryCount; j++)
+                                               read(fd_secure_input[0], &mapEntries[j], sizeof(VkSpecializationMapEntry));
+
+                                       specInfo->pMapEntries = mapEntries;
+                               }
+
+                               pStages[stage] = pStage;
+                       }
+
+                       /* Compile the shaders */
+                       VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
+                       radv_create_shaders(pipeline, device, NULL, &key, pStages, flags, NULL, stage_feedbacks);
+
+                       /* free memory allocated above */
+                       for (uint32_t set = 0; set < layout.num_sets; set++)
+                               free(layout.set[set].layout);
+
+                       for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
+                               if (!pStages[i])
+                                       continue;
+
+                               free((void *) pStages[i]->pName);
+                               free(radv_shader_module_from_handle(pStages[i]->module));
+                               if (pStages[i]->pSpecializationInfo) {
+                                       free((void *) pStages[i]->pSpecializationInfo->pData);
+                                       free((void *) pStages[i]->pSpecializationInfo->pMapEntries);
+                                       free((void *) pStages[i]->pSpecializationInfo);
+                               }
+                               free((void *) pStages[i]);
+                       }
+
+                       vk_free(&device->alloc, pipeline);
+
+                       sc_type = RADV_SC_TYPE_COMPILE_PIPELINE_FINISHED;
+                       write(fd_secure_output[1], &sc_type, sizeof(sc_type));
+
+               } else if (sc_type == RADV_SC_TYPE_DESTROY_DEVICE) {
+                       goto secure_compile_exit;
+               }
+       }
+
+secure_compile_exit:
+       close(fd_secure_input[1]);
+       close(fd_secure_input[0]);
+       close(fd_secure_output[1]);
+       close(fd_secure_output[0]);
+       _exit(0);
+}
+
+static void destroy_secure_compile_device(struct radv_device *device, unsigned process)
+{
+       int fd_secure_input = device->sc_state->secure_compile_processes[process].fd_secure_input;
+
+       enum radv_secure_compile_type sc_type = RADV_SC_TYPE_DESTROY_DEVICE;
+       write(fd_secure_input, &sc_type, sizeof(sc_type));
+
+       close(device->sc_state->secure_compile_processes[process].fd_secure_input);
+       close(device->sc_state->secure_compile_processes[process].fd_secure_output);
+
+       int status;
+       waitpid(device->sc_state->secure_compile_processes[process].sc_pid, &status, 0);
+}
+
+static VkResult fork_secure_compile_device(struct radv_device *device)
+{
+       device->sc_state = vk_zalloc(&device->alloc,
+                                    sizeof(struct radv_secure_compile_state),
+                                    8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+
+       mtx_init(&device->sc_state->secure_compile_mutex, mtx_plain);
+
+#define MAX_SC_PROCS 32
+       uint8_t sc_threads = device->instance->num_sc_threads;
+       int fd_secure_input[MAX_SC_PROCS][2];
+       int fd_secure_output[MAX_SC_PROCS][2];
+
+       /* create pipe descriptors (used to communicate between processes) */
+       for (unsigned i = 0; i < sc_threads; i++) {
+               if (pipe(fd_secure_input[i]) == -1 ||
+                   pipe(fd_secure_output[i]) == -1) {
+                       return VK_ERROR_INITIALIZATION_FAILED;
+               }
+       }
+
+       device->sc_state->secure_compile_processes = vk_zalloc(&device->alloc,
+                                                               sizeof(struct radv_secure_compile_process) * sc_threads, 8,
+                                                               VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+
+       for (unsigned process = 0; process < sc_threads; process++) {
+               if ((device->sc_state->secure_compile_processes[process].sc_pid = fork()) == 0) {
+                       device->sc_state->secure_compile_thread_counter = process;
+                       run_secure_compile_device(device, process, fd_secure_input[process], fd_secure_output[process]);
+               } else {
+                       if (device->sc_state->secure_compile_processes[process].sc_pid == -1)
+                               return VK_ERROR_INITIALIZATION_FAILED;
+
+                       /* Read the init result returned from the secure process */
+                       enum radv_secure_compile_type sc_type;
+                       read(fd_secure_output[process][0], &sc_type, sizeof(sc_type));
+
+                       if (sc_type == RADV_SC_TYPE_INIT_FAILURE) {
+                               close(fd_secure_input[process][0]);
+                               close(fd_secure_input[process][1]);
+                               close(fd_secure_output[process][1]);
+                               close(fd_secure_output[process][0]);
+                               int status;
+                               waitpid(device->sc_state->secure_compile_processes[process].sc_pid, &status, 0);
+
+                               /* Destroy any forks that were created sucessfully */
+                               for (unsigned i = 0; i < process; i++) {
+                                       destroy_secure_compile_device(device, i);
+                               }
+
+                               return VK_ERROR_INITIALIZATION_FAILED;
+                       } else {
+                               assert(sc_type == RADV_SC_TYPE_INIT_SUCCESS);
+                               device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input[process][1];
+                               device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output[process][0];
+                       }
+               }
+       }
+
+       return VK_SUCCESS;
+}
+
 VkResult radv_CreateDevice(
        VkPhysicalDevice                            physicalDevice,
        const VkDeviceCreateInfo*                   pCreateInfo,
@@ -2076,8 +2352,12 @@ VkResult radv_CreateDevice(
                radv_dump_enabled_options(device, stderr);
        }
 
-       device->keep_shader_info = keep_shader_info;
+       /* Temporarily disable secure compile while we create meta shaders, etc */
+       uint8_t sc_threads = device->instance->num_sc_threads;
+       if (sc_threads)
+               device->instance->num_sc_threads = 0;
 
+       device->keep_shader_info = keep_shader_info;
        result = radv_device_init_meta(device);
        if (result != VK_SUCCESS)
                goto fail;
@@ -2124,6 +2404,14 @@ VkResult radv_CreateDevice(
                        1 << util_logbase2(device->force_aniso));
        }
 
+       /* Fork device for secure compile as required */
+       device->instance->num_sc_threads = sc_threads;
+       if (radv_device_use_secure_compile(device->instance)) {
+               result = fork_secure_compile_device(device);
+               if (result != VK_SUCCESS)
+                       goto fail_meta;
+       }
+
        *pDevice = radv_device_to_handle(device);
        return VK_SUCCESS;
 
@@ -2180,6 +2468,16 @@ void radv_DestroyDevice(
        radv_destroy_shader_slabs(device);
 
        radv_bo_list_finish(&device->bo_list);
+
+       if (radv_device_use_secure_compile(device->instance)) {
+               for (unsigned i = 0; i < device->instance->num_sc_threads; i++ ) {
+                       destroy_secure_compile_device(device, i);
+               }
+       }
+
+       if (device->sc_state)
+               vk_free(&device->alloc, device->sc_state->secure_compile_processes);
+       vk_free(&device->alloc, device->sc_state);
        vk_free(&device->alloc, device);
 }