Fix driver sorting given driver failures and instrumentation (#288)

* Fix driver sorting given driver failures and instrumentation

- To avoid issues when sorting of the drivers can cause an invalid call
  during instrumentation after driver failures, then the driver sorting
has been moved to occur during the call to zeDriverGet or zeinitDrivers.
- In addition, the driver type is not init to UINT32_MAX in the enum
  such that all failing drivers are forced to the back of the sorted
list.

Signed-off-by: Neil R. Spruit <[email protected]>

* Create recursion guard and check for sorting only once

Signed-off-by: Neil R. Spruit <[email protected]>

* Disable sorting given program instrumentation is enabled

Signed-off-by: Neil R. Spruit <[email protected]>

* Updated documentation on limitations of the sorting support

Signed-off-by: Neil R. Spruit <[email protected]>

* Sort all drivers regardless of the flags set by the user to ensure sorting once

Signed-off-by: Neil R. Spruit <[email protected]>

* Update loader internal header template with changes

Signed-off-by: Neil R. Spruit <[email protected]>

---------

Signed-off-by: Neil R. Spruit <[email protected]>

Author: nrspruit

CHANGELOG.md

@@ -1,5 +1,7 @@
 # Level zero loader changelog
 
+## v1.21.2
+*  Fix driver sorting given driver failures and instrumentation
 ## v1.21.1
 * Fix stype assignment in zello_world
 * Given static Loader, allocate lib context_t always as dynamic and enable delayed destroy of context

CMakeLists.txt

@@ -13,7 +13,7 @@ if(MSVC AND (MSVC_VERSION LESS 1900))
 endif()
 
 # This project follows semantic versioning (https://semver.org/)
-project(level-zero VERSION 1.21.1)
+project(level-zero VERSION 1.21.2)
 
 include(GNUInstallDirs)
 

README.md

@@ -109,6 +109,7 @@ The ordering of the drivers reported to the user is based on the order of the en
 When additional driver types are added, they should be added to the end of the list to avoid reporting new device types
 before known device types.
 
+NOTE: Due to known issues with Program Instrumentation usecases, when ZET_ENABLE_PROGRAM_INSTRUMENTATION is enabled, driver sorting is not possible in the loader.
 
 # Contributing
 

scripts/templates/ldrddi.cpp.mako

@@ -59,6 +59,25 @@ namespace loader
         %elif re.match(r"\w+DriverGet$", th.make_func_name(n, tags, obj)) or re.match(r"\w+InitDrivers$", th.make_func_name(n, tags, obj)):
         uint32_t total_driver_handle_count = 0;
 
+        if (!loader::context->sortingInProgress.exchange(true) && !loader::context->instrumentationEnabled) {
+            %if namespace != "zes":
+            %if not re.match(r"\w+InitDrivers$", th.make_func_name(n, tags, obj)):
+            std::call_once(loader::context->coreDriverSortOnce, []() {
+                loader::context->driverSorting(&loader::context->zeDrivers, nullptr);
+            });
+            %else:
+            std::call_once(loader::context->coreDriverSortOnce, [desc]() {
+                loader::context->driverSorting(&loader::context->zeDrivers, desc);
+            });
+            %endif
+            %else:
+            std::call_once(loader::context->sysmanDriverSortOnce, []() {
+                loader::context->driverSorting(loader::context->sysmanInstanceDrivers, nullptr);
+            });
+            %endif
+            loader::context->sortingInProgress.store(false);
+        }
+
         %if namespace != "zes":
         for( auto& drv : loader::context->zeDrivers )
         %else:

scripts/templates/ze_loader_internal.h.mako

@@ -61,7 +61,7 @@ namespace loader
         dditable_t dditable = {};
         std::string name;
         bool driverInuse = false;
-        zel_driver_type_t driverType;
+        zel_driver_type_t driverType = ZEL_DRIVER_TYPE_FORCE_UINT32;
         ze_driver_properties_t properties;
         bool pciOrderingRequested = false;
     };
@@ -111,10 +111,15 @@ namespace loader
         ze_result_t init();
         ze_result_t init_driver(driver_t &driver, ze_init_flags_t flags, ze_init_driver_type_desc_t* desc, ze_global_dditable_t *globalInitStored, zes_global_dditable_t *sysmanGlobalInitStored, bool sysmanOnly);
         void add_loader_version();
+        bool driverSorting(driver_vector_t *drivers, ze_init_driver_type_desc_t* desc);
         ~context_t();
         bool intercept_enabled = false;
         bool debugTraceEnabled = false;
         bool tracingLayerEnabled = false;
+        std::once_flag coreDriverSortOnce;
+        std::once_flag sysmanDriverSortOnce;
+        std::atomic<bool> sortingInProgress = {false};
+        bool instrumentationEnabled = false;
         dditable_t tracing_dditable = {};
         std::shared_ptr<Logger> zel_logger;
     };

source/drivers/null/ze_null.cpp

@@ -8,6 +8,7 @@
  *
  */
 #include "ze_null.h"
+#include <cstring>
 
 namespace driver
 {
@@ -138,6 +139,10 @@ namespace driver
             ze_device_properties_t deviceProperties = {};
             deviceProperties.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
             deviceProperties.type = ZE_DEVICE_TYPE_GPU;
+            auto driver_type = getenv_string( "ZEL_TEST_NULL_DRIVER_TYPE" );
+            if (std::strcmp(driver_type.c_str(), "NPU") == 0) {
+                deviceProperties.type = ZE_DEVICE_TYPE_VPU;
+            }
 #if defined(_WIN32)
             strcpy_s( deviceProperties.name, "Null Device" );
 #else

source/loader/ze_ldrddi.cpp

@@ -55,6 +55,13 @@ namespace loader
 
         uint32_t total_driver_handle_count = 0;
 
+        if (!loader::context->sortingInProgress.exchange(true) && !loader::context->instrumentationEnabled) {
+            std::call_once(loader::context->coreDriverSortOnce, []() {
+                loader::context->driverSorting(&loader::context->zeDrivers, nullptr);
+            });
+            loader::context->sortingInProgress.store(false);
+        }
+
         for( auto& drv : loader::context->zeDrivers )
         {
             if(drv.initStatus != ZE_RESULT_SUCCESS)
@@ -131,6 +138,13 @@ namespace loader
 
         uint32_t total_driver_handle_count = 0;
 
+        if (!loader::context->sortingInProgress.exchange(true) && !loader::context->instrumentationEnabled) {
+            std::call_once(loader::context->coreDriverSortOnce, [desc]() {
+                loader::context->driverSorting(&loader::context->zeDrivers, desc);
+            });
+            loader::context->sortingInProgress.store(false);
+        }
+
         for( auto& drv : loader::context->zeDrivers )
         {
             if (!drv.dditable.ze.Global.pfnInitDrivers) {