Change default mutex to thread only (#104)

Most sysfs reads do not require cross-process level mutex, and writes to sysfs
should be protected by the kernel already.
Users can still switch to the old behavior by setting the environment variable AMDSMI_MUTEX_CROSS_PROCESS=1.


[ROCm/amdsmi commit: 8e74024b11]

projects/amdsmi/CHANGELOG.md

@@ -13,11 +13,11 @@ Full documentation for amd_smi_lib is available at [https://rocm.docs.amd.com/pr
 
 - **Added TVIOL_ACTIVE to `amd-smi monitor`**.  
 Added temperature violation active or not status to `amd-smi monitor`. TVIOL_ACTIVE will be displayed as below:
- - True if active
- - False if not active
- - N/A if not supported.
+  - True if active
+  - False if not active
+  - N/A if not supported.
 
- Example CLI output:
+  Example CLI output:
 ```shell
 $ amd-smi monitor --viol
 GPU  PVIOL  TVIOL  TVIOL_ACTIVE  PHOT_TVIOL  VR_TVIOL  HBM_TVIOL
@@ -31,7 +31,7 @@ GPU  PVIOL  TVIOL  TVIOL_ACTIVE  PHOT_TVIOL  VR_TVIOL  HBM_TVIOL
   7  100 %    0 %         False         0 %       0 %        0 %
 ```
 
-- **Added support for GPU metrics 1.7 to `amdsmi_get_gpu_metrics_info()`**  
+- **Added support for GPU metrics 1.7 to `amdsmi_get_gpu_metrics_info()`**.  
 Updated `amdsmi_get_gpu_metrics_info()` and structure `amdsmi_gpu_metrics_t` to include new fields for XGMI Link Status, graphics clocks below host limit (per XCP), and VRAM max bandwidth:  
   - `uint64_t vram_max_bandwidth` - VRAM max bandwidth at max memory clock (GB/s)
   - `uint16_t xgmi_link_status[MAX_NUM_XGMI_LINKS]` - XGMI link statis, 1=Up 0=Down
@@ -145,7 +145,11 @@ Python API now accepts `sensor_ind` as an optional argument, does not imapact pr
 
 - **Depricated enum `AMDSMI_NORMAL_STRING_LENGTH` in favor of `AMDSMI_MAX_STRING_LENGTH`**.  
 
-- **Changed amdsmi_vram_vendor_type_t enum names impacting amdsmi_vram_info_t structure**. 
+- **Changed to use thread local mutex by default**.  
+Most sysfs reads do not require cross-process level mutex, and writes to sysfs should be protected by the kernel already.  
+Users can still switch to the old behavior by setting the environment variable `AMDSMI_MUTEX_CROSS_PROCESS=1`.
+
+- **Changed `amdsmi_vram_vendor_type_t` enum names impacting `amdsmi_vram_info_t` structure**. 
 This also change impacts usage of the vram_vendor output of `amdsmi_get_gpu_vram_info()` 
 
 - **Changed `amdsmi_nps_caps_t` struct impacting `amdsmi_memory_partition_config_t`, `amdsmi_accelerator_partition_t`, `amdsmi_accelerator_partition_profile_config_t`**.  

projects/amdsmi/third_party/shared_mutex/shared_mutex.cc

@@ -44,7 +44,8 @@ THE SOFTWARE.
 #include "rocm_smi/rocm_smi_exception.h"
 #include "rocm_smi/rocm_smi_main.h"
 
-#define THREAD_ONLY_ENV_VAR "RSMI_MUTEX_THREAD_ONLY"
+// Default to thread only mutex unless export AMDSMI_MUTEX_CROSS_PROCESS=1
+#define PROCESS_CROSS_PROCESS_ENV_VAR "AMDSMI_MUTEX_CROSS_PROCESS"
 #define MUTEX_TIME_OUT_ENV_VAR "RSMI_MUTEX_TIMEOUT"
 #define DEFAULT_MUTEX_TIMEOUT_SECONDS 5
 
@@ -141,7 +142,7 @@ shared_mutex_t shared_mutex_init(const char *name, mode_t mode, bool retried) {
 
   amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
 
-  if (GetEnvVarUInteger(THREAD_ONLY_ENV_VAR) == 1 || smi.is_thread_only_mutex()) {
+  if (GetEnvVarUInteger(PROCESS_CROSS_PROCESS_ENV_VAR) != 1 || smi.is_thread_only_mutex()) {
     return init_thread_safe_only(name);
   }
 
@@ -296,7 +297,7 @@ shared_mutex_t shared_mutex_init(const char *name, mode_t mode, bool retried) {
 
 int shared_mutex_close(shared_mutex_t mutex) {
   amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
-  const bool is_thread_only = GetEnvVarUInteger(THREAD_ONLY_ENV_VAR) == 1 ||
+  const bool is_thread_only = GetEnvVarUInteger(PROCESS_CROSS_PROCESS_ENV_VAR) != 1 ||
           smi.is_thread_only_mutex();
   if (is_thread_only) {
     delete mutex.ptr;
@@ -317,7 +318,7 @@ int shared_mutex_close(shared_mutex_t mutex) {
 
 int shared_mutex_destroy(shared_mutex_t mutex) {
   amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
-  const bool is_thread_only = GetEnvVarUInteger(THREAD_ONLY_ENV_VAR) == 1 ||
+  const bool is_thread_only = GetEnvVarUInteger(PROCESS_CROSS_PROCESS_ENV_VAR) != 1 ||
           smi.is_thread_only_mutex();
   if ((errno = pthread_mutex_destroy(mutex.ptr))) {
     perror("pthread_mutex_destroy");