From a93b9d473de684d24fe46267e9dfe840ec164e07 Mon Sep 17 00:00:00 2001
From: "Pryor, Adam" <Adam.Pryor@amd.com>
Date: Tue, 7 Oct 2025 21:31:23 -0500
Subject: [PATCH] [SWDEV-558895] Fix rsmi monitor fds (#748)

Signed-off-by: adapryor <Adam.pryor@amd.com>

[ROCm/amdsmi commit: 346e1516afed1038f37d98878372edace0acebd0]
---
 projects/amdsmi/rocm_smi/src/rocm_smi.cc      | 33 ++++++++++++++-----
 .../amdsmi/rocm_smi/src/rocm_smi_io_link.cc   | 10 ++++--
 .../amdsmi/rocm_smi/src/rocm_smi_monitor.cc   |  4 ++-
 3 files changed, 36 insertions(+), 11 deletions(-)

diff --git a/projects/amdsmi/rocm_smi/src/rocm_smi.cc b/projects/amdsmi/rocm_smi/src/rocm_smi.cc
index ff48a7b5c4..1b193f9924 100644
--- a/projects/amdsmi/rocm_smi/src/rocm_smi.cc
+++ b/projects/amdsmi/rocm_smi/src/rocm_smi.cc
@@ -7252,6 +7252,11 @@ rsmi_event_notification_init(uint32_t dv_ind) {
   DEVICE_MUTEX
 
   std::lock_guard<std::mutex> guard(*smi.kfd_notif_evt_fh_mutex());
+
+  if (dev->evt_notif_anon_fd() > 0 && dev->evt_notif_anon_file_ptr() != nullptr) {
+    return RSMI_STATUS_SUCCESS;
+  }
+
   if (smi.kfd_notif_evt_fh() == -1) {
     assert(smi.kfd_notif_evt_fh_refcnt() == 0);
     int kfd_fd = open(kPathKFDIoctl, O_RDWR | O_CLOEXEC);
@@ -7666,15 +7671,27 @@ rsmi_status_t rsmi_event_notification_stop(uint32_t dv_ind) {
 
   std::lock_guard<std::mutex> guard(*smi.kfd_notif_evt_fh_mutex());
 
-  if (dev->evt_notif_anon_fd() == -1) {
-    return RSMI_STATUS_INVALID_ARGS;
-  }
-//  close(dev->evt_notif_anon_fd());
   FILE *anon_fp = smi.devices()[dv_ind]->evt_notif_anon_file_ptr();
-  fclose(anon_fp);
-  assert(errno == 0 || errno == EAGAIN);
-  dev->set_evt_notif_anon_file_ptr(nullptr);
-  dev->set_evt_notif_anon_fd(-1);
+  int   anon_fd = smi.devices()[dv_ind]->evt_notif_anon_fd();
+
+  // If nothing to close, success
+  if (!anon_fp && anon_fd <= 0) {
+    return RSMI_STATUS_SUCCESS;
+  }
+
+  // Clear state first so nobody else can race a second close
+  smi.devices()[dv_ind]->set_evt_notif_anon_file_ptr(nullptr);
+  smi.devices()[dv_ind]->set_evt_notif_anon_fd(-1);
+
+  if (anon_fp) {
+    if (fclose(anon_fp) != 0) {
+      return amd::smi::ErrnoToRsmiStatus(errno);
+    }
+  } else { // no FILE*, but fd was valid
+    if (close(anon_fd) != 0) {
+      return amd::smi::ErrnoToRsmiStatus(errno);
+    }
+  }
 
   if (smi.kfd_notif_evt_fh_refcnt_dec() == 0) {
     int ret = close(smi.kfd_notif_evt_fh());
diff --git a/projects/amdsmi/rocm_smi/src/rocm_smi_io_link.cc b/projects/amdsmi/rocm_smi/src/rocm_smi_io_link.cc
index c18b6ae82b..5d8c4f6a0d 100644
--- a/projects/amdsmi/rocm_smi/src/rocm_smi_io_link.cc
+++ b/projects/amdsmi/rocm_smi/src/rocm_smi_io_link.cc
@@ -196,7 +196,10 @@ static int DiscoverLinks(std::map<std::pair<uint32_t, uint32_t>,
     std::string link_path_root = LinkPathRoot(node_indx, directory);
 
     auto io_link_dir = opendir(link_path_root.c_str());
-    assert(io_link_dir != nullptr);
+
+    if (!io_link_dir) {
+      return errno ? errno : 1;
+    }
 
     auto dentry_io_link = readdir(io_link_dir);
     while (dentry_io_link != nullptr) {
@@ -264,7 +267,10 @@ static int DiscoverLinksPerNode(uint32_t node_indx, std::map<uint32_t,
   std::string link_path_root = LinkPathRoot(node_indx, directory);
 
   auto io_link_dir = opendir(link_path_root.c_str());
-  assert(io_link_dir != nullptr);
+
+  if (!io_link_dir) {
+    return errno ? errno : 1;
+  }
 
   auto dentry = readdir(io_link_dir);
   while (dentry != nullptr) {
diff --git a/projects/amdsmi/rocm_smi/src/rocm_smi_monitor.cc b/projects/amdsmi/rocm_smi/src/rocm_smi_monitor.cc
index 5d37b8a849..567089cc24 100644
--- a/projects/amdsmi/rocm_smi/src/rocm_smi_monitor.cc
+++ b/projects/amdsmi/rocm_smi/src/rocm_smi_monitor.cc
@@ -397,7 +397,9 @@ Monitor::setVoltSensorLabelMap(void) {
 static int get_supported_sensors(std::string dir_path, std::string fn_reg_ex,
                                               std::vector<uint64_t> *sensors) {
   auto hwmon_dir = opendir(dir_path.c_str());
-  assert(hwmon_dir != nullptr);
+  if (!hwmon_dir) {
+    return errno ? errno : 1;
+  }
   assert(sensors != nullptr);
 
   sensors->clear();