[26.04+ experimental] Add io_uring backend to improve I/O performance in general

cpp/CMakeLists.txt

@@ -153,6 +153,7 @@ set(SOURCES
     "src/file_utils.cpp"
     "src/mmap.cpp"
     "src/detail/env.cpp"
+    "src/detail/io_uring.cpp"
     "src/detail/nvtx.cpp"
     "src/detail/posix_io.cpp"
     "src/shim/cuda.cpp"
@@ -189,7 +190,7 @@ target_include_directories(
 target_link_libraries(
   kvikio
   PUBLIC Threads::Threads BS::thread_pool ${CMAKE_DL_LIBS} $<TARGET_NAME_IF_EXISTS:nvtx3::nvtx3-cpp>
-  PRIVATE $<TARGET_NAME_IF_EXISTS:CURL::libcurl>
+  PRIVATE $<TARGET_NAME_IF_EXISTS:CURL::libcurl> uring
 )
 
 target_compile_definitions(

cpp/include/kvikio/defaults.hpp

@@ -121,6 +121,8 @@ class defaults {
   std::vector<int> _http_status_codes;
   bool _auto_direct_io_read;
   bool _auto_direct_io_write;
+  bool _io_uring_enabled;
+  unsigned int _io_uring_queue_depth;
   bool _thread_pool_per_block_device;
 
   static unsigned int get_num_threads_from_env();
@@ -396,6 +398,13 @@ class defaults {
    */
   static void set_auto_direct_io_write(bool flag);
 
+  static bool io_uring_enabled();
+
+  static void set_io_uring_enabled(bool flag);
+
+  static unsigned int io_uring_queue_depth();
+
+  static void set_io_uring_queue_depth(unsigned int io_uring_queue_depth);
   /**
    * @brief Check if per-block-device thread pools are enabled.
    *

cpp/include/kvikio/detail/io_uring.hpp

@@ -0,0 +1,42 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2022-2025, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+#pragma once
+
+#include <liburing.h>
+#include <linux/io_uring.h>
+
+#include <kvikio/shim/cuda.hpp>
+
+namespace kvikio::detail {
+
+class IoUringManager {
+ private:
+  IoUringManager();
+  io_uring _ring{};
+  unsigned int _queue_depth{};
+  std::size_t _task_size{};
+
+ public:
+  static IoUringManager& get();
+  ~IoUringManager() noexcept;
+
+  IoUringManager(IoUringManager const&)            = delete;
+  IoUringManager& operator=(IoUringManager const&) = delete;
+  IoUringManager(IoUringManager&&)                 = delete;
+  IoUringManager& operator=(IoUringManager&&)      = delete;
+
+  io_uring* ring() noexcept;
+  unsigned int queue_depth() noexcept;
+  std::size_t task_size() noexcept;
+};
+
+bool is_io_uring_supported() noexcept;
+
+std::size_t io_uring_read_host(int fd, void* buf, std::size_t size, std::size_t file_offset);
+
+std::size_t io_uring_read_device(
+  int fd, void* buf, std::size_t size, std::size_t file_offset, CUstream stream);
+
+}  // namespace kvikio::detail

cpp/src/defaults.cpp

@@ -142,6 +142,11 @@ defaults::defaults()
     _auto_direct_io_write = getenv_or("KVIKIO_AUTO_DIRECT_IO_WRITE", true);
   }
 
+  {
+    _io_uring_enabled     = getenv_or("KVIKIO_IO_URING_ENABLED", false);
+    _io_uring_queue_depth = getenv_or("KVIKIO_IO_URING_QUEUE_DEPTH", 32U);
+  }
+
   // Determine the default value of `thread_pool_per_block_device`
   {
     _thread_pool_per_block_device = getenv_or("KVIKIO_THREAD_POOL_PER_BLOCK_DEVICE", false);
@@ -244,6 +249,17 @@ bool defaults::auto_direct_io_write() { return instance()->_auto_direct_io_write
 
 void defaults::set_auto_direct_io_write(bool flag) { instance()->_auto_direct_io_write = flag; }
 
+bool defaults::io_uring_enabled() { return instance()->_io_uring_enabled; }
+
+void defaults::set_io_uring_enabled(bool flag) { instance()->_io_uring_enabled = flag; }
+
+unsigned int defaults::io_uring_queue_depth() { return instance()->_io_uring_queue_depth; }
+
+void defaults::set_io_uring_queue_depth(unsigned int io_uring_queue_depth)
+{
+  instance()->_io_uring_queue_depth = io_uring_queue_depth;
+}
+
 bool defaults::thread_pool_per_block_device() { return instance()->_thread_pool_per_block_device; }
 
 void defaults::set_thread_pool_per_block_device(bool flag)