Improve Windows windowsaio engine performance
[fio.git] / engines / windowsaio.c
index 5c9d85e11b528c19dccfc1337a7fdf27d2982df4..78f7382e3aa2aa48bf78c90fe870a9db47fdac2b 100644 (file)
@@ -1,9 +1,8 @@
 /*
  * Native Windows async IO engine
- * Copyright (C) 2010 Bruce Cran <bruce@cran.org.uk>
+ * Copyright (C) 2011 Bruce Cran <bruce@cran.org.uk>
  */
 
-
 #include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 
 #include "../fio.h"
 
-BOOL windowsaio_debug = FALSE;
+typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
+
+struct fio_overlapped {
+       OVERLAPPED o;
+       struct io_u *io_u;
+       BOOL io_complete;
+    BOOL io_free;
+};
 
 struct windowsaio_data {
+       struct fio_overlapped *ovls;
        struct io_u **aio_events;
-       unsigned int ioFinished;
-       BOOL running;
-       BOOL stopped;
-       HANDLE hThread;
+       HANDLE iothread;
+       HANDLE iocomplete_event;
+       CANCELIOEX pCancelIoEx;
+       BOOL iothread_running;
 };
 
-typedef struct {
- OVERLAPPED o;
- struct io_u *io_u;
-} FIO_OVERLAPPED;
-
 struct thread_ctx {
-       HANDLE ioCP;
+       HANDLE iocp;
        struct windowsaio_data *wd;
 };
 
-static void PrintError(LPCSTR lpszFunction);
 static int fio_windowsaio_cancel(struct thread_data *td,
                               struct io_u *io_u);
-static DWORD GetEndCount(DWORD startCount, struct timespec *t);
-static BOOL TimedOut(DWORD startCount, DWORD endCount);
+static BOOL timeout_expired(DWORD start_count, DWORD end_count);
 static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
                                    unsigned int max, struct timespec *t);
 static struct io_u *fio_windowsaio_event(struct thread_data *td, int event);
@@ -49,327 +49,114 @@ static int fio_windowsaio_init(struct thread_data *td);
 static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f);
 static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f);
 
-static void PrintError(LPCSTR lpszFunction)
+int sync_file_range(int fd, off64_t offset, off64_t nbytes,
+                          unsigned int flags)
 {
-       // Retrieve the system error message for the last-error code
-
-       LPSTR lpMsgBuf;
-       DWORD dw = GetLastError();
-
-       FormatMessage(
-               FORMAT_MESSAGE_ALLOCATE_BUFFER |
-               FORMAT_MESSAGE_FROM_SYSTEM |
-               FORMAT_MESSAGE_IGNORE_INSERTS,
-               NULL,
-               dw,
-               MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
-               (LPTSTR)&lpMsgBuf,
-               0, NULL );
-
-       log_err("%s - %s", lpszFunction, lpMsgBuf);
-       LocalFree(lpMsgBuf);
+       errno = ENOSYS;
+       return -1;
 }
 
-static int fio_windowsaio_cancel(struct thread_data *td,
-                              struct io_u *io_u)
+static int fio_windowsaio_init(struct thread_data *td)
 {
-       BOOL bSuccess;
+       struct windowsaio_data *wd;
+       HANDLE hKernel32Dll;
        int rc = 0;
+       int i;
 
-       bSuccess = CancelIo(io_u->file->hFile);
-
-       if (!bSuccess)
+       wd = malloc(sizeof(struct windowsaio_data));
+       if (wd != NULL)
+               ZeroMemory(wd, sizeof(struct windowsaio_data));
+       else
                rc = 1;
 
-       return rc;
-}
-
-static DWORD GetEndCount(DWORD startCount, struct timespec *t)
-{
-       DWORD endCount = startCount;
-
-       if (t == NULL)
-               return 0;
-
-       endCount += (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
-       return endCount;
-}
-
-static BOOL TimedOut(DWORD startCount, DWORD endCount)
-{
-       BOOL expired = FALSE;
-       DWORD currentTime;
-
-       if (startCount == 0 || endCount == 0)
-               return FALSE;
-
-       currentTime = GetTickCount();
-
-       if ((endCount > startCount) && currentTime >= endCount)
-               expired = TRUE;
-       else if (currentTime < startCount && currentTime > endCount)
-               expired = TRUE;
-
-       if (windowsaio_debug)
-               printf("windowsaio: timedout = %d\n", expired);
-
-       return expired;
-}
-
-static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
-                                   unsigned int max, struct timespec *t)
-{
-       struct windowsaio_data *wd = td->io_ops->data;
-       struct flist_head *entry;
-       unsigned int dequeued = 0;
-       struct io_u *io_u;
-       DWORD startCount = 0, endCount = 0;
-       BOOL timedout = FALSE;
-       unsigned int r = 0;
-
-       if (windowsaio_debug)
-               printf("getevents (min %d, max %d)\n", min, max);
-
-       if (t != NULL) {
-               startCount = GetTickCount();
-               endCount = GetEndCount(startCount, t);
+       if (!rc) {
+               wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
+               if (wd->aio_events == NULL)
+                       rc = 1;
        }
 
-       while (dequeued < min && !timedout) {
-
-               flist_for_each(entry, &td->io_u_busylist) {
-                       io_u = flist_entry(entry, struct io_u, list);
-
-                       if (io_u->seen == 0)
-                               continue;
-
-                       dequeued++;
-
-                       wd->ioFinished--;
-                       wd->aio_events[r] = io_u;
-                       r++;
+       if (!rc) {
+               wd->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped));
+               if (wd->ovls == NULL)
+                       rc = 1;
+       }
 
-                       if (windowsaio_debug)
-                               printf("dequeued %d\n", dequeued);
+       if (!rc) {
+           for (i = 0; i < td->o.iodepth; i++) {
+               wd->ovls[i].io_free = TRUE;
+               wd->ovls[i].io_complete = FALSE;
 
-                       if (dequeued == max)
+                       wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
+                       if (wd->ovls[i].o.hEvent == NULL) {
+                               rc = 1;
                                break;
-               }
-
-               if (TimedOut(startCount, endCount))
-                       timedout = TRUE;
-
-               if (dequeued < min && !timedout)
-                       Sleep(250);
+                       }
+           }
        }
 
-       if (windowsaio_debug)
-               printf("leave getevents (%d)\n", dequeued);
-
-       return dequeued;
-}
-
-static struct io_u *fio_windowsaio_event(struct thread_data *td, int event)
-{
-       struct windowsaio_data *wd = td->io_ops->data;
-       return wd->aio_events[event];
-}
-
-static int fio_windowsaio_queue(struct thread_data *td,
-                             struct io_u *io_u)
-{
-       FIO_OVERLAPPED *fov;
-       DWORD ioBytes;
-       BOOL bSuccess = TRUE;
-       int rc;
-
-       fio_ro_check(td, io_u);
-
-       if (windowsaio_debug)
-               printf("enqueue enter\n");
-
-       fov = malloc(sizeof(FIO_OVERLAPPED));
-       ZeroMemory(fov, sizeof(FIO_OVERLAPPED));
-
-       io_u->seen = 0;
-
-       fov->o.Offset = io_u->offset & 0xFFFFFFFF;
-       fov->o.OffsetHigh = io_u->offset >> 32;
-       fov->o.hEvent = CreateEvent(NULL, FALSE, FALSE, NULL);
-       fov->io_u = io_u;
-
-       if (fov->o.hEvent == NULL) {
-               PrintError(__func__);
-               return 1;
+       if (!rc) {
+               /* Create an auto-reset event */
+               wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
+               if (wd->iocomplete_event == NULL)
+                       rc = 1;
        }
 
-       if (io_u->ddir == DDIR_WRITE)
-               bSuccess = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &ioBytes, &fov->o);
-       else if (io_u->ddir == DDIR_READ)
-               bSuccess = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &ioBytes, &fov->o);
-       else if (io_u->ddir == DDIR_SYNC     ||
-                        io_u->ddir == DDIR_DATASYNC ||
-                        io_u->ddir == DDIR_SYNC_FILE_RANGE)
-       {
-               FlushFileBuffers(io_u->file->hFile);
-               return FIO_Q_COMPLETED;
-       } else if (io_u->ddir == DDIR_TRIM) {
-               log_info("explicit TRIM isn't supported on Windows");
-               return FIO_Q_COMPLETED;
-       }
+       if (rc) {
+               if (wd != NULL) {
+                       if (wd->ovls != NULL)
+                               free(wd->ovls);
+                       if (wd->aio_events != NULL)
+                               free(wd->aio_events);
 
-       if (bSuccess) {
-               io_u->seen = 1;
-               io_u->resid = io_u->xfer_buflen - fov->o.InternalHigh;
-               io_u->error = 0;
-               rc = FIO_Q_COMPLETED;
-       } else if (!bSuccess && GetLastError() == ERROR_IO_PENDING) {
-               rc = FIO_Q_QUEUED;
-       } else {
-               PrintError(__func__);
-               io_u->error = GetLastError();
-               io_u->resid = io_u->xfer_buflen;
-               rc = FIO_Q_COMPLETED;
+                       free(wd);
+               }
        }
 
-       if (windowsaio_debug)
-               printf("enqueue - leave (offset %llu)\n", io_u->offset);
+       hKernel32Dll = GetModuleHandle("kernel32.dll");
+       wd->pCancelIoEx = GetProcAddress(hKernel32Dll, "CancelIoEx");
 
+       td->io_ops->data = wd;
        return rc;
 }
 
 static void fio_windowsaio_cleanup(struct thread_data *td)
 {
+       int i;
        struct windowsaio_data *wd;
 
-       if (windowsaio_debug)
-               printf("windowsaio: cleanup - enter\n");
-
        wd = td->io_ops->data;
-       wd->running = FALSE;
-
-       while (wd->stopped == FALSE)
-               Sleep(5);
 
        if (wd != NULL) {
-               CloseHandle(wd->hThread);
-               free(wd->aio_events);
-               wd->aio_events = NULL;
-               free(wd);
-               td->io_ops->data = NULL;
-       }
-
-       if (windowsaio_debug)
-               printf("windowsaio: cleanup - leave\n");
-}
+           wd->iothread_running = FALSE;
+       WaitForSingleObject(wd->iothread, INFINITE);
 
-static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
-{
-       OVERLAPPED *ovl;
-       FIO_OVERLAPPED *fov;
-       struct io_u *io_u;
-       struct windowsaio_data *wd;
+               CloseHandle(wd->iothread);
+               CloseHandle(wd->iocomplete_event);
 
-       struct thread_ctx *ctx;
-       ULONG_PTR ulKey = 0;
-       BOOL bSuccess;
-       DWORD bytes;
-
-
-       ctx = (struct thread_ctx*)lpParameter;
-       wd = ctx->wd;
-       bSuccess = TRUE;
-
-       if (windowsaio_debug)
-               printf("windowsaio: IoCompletionRoutine - enter\n");
-
-       while (ctx->wd->running) {
-               bSuccess = GetQueuedCompletionStatus(ctx->ioCP, &bytes, &ulKey, &ovl, 500);
-
-               if (windowsaio_debug)
-                       printf("GetQueuedCompletionStatus returned %d\n", bSuccess);
-
-               if (!bSuccess) {
-                       if (GetLastError() == WAIT_TIMEOUT) {
-                               continue;
-                       } else {
-                               PrintError(__func__);
-                               continue;
-                       }
-               }
-
-               fov = CONTAINING_RECORD(ovl, FIO_OVERLAPPED, o);
-               io_u = fov->io_u;
-
-               if (windowsaio_debug) {
-                       if (io_u->seen == 1)
-                               printf("IoCompletionRoutine - got already completed IO\n");
-                       else
-                               printf("IoCompletionRoutine - completed %d IO\n", ctx->wd->ioFinished);
+               for (i = 0; i < td->o.iodepth; i++) {
+                       CloseHandle(wd->ovls[i].o.hEvent);
                }
 
-               if (io_u->seen == 1)
-                       continue;
-
-               ctx->wd->ioFinished++;
-
-               if (ovl->Internal == ERROR_SUCCESS) {
-                       io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
-                       io_u->error = 0;
-               } else {
-                       io_u->resid = io_u->xfer_buflen;
-                       io_u->error = 1;
-               }
+               free(wd->aio_events);
+               free(wd->ovls);
+               free(wd);
 
-               io_u->seen = 1;
-               CloseHandle(ovl->hEvent);
-               free(ovl);
+               td->io_ops->data = NULL;
        }
-
-       bSuccess = CloseHandle(ctx->ioCP);
-       if (!bSuccess)
-               PrintError(__func__);
-
-       if (windowsaio_debug)
-               printf("windowsaio: IoCompletionRoutine - leave\n");
-
-       ctx->wd->stopped = TRUE;
-       free(ctx);
-       return 0;
 }
 
-static int fio_windowsaio_init(struct thread_data *td)
-{
-       int rc = 0;
-       struct windowsaio_data *wd;
-
-       if (windowsaio_debug)
-               printf("windowsaio: init\n");
-
-       wd = malloc(sizeof(struct windowsaio_data));
-
-       ZeroMemory(wd, sizeof(*wd));
-       wd->aio_events = malloc((td->o.iodepth + 1) * sizeof(struct io_u *));
-       ZeroMemory(wd->aio_events, (td->o.iodepth + 1) * sizeof(struct io_u *));
-
-       td->io_ops->data = wd;
-       return rc;
-}
 
 static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
 {
        int rc = 0;
        HANDLE hFile;
-       DWORD flags = FILE_ATTRIBUTE_TEMPORARY | FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
+       DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
        DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
        DWORD openmode = OPEN_ALWAYS;
        DWORD access;
 
        dprint(FD_FILE, "fd open %s\n", f->file_name);
 
-       if (windowsaio_debug)
-               printf("windowsaio: open file %s - enter\n", f->file_name);
-
        if (f->filetype == FIO_TYPE_PIPE) {
                log_err("fio: windowsaio doesn't support pipes\n");
                return 1;
@@ -385,19 +172,13 @@ static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
        if (td->o.sync_io)
                flags |= FILE_FLAG_WRITE_THROUGH;
 
-
        if (td->o.td_ddir == TD_DDIR_READ  ||
-               td->o.td_ddir == TD_DDIR_WRITE ||
-               td->o.td_ddir == TD_DDIR_RANDRW)
-       {
+               td->o.td_ddir == TD_DDIR_WRITE)
                flags |= FILE_FLAG_SEQUENTIAL_SCAN;
-       }
        else
-       {
                flags |= FILE_FLAG_RANDOM_ACCESS;
-       }
 
-       if (td_read(td) || read_only)
+       if (!td_write(td) || read_only)
                access = GENERIC_READ;
        else
                access = (GENERIC_READ | GENERIC_WRITE);
@@ -410,58 +191,244 @@ static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
        f->hFile = CreateFile(f->file_name, access, sharemode,
                NULL, openmode, flags, NULL);
 
-       if (f->hFile == INVALID_HANDLE_VALUE) {
-               log_err("Failed to open %s\n", f->file_name);
-               PrintError(__func__);
+       if (f->hFile == INVALID_HANDLE_VALUE)
                rc = 1;
-       }
 
        /* Only set up the competion port and thread if we're not just
         * querying the device size */
-       if (!rc && td->io_ops->data != NULL) {
-               struct windowsaio_data *wd;
+    if (!rc && td->io_ops->data != NULL) {
                struct thread_ctx *ctx;
+        struct windowsaio_data *wd;
                hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
 
-               wd = td->io_ops->data;
-               wd->running = TRUE;
-               wd->stopped = FALSE;
+        wd = td->io_ops->data;
 
-               ctx = malloc(sizeof(struct thread_ctx));
-               ctx->ioCP = hFile;
-               ctx->wd = wd;
+               wd->iothread_running = TRUE;
 
-               wd->hThread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
+               if (!rc) {
+                       ctx = malloc(sizeof(struct thread_ctx));
+                       ctx->iocp = hFile;
+                       ctx->wd = wd;
 
-               if (wd->hThread == NULL) {
-                       PrintError(__func__);
-                       rc = 1;
+                       wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
                }
-       }
 
-       if (windowsaio_debug)
-               printf("windowsaio: open file - leave (%d)\n", rc);
+               if (rc || wd->iothread == NULL)
+                       rc = 1;
+       }
 
        return rc;
 }
 
 static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 {
-       BOOL bSuccess;
-
-       if (windowsaio_debug)
-               printf("windowsaio: close file\n");
+       int rc = 0;
+       
+       dprint(FD_FILE, "fd close %s\n", f->file_name);
 
        if (f->hFile != INVALID_HANDLE_VALUE) {
-               bSuccess = CloseHandle(f->hFile);
-               if (!bSuccess)
-                       PrintError(__func__);
+               if (!CloseHandle(f->hFile))
+                       rc = 1;
        }
 
        f->hFile = INVALID_HANDLE_VALUE;
+       return rc;
+}
+
+static BOOL timeout_expired(DWORD start_count, DWORD end_count)
+{
+       BOOL expired = FALSE;
+       DWORD current_time;
+
+       current_time = GetTickCount();
+
+       if ((end_count > start_count) && current_time >= end_count)
+               expired = TRUE;
+       else if (current_time < start_count && current_time > end_count)
+               expired = TRUE;
+
+       return expired;
+}
+
+static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
+{
+       struct windowsaio_data *wd = td->io_ops->data;
+       return wd->aio_events[event];
+}
+
+static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
+                                   unsigned int max, struct timespec *t)
+{
+       struct windowsaio_data *wd = td->io_ops->data;
+       struct flist_head *entry;
+       unsigned int dequeued = 0;
+       struct io_u *io_u;
+       struct fio_overlapped *fov;
+       DWORD start_count = 0;
+       DWORD end_count = 0;
+       DWORD status;
+       DWORD mswait = 250;
+
+       if (t != NULL) {
+               mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
+               start_count = GetTickCount();
+               end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
+       }
+
+       do {
+               flist_for_each(entry, &td->io_u_busylist) {
+                       io_u = flist_entry(entry, struct io_u, list);
+                       fov = (struct fio_overlapped*)io_u->engine_data;
+
+                       if (fov->io_complete) {
+                fov->io_complete = FALSE;
+                               fov->io_free  = TRUE;
+                               wd->aio_events[dequeued] = io_u;
+                               dequeued++;
+                       }
+
+                       if (dequeued >= min)
+                               break;
+               }
+
+        if (dequeued < min) {
+                       status = WaitForSingleObject(wd->iocomplete_event, mswait);
+                       if (status != WAIT_OBJECT_0 && dequeued > 0)
+                           break;
+               }
+
+               if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
+                       break;
+       } while (1);
+
+       return dequeued;
+}
+
+static int fio_windowsaio_queue(struct thread_data *td,
+                             struct io_u *io_u)
+{
+    LPOVERLAPPED lpOvl = NULL;
+       struct windowsaio_data *wd;
+       DWORD iobytes;
+       BOOL success;
+       int index;
+       int rc = FIO_Q_COMPLETED;
+
+       fio_ro_check(td, io_u);
+
+       wd = td->io_ops->data;
+
+    for (index = 0; index < td->o.iodepth; index++) {
+        if (wd->ovls[index].io_free) {
+            wd->ovls[index].io_free = FALSE;
+            ResetEvent(wd->ovls[index].o.hEvent);
+            break;
+        }
+    }
+
+    assert(index < td->o.iodepth);
+
+    lpOvl = &wd->ovls[index].o;
+    wd->ovls[index].io_u = io_u;
+       lpOvl->Internal = STATUS_PENDING;
+       lpOvl->InternalHigh = 0;
+       lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
+       lpOvl->OffsetHigh = io_u->offset >> 32;
+       lpOvl->Pointer = NULL;
+    io_u->engine_data = &wd->ovls[index];
+
+       switch (io_u->ddir) {
+    case DDIR_WRITE:
+               success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
+               break;
+       case DDIR_READ:
+               success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
+               break;
+       case DDIR_SYNC:
+       case DDIR_DATASYNC:
+       case DDIR_SYNC_FILE_RANGE:
+               success = FlushFileBuffers(io_u->file->hFile);
+               if (!success)
+                   io_u->error = GetLastError();
+
+               return FIO_Q_COMPLETED;
+               break;
+       case DDIR_TRIM:
+               log_err("manual TRIM isn't supported on Windows");
+               io_u->error = 1;
+               io_u->resid = io_u->xfer_buflen;
+               return FIO_Q_COMPLETED;
+               break;
+       default:
+               assert(0);
+       }
+
+    if (success || GetLastError() == ERROR_IO_PENDING) {
+               rc = FIO_Q_QUEUED;
+       } else {
+               io_u->error = GetLastError();
+               io_u->resid = io_u->xfer_buflen;
+       }
+
+       return rc;
+}
+
+/* Runs as a thread and waits for queued IO to complete */
+static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
+{
+       OVERLAPPED *ovl;
+       struct fio_overlapped *fov;
+       struct io_u *io_u;
+       struct windowsaio_data *wd;
+       struct thread_ctx *ctx;
+       ULONG_PTR ulKey = 0;
+       DWORD bytes;
+
+       ctx = (struct thread_ctx*)lpParameter;
+       wd = ctx->wd;
+
+       do {
+               if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
+                       continue;
+
+               fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
+               io_u = fov->io_u;
+
+               if (ovl->Internal == ERROR_SUCCESS) {
+                       io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
+                       io_u->error = 0;
+               } else {
+                       io_u->resid = io_u->xfer_buflen;
+                       io_u->error = ovl->Internal;
+               }
+
+        fov->io_complete = TRUE;
+               SetEvent(wd->iocomplete_event);
+       } while (ctx->wd->iothread_running);
+
+       CloseHandle(ctx->iocp);
+       free(ctx);
        return 0;
 }
 
+static int fio_windowsaio_cancel(struct thread_data *td,
+                              struct io_u *io_u)
+{
+       int rc = 0;
+
+       struct windowsaio_data *wd = td->io_ops->data;
+
+       /* If we're running on Vista or newer, we can cancel individual IO requests */
+       if (wd->pCancelIoEx != NULL) {
+               struct fio_overlapped *ovl = io_u->engine_data;
+               if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
+                       rc = 1;
+       } else
+               rc = 1;
+
+       return rc;
+}
+
 static struct ioengine_ops ioengine = {
        .name           = "windowsaio",
        .version        = FIO_IOOPS_VERSION,