Merge remote-tracking branch 'bonzini/threadpool' into staging

* bonzini/threadpool: (39 commits)
  raw-win32: implement native asynchronous I/O
  raw-posix: move linux-aio.c to block/
  raw-win32: add emulated AIO support
  raw-posix: rename raw-posix-aio.h, hide unavailable prototypes
  raw: merge posix-aio-compat.c into block/raw-posix.c
  block: switch posix-aio-compat to threadpool
  threadpool: do not take lock in event_notifier_ready
  aio: add generic thread-pool facility
  qemu-thread: add QemuSemaphore
  linux-aio: use event notifiers
  aio: clean up now-unused functions
  main-loop: use aio_notify for qemu_notify_event
  main-loop: use GSource to poll AIO file descriptors
  aio: call aio_notify after setting I/O handlers
  aio: add aio_notify
  aio: make AioContexts GSources
  aio: add Win32 implementation
  aio: prepare for introducing GSource-based dispatch
  aio: add non-blocking variant of aio_wait
  aio: test node->deleted before calling io_flush
  ...

Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>

1 files changed, 297 insertions, 11 deletions

diff --git a/block/raw-posix.c b/block/raw-posix.c
index 28d439fa81..f2f0404f6f 100644
--- a/block/raw-posix.c
+++ b/block/raw-posix.c
@@ -27,7 +27,10 @@
 #include "qemu-log.h"
 #include "block_int.h"
 #include "module.h"
-#include "block/raw-posix-aio.h"
+#include "trace.h"
+#include "thread-pool.h"
+#include "iov.h"
+#include "raw-aio.h"
 
 #if defined(__APPLE__) && (__MACH__)
 #include <paths.h>
@@ -149,6 +152,20 @@ typedef struct BDRVRawReopenState {
 static int fd_open(BlockDriverState *bs);
 static int64_t raw_getlength(BlockDriverState *bs);
 
+typedef struct RawPosixAIOData {
+    BlockDriverState *bs;
+    int aio_fildes;
+    union {
+        struct iovec *aio_iov;
+        void *aio_ioctl_buf;
+    };
+    int aio_niov;
+    size_t aio_nbytes;
+#define aio_ioctl_cmd   aio_nbytes /* for QEMU_AIO_IOCTL */
+    off_t aio_offset;
+    int aio_type;
+} RawPosixAIOData;
+
 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 static int cdrom_reopen(BlockDriverState *bs);
 #endif
@@ -266,14 +283,10 @@ static int raw_open_common(BlockDriverState *bs, const char *filename,
     }
     s->fd = fd;
 
-    /* We're falling back to POSIX AIO in some cases so init always */
-    if (paio_init() < 0) {
-        goto out_close;
-    }
-
 #ifdef CONFIG_LINUX_AIO
     if (raw_set_aio(&s->aio_ctx, &s->use_aio, bdrv_flags)) {
-        goto out_close;
+        qemu_close(fd);
+        return -errno;
     }
 #endif
 
@@ -284,10 +297,6 @@ static int raw_open_common(BlockDriverState *bs, const char *filename,
 #endif
 
     return 0;
-
-out_close:
-    qemu_close(fd);
-    return -errno;
 }
 
 static int raw_open(BlockDriverState *bs, const char *filename, int flags)
@@ -434,6 +443,283 @@ static int qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
     return 1;
 }
 
+static ssize_t handle_aiocb_ioctl(RawPosixAIOData *aiocb)
+{
+    int ret;
+
+    ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
+    if (ret == -1) {
+        return -errno;
+    }
+
+    /*
+     * This looks weird, but the aio code only considers a request
+     * successful if it has written the full number of bytes.
+     *
+     * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
+     * so in fact we return the ioctl command here to make posix_aio_read()
+     * happy..
+     */
+    return aiocb->aio_nbytes;
+}
+
+static ssize_t handle_aiocb_flush(RawPosixAIOData *aiocb)
+{
+    int ret;
+
+    ret = qemu_fdatasync(aiocb->aio_fildes);
+    if (ret == -1) {
+        return -errno;
+    }
+    return 0;
+}
+
+#ifdef CONFIG_PREADV
+
+static bool preadv_present = true;
+
+static ssize_t
+qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
+{
+    return preadv(fd, iov, nr_iov, offset);
+}
+
+static ssize_t
+qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
+{
+    return pwritev(fd, iov, nr_iov, offset);
+}
+
+#else
+
+static bool preadv_present = false;
+
+static ssize_t
+qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
+{
+    return -ENOSYS;
+}
+
+static ssize_t
+qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
+{
+    return -ENOSYS;
+}
+
+#endif
+
+static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
+{
+    ssize_t len;
+
+    do {
+        if (aiocb->aio_type & QEMU_AIO_WRITE)
+            len = qemu_pwritev(aiocb->aio_fildes,
+                               aiocb->aio_iov,
+                               aiocb->aio_niov,
+                               aiocb->aio_offset);
+         else
+            len = qemu_preadv(aiocb->aio_fildes,
+                              aiocb->aio_iov,
+                              aiocb->aio_niov,
+                              aiocb->aio_offset);
+    } while (len == -1 && errno == EINTR);
+
+    if (len == -1) {
+        return -errno;
+    }
+    return len;
+}
+
+/*
+ * Read/writes the data to/from a given linear buffer.
+ *
+ * Returns the number of bytes handles or -errno in case of an error. Short
+ * reads are only returned if the end of the file is reached.
+ */
+static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
+{
+    ssize_t offset = 0;
+    ssize_t len;
+
+    while (offset < aiocb->aio_nbytes) {
+        if (aiocb->aio_type & QEMU_AIO_WRITE) {
+            len = pwrite(aiocb->aio_fildes,
+                         (const char *)buf + offset,
+                         aiocb->aio_nbytes - offset,
+                         aiocb->aio_offset + offset);
+        } else {
+            len = pread(aiocb->aio_fildes,
+                        buf + offset,
+                        aiocb->aio_nbytes - offset,
+                        aiocb->aio_offset + offset);
+        }
+        if (len == -1 && errno == EINTR) {
+            continue;
+        } else if (len == -1) {
+            offset = -errno;
+            break;
+        } else if (len == 0) {
+            break;
+        }
+        offset += len;
+    }
+
+    return offset;
+}
+
+static ssize_t handle_aiocb_rw(RawPosixAIOData *aiocb)
+{
+    ssize_t nbytes;
+    char *buf;
+
+    if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
+        /*
+         * If there is just a single buffer, and it is properly aligned
+         * we can just use plain pread/pwrite without any problems.
+         */
+        if (aiocb->aio_niov == 1) {
+             return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
+        }
+        /*
+         * We have more than one iovec, and all are properly aligned.
+         *
+         * Try preadv/pwritev first and fall back to linearizing the
+         * buffer if it's not supported.
+         */
+        if (preadv_present) {
+            nbytes = handle_aiocb_rw_vector(aiocb);
+            if (nbytes == aiocb->aio_nbytes ||
+                (nbytes < 0 && nbytes != -ENOSYS)) {
+                return nbytes;
+            }
+            preadv_present = false;
+        }
+
+        /*
+         * XXX(hch): short read/write.  no easy way to handle the reminder
+         * using these interfaces.  For now retry using plain
+         * pread/pwrite?
+         */
+    }
+
+    /*
+     * Ok, we have to do it the hard way, copy all segments into
+     * a single aligned buffer.
+     */
+    buf = qemu_blockalign(aiocb->bs, aiocb->aio_nbytes);
+    if (aiocb->aio_type & QEMU_AIO_WRITE) {
+        char *p = buf;
+        int i;
+
+        for (i = 0; i < aiocb->aio_niov; ++i) {
+            memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
+            p += aiocb->aio_iov[i].iov_len;
+        }
+    }
+
+    nbytes = handle_aiocb_rw_linear(aiocb, buf);
+    if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
+        char *p = buf;
+        size_t count = aiocb->aio_nbytes, copy;
+        int i;
+
+        for (i = 0; i < aiocb->aio_niov && count; ++i) {
+            copy = count;
+            if (copy > aiocb->aio_iov[i].iov_len) {
+                copy = aiocb->aio_iov[i].iov_len;
+            }
+            memcpy(aiocb->aio_iov[i].iov_base, p, copy);
+            p     += copy;
+            count -= copy;
+        }
+    }
+    qemu_vfree(buf);
+
+    return nbytes;
+}
+
+static int aio_worker(void *arg)
+{
+    RawPosixAIOData *aiocb = arg;
+    ssize_t ret = 0;
+
+    switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
+    case QEMU_AIO_READ:
+        ret = handle_aiocb_rw(aiocb);
+        if (ret >= 0 && ret < aiocb->aio_nbytes && aiocb->bs->growable) {
+            iov_memset(aiocb->aio_iov, aiocb->aio_niov, ret,
+                      0, aiocb->aio_nbytes - ret);
+
+            ret = aiocb->aio_nbytes;
+        }
+        if (ret == aiocb->aio_nbytes) {
+            ret = 0;
+        } else if (ret >= 0 && ret < aiocb->aio_nbytes) {
+            ret = -EINVAL;
+        }
+        break;
+    case QEMU_AIO_WRITE:
+        ret = handle_aiocb_rw(aiocb);
+        if (ret == aiocb->aio_nbytes) {
+            ret = 0;
+        } else if (ret >= 0 && ret < aiocb->aio_nbytes) {
+            ret = -EINVAL;
+        }
+        break;
+    case QEMU_AIO_FLUSH:
+        ret = handle_aiocb_flush(aiocb);
+        break;
+    case QEMU_AIO_IOCTL:
+        ret = handle_aiocb_ioctl(aiocb);
+        break;
+    default:
+        fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
+        ret = -EINVAL;
+        break;
+    }
+
+    g_slice_free(RawPosixAIOData, aiocb);
+    return ret;
+}
+
+static BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
+        int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
+        BlockDriverCompletionFunc *cb, void *opaque, int type)
+{
+    RawPosixAIOData *acb = g_slice_new(RawPosixAIOData);
+
+    acb->bs = bs;
+    acb->aio_type = type;
+    acb->aio_fildes = fd;
+
+    if (qiov) {
+        acb->aio_iov = qiov->iov;
+        acb->aio_niov = qiov->niov;
+    }
+    acb->aio_nbytes = nb_sectors * 512;
+    acb->aio_offset = sector_num * 512;
+
+    trace_paio_submit(acb, opaque, sector_num, nb_sectors, type);
+    return thread_pool_submit_aio(aio_worker, acb, cb, opaque);
+}
+
+static BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
+        unsigned long int req, void *buf,
+        BlockDriverCompletionFunc *cb, void *opaque)
+{
+    RawPosixAIOData *acb = g_slice_new(RawPosixAIOData);
+
+    acb->bs = bs;
+    acb->aio_type = QEMU_AIO_IOCTL;
+    acb->aio_fildes = fd;
+    acb->aio_offset = 0;
+    acb->aio_ioctl_buf = buf;
+    acb->aio_ioctl_cmd = req;
+
+    return thread_pool_submit_aio(aio_worker, acb, cb, opaque);
+}
+
 static BlockDriverAIOCB *raw_aio_submit(BlockDriverState *bs,
         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
         BlockDriverCompletionFunc *cb, void *opaque, int type)