Patchwork [6/6] qemu-file: do not use stdio for qemu_fdopen

login
register
mail settings
Submitter Paolo Bonzini
Date March 27, 2013, 4:36 p.m.
Message ID <1364402192-18169-7-git-send-email-pbonzini@redhat.com>
Download mbox | patch
Permalink /patch/231760/
State New
Headers show

Comments

Paolo Bonzini - March 27, 2013, 4:36 p.m.
This uses system calls directly for Unix file descriptors, so that the
efficient writev_buffer can be used.  Pay attention to the possibility
of partial writes in writev.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 savevm.c | 100 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 89 insertions(+), 11 deletions(-)
Juan Quintela - April 9, 2013, 1:30 p.m.
Paolo Bonzini <pbonzini@redhat.com> wrote:
> This uses system calls directly for Unix file descriptors, so that the
> efficient writev_buffer can be used.  Pay attention to the possibility
> of partial writes in writev.
>
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Reviewed-by: Juan Quintela <quintela@redhat.com>

> +static ssize_t unix_writev_buffer(void *opaque, struct iovec *iov, int iovcnt)
> +{
> +    QEMUFileSocket *s = opaque;
> +    ssize_t len, offset;
> +    ssize_t size = iov_size(iov, iovcnt);
> +    ssize_t total = 0;
> +
> +    assert(iovcnt > 0);
> +    offset = 0;
> +    while (size > 0) {
> +        /* Find the next start position; skip all full-sized vector elements  */
> +        while (offset >= iov[0].iov_len) {
> +            offset -= iov[0].iov_len;
> +            iov++, iovcnt--;
> +        }
> +
> +        /* skip `offset' bytes from the (now) first element, undo it on exit */
> +        assert(iovcnt > 0);
> +        iov[0].iov_base += offset;
> +        iov[0].iov_len -= offset;
> +
> +        do {
> +            len = writev(s->fd, iov, iovcnt);
> +        } while (len == -1 && errno == EINTR);
> +        if (len == -1) {
> +            return -errno;
> +        }
> +
> +        /* Undo the changes above */
> +        iov[0].iov_base -= offset;
> +        iov[0].iov_len += offset;
> +
> +        /* Prepare for the next iteration */
> +        offset += len;
> +        total += len;
> +        size -= len;
> +    }
> +
> +    return total;
> +}

This code is very similar to the one in the iov_send_recv(),  but I
can't think on a trivial way to share it :p

Later,  Juan.
Anthony Liguori - April 16, 2013, 9:09 p.m.
Paolo Bonzini <pbonzini@redhat.com> writes:

> This uses system calls directly for Unix file descriptors, so that the
> efficient writev_buffer can be used.  Pay attention to the possibility
> of partial writes in writev.
>
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Needs rebasing:

  CC    i386-softmmu/savevm.o
/home/aliguori/git/qemu/savevm.c:439:5: error: initialization from incompatible pointer type [-Werror]
/home/aliguori/git/qemu/savevm.c:439:5: error: (near initialization for ‘unix_write_ops.writev_buffer’) [-Werror]
cc1: all warnings being treated as errors
make[1]: *** [savevm.o] Error 1
make: *** [subdir-i386-softmmu] Error 2
d13ed415c0ab9f7cf29226022f8c24d1b8dc794d is the first bad commit
commit d13ed415c0ab9f7cf29226022f8c24d1b8dc794d
Author: Paolo Bonzini <pbonzini@redhat.com>
Date:   Wed Mar 27 17:36:32 2013 +0100

    qemu-file: do not use stdio for qemu_fdopen

Regards,

Anthony Liguori

> ---
>  savevm.c | 100 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++-------
>  1 file changed, 89 insertions(+), 11 deletions(-)
>
> diff --git a/savevm.c b/savevm.c
> index 0415830..8eb5aab 100644
> --- a/savevm.c
> +++ b/savevm.c
> @@ -356,9 +356,93 @@ static const QEMUFileOps stdio_file_write_ops = {
>      .close =      stdio_fclose
>  };
>  
> +static ssize_t unix_writev_buffer(void *opaque, struct iovec *iov, int iovcnt)
> +{
> +    QEMUFileSocket *s = opaque;
> +    ssize_t len, offset;
> +    ssize_t size = iov_size(iov, iovcnt);
> +    ssize_t total = 0;
> +
> +    assert(iovcnt > 0);
> +    offset = 0;
> +    while (size > 0) {
> +        /* Find the next start position; skip all full-sized vector elements  */
> +        while (offset >= iov[0].iov_len) {
> +            offset -= iov[0].iov_len;
> +            iov++, iovcnt--;
> +        }
> +
> +        /* skip `offset' bytes from the (now) first element, undo it on exit */
> +        assert(iovcnt > 0);
> +        iov[0].iov_base += offset;
> +        iov[0].iov_len -= offset;
> +
> +        do {
> +            len = writev(s->fd, iov, iovcnt);
> +        } while (len == -1 && errno == EINTR);
> +        if (len == -1) {
> +            return -errno;
> +        }
> +
> +        /* Undo the changes above */
> +        iov[0].iov_base -= offset;
> +        iov[0].iov_len += offset;
> +
> +        /* Prepare for the next iteration */
> +        offset += len;
> +        total += len;
> +        size -= len;
> +    }
> +
> +    return total;
> +}
> +
> +static int unix_get_buffer(void *opaque, uint8_t *buf, int64_t pos, int size)
> +{
> +    QEMUFileSocket *s = opaque;
> +    ssize_t len;
> +
> +    for (;;) {
> +        len = read(s->fd, buf, size);
> +        if (len != -1) {
> +            break;
> +        }
> +        if (errno == EAGAIN) {
> +            yield_until_fd_readable(s->fd);
> +        } else if (errno != EINTR) {
> +            break;
> +        }
> +    }
> +
> +    if (len == -1) {
> +        len = -errno;
> +    }
> +    return len;
> +}
> +
> +static int unix_close(void *opaque)
> +{
> +    QEMUFileSocket *s = opaque;
> +    close(s->fd);
> +    g_free(s);
> +    return 0;
> +}
> +
> +static const QEMUFileOps unix_read_ops = {
> +    .get_fd =     socket_get_fd,
> +    .get_buffer = unix_get_buffer,
> +    .close =      unix_close
> +};
> +
> +static const QEMUFileOps unix_write_ops = {
> +    .get_fd =     socket_get_fd,
> +    .writev_buffer = unix_writev_buffer,
> +    .close =      unix_close
> +};
> +
>  QEMUFile *qemu_fdopen(int fd, const char *mode)
>  {
> -    QEMUFileStdio *s;
> +    QEMUFileSocket *s;
>  
>      if (mode == NULL ||
>  	(mode[0] != 'r' && mode[0] != 'w') ||
> @@ -367,21 +451,15 @@ QEMUFile *qemu_fdopen(int fd, const char *mode)
>          return NULL;
>      }
>  
> -    s = g_malloc0(sizeof(QEMUFileStdio));
> -    s->stdio_file = fdopen(fd, mode);
> -    if (!s->stdio_file)
> -        goto fail;
> +    s = g_malloc0(sizeof(QEMUFileSocket));
> +    s->fd = fd;
>  
>      if(mode[0] == 'r') {
> -        s->file = qemu_fopen_ops(s, &stdio_file_read_ops);
> +        s->file = qemu_fopen_ops(s, &unix_read_ops);
>      } else {
> -        s->file = qemu_fopen_ops(s, &stdio_file_write_ops);
> +        s->file = qemu_fopen_ops(s, &unix_write_ops);
>      }
>      return s->file;
> -
> -fail:
> -    g_free(s);
> -    return NULL;
>  }
>  
>  static const QEMUFileOps socket_read_ops = {
> -- 
> 1.8.1.4

Patch

diff --git a/savevm.c b/savevm.c
index 0415830..8eb5aab 100644
--- a/savevm.c
+++ b/savevm.c
@@ -356,9 +356,93 @@  static const QEMUFileOps stdio_file_write_ops = {
     .close =      stdio_fclose
 };
 
+static ssize_t unix_writev_buffer(void *opaque, struct iovec *iov, int iovcnt)
+{
+    QEMUFileSocket *s = opaque;
+    ssize_t len, offset;
+    ssize_t size = iov_size(iov, iovcnt);
+    ssize_t total = 0;
+
+    assert(iovcnt > 0);
+    offset = 0;
+    while (size > 0) {
+        /* Find the next start position; skip all full-sized vector elements  */
+        while (offset >= iov[0].iov_len) {
+            offset -= iov[0].iov_len;
+            iov++, iovcnt--;
+        }
+
+        /* skip `offset' bytes from the (now) first element, undo it on exit */
+        assert(iovcnt > 0);
+        iov[0].iov_base += offset;
+        iov[0].iov_len -= offset;
+
+        do {
+            len = writev(s->fd, iov, iovcnt);
+        } while (len == -1 && errno == EINTR);
+        if (len == -1) {
+            return -errno;
+        }
+
+        /* Undo the changes above */
+        iov[0].iov_base -= offset;
+        iov[0].iov_len += offset;
+
+        /* Prepare for the next iteration */
+        offset += len;
+        total += len;
+        size -= len;
+    }
+
+    return total;
+}
+
+static int unix_get_buffer(void *opaque, uint8_t *buf, int64_t pos, int size)
+{
+    QEMUFileSocket *s = opaque;
+    ssize_t len;
+
+    for (;;) {
+        len = read(s->fd, buf, size);
+        if (len != -1) {
+            break;
+        }
+        if (errno == EAGAIN) {
+            yield_until_fd_readable(s->fd);
+        } else if (errno != EINTR) {
+            break;
+        }
+    }
+
+    if (len == -1) {
+        len = -errno;
+    }
+    return len;
+}
+
+static int unix_close(void *opaque)
+{
+    QEMUFileSocket *s = opaque;
+    close(s->fd);
+    g_free(s);
+    return 0;
+}
+
+static const QEMUFileOps unix_read_ops = {
+    .get_fd =     socket_get_fd,
+    .get_buffer = unix_get_buffer,
+    .close =      unix_close
+};
+
+static const QEMUFileOps unix_write_ops = {
+    .get_fd =     socket_get_fd,
+    .writev_buffer = unix_writev_buffer,
+    .close =      unix_close
+};
+
 QEMUFile *qemu_fdopen(int fd, const char *mode)
 {
-    QEMUFileStdio *s;
+    QEMUFileSocket *s;
 
     if (mode == NULL ||
 	(mode[0] != 'r' && mode[0] != 'w') ||
@@ -367,21 +451,15 @@  QEMUFile *qemu_fdopen(int fd, const char *mode)
         return NULL;
     }
 
-    s = g_malloc0(sizeof(QEMUFileStdio));
-    s->stdio_file = fdopen(fd, mode);
-    if (!s->stdio_file)
-        goto fail;
+    s = g_malloc0(sizeof(QEMUFileSocket));
+    s->fd = fd;
 
     if(mode[0] == 'r') {
-        s->file = qemu_fopen_ops(s, &stdio_file_read_ops);
+        s->file = qemu_fopen_ops(s, &unix_read_ops);
     } else {
-        s->file = qemu_fopen_ops(s, &stdio_file_write_ops);
+        s->file = qemu_fopen_ops(s, &unix_write_ops);
     }
     return s->file;
-
-fail:
-    g_free(s);
-    return NULL;
 }
 
 static const QEMUFileOps socket_read_ops = {