Re: [Qemu-devel] [PATCH 01/10] util, qga: drop guest_file_toggle_flags

[Michael Roth <mdroth@linux.vnet.ibm.com>]

Quoting Denis V. Lunev (2015-07-08 17:40:30)
> On 09/07/15 00:16, Michael Roth wrote:
> > Quoting Denis V. Lunev (2015-06-30 05:25:14)
> >> From: Olga Krishtal <okrishtal@virtuozzo.com>
> >>
> >> guest_file_toggle_flags is a copy from semi-portable qemu_set_nonblock.
> >> The latter is not working properly for Windows due to reduced Windows
> >> Posix implementation.
> >>
> >> On Windows OS there is a separate API for changing flags of file, pipes
> >> and sockets. Portable way to change file descriptor flags requires
> >> to detect file descriptor type and proper actions depending of that
> >> type. The patch adds wrapper qemu_set_fd_nonblocking into Windows specific
> >> code to handle this stuff properly.
> >>
> >> The only problem is that qemu_set_nonblock is void but this should not
> >> be a problem.
> >>
> >> Signed-off-by: Olga Krishtal <okrishtal@virtuozzo.com>
> >> Signed-off-by: Denis V. Lunev <den@openvz.org>
> >> CC: Eric Blake <eblake@redhat.com>
> >> CC: Michael Roth <mdroth@linux.vnet.ibm.com>
> >> ---
> >>   qga/commands-posix.c | 27 ++-------------------------
> >>   util/oslib-win32.c   | 52 +++++++++++++++++++++++++++++++++++++++++++++++-----
> >>   2 files changed, 49 insertions(+), 30 deletions(-)
> >>
> >> diff --git a/qga/commands-posix.c b/qga/commands-posix.c
> >> index befd00b..40dbe25 100644
> >> --- a/qga/commands-posix.c
> >> +++ b/qga/commands-posix.c
> >> @@ -28,6 +28,7 @@
> >>   #include "qapi/qmp/qerror.h"
> >>   #include "qemu/queue.h"
> >>   #include "qemu/host-utils.h"
> >> +#include "qemu/sockets.h"
> >>
> >>   #ifndef CONFIG_HAS_ENVIRON
> >>   #ifdef __APPLE__
> >> @@ -376,27 +377,6 @@ safe_open_or_create(const char *path, const char *mode, Error **errp)
> >>       return NULL;
> >>   }
> >>
> >> -static int guest_file_toggle_flags(int fd, int flags, bool set, Error **err)
> >> -{
> >> -    int ret, old_flags;
> >> -
> >> -    old_flags = fcntl(fd, F_GETFL);
> >> -    if (old_flags == -1) {
> >> -        error_setg_errno(err, errno, QERR_QGA_COMMAND_FAILED,
> >> -                         "failed to fetch filehandle flags");
> >> -        return -1;
> >> -    }
> >> -
> >> -    ret = fcntl(fd, F_SETFL, set ? (old_flags | flags) : (old_flags & ~flags));
> >> -    if (ret == -1) {
> >> -        error_setg_errno(err, errno, QERR_QGA_COMMAND_FAILED,
> >> -                         "failed to set filehandle flags");
> >> -        return -1;
> >> -    }
> >> -
> >> -    return ret;
> >> -}
> >> -
> >>   int64_t qmp_guest_file_open(const char *path, bool has_mode, const char *mode,
> >>                               Error **errp)
> >>   {
> >> @@ -417,10 +397,7 @@ int64_t qmp_guest_file_open(const char *path, bool has_mode, const char *mode,
> >>       /* set fd non-blocking to avoid common use cases (like reading from a
> >>        * named pipe) from hanging the agent
> >>        */
> >> -    if (guest_file_toggle_flags(fileno(fh), O_NONBLOCK, true, errp) < 0) {
> >> -        fclose(fh);
> >> -        return -1;
> >> -    }
> >> +    qemu_set_nonblock(fileno(fh));
> >>
> >>       handle = guest_file_handle_add(fh, errp);
> >>       if (handle < 0) {
> >> diff --git a/util/oslib-win32.c b/util/oslib-win32.c
> >> index 730a670..1a6ae72 100644
> >> --- a/util/oslib-win32.c
> >> +++ b/util/oslib-win32.c
> >> @@ -119,17 +119,59 @@ struct tm *localtime_r(const time_t *timep, struct tm *result)
> >>       return p;
> >>   }
> >>
> >> -void qemu_set_block(int fd)
> >> +static void qemu_set_fd_nonblocking(int fd, bool nonblocking)
> >>   {
> >> -    unsigned long opt = 0;
> >> -    WSAEventSelect(fd, NULL, 0);
> >> +    HANDLE handle;
> >> +    DWORD file_type, pipe_state;
> >> +
> >> +    handle = (HANDLE)_get_osfhandle(fd);
> >> +    if (handle == INVALID_HANDLE_VALUE) {
> >> +        return;
> >> +    }
> >> +
> >> +    file_type = GetFileType(handle);
> >> +    if (file_type != FILE_TYPE_PIPE) {
> >> +        return;
> >> +    }
> >> +
> >> +    /* If file_type == FILE_TYPE_PIPE, according to msdn
> >> +     * the specified file is socket or named pipe */
> >> +    if (GetNamedPipeHandleState(handle, &pipe_state, NULL,
> >> +                                NULL, NULL, NULL, 0)) {
> >> +        /* The fd is named pipe fd */
> >> +        if (!nonblocking == !(pipe_state & PIPE_NOWAIT)) {
> >> +            /* In this case we do not need perform any operation, because
> >> +             * nonblocking = true and PIPE_NOWAIT is already set or
> >> +             * nonblocking = false and PIPE_NOWAIT is not set */
> >> +            return;
> >> +        }
> >> +
> >> +        if (nonblocking) {
> >> +            pipe_state |= PIPE_NOWAIT;
> >> +        } else {
> >> +            pipe_state &= ~PIPE_NOWAIT;
> >> +        }
> >> +
> >> +        SetNamedPipeHandleState(handle, &pipe_state, NULL, NULL);
> >> +        return;
> >> +    }
> >> +
> >> +    /* The fd is socket fd */
> >> +    unsigned long opt = (unsigned long)nonblocking;
> >> +    if (!nonblocking) {
> >> +        WSAEventSelect(fd, NULL, 0);
> >> +    }
> >>       ioctlsocket(fd, FIONBIO, &opt);
> >>   }
> >>
> >> +void qemu_set_block(int fd)
> >> +{
> >> +    qemu_set_fd_nonblocking(fd, false);
> >> +}
> >> +
> >>   void qemu_set_nonblock(int fd)
> >>   {
> >> -    unsigned long opt = 1;
> >> -    ioctlsocket(fd, FIONBIO, &opt);
> >> +    qemu_set_fd_nonblocking(fd, true);
> >>       qemu_fd_register(fd);
> > We wouldn't want to pass a non-socket FD to qemu_fd_register(), so
> > this should get moved to qemu_set_fd_nonblocking() at least.
> >
> > I think it's confusing to have a qemu_set_nonblock() that's limited
> > to pipes and sockets. At least with the current code it's fairly
> > obvious it's intended for sockets.
> >
> > It might be worthwhile if we could at least cover disk files, but
> > I seem to recall that being non-trivial on w32 and generally
> > requiring non-posix functions that support overlapped IO. So, if
> > we're only extending this to support pipes it should probably just
> > be a separate helper in qga/.
> >
> at my opinion you are a little bit wrong here. First of all, from
> Linux point of view, O_NONBLOCK for ordinary file is noop, i.e.
> there is no difference in behavior with and without this flag
> for such descriptors.

Yes, I think you're right... I'd forgotten that the reason we use it
in guest-file-open was for dealing with pipes/chardevs/sockets rather
than anything to do with files.

> 
> This means that the write will take some time and will finish
> without indefinite timeout. Sockets and pipes are different,
> they can hang in read/write forever and thus the flag has
> some value to avoid this.
> 
> The same applies for Windows platform too. Setting nonblock
> here we want to avoid infinite hang and we do not want to
> implement asynchronous IO, which is implemented using
> overlapped API.
> 
> There are only 2 primitives in QGA which subjects to the rules
> above - sockets and pipes. Thus this common helper makes
> sense. If we will need more primitives - we will add proper
> distinction into these calls.

There's still an important invariant here in that O_NONBLOCK ensures
EAGAIN in cases where you might otherwise wait indefinitely. That
covers named/anonymous pipes, sockets, files, chardevs, any fd.
To take the socket-only w32 implementation and only tack on pipes
seems somewhat arbitrary in comparison, but I do understand your
logic from a "FDs that make sense to use it with" standpoint. I think
you'll need to Cc: our w32 maintainer Stefan Weil <sw@weilnetz.de>
for any such change in oslib-win32.c though.

Personally I think set_pipe_nonblock() in qga/commands-win32.c is
clearer, and it seems unlikely we'll ever end up with a 2nd w32
user of it so we don't stand to gain much from a generalized
function.

> 
> Den
>