Re: [PATCH v2 2/7] virtiofds: Changed allocations of iovec to GLib's functions

[Mahmoud Mandour <ma.mandourr@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 7784 bytes --]

On Tue, Apr 27, 2021 at 12:25 PM Dr. David Alan Gilbert <dgilbert@redhat.com>
wrote:

> * Mahmoud Mandour (ma.mandourr@gmail.com) wrote:
> > Replaced the calls to malloc()/calloc() and their respective
> > calls to free() of iovec structs with GLib's allocation and
> > deallocation functions.
> >
> > Also, in one instance, used g_new0() instead of a calloc() call plus
> > a null-checking assertion.
> >
> > iovec structs were created locally and freed as the function
> > ends. Hence, I used g_autofree and removed the respective calls to
> > free().
> >
> > In one instance, a struct fuse_ioctl_iovec pointer is returned from a
> > function, namely, fuse_ioctl_iovec_copy. There, I used g_steal_pointer()
> > in conjunction with g_autofree, this gives the ownership of the pointer
> > to the calling function and still auto-frees the memory when the calling
> > function finishes (maintaining the symantics of previous code).
> >
> > Signed-off-by: Mahmoud Mandour <ma.mandourr@gmail.com>
> > Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> >  tools/virtiofsd/fuse_lowlevel.c | 19 +++++++------------
> >  tools/virtiofsd/fuse_virtio.c   |  6 +-----
> >  2 files changed, 8 insertions(+), 17 deletions(-)
> >
> > diff --git a/tools/virtiofsd/fuse_lowlevel.c
> b/tools/virtiofsd/fuse_lowlevel.c
> > index 812cef6ef6..f965299ad9 100644
> > --- a/tools/virtiofsd/fuse_lowlevel.c
> > +++ b/tools/virtiofsd/fuse_lowlevel.c
> > @@ -217,9 +217,9 @@ static int send_reply(fuse_req_t req, int error,
> const void *arg,
> >  int fuse_reply_iov(fuse_req_t req, const struct iovec *iov, int count)
> >  {
> >      int res;
> > -    struct iovec *padded_iov;
> > +    g_autofree struct iovec *padded_iov;
> >
> > -    padded_iov = malloc((count + 1) * sizeof(struct iovec));
> > +    padded_iov = g_try_new(struct iovec, count + 1);
> >      if (padded_iov == NULL) {
> >          return fuse_reply_err(req, ENOMEM);
> >      }
> > @@ -228,7 +228,6 @@ int fuse_reply_iov(fuse_req_t req, const struct
> iovec *iov, int count)
> >      count++;
> >
> >      res = send_reply_iov(req, 0, padded_iov, count);
> > -    free(padded_iov);
> >
> >      return res;
> >  }
>
> OK.
>
> > @@ -565,10 +564,10 @@ int fuse_reply_bmap(fuse_req_t req, uint64_t idx)
> >  static struct fuse_ioctl_iovec *fuse_ioctl_iovec_copy(const struct
> iovec *iov,
> >                                                        size_t count)
> >  {
> > -    struct fuse_ioctl_iovec *fiov;
> > +    g_autofree struct fuse_ioctl_iovec *fiov;
> >      size_t i;
> >
> > -    fiov = malloc(sizeof(fiov[0]) * count);
> > +    fiov = g_try_new(fuse_ioctl_iovec, count);
> >      if (!fiov) {
> >          return NULL;
> >      }
> > @@ -578,7 +577,7 @@ static struct fuse_ioctl_iovec
> *fuse_ioctl_iovec_copy(const struct iovec *iov,
> >          fiov[i].len = iov[i].iov_len;
> >      }
> >
> > -    return fiov;
> > +    return g_steal_pointer(&fiov);
> >  }
>
> This is OK, but doesn't gain anything - marking it as g_autofree'ing and
> always stealing is no benefit.
>
> >
> >  int fuse_reply_ioctl_retry(fuse_req_t req, const struct iovec *in_iov,
> > @@ -629,9 +628,6 @@ int fuse_reply_ioctl_retry(fuse_req_t req, const
> struct iovec *in_iov,
> >
> >      res = send_reply_iov(req, 0, iov, count);
> >  out:
> > -    free(in_fiov);
> > -    free(out_fiov);
> > -
>
> I don't think you can do that - I think you're relying here on the
> g_autofree from fuse_ioclt_iovec_copy - but my understanding is that
> doesn't work; g_autofree is scoped, so it's designed to free at the end
> of fuse_ioctl_iovec_copy, fuse_reply_ioctl_retry doesn't know that the
> ion_fiov were allocated that way, so it won't get autocleaned up.
>
>
In GLib's documentation, it is clarified (w.r.t. g_autoptr but I think
similar logic applies to g_autofree)
that g_steal_pointer() "This can be very useful when combined with
g_autoptr() to prevent
the return value of a function from being automatically freed."
I think, but not 100% clear of course, that this means that the
g_autoptr-annotated memory
does not get freed at the end of the current scope, and  its "scope" is
migrated to the calling
function(to be honest I don't know how would they implement that but maybe
this is the case).
Otherwise why bother with g_autoptr'ing memory that we don't want to free
automatically and
would like to return to the calling function?

The first example in Memory Allocation: GLib Reference Manual (gnome.org)
<https://developer.gnome.org/glib/stable/glib-Memory-Allocation.html#g-steal-pointer>
does
annotate
the memory as g_autoptr and then returns it through g_steal_pointer. With
your logic, I think that
this example would be wrong(?)

Mr. Hajnoczi already reviewed this patch  Re: [PATCH 2/8] virtiofds:
Changed allocations of iovec to GLib's functi
<https://lists.gnu.org/archive/html/qemu-devel/2021-03/msg08459.html>
in a previous version and this v2 patch series is supposed to only contain
already-reviewed patches and
remove bad ones


> >      return res;
> >
> >  enomem:
> > @@ -663,11 +659,11 @@ int fuse_reply_ioctl(fuse_req_t req, int result,
> const void *buf, size_t size)
> >  int fuse_reply_ioctl_iov(fuse_req_t req, int result, const struct iovec
> *iov,
> >                           int count)
> >  {
> > -    struct iovec *padded_iov;
> > +    g_autofree struct iovec *padded_iov;
> >      struct fuse_ioctl_out arg;
> >      int res;
> >
> > -    padded_iov = malloc((count + 2) * sizeof(struct iovec));
> > +    padded_iov = g_try_new(struct iovec, count + 2);
> >      if (padded_iov == NULL) {
> >          return fuse_reply_err(req, ENOMEM);
> >      }
> > @@ -680,7 +676,6 @@ int fuse_reply_ioctl_iov(fuse_req_t req, int result,
> const struct iovec *iov,
> >      memcpy(&padded_iov[2], iov, count * sizeof(struct iovec));
> >
> >      res = send_reply_iov(req, 0, padded_iov, count + 2);
> > -    free(padded_iov);
> >
> >      return res;
> >  }
>
> OK
>
> > diff --git a/tools/virtiofsd/fuse_virtio.c
> b/tools/virtiofsd/fuse_virtio.c
> > index 3e13997406..07e5d91a9f 100644
> > --- a/tools/virtiofsd/fuse_virtio.c
> > +++ b/tools/virtiofsd/fuse_virtio.c
> > @@ -347,8 +347,7 @@ int virtio_send_data_iov(struct fuse_session *se,
> struct fuse_chan *ch,
> >       * Build a copy of the the in_sg iov so we can skip bits in it,
> >       * including changing the offsets
> >       */
> > -    struct iovec *in_sg_cpy = calloc(sizeof(struct iovec), in_num);
> > -    assert(in_sg_cpy);
> > +    g_autofree struct iovec *in_sg_cpy = g_new0(struct iovec, in_num);
> >      memcpy(in_sg_cpy, in_sg, sizeof(struct iovec) * in_num);
> >      /* These get updated as we skip */
> >      struct iovec *in_sg_ptr = in_sg_cpy;
> > @@ -386,7 +385,6 @@ int virtio_send_data_iov(struct fuse_session *se,
> struct fuse_chan *ch,
> >              ret = errno;
> >              fuse_log(FUSE_LOG_DEBUG, "%s: preadv failed (%m) len=%zd\n",
> >                       __func__, len);
> > -            free(in_sg_cpy);
> >              goto err;
> >          }
> >          fuse_log(FUSE_LOG_DEBUG, "%s: preadv ret=%d len=%zd\n",
> __func__,
> > @@ -410,13 +408,11 @@ int virtio_send_data_iov(struct fuse_session *se,
> struct fuse_chan *ch,
> >          if (ret != len) {
> >              fuse_log(FUSE_LOG_DEBUG, "%s: ret!=len\n", __func__);
> >              ret = EIO;
> > -            free(in_sg_cpy);
> >              goto err;
> >          }
> >          in_sg_left -= ret;
> >          len -= ret;
> >      } while (in_sg_left);
> > -    free(in_sg_cpy);
>
> Yes, this is where the autofree really helps; getting rid of a few
> free's.
>
> Dave
>
> >      /* Need to fix out->len on EOF */
> >      if (len) {
> > --
> > 2.25.1
> >
> --
> Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
>
>
Thanks,
Mahmoud

[-- Attachment #2: Type: text/html, Size: 10162 bytes --]