[Libguestfs] [PATCH 4/5] fuse: Fix read for empty files.

[Jim Meyering]

Matthew Booth wrote:
> On 17/11/09 17:07, Richard W.M. Jones wrote:
>>> From 3c5d63be17d281634f036f53aa32103ad4ab1c41 Mon Sep 17 00:00:00 2001
>> From: Richard Jones<rjones at redhat.com>
>> Date: Tue, 17 Nov 2009 17:02:45 +0000
>> Subject: [PATCH 4/5] fuse: Fix read for empty files.
>>
>> Error handling for the guestfs_pread call was incorrect, which
>> meant that empty files could produce spurious error messages.
>> ---
>>   fuse/guestmount.c |    8 +++++++-
>>   1 files changed, 7 insertions(+), 1 deletions(-)
>>
>> diff --git a/fuse/guestmount.c b/fuse/guestmount.c
>> index 9d16cef..669bf80 100644
>> --- a/fuse/guestmount.c
>> +++ b/fuse/guestmount.c
>> @@ -621,8 +621,14 @@ fg_read (const char *path, char *buf, size_t size, off_t offset,
>>     if (size>  limit)
>>       size = limit;
>>
>> +  /* Note the correct error handling here is tricky, because in the
>> +   * case where the call returns a zero-length buffer, it might return
>> +   * NULL.  However it won't adjust rsize along the error path, so we
>> +   * can set rsize to something beforehand and use that as a flag.
>> +   */
>> +  rsize = 1;
>>     r = guestfs_pread (g, path, size, offset,&rsize);
>> -  if (r == NULL)
>> +  if (rsize == 1&&  r == NULL)
>>       return error ();
>>
>>     /* This should never happen, but at least it stops us overflowing
>> -- 1.6.5.2
>
> I've stared at this til my head hurts, and I don't get it. Why is 1
> not a valid returned size from guestfs_pread? Why is NULL not an
> indicator of failure?

1 is a valid result size, but the only way to get
rsize of 1 (assuming the prior initialization) and
*also* r == NULL is in this failure case.

>> +  if (rsize == 1&&  r == NULL)
(oops, spacing: should be
      if (rsize == 1 && r == NULL)

> The documentation says:
>
>        This function returns a buffer, or NULL on error.  The size of the
>        returned buffer is written to *size_r.  The caller must free the
>        returned buffer after use.
>
> If NULL isn't sufficient to indicate error, guestfs_pread needs to be
> fixed, not this funtion.

Right.  The API is too hard to use properly.

An alternative is to make it so guestfs_pread returns a non-NULL
pointer to one malloc'd (and unused) byte for an empty read,
and also sets *rsize to 0.  *THEN* a NULL return value could
suffice to indicate failure.