Portability of apr_file_writev_full?

[Joe Schaefer <jo...@sunstarsys.com>]

apreq has a similar function, and some questions
came up yesterday regarding its implementation. So
I took a look at the recent apr implementation,
and I have a few questions of my own regarding this loop:

    do {
        apr_size_t i, amt;
        status = apr_file_writev(thefile, vec, nvec, &amt);
 
       /* We assume that writev will only write complete iovec areas.
        * Incomplete writes inside a single area are not supported.
        * This should be safe according to SuS v2. 
        */
        for (i = 0; i < nvec; i++) {
            total += vec[i].iov_len;
            if (total >= amt) {
                vec = &vec[i+1];
                nvec -= i+1;
                break;
            }
        }
    } while (status == APR_SUCCESS && nvec > 0);


First, this is the only implementation I could find in apr,
so I assume its being used on non-unix platforms where SuS v2
does not apply, particularly wrt writev() itself.

Second, I disagree that SuS v2 makes any guarantees about writev()
only writing "complete iovec areas".  I assume a conforming writev() 
implementation can hand the single-iovec case off to write().  So if 
there's more data in the iovec than the target device can accept,
it winds up doing a partial write, successfully.  

IMO the statement in SuS: "the writev() function shall always write a 
complete area before proceeding to the next" should be interpreted
conservatively: the implementation isn't allowed to move on to the
next iovec, unless it can write all of the current one.

-- 
Joe Schaefer

Re: Portability of apr_file_writev_full?

[Paul Querna <ch...@force-elite.com>]

Joe Orton wrote:
> On Thu, Mar 10, 2005 at 10:01:37AM -0500, Joe Schaefer wrote:
> ...
> 
>>Second, I disagree that SuS v2 makes any guarantees about writev()
>>only writing "complete iovec areas".  I assume a conforming writev() 
>>implementation can hand the single-iovec case off to write().  So if 
>>there's more data in the iovec than the target device can accept,
>>it winds up doing a partial write, successfully.  
> 
> 
> Yes it's broken as you indicate, I pointed this out when it was
> committed... this had dropped off my radar at least.  It needs to call
> apr_file_write_full to complete writing any partially-written buffer, or
> something like that.

Yep, my fault.  It dropped off my radar too.  I will try to fix it this 
weekend.

-Paul

Re: Portability of apr_file_writev_full?

[Joe Orton <jo...@redhat.com>]

On Thu, Mar 10, 2005 at 10:01:37AM -0500, Joe Schaefer wrote:
...
> Second, I disagree that SuS v2 makes any guarantees about writev()
> only writing "complete iovec areas".  I assume a conforming writev() 
> implementation can hand the single-iovec case off to write().  So if 
> there's more data in the iovec than the target device can accept,
> it winds up doing a partial write, successfully.  

Yes it's broken as you indicate, I pointed this out when it was
committed... this had dropped off my radar at least.  It needs to call
apr_file_write_full to complete writing any partially-written buffer, or
something like that.

joe