Re: [BUG] drivers/tty: read() on a noncanonical blocking tty randomly fails when VMIN > received >= buf

[Peter Hurley <peter@hurleysoftware.com>]

On 05/05/2016 03:08 AM, One Thousand Gnomes wrote:
> On Wed, 4 May 2016 16:07:44 -0700
> Peter Hurley <peter@hurleysoftware.com> wrote:
> 
>> Hi Julio,
>>
>> On 05/04/2016 04:00 PM, Julio Guerra wrote:
>>> Hi,
>>>
>>> When a tty (here a slave pty) is set in noncanonical input and blocking read modes, a read() randomly blocks when:
>>> "VMIN > kernel received >= user buffer size > 0".
>>>
>>> The standard says that read() should block until VMIN bytes are received [1][2]. Whether this is an implementation defined case not really specified by POSIX or not, it should not behave randomly (otherwise it really should be documented in termios manpage).  
>>
>> This is not a bug.
>>
>> >From the termios(3) man page:  
>>
>>        * MIN > 0; TIME == 0: read(2) blocks until the lesser of MIN bytes or the number of bytes requested are avail‐
>>          able, and returns the lesser of these two values.
> 
> The standard says
> 
> 	Case B: MIN>0, TIME=0
> 
> 	In case B, since the value of TIME is zero, the timer plays no
> 	role and only MIN is significant. A pending read shall not be
> 	satisfied until MIN bytes are received (that is, the pending read
> 	shall block until MIN bytes are received), or a signal is
> 	received. A program that uses case B to read record-based
> 	terminal I/O may block indefinitely in the read operation.
> 
> That is if you do 
> 
> 
> 	read(fd, buf, 3)
> 
> and MIN is 5, the read should not return until there are 5 bytes in the
> queue. The following code is guaranteed to work reliably by the standard
> with TIME 0 MIN 5 (ignoring signals for the moment)
> 
> 
> 	read(fd, buf, 3);
> 	fcntl(fd, F_SETFL, FNDELAY);
> 	assert(read(fd, buf, 2) == 2);
> 
> Historically this behaviour was useful for things like block transfer
> protocols, especially with offloaded serial processing.
> 
> So actually I think we do have a bug, the behaviuour is not standards
> compliant, and the man page documents the erroneous behaviour.

I disagree; I think SUSv4 fails to address this degenerate condition at all.
For example, SUSv4 specifically states that there is no precedence of
MIN/TIME with O_NONBLOCK. IOW, the standard does _not_ guarantee that your
code fragment above won't block on the subsequent read anyway since it fails
to meet the new MIN 5 watermark.

But I have no problem fixing a bona fide regression; what's broken?