Re: cvs: php4 /ext/standard file.c

Hi! >>>>> "Zeev" == Zeev Suraski <zeev@zend.com> writes: Zeev> At 14:43 23/12/2000, Michael Widenius wrote: >> >> This code is a good example of the write() API problem. If >> >> write() does not return Count, the code assumes that errno >> >> was set by write(). If errno was not set by write(), the >> >> code will simply take the error code from a previous system >> >> call and reinterprete it in this context. That is pretty >> >> much wrong, considering the following three paragraphs: >> >> >> >> The above is not really right; If you check the code of my_write(), >> it only uses errno if write() returns 0 or -1, in which case >> errno should contain the right error code. Zeev> I can't see how it does that... What if Count equals 1024, but write() Zeev> returns only 512? It won't reach the 'break' statement, and will fall Zeev> through and use errno. If it falls through, it will just schedule another write without using errno (except in the case where write() returned 0 or -1, in which case it's ok to check errno) Zeev> Sascha's main concern is this: Zeev> (a) according to the documentation, errno is only set if write() returns -1 Zeev> - a 'complete' error code. In most systems it's set if write() returns -1 or 0; Note that according to the specification write() should never return 0 except if Count is 0, but in practice this happens! Zeev> (b) when reaching the end of the disk space, we may have a situation where Zeev> a write() will succeed partially, i.e., it will write bytes, but not as Zeev> many as we asked. In this case, write() will return a positive integer, Zeev> that is lower than the count we sent it. No, it will not; If write() returns a lower number my_write() will return with another write() until this returns 0 or -1 Zeev> The big question is - whether Zeev> errno will be set or not. The documentation says nothing about it. It's Zeev> extremely odd if it won't be set though, because it would then essentially Zeev> mean that there's no way to properly handle this situation. It will not be set :( (I agree that this is a stupid specification) On the other hand, my_write() handles this case without any problems! To satisfy Sascha, I changed the test of ENOSCP to: if (my_errno == ENOSPC && (MyFlags & MY_WAIT_IF_FULL) && (uint) writenbytes != (uint) -1) this will take care of a theoretical bug when errno contained ENOSPC from an old call. Just note that the old code worked, even if it was not optimal in this case. Zeev> Question is - do you have real world experience with this? I know that my_write() has worked perfectly for us during the last 10 years, with and without threads... Regards, Monty