Re: closefd: closes a file of any process

Tigran Aivazian wrote:
> to a different inode. The real inode is released. Do you think this is the
> wrong direction too?

What I don't quite understand about your close mechanism is for what type
of situation it is intended. It seems to me that only very few programs
can be expected to continue correct operation if you just break their
file descriptors. If in doubt, you have to exit or kill the program
anyway.

The only obvious exceptions are programs that are known not to do anything 
with files anyway (so you can take away root, cwd, and std*), and programs 
that do some background jobs and don't really need to interact over their 
tty (take away std*). Is this what you're after ?

What your approach doesn't seem to solve is the much nastier problem of 
hanging somewhere in the kernel in an uninterruptible sleep on something 
underneath a file descriptor. Obviously, trying to close it can't work in 
this case either. 

One way to solve this is to find any cases where there is an
uninterruptible sleep that could be safely turned into an interruptible 
one. In cases where interruption is impossible either because the API 
doesn't allow it (e.g. getpid(2)), or because interruption is not 
acceptable for temporary failures and the kernel can't distinguish them 
from permanent failures (e.g. some important metadata updates on a failing 
disk), one could maybe introduce some SIGKILL_DESPERATELY (*) which gets 
checked even in an uninterruptible sleep and takes whatever desperate
action is necessary in that case. Then, common problem spots could be
converted accordingly.

Of course, part of the problem may also be that NFS doesn't offer enough
granularity, i.e. you either get a rigidly uninterruptible system or
something fragile that happily aborts operations even if you brush ^C by
accident. Again, something like SIGKILL_DESPERATELY would help here.

(* perferrably the whole name would have to be typed into kill. It makes
   me sick whenever I see some id^Hnsufficiently educated user kill -9
   everything in sight just because somebody told them that kill -9
   always works.)

- Werner

-- 
  _________________________________________________________________________
 / Werner Almesberger, ICA, EPFL, CH       [e-mail] /
/_IN_N_032__Tel_+41_21_693_6621__Fax_+41_21_693_6610_____________________/