Alan Cox wrote:
> I'd disagree. The remove and attach technique is very very commonly used
> by other applications in order to ensure the segment is deleted if a program
> crashed - directly analogous to open/unlink/read.
The problem is that shm segments are like temporary files in a lot of
way. You can't do open/unlink pass the name of a tmp file to another
program either -- so why should you be able to do it with shm segments?
To do open/unlink with multiple programs, you need some way to pass a
file descriptor around.
Besides, the creat/remove technique _still_ leaks from time to time. I
have seen it -- more and more shm segments being added because a program
crashed between the create and the remove.
> Gimp wont be the only thing that breaks. So the old behaviour needs to work
> at least for SYS5 IPC, what posix ipc does is another matter
I propose we wait and see what else breaks.
I would really like to see programs motivated to use a file descriptor
passing method for X. Currently these things can go wrong with X SHM:
- The X connection is local, so you create a SHM segment. But the
server is remote via a local proxy (an increasingly common
arrangement), so the SHM fails. Some programs fall over at this
point, but they can be fixed to fall back.
- Same scenario but the server finds a SHM segment on it's system
with the id you passed. Very bad: you think it succeeded, but the
wrong images appear. It can even be a security hole.
- Same again but you're reading images from X. E.g. to read font
bitmaps. Very bad: you just trashed another applications shm
- Segments tend to get leaked because programs sometimes crash
before IPC_RMID. On some systems, the total space is limited too,
so you can't afford to have lots of them. Maybe this happens less
with Linux. Typical fix: a script to delete bogus shms is run at
the start of the application. Result: application occasionally
breaks other applications or instances of itself.
These are typical, but easily fixed application problems:
- The server is local but on a different IP to your standard
address. You neglect to check all alias addresses so you don't
use SHM even though you probably should.
- You check for SHM extension, find it, try SHM and it fails. You
crash. Yes I've seen this.
All the mess I can think of, including the "maps wrong SHM segment"
problem are all fixed nicely by using fd passing over a unix socket.
That, IMO, is the friendly way to establish a shm area with the X
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Wed Mar 15 2000 - 21:00:21 EST