lkml.org 
[lkml]   [2006]   [Mar]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: Remapping pages mapped to userspace
From
Date
    Linus> Generally, replacing the mmap with an anonymous zero-mapped
Linus> mapping would be a horribly bad idea.

Linus> The fact is, you can't avoid the race of seeing the removed
Linus> state (which _usually_ means that you will read 0xffffffff
Linus> from the bus - normal PC's won't result in bus errors
Linus> etc). Whatever the kernel does, it can do only after the
Linus> device has already been removed - we no longer live in a
Linus> world where the administrator can tell the system
Linus> before-hand that something will go away.
Linus> Replacing the MMIO map with a zero map would be absolutely
Linus> horrible. It would be inconsistent, and not even help the
Linus> fact that the user will haev seen the removed state.

Linus> In fact, I think even "revert" is pretty useless. You're
Linus> much better off just sending a perfectly good signal -
Linus> something that the app will get regardless of whether it
Linus> reads the MMIO space at that point in time or not. After
Linus> all, the only thing the "revert" would really do is to send
Linus> a signal, but then only if the user is trying to access the
Linus> device.

Hmm, you're probably right for the hot unplug case. However, there
are a couple of other related situations I've thought of in this context:

- Module remove: userspace has PCI memory mmap()ed. The module is
removed. Userspace still has access to PCI memory. And if the
module is reloaded, userspace still has access to the device that
the driver doesn't know about, so the driver may hand off the same
access to another process.

The obvious solution here is to just take a module reference when
userspace mmap()s the device. However unprivileged processes can
get direct access to IB devices, and it may not be so nice for
unprivileged processes to be able to hold off module removal.

- PCI error recovery (or internal device error recovery): if the
driver detects a problem with the PCI bus or device, it might have
to reset the device. The most natural way to model this is as hot
unplug followed by hot plug. However we don't want userspace
processes to keep their BAR mapping across the device reset,
because the device is probably not set up to handle it right after reset.

Really in this case at least, revoking an mmap() seems the cleanest
solution to me.

Any further thoughts?

Thanks,
Roland
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-03-18 00:00    [W:0.161 / U:0.064 seconds]
©2003-2011 Jasper Spaans. Advertise on this site