USB not processing APM suspend event properly?
Quoted 5 lines
This was caused by the PCMCIA network interface being off, and the apmd scripts trying to fuser -k and unmount an in-use NFS partition, which in turn wanted to generate NFS traffic to the server over the shut down PCMCIA network interface...
So your problem was that PCMCIA got shut down too early.
Quoted 5 lines
Basically, I've changed the order that things happen on suspend, such that we always pass the event to apmd, and any other listeners on
/dev/apm_bios. Once everyone has replied (subject to the rules of course), it then sends thePM_SUSPENDto the devices.
I suppose that the reason why it is presently coded as it is is that the drivers can veto the suspend whereas the "listeners" can't, and the apm driver wants to tell the listeners about the event only if it isn't vetoed. With your change, the apm driver really should ignore a veto from the drivers since it has already told the listeners about the event. Therefore, while your patch is an improvement, it still isn't the whole solution.
One possible solution is for the apm driver first to _ask_ the device drivers whether a suspend is allowed; then to tell the listeners; then to tell the device drivers to suspend. This solution could be implemented by adding the "ask" function to each driver with pm support. The driver's opportunity to veto the suspend would be when it is asked, not when it is told to suspend.
Quoted 5 lines
The original code sent
PM_SUSPENDon receiving the original request, and then multiple times each time a listener on/dev/apm_biosreplied. Not good if apmd wants to unmount NFS, and your NFS mounts require traffic over your PCMCIA ether card!
Actually, it's not true that the original code sends PM_SUSPEND each time a listener on /dev/apm_bios replies. The code fragment is:
if (as->suspends_read > 0) {
as->suspends_read--;
as->suspends_pending--;
suspends_pending--;
} else if (!send_event(APM_USER_SUSPEND))
return -EAGAIN;
else
queue_event(APM_USER_SUSPEND, as);This only calls send_event() if the ioctl(suspend) is NOT a reply to a notification that was earlier read from the queue.
Quoted 8 lines · Thomas Hood
I suppose that the reason why it is presently coded as it is is that the drivers can veto the suspend whereas the "listeners" can't, and the apm driver wants to tell the listeners about the event only if it isn't vetoed. With your change, the apm driver really should ignore a veto from the drivers since it has already told the listeners about the event. Therefore, while your patch is an improvement, it still isn't the whole solution.
I think it does perform acceptable behaviour however - if the drivers (in the unlikely event) refuse the event after we've told user space, we tell user space we resumed.
Quoted 4 lines
One possible solution is for the apm driver first to _ask_ the device drivers whether a suspend is allowed; then to tell the listeners; then to tell the device drivers to suspend.
What if the condition of the suspend gets changed by the act of paging in memory, or some other change that occured in user space? It's a chicken and egg problem. 8(
Quoted 11 lines
Actually, it's not true that the original code sends
PM_SUSPENDeach time a listener on/dev/apm_biosreplies. The code fragment is:if (as->suspends_read > 0) { as->suspends_read--; as->suspends_pending--; suspends_pending--; } else if (!send_event(APM_USER_SUSPEND)) return -EAGAIN; else queue_event(APM_USER_SUSPEND, as);
Quoted 2 lines
This only calls
send_event()if the ioctl(suspend) is NOT a reply to a notification that was earlier read from the queue.
It is a minor point, but an unwanted one that just obfuscates the operation of the APM system - only the first PM SUSPEND event causes a change.
Sending 'n' PM_SUSPEND events because you have 'n' listeners on /dev/apm_bios seems to be a waste of time when it could be handled more cleanly.
--
Russell King ([email protected]) The developer of ARM Linux
http://www.arm.linux.org.uk/personal/aboutme.html