In most cases after vchan close program terminates (or reexec itself to wait
for next connection), so cleanup isn't needed (not sure how about evtchn and
shared pages...).
But in case of stubdom it is needed to cleanup before waiting for new
connection - we can't reexec agent there.
Use new u2mfn_fd for each ring_init call - each open fd to u2mfn is capable to
do only one u2mfn_alloc_kpage, so it can't be reused at the next ring_init.
In case of evtchn buffer overflow (received more than 1024 events between
libvchan_wait calls) further reads returns ERROR_IO_DEVICE. The only way to
recover from that is to reset the buffer. Because vchan code doesn't take care
of number of fired events - only the fact that some event was fired - lost
events here shouldn't break anything. Events reported _after_ libvchan_wait
call will be collected and reported correctly.
Some more comments in the code (here and in qrexec-agent in the next commit).
Windows evtchn driver can double fire on one event (details in
xenpci/evtchn_device_interface.c:EvtChn_EvtIoWrite). Less intrusive way of
fixing it is just remove all events from queue on libvchan_wait (to prevent
queue fill).
This won't hurt because we are interested in events only sent
after xc_evtchn_unmask call which is one line below.
This reverts commit f7ce10db8f76f752cf9d65b420de352e70bdb048.
This should be fixed by doing 0-length reads in qrexec main loop (instead of
real read of fired port number).
Use features macros (QREXEC_RING_V2 and ASYNC_INIT) instead of directly
depending on building environment. The macros are turned on (when required) in
libvchan.h based on target system.
