It was possible for a registered eloop socket handler to be unregistered
and re-registered for a re-opened socket with the same fd from a timeout
or signal handler. If such a case happened with the old socket having a
pending event waiting for processing, some eloop combinations could end
up calling the new handler function with the new socket and get stuck
waiting for an event that has not yet happened on the new socket. This
happened with timeout and signal handlers with all eloop.c types. In
addition to that, the epoll case could also trigger this when a socket
handler re-registered a re-opened socket.
Fix these by checking whether there has been socket handler changes
during processing and break the processing round by going back to
select/poll/epoll for an updated result if any changes are done during
the eloop handler calls before processing the old socket results.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
It was possible for the SIGINT/SIGTERM signal to be received while
processing a pending timeout/socket/signal event and then get stuck in
the following select() call before processing the signal event. If no
other events show up within the two second SIGALRM trigger, process will
be terminated forcefully even though there would have been possibility
to do clean termination assuming no operationg blocked for that two
second time.
Handle this more cleanly by checking for eloop.pending_terminate before
starting the select()/poll()/epoll_wait() wait for the following event.
Terminate the loop if pending signal handling requests termination.
In addition, make eloop_terminated() return 1 on eloop.pending_terminate
in addition to eloop.terminate since the process will be terminated
shortly and there is no point in starting additional processing.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The socket reference tracking entries need to be restored in case
os_realloc_array() fails when adding a new eloop socket.
Signed-off-by: Jouni Malinen <j@w1.fi>
This patch adds epoll option for the eloop implementation. This can be
selected with the CONFIG_ELOOP_EPOLL=y build option.
[merit]
See Table1.
Table1. comparison table
+--------+--------+-----------+------------+-------------+
| | add fd | remove fd | prepare fd | dispatch fd |
+--------+--------+-----------+------------+-------------+
| select | O(1) | O(1) | O(N) | O(N) |
+--------+--------+-----------+------------+-------------+
| poll | O(1) | O(1) | O(N) | O(N) |
+--------+--------+-----------+------------+-------------+
| epoll | O(1) | O(1) | 0 | O(M) |
+--------+--------+-----------+------------+-------------+
"add fd" is addition of fd by eloop_sock_table_add_sock().
"remove fd" is removal of fd by eloop_sock_table_remove_sock().
"prepare fd" is preparation of fds before wait in eloop_run().
"dispatch fd" is dispatchment of fds by eloop_sock_table_dispatch().
"N" is all watching fds.
"M" is fds which could be dispatched after waiting.
As shown in Table1, epoll option has better performance on "prepare fd" column.
Because select/poll option requires setting fds before every select()/poll().
But epoll_wait() doesn't need it.
And epoll option has also better performance on "dispatch fd" column.
Because select/poll option needs to check all registered fds to find out
dispatchable fds. But epoll option doesn't require checking all registered fds.
Because epoll_wait() returns dispatchable fd set.
So epoll option is effective for GO/AP functionality.
[demerit]
The epoll option requires additional heap memory. In case of P2P GO, it is
about 8K bytes.
Signed-off-by: Masashi Honma <masashi.honma@gmail.com>
Commit e2f5a9889a was supposed to prevent
new scan request from pushing out the old one. However, it did not
really do that since eloop_deplete_timeout() returned 0 both for the
case where the old timeout existed (and was sooner) and if the old
timeout did not exist. It returned 1 only for the case where an old
timeout did exist and was larger than the new requested value. That case
used to result in wpa_supplicant_req_scan() rescheduling the timeout,
but hew code in eloop_deplete_timeout() did the exact same thing and as
such, did not really change anything apart from the debug log message.
Extend the eloop_deplete_timeout() (and eloop_replenish_timeout() for
that matter since it is very similar) to return three different values
based on whether the timeout existed or not and if yes, whether it was
modified. This allows wpa_supplicant_req_scan() to schedule a new
timeout only in the case there was no old timeout.
Signed-hostap: Jouni Malinen <j@w1.fi>
Relative time shouldn't be calculated based on gettimeofday
because that clock can jump (e.g., when the time is adjusted
by the system administrator.)
On systems where that is available, use CLOCK_BOOTTIME (on
fairly recent Linux systems, this clock takes into account
the time spend suspended) or CLOCK_MONOTONIC (on Linux and
some POSIX systems, this clock is just freely running with
no adjustments.)
Reported-by: Holger Schurig <holgerschurig@gmail.com>
Signed-hostap: Johannes Berg <johannes.berg@intel.com>
eloop_replenish_timeout() finds a registered matching
<handler,eloop_data,user_data> timeout. If found, replenishes
the timeout if remaining time is less than the requested time.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
DFS implementation requires to run an eventloop while monitoring
the Channel Availability Check (CAC). After that, the "real" event
loop is started, and should not fail doing so.
Signed-hostap: Simon Wunderlich <siwu@hrz.tu-chemnitz.de>
Without this, we can get into a tight loop because the
code in general doesn't add eloop exception handlers,
so socket reporting the POLLERR would never be read.
With this change, any socket with POLLERR or POLLHUP
asserted will be handled by the read logic.
Signed-hostap: Ben Greear <greearb@candelatech.com>
eloop_sock_table_add_sock() needs to fail if pollfd array allocation
fails instead of returning success and leaving behind no buffer.
Signed-hostap: Jouni Malinen <j@w1.fi>
When using more than around 200 virtual stations, we start hitting the
max number of file descriptors supported by select(). This patch adds
support for poll(), which has no hard upper limit.
Signed-hostap: Ben Greear <greearb@candelatech.com>
If the os_time_t variable used for the expiration time (seconds)
overflows when the registered timeout value is being added,
assume that the event would happen after an infinite time, i.e.,
would not really happen in practice. This fixes issues with
long key timeouts getting converted to immediate expiration due
to the overflow.
Better make sure the eloop_timeout data gets fully initialized. The
current code is filling in all the fields, but it is clearer to just
zero the buffer to make sure any new field added to the structure gets
initialized.
This WPA_TRACE=y additions allows components to register active references
to memory that has been provided to them as a pointer. If such an actively
referenced memory area is freed, tracer will report this as an error and
backtraces of both the invalid free and the location where this pointer
was marked referenced are shown.
WPA_TRACE=y can now be used to enable internal backtrace support that
will provide more details about implementation errors, e.g., when some
resources are not released correctly. In addition, this will print out
a backtrace automatically if SIGSEGV is received.
In situations where the driver does background scanning and sends a
steady stream of scan results, wpa_supplicant would continually
reschedule the scan. This resulted in specific SSID scans never
happening for a hidden AP, and the supplicant never connecting to the AP
because it never got found. Instead, if there's an already scheduled
scan, and a request comes in to reschedule it, and there are enabled
scan_ssid=1 network blocks, let the scan happen anyway so the hidden
SSID has a chance to be found.
