Change is_timeout() to sample current time before invoking poller
infrascructure to avoid occasional false timeout conditions.
Consider the following timeout loop (can be found in wait_on_timeout):
while (!(condition)) {
if (is_timeout(...)) {
return -ETIMEDOUT;
}
}
...
in the original logic of is_timeout() it was possible to end up in the
following situation:
1. Immediate check of of "condition" yeilds false (not enough
time for it to become true has passed yet)
2. is_timeout is called which, in turn, calls poller_call()
3. All registerd pollers take more than specified timeout time
to execute.
4. Sometime during poller_call() "contition" becomes true
5. As a result of #3 is_timeout() returns "true"
6. Code bails out with -ETIMEDOUT early even though timeout
condition didn't really occur.
One concrete example of this problem was discovered on ZII RDU1
board (poller_call() is long due to a serdev) when doing large data
transfers over SPI to attached DataFlash chip.
This commit changes is_timeout() to sample the value of time before
calling poller_call(). This way first call to is_timeout() will almost
always return false thus checking the condition at least twice before
declaring a timeout.
Signed-off-by: Andrey Smirnov <andrew.smirnov@xxxxxxxxx>
---
common/clock.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/common/clock.c b/common/clock.c
index 7f5cff232..2b218fb6a 100644
--- a/common/clock.c
+++ b/common/clock.c
@@ -181,10 +181,12 @@ EXPORT_SYMBOL(is_timeout_non_interruptible);
int is_timeout(uint64_t start_ns, uint64_t time_offset_ns)
{
+ int ret = is_timeout_non_interruptible(start_ns, time_offset_ns);
+
if (time_offset_ns >= 100 * USECOND)
poller_call();
- return is_timeout_non_interruptible(start_ns, time_offset_ns);
+ return ret;
}
EXPORT_SYMBOL(is_timeout);
--
2.17.1
_______________________________________________
barebox mailing list
barebox@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/barebox
