On 04/20/2016 06:24 PM, Matt Wilmas wrote:
Hi Dmitry,
----- Original Message -----
From: "Dmitry Stogov"
Sent: Wednesday, April 20, 2016
Hi,
It's a well known PHP problem, that exceeding of execution time-out
(max_execution_time) may lead to unexpected crashes.
They occur because PHP may be interrupted in inconsistent state, and
attempt
to release allocated by request resources leads to failure.
Almost any big site sees these crashes from time to time.
I propose to delay actual request termination until a "safe" point in
interpreter.
Signal handler will just set EG(timed_out) flag.
Interpreter will check this time from time to time (on jumps and
calls that
may make loops or recursion) and perform the actual termination.
This approach already works in PHP for Windows.
I was thinking about this, checking for things like EG(exception)
"constantly," a few months ago for another reason...
This is a bit different problem. We can't delay EG(exception) checks. I
thought about a better way of exception handing but didn't find anything
usable and portable.
What about instead of adding additional checks in the same place(s) in
VM, we just limit it to 1 check, for multiple things? Just have
EG(something_unexpected_to_check), and behind that (or in a function,
I guess), the actual rare/unexpected thing gets checked: timed_out,
exception, etc.
Yes, I have the same idea in background. I even wrote: The same
"interrupt" handling mechanism in the future may be reused for TICK
and signal handling.
It seems Bob had a similar idea in the PR comment, except literally
using exceptions...
In addition I introduce hard_timeout (default value 2 seconds).
In case the "soft" timeout wasn't handled "safely" in that 2 seconds
(because of long running internal function), PHP process will be
terminated
without attempt to free any resources.
ZTS build will ignore "hard_timeout" (in the same way as PHP on
Windows do).
The PR: https://github.com/php/php-src/pull/1876
It removes "exit_on_timeout" ini directive, and introduces
"hard_timeout"
instead.
Additional checks in VM make 0.5-1% slowdown in term of instruction
retired
reported by callgrind.
A single check would save those additional instructions and branches,
and would actually improve things on Windows (since this PR doesn't
change anything there).
If you or Bob show me a better working solution (or just PoC), I'll be
only happy with this.
Thanks. Dmitry.
I think we don't need RFC for this. This is a long time desired fix.
The same "interrupt" handling mechanism in the future may be reused
for TICK
and signal handling.
Thanks. Dmitry.
- Matt
--
PHP Internals - PHP Runtime Development Mailing List
To unsubscribe, visit: http://www.php.net/unsub.php
