alert: ... pread() read only

[Maxim Dounin]

Hello!

On Tue, Nov 19, 2013 at 09:39:44AM +0200, Nikolaos Milas wrote:

[...]

> The problem is that there is a repeating error of the form (I have
> changed real host name and web root path, as well as client IP
> address):
> 
> 2013/11/17 12:39:14 [alert] 20709#0: *9059 pread() read only 38605
> of 39107 from "/path/to/web/root/HTML/gmap/gmapv3_auto_el.html"
> while sending response to client, client: ::ffff:xxx.xxx.241.42,
> server: www.example.com, request: "GET
> /HTML/gmap/gmapv3_auto_el.html HTTP/1.1", host: "www.example.com"
> 
> I found here:
> 
> http://translate.google.com/translate?sl=ru&tl=en&js=n&prev=_t&hl=pt-PT&ie=UTF-8&u=http%3A%2F%2Fforum.nginx.org%2Fread.php%3F21%2C9856&act=url
> 
> ...that this is probably related to the "open_file_cache" directive,
> and in fact I do use (based on advice found on the Internet):

It's not "related" to the "open_file_cache" directive, but the 
"open_file_cache" directive makes the underlying problem more 
visible.

The root cause of the messages in question is non-atomic file 
update.  Somebody on your system edited the file in question 
in-place, instead of re-creating it with a temporary name and then 
using "mv" to atomically update the file.

Non-atomic updates create a race: a file which is opened by nginx 
(and stat()'ed for nginx to know it's length) suddenly changes.  
This can happen in the middle of a response, and results in a 
corrupted response - first part of a response is from original 
file, and second is from updated one.  If nginx is able to detect 
the problem due to file size mismatch - it logs the message in 
question.

The only correct solution is to update files atomically, i.e., 
create a new file and then rename to a desired name.

However, the "open_file_cache" directive makes the race window 
bigger by keeping files open for a long time.  Switching it off is 
a good idea if you can't eliminate non-atomic updates for some 
reason.

[...]

> Should I disable open_file_cache or not? (I do not entirely
> understand its implications.)
> 
> How should we determine the directive benefits?
> 
> Note: The aim is to be able to serve a few thousand requests per sec
> at peak, while normal traffic is < 20 reqs/sec.

I don't think that open_file_cache results in a measurable 
difference in your case.  I would recommend disabling it unless 
you have good reasons to enable it, just to simplify maintenance.

-- 
Maxim Dounin
http://nginx.org/en/donation.html