[squid-dev] Issue detected in squid 2.7 version

scanty babu scantybabu at gmail.com
Fri Oct 31 05:20:05 UTC 2014


Hi,

We are using squid 2.7.7 version in our production environment and
recently faced an issue with this version.

The scenario:
1. Slow Client side connection
2. Fast Server side connection
3. Client sends a range request from the end eg. 10000 byte Content
length and the client requests 0-9500 bytes.

Client connection being slow, the server side reads get automatically
stopped when the data to be written to the client is more than the
ReadAhead gap(16KB, default in our case).

The server side reads are resumed when all the data in memory has been
written to the client.
The mechanism of defer and resume keeps occurring during the length of
the download.

This mechanism works properly in most of the cases, except where Range
request are involved.
The client gets the requested range of bytes, but does not flip
DEFER_READ flag on the server side FD. So, server reads wait for about
15 minutes(timeout) and connection is closed. The HTTP object is
released from the store even though the server has sent the complete
object.


The solution:
fwdDeferReadCheck function checks if the server side reads have to be
resumed/deferred. As DEFER_READ flag is still set on the FD reads will
never resume.So, we put a check


-    if (EBIT_TEST(e->flags, ENTRY_DEFER_READ)) {
+       Server-side fd stuck in defer state after client has received
the byte range requested by it.
+       Unblocking the server side connection, if there is no pending
client waiting on a store entry object.
+       If the pending data is more than quick_abort_max, then the
connection will be aborted by squid
+    */
+    if (EBIT_TEST(e->flags, ENTRY_DEFER_READ) &&
+         storePendingNClients(e)!=0 ) {
   commDeferFD(mem->serverfd);
   return 1;
     }

If there is more data to be read from the server, the quick_abort_max
configuration will make sure squid aborts the connection

Just wanted to bring to notice this issue in 2.7 version, so that it
can be fixed in latest squid versions 3.x (if present).

Also, please advise if we are doing it correctly (or if it could be
done otherwise).

Thanks,
scanty


More information about the squid-dev mailing list