So I started to work on back-patching the no-pallocs change in xml2.
It seems to work fine in 8.4, but it crashes in 8.3. The reason is
that we never back-patched the 8.4 change to make utils/adt/xml.c
not change the libxml2 memory allocation functions. That means that
if you try to intermix core xml operations with contrib/xml2 operations,
it fails because libxml2 is still trying to use the core's substituted
memory alloc functions, and the contrib module isn't doing what's
needful to make those actually work.
This may explain the previous observations that 8.3's contrib/xml2
didn't crash in as many cases as 8.4's does. If core and contrib code
both replace the allocation hooks, they're somewhat independent of each
other. When only one does, big trouble is what you've got.
It seems like the most rational response to this is to go ahead and
back-patch the 8.4 changes, specifically this patch
into 8.3 so that we can also fix xml2. I was afraid to do that back
in May when the patch was committed, but by now we have enough field
testing to suggest that 8.4 is no worse than 8.3 as far as the core
xml operations go.
regards, tom lane
pgsql-hackers by date
|Next:||From: Greg Stark||Date: 2010-03-01 00:04:06|
|Subject: pgsql: add EPERM to the list of return codes to expect from opening |
|Previous:||From: Andrew Dunstan||Date: 2010-02-28 22:58:42|
|Subject: Re: remove contrib/xml2|