[
https://issues.apache.org/jira/browse/TS-3883?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14985975#comment-14985975
]
Alan M. Carroll commented on TS-3883:
-------------------------------------
This change was causing the regression tests to fail with {{--enable-debug}}.
The problem is the IO buffers are for some reason set to use a default
alignment of 8192 ({{DEFAULT_BUFFER_ALIGNMENT}}). Prior to this change the
memory allocation call would align to this value (via {{f->alignment}}). Now,
however, the allocation is always page boundary aligned. If the page size is
less than the request alignment the alignment assertion in {{freelist_new}} can
fail page size alignment is insufficiently strict. The change I made was to
clip the alignment requirement for IO buffers to the page size as, IMHO, it is
never useful to require alignment more strict than the page size.
> ats_madvise() has no effect on Linux (including MADV_DONTDUMP)
> --------------------------------------------------------------
>
> Key: TS-3883
> URL: https://issues.apache.org/jira/browse/TS-3883
> Project: Traffic Server
> Issue Type: Bug
> Components: Core
> Reporter: Gancho Tenev
> Assignee: Phil Sorber
> Labels: A
> Fix For: 6.1.0
>
>
> While investigating unrelated issue with truncated core dumps on Linux,
> noticed that we run out of space on a few machines because of huge core dumps
> which were tending to the ATS process virtual memory size (reported by
> /proc/<pid>/status:VmSize on Linux).
> It looked like MADV_DONTDUMP memory use advice was not set properly.
> Further debugging showed that we have the following code in ats_madvise():
> {code}
> #if defined(linux)
> (void)addr;
> (void)len;
> (void)flags;
> return 0;
> #else . . .
> {code}
> Which would lead ats_madvise() to have no effect when "defined(linux)" is
> true, and would skip the necessary madvise() call to set MADV_DONTDUMP.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
