Web lists-archives.com

Consider exposing mmap_is_attached_or_noreserve




Hello,

A very technical request regarding Cygwin internals: In mmap.c there
is a function mmap_is_attached_or_noreserve(void *addr, size_t len)
which is called from Cygwin's exception handler in the case of a
STATUS_ACCESS_VIOLATION.

This is called in case an access violation occurs in memory that was
allocated with Cygwin's mmap() with the MAP_NORESERVE flag, and allows
us to commit the relevant pages when they are accessed.

After a successful call of mmap_is_attached_or_noreserve(), the Cygwin
exception handler returns with ExceptionContinueExecution.
Unfortunately, if the application happens to have a Vectored Continue
Handler registered which happens to do something in the case of
STATUS_ACCESS_VIOLATION (see [1]) there is no obvious way to tell if
we're handling this sort of case.

Normally this isn't too much of a problem: E.g. we could just check
the address that caused the access violation and see if its status is
now MEM_COMMIT (i.e. Cygwin ran its exception handler and all is
good).  However, due to the bug described in [1], if an exception
occurs in code running on a sigaltstack, the Cygwin exception handler
isn't run.

This makes for a tricky to handle use case:  What if some code in a
signal handler function tries to access uncommitted memory in a
MAP_NORESERVE mmap?  It's probably an unusual, undesirable case, and I
haven't personally encountered it *yet*, but I could imagine some
cases where it might happen.

In order to handle such a case it might be nice if
mmap_is_attached_or_noreserve were able to be called by user code,
perhaps as a new cygwin_internal(...) call.  I'd happily provide a
patch, but I fear this might be an X/Y problem that I'm not seeing.

Thanks,
Madison



[1] https://cygwin.com/ml/cygwin/2018-05/msg00333.html

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple