ghc runtime by default (otherwise depending on a "configure" option)
does memory allocation on their own by first mmapping a 1 TB range of
memory into the process and then parceling out chunks from it.

If one of the chunks is not needed, the kernel needs to be informed -
otherwise the system would quickly run out of available RAM.

ghc does that via madvise(2).

There are two options when doing this informing:

MADV_FREE - Means "I don't need this range or the data in it any more". 
Kernel promises to fail later accesses to it.

MADV_DONTNEED - Means "I don't need this range right now - and I don't
need the data in it anymore". Kernel promises to make later accesses to
it succeed (if necessary by providing a new page initialized with zeroes).

MADV_FREE was introduced in Linux 4.5.
glibc 2.25 and later always define MADV_FREE.

Unpatched ghc 8.0.2 will use either MADV_FREE or MADV_DONTNEED, determined
at ghc compile time.  Which of them will actually succeed is determined
by the Linux kernel at run time.

This patch makes ghc try MADV_FREE.  If it doesn't work, it falls back to
MADV_DONTNEED.

The end result is that ghc programs free their memory with Linux < 4.5 again.

See https://git.haskell.org/ghc.git/commitdiff/6576bf83cdf4eac05eb88a24aa934a736c91e3da for more information.
--- a/rts/posix/OSMem.c
+++ b/rts/posix/OSMem.c
@@ -541,11 +541,24 @@ void osDecommitMemory(void *at, W_ size)
 
 #ifdef MADV_FREE
     // Try MADV_FREE first, FreeBSD has both and MADV_DONTNEED
-    // just swaps memory out
+    // just swaps memory out. Linux >= 4.5 has both DONTNEED and FREE; either
+    // will work as they both allow the system to free anonymous pages.
+    // It is important that we try both methods as the kernel which we were
+    // built on may differ from the kernel we are now running on.
     r = madvise(at, size, MADV_FREE);
-#else
-    r = madvise(at, size, MADV_DONTNEED);
+    if(r < 0) {
+        if (errno == EINVAL) {
+            // Perhaps the system doesn't support MADV_FREE; fall-through and
+            // try MADV_DONTNEED.
+        } else {
+            sysErrorBelch("unable to decommit memory");
+        }
+    } else {
+        return;
+    }
 #endif
+
+    r = madvise(at, size, MADV_DONTNEED);
     if(r < 0)
         sysErrorBelch("unable to decommit memory");
 }
lass='left'>Commit message (<a href='/guix/log/tests/grafts.scm?id=39ad23ec9a8625eae3d23618e902e83608476b36&amp;showmsg=1'>Expand</a>)</th><th class='left'>Author</th></tr>
<tr><td><span title='2023-10-28 00:17:23 +0200'>2023-10-28</span></td><td><a href='/guix/commit/tests/grafts.scm?id=67effc1560fc175dfbcb58ef5b965b08b3942d6c'>grafts: Fix corner case involving multiple-output derivations.</a><span class='msg-avail'>...<span class='msg-tooltip'>Fixes a bug that would occur with references to two outputs of the same
derivation, with one of them referring to the other one.

For example, the references of libreoffice include both mariadb:dev and
mariadb:lib; additionally, mariadb:dev refers to mariadb:lib.  In this
case, the glibc graft would not be applied on one of the mariadb paths,
and both the grafted and ungrafted glibc would end up in the closure of
libreoffice.

Fixes &lt;https://issues.guix.gnu.org/66662&gt;.

* guix/grafts.scm (non-self-references): Simplify and include references
to outputs of DRV other than OUTPUTS.
(reference-origins): Simplify and possibly return outputs of DRV itself.
(cumulative-grafts)[graft-origin?]: Add OUTPUT parameter and honor it.
[dependency-grafts]: Adjust accordingly.
* tests/grafts.scm ("graft-derivation, multiple outputs need to be replaced"):
New test.

Change-Id: Iac2005024ab7049037537b3af55298696ec90e3c
</span></span></td><td>Ludovic Courtès</td></tr>