os, dix, mi, render: optimize RAM consumption in background when idle by GermanAizek · Pull Request #3961 · X11Libre/xserver

GermanAizek · 2026-10-13T19:24:37Z

@metux, I’d like to raise a pressing issue about optimizing memory using malloc_trim by default (https://man7.org/linux/man-pages/man3/malloc_trim.3.html)

malloc_trim function first appeared in 1996 as part of the ptmalloc memory allocator
(Wolfram Glogger’s multithreaded extension of Doug Lea’s dlmalloc).

The function is declared with 1997 year in <malloc.h> and is a GNU-specific extension.
Its purpose is to attempt to release unused memory from the top of the heap (via sbrk) or from arenas
back to the operating system.

Regressions on x11perf not found.

The prices of RAM and flash memory continue to rise and are once again breaking records.

https://nl.pcpartpicker.com/trends/price/memory/#ram.ddr5.6000.2x16384

Solution for low RAM specs from Meta

In this regard, to optimize costs for purchasing RAM, Meta has decided to optimize compression instead of using ZRAM and ZSWAP.

https://www.phoronix.com/news/Linux-CRAM-Compressed-RAM

I think we also need to value every user megabyte, because it’s unclear when prices for components will start to fall, and I hope you agree that x11libre also supports reuse electronic waste for ECO environmental purposes.

My benchmarks

Changes:

Implement background memory trimming and heap compaction when the X server
is idle, without changing the default glibc mem allocator or breaking ABI:

Allocator tuning (os/osinit.c):
- Set M_TRIM_THRESHOLD (128 KB) and M_ARENA_MAX (2) at startup.
- Explicitly pin M_MMAP_THRESHOLD (128 KB) to prevent dynamic threshold
  escalation that causes heap retention and fragmentation.
Background idle trim timer (os/WaitFor.c):
- Trigger an idle trim callback after 5 seconds of inactivity.
- Disarm on active client requests (0% active rendering overhead).
- Re-arm periodically (every 60s) and hook into screen saver / DPMS events.
Buffer and cache reclamation:
- Free empty input (oc->input) and output (oc->output) buffers for idle
  clients, along with AvailableInput and buffer pools (os/io.c).
- Purge font pattern caches on idle (dix/dixfonts.c).
- Downsize client resource hash tables back to INITBUCKETS (dix/resource.c).
- Shrink the mi event queue back to its initial size (mi/mieq.c).
- Compact global glyph hash tables and clear tombstones (Xext/render/glyph.c).
- Release the byte-swapped event staging buffer (dix/events.c).
Clean non-core library page reclamation (os/WaitFor.c):
- Issue madvise(MADV_DONTNEED) exclusively on clean executable (r-xp)
  and offset-0 rodata mappings of heavy non-core libraries (libLLVM,
  libz3, libstdc++, libxml2, and Mesa DRI drivers).
- Strictly preserve writable and GNU_RELRO segments to protect relocations.
- Invoke malloc_trim(0) to return heap pages to the kernel.

metux · 2026-10-15T10:29:05Z

+                      if [ "${{ github.repository }}" = "X11Libre/xserver" ] && [ "${{ github.event_name }}" != "pull_request" ]; then
+                          docker push "$img:$tag"
+                          if [ "${{ github.ref }}" = "refs/heads/master" ]; then
+                              docker tag "$img:$tag" "$img:latest"
+                              docker push "$img:latest"
+                          fi
+                      else
+                          echo "Skipping docker push: write permissions to GHCR unavailable on pull requests or forks"
                      fi


looks unrelated to this PR - should be slit out to separate one.

I accidentally pulled in local changes

metux · 2026-10-15T12:06:08Z

+void
+dixTrimEvents(void)
+{
+    if (swapEvent) {
+        free(swapEvent);
+        swapEvent = NULL;
+        swapEventLen = 0;
+    }
+}
+


maybe we don't even need to keep that buffer global. it's not used often and could be allocated / freed when needed.

Fix it 1b5cf6a

github-actions · 2026-10-16T12:10:01Z

Merge Conflict found

metux · 2026-10-19T11:08:01Z

@GermanAizek has conflicts that need to be resolved.

When the X server runs for extended periods or handles bursty client workloads, significant memory bloat is retained across heap arenas, unshrunk resource tables, event queues, client I/O buffers, and dormant shared libraries even after all client resources are destroyed. Implement comprehensive background idle memory reclamation without changing memory allocators and preserving full ABI compatibility: - Startup allocator tuning: * Configure M_TRIM_THRESHOLD (128 KB) and M_ARENA_MAX (2). * Configure M_MMAP_THRESHOLD (128 KB) to prevent runaway dynamic threshold escalation that causes heap fragmentation. - Idle background trim timer: * Schedule an ABI-safe idle callback after 5 seconds of inactivity. * Disarm the timer on incoming client traffic so active workloads suffer zero overhead or latency penalty. * Re-arm periodically (every 60 seconds) during sustained idle states, and trigger trim on screensaver / DPMS power-saving events. - Buffer and cache reclamation: * Reclaim idle client input (oc->input) and output (oc->output) buffers, along with AvailableInput and cached free buffer pools. * Empty dix font pattern caches (patternCache and fontPatternCache). * Downsize client resource bucket tables back to INITBUCKETS when client resource counts shrink (preserving serverClient index 0). * Compact mi input event queue (miEventQueue) back to initial size. * Compact global glyph hash tables and clear deleted glyph tombstones. - Clean non-core library page reclamation: * Inspect /proc/self/maps on Linux and issue madvise(MADV_DONTNEED) on purely clean file-backed segments (executable r-xp code and file-offset zero rodata) of heavy dormant libraries (libLLVM, libz3, libstdc++, libxml2, and Mesa DRI drivers). * Strictly avoid writable or GNU_RELRO segments to preserve copy-on-write dynamic relocations. * Invoke malloc_trim(0) to release freed heap pages to the kernel. - Robust listener socket binding: * In _XSERVTransSocketUNIXCreateListener(), attempt direct bind() first and only inspect/unlink stale files on EADDRINUSE to prevent race conditions between concurrent servers under -displayfd. - CI compatibility: * Guard M_ARENA_MAX and HAVE_MALLOPT for Cygwin. Signed-off-by: Herman Semenoff <GermanAizek@yandex.ru>

GermanAizek · 2026-10-20T23:19:06Z

@GermanAizek has conflicts that need to be resolved.

Fix it 1b5cf6a

GermanAizek · 2026-11-03T19:44:40Z

@metux Haven’t you reviewed this PR yet? I specifically made all the other PRs drafts so that I could finish the work with the active ones first.

GermanAizek · 2026-11-04T05:34:38Z

@metux origin story malloc_trim() and algorithm

In Unix systems, memory trimming is already available and complies with POSIX standards — it has been so since the end 90s. It’s just that many developers have forgotten about this technology. Back then, the price of RAM per 1 MB was high, just like it is now. That’s why they came up with the idea of manually freeing up memory earlier call signal to OS memallocator and using it for other processes without losing performance — but this requires manual work from programmer.

cepelinas9000 · 2026-11-05T22:50:38Z

-     * Don't blindly unlink a pre-existing socket file: it may belong to a
-     * server that's still running (see unix_socket_is_live()). Only clean
-     * up and rebind if nobody answers there.
+     * Try to bind directly without unlinking first. If the socket file does


I think this should go to other commit, or even better other PR.

cepelinas9000 · 2026-11-05T22:56:29Z

+#endif
+
+static CARD32
+IdleTrimCallback(OsTimerPtr timer, CARD32 now, void *arg)


I'm afraid there are chance with timer callback like this to be source of unexplained latency spikes.

I don't know how it should be approached, one of suggestion to hook it on new opcode in xtest extension (the only few calls defined) and allow to be called manually

GermanAizek force-pushed the opt-ram-trim-alloc branch 3 times, most recently from ab524e9 to 2968091 Compare October 14, 2026 00:35

metux reviewed Oct 15, 2026

View reviewed changes

metux requested a review from a team October 15, 2026 12:03

metux force-pushed the opt-ram-trim-alloc branch from 2968091 to 81f1357 Compare October 15, 2026 12:04

metux reviewed Oct 15, 2026

View reviewed changes

GermanAizek force-pushed the opt-ram-trim-alloc branch from 81f1357 to 1b5cf6a Compare October 20, 2026 23:16

metux force-pushed the master branch 5 times, most recently from eae2c21 to 483a172 Compare October 30, 2026 11:57

cepelinas9000 reviewed Nov 5, 2026

View reviewed changes

Navigation Menu

Repository navigation

os, dix, mi, render: optimize RAM consumption in background when idle - #3961

os, dix, mi, render: optimize RAM consumption in background when idle#3961
GermanAizek wants to merge 1 commit into
X11Libre:masterfrom
GermanAizek:opt-ram-trim-alloc

GermanAizek commented Oct 13, 2026 •

edited

Loading

Uh oh!

metux Oct 15, 2026

Uh oh!

GermanAizek Oct 15, 2026

Uh oh!

metux Oct 15, 2026

Uh oh!

GermanAizek Oct 20, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

Uh oh!

metux commented Oct 19, 2026

Uh oh!

GermanAizek commented Oct 20, 2026

Uh oh!

GermanAizek commented Nov 3, 2026

Uh oh!

GermanAizek commented Nov 4, 2026

Uh oh!

cepelinas9000 Nov 5, 2026

Uh oh!

cepelinas9000 Nov 5, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

GermanAizek commented Oct 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Solution for low RAM specs from Meta

My benchmarks

Changes:

Uh oh!

metux Oct 15, 2026

Choose a reason for hiding this comment

Uh oh!

GermanAizek Oct 15, 2026

Choose a reason for hiding this comment

Uh oh!

metux Oct 15, 2026

Choose a reason for hiding this comment

Uh oh!

GermanAizek Oct 20, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Oct 16, 2026

Uh oh!

metux commented Oct 19, 2026

Uh oh!

GermanAizek commented Oct 20, 2026

Uh oh!

GermanAizek commented Nov 3, 2026

Uh oh!

GermanAizek commented Nov 4, 2026

Uh oh!

cepelinas9000 Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

cepelinas9000 Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

GermanAizek commented Oct 13, 2026 •

edited

Loading