GNU bug report logs - #77478
Fixes a crash in the Haiku font driver for daemon mode

Previous Next

Full log

View this message in rfc822 format

From: "Kyle Ambroff-Kao" <kyle <at> ambroffkao.com>
To: 77478 <at> debbugs.gnu.org
Subject: bug#77478: Details
Date: Thu, 03 Apr 2025 07:10:53 +0000

This fixes double-free bug in Emacs daemon mode on Haiku. To reproduce:

1. Start emacs with "emacs --daemon"
2. Create a new frame with "emacsclient -c" and then close it.
3. Create a new frame with "emacsclient -c"

Step 3 will cause the Emacs daemon to crash.

  KERN: debug_server: Thread 3616 entered the debugger: Debugger call:
  `tried to free 0xb960bc9fd0 which points at page 232 which is not an
  allocation first page'

The backtrace from Emacs:
   heap_free(void*) + 0x35 
   BFont_close + 0x4d 
   haikufont_close(font*) + 0x29 (/Code/emacs/src/haikufont.c:893)
   sweep_vectors(void) + 0x1af (/Code/emacs/src/alloc.c:3242)
   garbage_collect(void) + 0x7b3 (/Code/emacs/src/alloc.c:7247)
   Ffuncall(ptrdiff_t, Lisp_Object*) + 0x194 (/Code/emacs/src/eval.c:3084)
   internal_condition_case_n(*, ptrdiff_t, Lisp_Object*, Lisp_Object, *)
   + 0x6c (/Code/emacs/src/eval.c:1699)
   safe_funcall(ptrdiff_t, Lisp_Object*) + 0x50 (/Code/emacs/src/eval.c:3114)
   map_keymap_canonical(Lisp_Object,map_keymap_function_t,Lisp_Object,void*)
   + 0x2b (/Code/emacs/src/keymap.c:608)
   ...

It appears that the BFont has already been closed. I think that the
driver is holding on to the pointer to the freed BFont
(into->be_font). This patch addresses this by setting be_font to NULL so
that this pointer will not be freed again.

The same thing applies to info->metrics and info->glyphs, since just
making this change to be_font wasn't enough to avoid crashes.

With this patch I can open and close as many frames as I want without
crashing.

I don't totally understand the interactions here, and I see there are
similar bugs in other font drivers with different workarounds. For
example, in Bug#16069 which I found from xfont.c:xfont_close, it seems
like there is an attempt to just not free the fonts when GC is invoked.

I think the solution in this patch seems a little simpler, but possibly
means that the fonts are initialized every time the frame count goes
from 0 to 1 or more instead of just once for the life of the daemon.

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.