From 9a78ebba126036f6ebdd15ffbe50ffbd8ccd84dd Mon Sep 17 00:00:00 2001
From: Xi Ruoyao <xry111@xry111.site>
Date: Mon, 12 Aug 2024 18:23:47 +0800
Subject: [PATCH 108/123] LoongArch: Fix DT_RELR and relaxation interaction

Due to the way BFD implements DT_RELR as a part of relaxation, if in a
relax trip size_relative_relocs has changed the layout, when
relax_section runs only the vma of the section being relaxed is
guaranteed to be updated.  Then bad thing can happen.  For example, when
linking an auxilary program _freeze_module in the Python 3.12.4 building
system (with PGO + LTO), before sizing the .relr.dyn section, the vma of
.text is 30560 and the vma of .rodata is 2350944; in the final
executable the vma of .text is 36704 and the vma of .rodata is 2357024.
The vma increase is expected because .relr.dyn is squashed somewhere
before .text.  But size_relative_relocs may see the state in which .text
is at 36704 but .rodata "is" still at 2350944.  Thus the distance
between .text and .rodata can be under-estimated and overflowing
R_LARCH_PCREL20_S2 reloc can be created.

To avoid this issue, if size_relative_relocs is updating the size of
.relr.dyn, just supress the normal relaxation in this relax trip.  In
this situation size_relative_relocs should have been demending the next
relax trip, so the normal relaxation can happen in the next trip.

I tried to make a reduced test case but failed.

Signed-off-by: Xi Ruoyao <xry111@xry111.site>
---
 bfd/elfnn-loongarch.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index 73eea0f9..0c499c47 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -121,6 +121,12 @@ struct loongarch_elf_link_hash_table
 
   /* Layout recomputation count.  */
   bfd_size_type relr_layout_iter;
+
+  /* In BFD DT_RELR is implemented as a "relaxation."  If in a relax trip
+     size_relative_relocs is updating the layout, relax_section may see
+     a partially updated state (some sections have vma updated but the
+     others do not), and it's unsafe to do the normal relaxation.  */
+  bool layout_mutating_for_relr;
 };
 
 struct loongarch_elf_section_data
@@ -2210,6 +2216,8 @@ loongarch_elf_size_relative_relocs (struct bfd_link_info *info,
          *need_layout = false;
        }
     }
+
+  htab->layout_mutating_for_relr = *need_layout;
   return true;
 }
 
@@ -5256,6 +5264,13 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
     return true;
 
   struct loongarch_elf_link_hash_table *htab = loongarch_elf_hash_table (info);
+
+  /* It may happen that some sections have updated vma but the others do
+     not.  Go to the next relax trip (size_relative_relocs should have
+     been demending another relax trip anyway).  */
+  if (htab->layout_mutating_for_relr)
+    return true;
+
   if (bfd_link_relocatable (info)
       || sec->sec_flg0
       || (sec->flags & SEC_RELOC) == 0
-- 
2.33.0