[PATCH 0/3] PE delay import fixes

[Jan Beulich]

On 22.04.2025 19:45, Jeremy Drake wrote:
> Ping after 1 week.

Just to assure you - this is on my list of things to look at (if no-one else
beats me), but I can't promise when I'll be able to get to it. One further
delaying factor is that I'll first need to read up on what the spec actually
permits (or even requires) to be done for delay imports.

Jan

>  My mail client doesn't like to keep In-Reply-To on
> subject changes, so here are links to the 3 patches:
> 
> https://inbox.sourceware.org/binutils/9e48774c-4b02-b763-aaf7-37133399e1fe@jdrake.com/T/#u
> https://inbox.sourceware.org/binutils/ea41e347-1fe8-2ca4-9d74-3bbe9ebf7017@jdrake.com/T/#u
> https://inbox.sourceware.org/binutils/346348f7-5f9e-de25-6ada-a8c9e59df702@jdrake.com/T/#u
> 
> On Mon, 14 Apr 2025, Jeremy Drake wrote:
> 
>> This patch series fixes delay import functionality, which was broken in
>> 2.44 when the .idata section was made read-only.  The first patch, by LIU
>> Hao, moved all of the delay import data that was previously in .idata into
>> .didat.  My followup patch moves the read-only portions of this data into
>> .rdata in the linker script, leaving only .didat$5 in the new .didat
>> section.  This,  conveniently, is compatible with the
>> IMAGE_GUARD_PROTECT_DELAYLOAD_IAT|IMAGE_GUARD_DELAYLOAD_IAT_IN_ITS_OWN_SECTION
>> flags in the load config .GuardFlags member.  The third patch fills in the
>> delay import directory in the PE headers.
>>
>> There was some debate as to whether it would be better to squash these
>> patches.  I feel it's easier to squash patches later than to un-squash
>> them, so I figured to send them separate and if requested I can squash
>> them for a v2.
>>
>> To test the delayload guard flags:
>> == main.c ==
>> #include <assert.h>
>> #include <stdio.h>
>> #include <winsock2.h>
>> #include <windows.h>
>>
>> __attribute__((__used__))
>> const IMAGE_LOAD_CONFIG_DIRECTORY _load_config_used = {
>>   .Size = sizeof(_load_config_used),
>>   .GuardFlags = 0x3000, /*IMAGE_GUARD_PROTECT_DELAYLOAD_IAT|IMAGE_GUARD_DELAYLOAD_IAT_IN_ITS_OWN_SECTION*/
>> };
>>
>> int global_var = 42;
>>
>> int main(void) {
>>   unsigned short ret;
>>   setvbuf(stdout, NULL, _IONBF, 0);
>>   printf("MAIN: starting\n");
>>   printf("MAIN: incremented global var = %d\n", ++global_var);
>>   printf("MAIN: about to call htons\n");
>>   ret = htons(0x1234);
>>   printf("MAIN: htons returned %hx\n", ret);
>>   printf("MAIN: incremented global var = %d\n", ++global_var);
>>   printf("MAIN: OK\n");
>>   return 0;
>> }
>>
>> == WS2_32.def (x86_64) ==
>> LIBRARY WS2_32.DLL
>> EXPORTS
>> htons
>>
>> == build ==
>> dlltool --input-def "ws2_32.def" --output-delaylib ws2_32.lib
>> gcc -o main.exe main.c ws2_32.lib -ldloadhelper
>> ./main.exe
>>
>> (-ldloadhelper is part of mingw-w64, and hooks up the delay load helper
>> with the kernel32 ResolveDelayLoadedAPI function)
>> (for i386 you'd need htons@4, and maybe the --kill-at flag for dlltool?)
>>
>> The first patch actually fixes PR 14339 and PR 32675.
>>
>> Jeremy Drake (2):
>>   move read-only delayimp data into .rdata
>>   bfd: populate delay import directory in PE header.
>>
>> LIU Hao (1):
>>   bfd,ld,dlltool: Emit delay-load import data into its own section
>>
>>  bfd/coffgen.c         |  1 +
>>  bfd/pe-aarch64.c      |  2 +
>>  bfd/pe-arm.c          |  2 +
>>  bfd/pe-i386.c         |  2 +
>>  bfd/pe-x86_64.c       |  2 +
>>  bfd/peXXigen.c        | 53 ++++++++++++++++++++++-
>>  bfd/pei-aarch64.c     |  2 +
>>  bfd/pei-arm.c         |  2 +
>>  bfd/pei-i386.c        |  2 +
>>  bfd/pei-loongarch64.c |  2 +
>>  bfd/pei-riscv64.c     |  2 +
>>  bfd/pei-x86_64.c      |  2 +
>>  bfd/syms.c            |  1 +
>>  binutils/dlltool.c    | 98 ++++++++++++++++++++++++++++++++++++-------
>>  ld/scripttempl/pe.sc  | 29 ++++++++++++-
>>  ld/scripttempl/pep.sc | 31 +++++++++++++-
>>  16 files changed, 214 insertions(+), 19 deletions(-)
>>
>>