Re: [PATCH] x86/build32: Discard all orphaned sections

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 13/05/2020 16:15, Jan Beulich wrote:
>>> But it would be really nice if we had a way to
>>> flag the need to play with the linker script. Hence perhaps
>>> on new enough tool chains we indeed may want to make use of
>>> --orphan-handling= ? And then discard just .note and .note.*
>>> here?
>> The only valid option would be =error, but experimenting with that yields
>>
>> ld: error: unplaced orphan section `.comment' from `cmdline.o'
>> ld: error: unplaced orphan section `.note.GNU-stack' from `cmdline.o'
>> ld: error: unplaced orphan section `.note.gnu.property' from `cmdline.o'
>> ld: error: unplaced orphan section `.rel.got' from `cmdline.o'
>> ld: error: unplaced orphan section `.got' from `cmdline.o'
>> ld: error: unplaced orphan section `.got.plt' from `cmdline.o'
>> ld: error: unplaced orphan section `.iplt' from `cmdline.o'
>> ld: error: unplaced orphan section `.rel.iplt' from `cmdline.o'
>> ld: error: unplaced orphan section `.igot.plt' from `cmdline.o'
>>
>> which I think is going to get us massively bogged down in toolchain
>> specifics.  I'm not entirely convinced this would be a good move.
> That's ugly indeed; especially the .rel.* sections are worrying to
> appear there.

What is even more curious, most of them don't exist in cmdine.o

Section Headers:
  [Nr] Name              Type            Addr     Off    Size   ES Flg
Lk Inf Al
  [ 0]                   NULL            00000000 000000 000000 00     
0   0  0
  [ 1] .group            GROUP           00000000 000034 000008 04    
14  38  4
  [ 2] .group            GROUP           00000000 00003c 000008 04    
14  40  4
  [ 3] .text             PROGBITS        00000000 000044 00094a 00  AX 
0   0  1
  [ 4] .rel.text         REL             00000000 000e88 0001e8 08   I
14   3  4
  [ 5] .data             PROGBITS        00000000 00098e 000000 00  WA 
0   0  1
  [ 6] .bss              NOBITS          00000000 00098e 000000 00  WA 
0   0  1
  [ 7] .rodata           PROGBITS        00000000 000990 0000f3 00   A 
0   0  4
  [ 8] .rel.rodata       REL             00000000 001070 000120 08   I
14   7  4
  [ 9] .text.__x86.get_pc_thunk.ax PROGBITS        00000000 000a83
000004 00 AXG  0   0  1
  [10] .text.__x86.get_pc_thunk.bx PROGBITS        00000000 000a87
000004 00 AXG  0   0  1
  [11] .comment          PROGBITS        00000000 000a8b 00002d 01  MS 
0   0  1
  [12] .note.GNU-stack   PROGBITS        00000000 000ab8 000000 00     
0   0  1
  [13] .note.gnu.property NOTE            00000000 000ab8 00001c 00   A 
0   0  4
  [14] .symtab           SYMTAB          00000000 000ad4 000290 10    
15  36  4
  [15] .strtab           STRTAB          00000000 000d64 000124 00     
0   0  1
  [16] .shstrtab         STRTAB          00000000 001190 0000a7 00     
0   0  1

I suspect they are inserted by default as part of processing the
relocations, and end up empty.

With =warn rather than =error, we instead get

ld: warning: orphan section `.comment' from `cmdline.o' being placed in
section `.comment'
ld: warning: orphan section `.note.GNU-stack' from `cmdline.o' being
placed in section `.note.GNU-stack'
ld: warning: orphan section `.note.gnu.property' from `cmdline.o' being
placed in section `.note.gnu.property'
ld: warning: orphan section `.rel.got' from `cmdline.o' being placed in
section `.rel.dyn'
ld: warning: orphan section `.got' from `cmdline.o' being placed in
section `.got'
ld: warning: orphan section `.got.plt' from `cmdline.o' being placed in
section `.got.plt'
ld: warning: orphan section `.iplt' from `cmdline.o' being placed in
section `.iplt'
ld: warning: orphan section `.rel.iplt' from `cmdline.o' being placed in
section `.rel.dyn'
ld: warning: orphan section `.igot.plt' from `cmdline.o' being placed in
section `.igot.plt'

and

Section Headers:
  [Nr] Name              Type            Addr     Off    Size   ES Flg
Lk Inf Al
  [ 0]                   NULL            00000000 000000 000000 00     
0   0  0
  [ 1] .note.gnu.property NOTE            00000000 0000b4 00001c 00   A 
0   0  4
  [ 2] .text             PROGBITS        0000001c 0000d0 000a47 00 WAX 
0   0  4
  [ 3] .got.plt          PROGBITS        00000a64 000b18 00000c 04  WA 
0   0  4
  [ 4] .comment          PROGBITS        00000000 000b24 00002c 01  MS 
0   0  1
  [ 5] .symtab           SYMTAB          00000000 000b50 000230 10     
6  31  4
  [ 6] .strtab           STRTAB          00000000 000d80 000124 00     
0   0  1
  [ 7] .shstrtab         STRTAB          00000000 000ea4 000046 00     
0   0  1

in cmdline.lnk, so the .rel.* sections have been dropped overall.  I
think the =error logic is simply at an unhelpful position during processing.

> Hence patch as is Acked-by: Jan Beulich <jbeulich@xxxxxxxx>

Thanks,

As I say, I have a plan to replace all of this completely when a bit
more of kbuild is in place.

~Andrew