| Message ID | 20180205151300.539874-2-arnd@arndb.de |
|---|---|
| State | New |
| Headers | show |
| Series | LTO: two more fixes | expand |
On Mon, Feb 05, 2018 at 04:12:52PM +0100, Arnd Bergmann wrote: > WARNING: vmlinux.o(.data+0x12e0): Section mismatch in reference from the variable pfkey_net_ops.lto_priv.2992 to the function .init.text:pfkey_net_init.lto_priv.2977() > The variable pfkey_net_ops.lto_priv.2992 references > the function __init pfkey_net_init.lto_priv.2977() > If the reference is valid then annotate the > variable with __init* or __refdata (see linux/init.h) or name the variable: > *_template, *_timer, *_sht, *_ops, *_probe, *_probe_one, *_console A better fix would be to ensure modpost always runs on already LD_FINALed objects, so it never sees LTO. Otherwise you would need to teach modpost about the LTO symbol table and some other magic. I did that once, but it turned out to be very ugly. -Andi
On Mon, Feb 5, 2018 at 4:50 PM, Andi Kleen <ak@linux.intel.com> wrote: > On Mon, Feb 05, 2018 at 04:12:52PM +0100, Arnd Bergmann wrote: >> WARNING: vmlinux.o(.data+0x12e0): Section mismatch in reference from the variable pfkey_net_ops.lto_priv.2992 to the function .init.text:pfkey_net_init.lto_priv.2977() >> The variable pfkey_net_ops.lto_priv.2992 references >> the function __init pfkey_net_init.lto_priv.2977() >> If the reference is valid then annotate the >> variable with __init* or __refdata (see linux/init.h) or name the variable: >> *_template, *_timer, *_sht, *_ops, *_probe, *_probe_one, *_console > > A better fix would be to ensure modpost always runs on already LD_FINALed > objects, so it never sees LTO. > > Otherwise you would need to teach modpost about the LTO symbol table > and some other magic. I did that once, but it turned out to be very > ugly. I'm not sure I understand what that means. Do you mean that with LD_FINAL, those symbol references are completely eliminated so we don't need to worry about them any more? I got about a dozen section mismatch errors with LTO in cases where the calling function gets a specialized version of a structure, and my oneline patch above addresses them all. Arnd
On Mon, Feb 05, 2018 at 04:59:32PM +0100, Arnd Bergmann wrote: > On Mon, Feb 5, 2018 at 4:50 PM, Andi Kleen <ak@linux.intel.com> wrote: > > On Mon, Feb 05, 2018 at 04:12:52PM +0100, Arnd Bergmann wrote: > >> WARNING: vmlinux.o(.data+0x12e0): Section mismatch in reference from the variable pfkey_net_ops.lto_priv.2992 to the function .init.text:pfkey_net_init.lto_priv.2977() > >> The variable pfkey_net_ops.lto_priv.2992 references > >> the function __init pfkey_net_init.lto_priv.2977() > >> If the reference is valid then annotate the > >> variable with __init* or __refdata (see linux/init.h) or name the variable: > >> *_template, *_timer, *_sht, *_ops, *_probe, *_probe_one, *_console > > > > A better fix would be to ensure modpost always runs on already LD_FINALed > > objects, so it never sees LTO. > > > > Otherwise you would need to teach modpost about the LTO symbol table > > and some other magic. I did that once, but it turned out to be very > > ugly. > > I'm not sure I understand what that means. Do you mean that with LD_FINAL, > those symbol references are completely eliminated so we don't need to worry > about them any more? Yes there should not be any .lto_priv references after LDFINAL, it will be just like any other ELF object. We also need to do the same thing before objtool. > > I got about a dozen section mismatch errors with LTO in cases where the > calling function gets a specialized version of a structure, and my oneline > patch above addresses them all. Not sure I follow here. .lto_priv should be anything LTO, not just some specific optimizations. -Andi
On Mon, Feb 5, 2018 at 5:22 PM, Andi Kleen <ak@linux.intel.com> wrote: > On Mon, Feb 05, 2018 at 04:59:32PM +0100, Arnd Bergmann wrote: >> On Mon, Feb 5, 2018 at 4:50 PM, Andi Kleen <ak@linux.intel.com> wrote: >> > On Mon, Feb 05, 2018 at 04:12:52PM +0100, Arnd Bergmann wrote: >> >> WARNING: vmlinux.o(.data+0x12e0): Section mismatch in reference from the variable pfkey_net_ops.lto_priv.2992 to the function .init.text:pfkey_net_init.lto_priv.2977() >> >> The variable pfkey_net_ops.lto_priv.2992 references >> >> the function __init pfkey_net_init.lto_priv.2977() >> >> If the reference is valid then annotate the >> >> variable with __init* or __refdata (see linux/init.h) or name the variable: >> >> *_template, *_timer, *_sht, *_ops, *_probe, *_probe_one, *_console >> > >> > A better fix would be to ensure modpost always runs on already LD_FINALed >> > objects, so it never sees LTO. >> > >> > Otherwise you would need to teach modpost about the LTO symbol table >> > and some other magic. I did that once, but it turned out to be very >> > ugly. >> >> I'm not sure I understand what that means. Do you mean that with LD_FINAL, >> those symbol references are completely eliminated so we don't need to worry >> about them any more? > > Yes there should not be any .lto_priv references after LDFINAL, it will > be just like any other ELF object. We also need to do the same > thing before objtool. >> I got about a dozen section mismatch errors with LTO in cases where the >> calling function gets a specialized version of a structure, and my oneline >> patch above addresses them all. > > Not sure I follow here. .lto_priv should be anything LTO, not just > some specific optimizations. I think it's a bit different from my first interpretation, this is what I see in the output of 'objdump -Dr vmlinux' after the final LTO link: b1da26f0 <amba_console.lto_priv.21118>: b1da26f0: 74747941 cmnmi r9, r4, ror r4 b1da26f4: 4d000000 andeq r0, r0, sp, asr #32 ... b1da2700: b0bdfa08 ldmeq sl!, {r4, r5, r7, r8, sl, fp, ip, sp, pc}^ b1da2704: 00000000 andeq r0, r0, r0 b1da2708: b0bcc790 strhls fp, [r7], #192 ; 0xc0 b1da270c: 00000000 andeq r0, r0, r0 b1da2710: b1a2bb18 ldmne fp!, {r0, r4, r5, r7, r9, sp, pc} b1da2714: 00000000 andeq r0, r0, r0 b1da2718: 0001ffff ; <UNDEFINED> instruction: 0xffff0100 b1da271c: 00000000 andeq r0, r0, r0 b1da2720: b1da0938 stmdacc r9, {r0, r4, r5, r7, r9, fp, ip, lr, pc} b1da2724: 00000000 andeq r0, r0, r0 b1da2728 <amba_console.lto_priv.21119>: b1da2728: 74747941 cmnmi r9, r4, ror r4 b1da272c: 4d410000 andeq r4, r0, sp, asr #2 ... b1da2738: b0be163c ldccc 14, cr11, [r6], {176} ; 0xb0 b1da273c: 00000000 andeq r0, r0, r0 b1da2740: b0bcc790 strhls fp, [r7], #192 ; 0xc0 b1da2744: 00000000 andeq r0, r0, r0 b1da2748: b1a2b8b8 ldmlt r8!, {r0, r4, r5, r7, r9, sp, pc} b1da274c: b1a2e6d8 stmiale r6!, {r0, r4, r5, r7, r9, sp, pc}^ b1da2750: 0011ffff ; <UNDEFINED> instruction: 0xffff1100 b1da2754: 00000000 andeq r0, r0, r0 b1da2758: b1d9fed8 ldmle lr!, {r0, r4, r5, r7, r8, fp, ip, lr, pc}^ b1da275c: 00000000 andeq r0, r0, r0 I got a warning from scripts/mod/modpost about vmlinux.o before it, which has relocations from each of these two into an init function. My interpretation now is that LTO adds the .lto_priv.21118 suffix for the two structures to disambiguate them, as one comes from amba-pl010.c and the other one is from amba-pl011.c. This means that the whitelist in modpost no longer works, as it checks the name against the "*_console" glob string. Almost all other symbols are not affected because the names of the static symbols are unique, and LTO gets to drop the suffix. Arnd
diff --git a/scripts/mod/modpost.c b/scripts/mod/modpost.c index f75224d59294..8a8a473cee08 100644 --- a/scripts/mod/modpost.c +++ b/scripts/mod/modpost.c @@ -1027,7 +1027,7 @@ static const struct sectioncheck sectioncheck[] = { .bad_tosec = { INIT_SECTIONS, NULL }, .mismatch = DATA_TO_ANY_INIT, .symbol_white_list = { - "*_template", "*_timer", "*_sht", "*_ops", + "*_template", "*_timer", "*_sht", "*_ops", "*_ops.lto_priv*", "*_probe", "*_probe_one", "*_console", NULL }, },
WARNING: vmlinux.o(.data+0x12e0): Section mismatch in reference from the variable pfkey_net_ops.lto_priv.2992 to the function .init.text:pfkey_net_init.lto_priv.2977() The variable pfkey_net_ops.lto_priv.2992 references the function __init pfkey_net_init.lto_priv.2977() If the reference is valid then annotate the variable with __init* or __refdata (see linux/init.h) or name the variable: *_template, *_timer, *_sht, *_ops, *_probe, *_probe_one, *_console Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- scripts/mod/modpost.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) -- 2.9.0