Details

Reviewers

dchagin
emaste

Group Reviewers

Linux Emulation

Commits

rS351783: Unbreak Linux binaries linked against new glibc, such as the ones

Summary

Fix running linux binaries linked against newer glibc. Without
it they segfault on startup in a hard to reproduce way - try running
"/bin/echo 1 1 1 1 1 1" with varying amounts of "1"s; first do
"sysctl compat.linux.osrelease='5.0.0'".

The reason for this is not known; it's more of a hack than a proper fix.
There's a new compat.linux.align_stack sysctl to disable it, just in case.

Diff Detail

Repository

rS FreeBSD src repository - subversion

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

trasz created this revision.Jun 18 2019, 7:56 PM

Herald added a subscriber: imp. · View Herald TranscriptJun 18 2019, 7:56 PM

Harbormaster completed remote builds in B24922: Diff 58773.Jun 18 2019, 7:56 PM

trasz added reviewers: dchagin, Linux Emulation.Jun 18 2019, 7:57 PM

trasz edited the summary of this revision. (Show Details)Jun 18 2019, 8:00 PM

trasz added inline comments.Jun 18 2019, 8:06 PM

sys/amd64/linux/linux_sysvec.c
349 ↗	(On Diff #58773)	FWIW, 8 instead of 24 works too.

val_packett.cool added a subscriber: val_packett.cool.Jun 18 2019, 9:19 PM

trasz mentioned this in D20757: Bump Linux version to 3.2.0.Jun 25 2019, 2:52 PM

Do you have more details about working/non-working versions? I.e., what glibc version works?

Also is it the case that /bin/echo 1 1 1 1 1... consistently fails with some number of 1s and runs successfully with other cases? Like 1 1 1 works, 1 1 1 1 fails, 1 1 1 1 1 works?

In D20687#448843, @emaste wrote:

Do you have more details about working/non-working versions? I.e., what glibc version works?

Not really, I'm afraid - the one from CentOS 7 (latest one) works, the one from Ubuntu doesn't.

In D20687#448848, @emaste wrote:

Also is it the case that /bin/echo 1 1 1 1 1... consistently fails with some number of 1s and runs successfully with other cases? Like 1 1 1 works, 1 1 1 1 fails, 1 1 1 1 1 works?

Exactly. Basically: the vectp must not end in 0x0.

trasz edited the summary of this revision. (Show Details)Sep 2 2019, 5:45 PM

Prettify, add sysctl.

Harbormaster completed remote builds in B26231: Diff 61560.Sep 2 2019, 5:55 PM

trasz mentioned this in rS351723: Bump Linux version to 3.2.0. Without it, binaries linked against.Sep 2 2019, 6:10 PM

I suspect I know what's going on. If vectp % 16 == 8 then when long argc is stuck in front of it per the ABI requirements, the stack is aligned to 16 bytes. I suspect a change was made to require 16-byte alignment as recommend by Intel. A possibly relevant thread (I've only skimmed it) https://lore.kernel.org/patchwork/patch/749568/

Based on @brooks' comment let's drop the sysctl.

What do you think of a formulation like ((vectp + 8) & ~0xF) - 8

Apply comments from brooks and emaste.

Harbormaster completed remote builds in B26244: Diff 61576.Sep 3 2019, 11:18 AM

brooks added inline comments.Sep 3 2019, 11:41 AM

sys/amd64/linux/linux_sysvec.c
347 ↗	(On Diff #61576)	This should probably say something like: ...glibc depends on a 16-byte stack alignment. One "long argc" will be prepended later.