This is an archive of the discontinued LLVM Phabricator instance.

[Sanitizers] Get link map on FreeBSD via documented API
ClosedPublic

Authored by dim on Feb 4 2020, 1:05 PM.

Details

Summary

Instead of hand-crafting an offset into the structure returned by
dlopen(3) to get at the link map, use the documented API. This is
described in dlinfo(3): by calling it with RTLD_DI_LINKMAP, the
dynamic linker ensures the right address is returned.

Event Timeline

dim created this revision.Feb 4 2020, 1:05 PM
Herald added projects: Restricted Project, Restricted Project. · View Herald TranscriptFeb 4 2020, 1:05 PM
Herald added a subscriber: Restricted Project. · View Herald Transcript
vitalybuka added inline comments.
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.cpp
96

should be

void *__sanitizer_get_link_map_by_dlopen_handle(void* handle) {

clang-format it

dim updated this revision to Diff 242425.Feb 4 2020, 1:54 PM

Cast return value to link_map to avoid call site errors.

dim updated this revision to Diff 242426.Feb 4 2020, 1:55 PM

Formatting.

Harbormaster completed remote builds in B45720: Diff 242426.

I will mirror it in NetBSD whatever will be decided here.

MaskRay added inline comments.Feb 4 2020, 3:15 PM
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.cpp
96

Should be

void *__sanitizer_get_link_map_by_dlopen_handle(void *handle) {

It will be nice to include in the description what systems were not supported by hard coding 560, if such information can be obtained relatively easily.

dim marked an inline comment as done.Feb 4 2020, 10:12 PM
dim added inline comments.
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.cpp
96

Eh, what do you mean? It's already that form?

systems were not supported by hard coding 560

In the NetBSD case this offset changes few times a year.

emaste added a comment.Feb 6 2020, 5:36 AM

It will be nice to include in the description what systems were not supported by hard coding 560, if such information can be obtained relatively easily.

I don't quite understand this question; hard coding is definitely a hack and we should prefer the supported method for finding the information, if the system provides it.

emaste added inline comments.Feb 6 2020, 5:37 AM
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.h
29–30

I don't think the comment is applicable now

dim marked an inline comment as done.Feb 6 2020, 6:07 AM
dim added inline comments.
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.h
29–30

Well, it is still informative in the sense that it tells you the handle parameter (which is apparently the return value from dlopen) does not directly point to the desired link map information

Would you rather delete the comment, or replace it with something else? If the latter, what? :)

dim added a comment.Feb 6 2020, 6:11 AM

Btw, note that according to the Linux dlinfo() manpage, it also has RTLD_DI_LINKMAP as a documented way to get at the link map. Maybe we should generalize this in a follow-up review?

krytarowski added inline comments.Feb 6 2020, 6:11 AM
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.cpp
97

Actually please add internal_dlinfo() alongside internal_pid() and others.

It's sufficient to redirect it for all supported OSs (Linux, FreeBSD, Solaris, NetBSD) to dlinfo() for now and stub /* unimplemented */ for others.

On NetBSD we need to use it indirectly and handle it later.

dim marked an inline comment as done.Feb 6 2020, 6:35 AM
dim added inline comments.
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.cpp
97

Hm, I don't see any internal_pid() anywhere, but there are quite a lot of internal_ functions in compiler-rt/lib/sanitizer_common/sanitizer_libc.cpp. Is that the right location? It seems that these are all more like hand-rolled versions of known libc functions, not some sort of 'forwarders'.

krytarowski added inline comments.Feb 6 2020, 7:52 AM
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.cpp
97

Right, I was thinking about internal_getpid(). We wrap other functions too like sysctl().

dim updated this revision to Diff 243005.Feb 6 2020, 2:20 PM
  • Rebase onto reformatted files.
  • Add internal_dlinfo() for FreeBSD and NetBSD, with stubs for other platforms.
  • Add GET_LINK_MAP_BY_DLOPEN_HANDLE() macro for NetBSD.
krytarowski added inline comments.Feb 7 2020, 12:22 AM
compiler-rt/lib/sanitizer_common/sanitizer_netbsd.cpp
269 ↗(On Diff #243005)

So as we implemented it for NetBSD here. Please put it in the expected final form.

	​  DEFINE__REAL(int, dlinfo, void *a, int b, void *c);
	​  return _REAL(dlinfo, handle, request, p);
emaste added inline comments.Feb 7 2020, 6:03 AM
compiler-rt/lib/sanitizer_common/sanitizer_platform_limits_freebsd.h
29–30

I would delete it - the macro name is pretty explicit about obtaining the link map from the dlopen handle :-)

dim updated this revision to Diff 243225.Feb 7 2020, 10:29 AM

Add better NetBSD implementation of internal_dlinfo().

dim updated this revision to Diff 243227.Feb 7 2020, 10:32 AM

Remove stray newline.

Harbormaster completed remote builds in B45959: Diff 243227.
dim added a comment.Feb 9 2020, 11:12 PM

Ping, can we get this committed then? :)

krytarowski accepted this revision.Feb 10 2020, 3:14 AM
This revision is now accepted and ready to land.Feb 10 2020, 3:14 AM
This revision was automatically updated to reflect the committed changes.
plotfi added a subscriber: plotfi.Feb 10 2020, 2:02 PM

This breaks in our downstream runtimes builds for armv7-none-linux-androideabi16. FYI

dim added a comment.Feb 10 2020, 2:14 PM

This breaks in our downstream runtimes builds for armv7-none-linux-androideabi16. FYI

Can you point me to some build logs? (I didn't receive any failure mails from bots yet.)

In D73990#1868085, @dim wrote:

This breaks in our downstream runtimes builds for armv7-none-linux-androideabi16. FYI

Can you point me to some build logs? (I didn't receive any failure mails from bots yet.)

Our bots are not public, but I have a patch that I can post shortly. If you check the android NDK you'll see that it does not implement dlinfo.

In D73990#1868085, @dim wrote:

This breaks in our downstream runtimes builds for armv7-none-linux-androideabi16. FYI

Can you point me to some build logs? (I didn't receive any failure mails from bots yet.)

Posted a fix at D74358

dim added a comment.Feb 10 2020, 2:47 PM
In D73990#1868085, @dim wrote:

This breaks in our downstream runtimes builds for armv7-none-linux-androideabi16. FYI

Can you point me to some build logs? (I didn't receive any failure mails from bots yet.)

Posted a fix at D74358

I found this lab builder, which gave a similar error: http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-android

Committed a slightly different version in rG52f2df1ecdd79cc550b694ab280f3b0396d7cf9a, thanks for the heads up!