Problem with constified xfrm_policy_afinfo structure

[marcin1j]

Hi

I'm running IPSec in tunnel mode on 3.10.1 box (Gentoo Hardened, amd64/SMP). When constify plugin is enabled the kernel crashes under heavy load.
It's pretty easy to spot that it's caused by null pointer dereference in function xfrm4_garbage_collect. Static structure xfrm4_policy_afinfo field garbage_collect seems to be null.

Now, what's interesting, field garbage_collect is initialized in function xfrm_policy_register_afinfo. I checked that adding some printks. After calling xfrm_policy_register_afinfo structure has the following values:

Code: Select all

[    0.358427] xfrm_policy_register_afinfo: Struct xfrm_policy_afinfo: family=2, dst_ops=8183c4c0, garbage_collect=81416ee0, dst_lookup=81415090, get_saddr=81415020, decode_session=814152e0, get_tos=81414df0, init_dst=0, init_path=81414e10, fill_dst=81414e20, backhole_router=813bfec0


Even after initialization has completed, each xfrm_alloc_dst call this structure retains all the field values:

Code: Select all

[  120.876704] xfrm_alloc_dst:              Struct xfrm_policy_afinfo: family=2, dst_ops=8183c4c0, garbage_collect=81416ee0, dst_lookup=81415090, get_saddr=81415020, decode_session=814152e0, get_tos=81414df0, init_dst=0, init_path=81414e10, fill_dst=81414e20, backhole_router=813bfec0


Unfortunately, when it comes to xfrm4_garbage_collect garbage_collect field seems to have zero value:

Code: Select all

[  176.176290] xfrm4_garbage_collect:       Struct xfrm_policy_afinfo: family=2, dst_ops=8183c4c0, garbage_collect=00000000, dst_lookup=81415090, get_saddr=81415020, decode_session=814152e0, get_tos=81414df0, init_dst=0, init_path=81414e10, fill_dst=81414e20, backhole_router=813bfec0


What's even more interesting, when I add condition bypassing call to garbage_collect if it's null, which prevents kernel panic, it seems that garbage_collect value in function xfrm_alloc_dst is still intact:

Code: Select all

[  301.313524] xfrm_alloc_dst:              Struct xfrm_policy_afinfo: family=2, dst_ops=8183c4c0, garbage_collect=81416ee0, dst_lookup=81415090, get_saddr=81415020, decode_session=814152e0, get_tos=81414df0, init_dst=0, init_path=81414e10, fill_dst=81414e20, backhole_router=813bfec0


Why is this happening? Is this some sort of CPU cache issue?

[PaX Team]

what happens here is that gcc optimizes the indirect call through xfrm4_policy_afinfo.garbage_collect into a direct call to its initialized value (NULL) because the structure is constified and gcc assumes that it cannot change later at runtime (a fair assumption in general but in this case this assumption doesn't hold). this whole xfrm code has some decade old sillyness in it (e.g., xfrm[46]_garbage_collect are inline but their only use is as an initializer so they won't ever be inlined anywhere) so i decided to rewrite it properly and still keep these structures constified. i'll also look through the code for similar instances, apparently some things fell through the cracks when i last tried.