On Sun, Apr 29, 2018 at 02:05:07PM +0100, Marc Zyngier wrote: > On Sun, 29 Apr 2018 14:34:32 +0200 > Christoffer Dall <christoffer.dall@xxxxxxx> wrote: > > > On Fri, Apr 27, 2018 at 03:51:02PM +0100, Marc Zyngier wrote: > > > Proxying the cpuif accesses at EL2 makes use of vcpu_data_guest_to_host > > > and co, which check the endianness, which call into vcpu_read_sys_reg... > > > which isn't mapped at EL2 (it was inlined before, and got moved OoL > > > with the VHE optimizations). > > > > I thought we relied on static inline functions to always be inlined, but > > apparently not? Does this mean we have potential other bugs looming > > depending on the mood of the compiler, or was there something special > > that went wrong here? > > We do rely on that behaviour. And that was the case until you moved > vcpu_read_sys_reg() to be entirely out of line (see d47533dab9f5). ah, now I see, kvm_vcpu_is_be() calls vcpu_read_sys_reg(), I somehow missed this even though you spelled it out in the commit message. Sorry. > > At that point, kvm_vcpu_is_be() becomes a death trap. > > We missed it for two reasons: > - It was only indirectly called, making it quite hard to notice the > potential breakage > - Nobody gives a damn about 64k pages, specially on something like Juno > > What we'd need is a way to find cross-section calls (text -> hyp-text > should allowed, but not the reverse). We already have similar things in > the kernel, it is probably only a matter of reusing the infrastructure > for our own purpose. > That would be good indeed... Thanks for the further explanation. -Christoffer
