On Tue, Mar 24, 2015 at 5:12 PM, Dave Hansen <dave.han...@intel.com> wrote: > On 03/24/2015 04:52 PM, Andy Lutomirski wrote: >> On Tue, Mar 24, 2015 at 4:42 PM, Dave Hansen <dave.han...@intel.com> wrote: >>> On 03/24/2015 03:28 PM, Andy Lutomirski wrote: >>>> Your function appears to be getting it for write (I assume that's what >>>> the unlazy_fpu is for), so I'd rather have it called >>>> tsk_get_xsave_field_for_write or something like that. >>> >>> It should be entirely read-only. >>> >>> For MPX (the only user of get_xsave_addr() iirc), we are only worried >>> about getting the status codes (and addresses) out of the bndstatus >>> register and making sure that the kernel-recorded bounds directory >>> address matches the bndcfgu (configuration) register. >>> >>> We don't ever write to the registers. >> >> So why are you unlazying it? > > Oleg actually suggested it. > >> IIUC, the xstae for current can be in one of three logical states: >> >> 1. Live in CPU regs. The in-memory copy is garbage and the state is >> in CPU regs. >> 2. Lazy. The in-memory copy and the CPU regs match. Writing to >> either copy is illegal. >> 3. In memory only. Writing to the in-memory copy is safe. >> >> IIUC, you want to read the xstate, do you're okay with #2 or #3. This >> would be tsk_get_xsave_field_for_read in my terminology. >> >> If you want to write the xstate, you'd need to be in state #3, which >> would be tsk_get_xsave_field_for_write. >> >> IIUC, unlazy_fpu just moves from from state 2 to 3. > > I won't completely claim to understand what's going on with the FPU > code, but I think your analysis is a bit off. > > unlazy_fpu() does __save_init_fpu() which (among other things) calls > xsave to dump the CPU registers to memory. That doesn't make any sense > to do if "The in-memory copy and the CPU regs match." > > IOW, unlazy_fpu() is called when the in-memory copy is garbage and takes > us to a state where we can look at the in-memory copy.
I think that __save_init_fpu (called by unlazy_fpu) does that, but __thread_fpu_end calls __thread_clear_has_fpu, which, in turn, zaps fpu_owner_task, which will force an unnecessary xrstor. Or maybe not if we have further bugs. Holy crap these functions are poorly named. Also, what, if anything, guarantees that fpu_owner_task is set on entry to userspace? Do we even need it to be set? Oleg, help? --Andy -- Andy Lutomirski AMA Capital Management, LLC -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/