On Wed, Jul 15, 2020 at 11:13:37AM +0200, Joerg Roedel wrote:> On Wed, Jul 15, 2020 at 10:47:52AM +0200, Peter Zijlstra wrote: > > On Tue, Jul 14, 2020 at 02:09:05PM +0200, Joerg Roedel wrote: > > > > > @@ -1028,6 +1036,16 @@ DEFINE_IDTENTRY_VC_SAFE_STACK(exc_vmm_communication) > > > struct ghcb *ghcb; > > > > > > lockdep_assert_irqs_disabled(); > > > + > > > + /* > > > + * #DB is special and needs to be handled outside of the intrumentation_begin()/end(). > > > + * Otherwise the #VC handler could be raised recursivly. > > > + */ > > > + if (error_code == SVM_EXIT_EXCP_BASE + X86_TRAP_DB) { > > > + vc_handle_trap_db(regs); > > > + return; > > > + } > > > + > > > instrumentation_begin(); > > > > Wait what?! That makes no sense what so ever. > > Then my understanding of intrumentation_begin/end() is wrong, I thought > that the kernel will forbid setting breakpoints before > instrumentation_begin(), which is necessary here because a break-point > in the #VC handler might cause recursive #VC-exceptions when #DB is > intercepted. > Maybe you can elaborate on why this makes no sense?Kernel avoids breakpoints in any noinstr text, irrespective of instrumentation_begin(). instrumentation_begin() merely allows one to call !noinstr functions.
On Wed, Jul 15, 2020 at 11:51:36AM +0200, Peter Zijlstra wrote:> On Wed, Jul 15, 2020 at 11:13:37AM +0200, Joerg Roedel wrote: > > Then my understanding of intrumentation_begin/end() is wrong, I thought > > that the kernel will forbid setting breakpoints before > > instrumentation_begin(), which is necessary here because a break-point > > in the #VC handler might cause recursive #VC-exceptions when #DB is > > intercepted. > > Maybe you can elaborate on why this makes no sense? > > Kernel avoids breakpoints in any noinstr text, irrespective of > instrumentation_begin(). > > instrumentation_begin() merely allows one to call !noinstr functions.Right, but the handler calls into various other functions. I actually started to annotate them all with noinstr, but that was a can of worms when calling into generic kernel functions. And the only problem with intrumentation in the #VC handler is the #VC-for-#DB exit-code, so I decided to only handle this one with instrumentation forbidden and allow it for the rest of the handler. Regards, Joerg
On Wed, Jul 15, 2020 at 12:08:08PM +0200, Joerg Roedel wrote:> On Wed, Jul 15, 2020 at 11:51:36AM +0200, Peter Zijlstra wrote: > > On Wed, Jul 15, 2020 at 11:13:37AM +0200, Joerg Roedel wrote: > > > Then my understanding of intrumentation_begin/end() is wrong, I thought > > > that the kernel will forbid setting breakpoints before > > > instrumentation_begin(), which is necessary here because a break-point > > > in the #VC handler might cause recursive #VC-exceptions when #DB is > > > intercepted. > > > Maybe you can elaborate on why this makes no sense? > > > > Kernel avoids breakpoints in any noinstr text, irrespective of > > instrumentation_begin(). > > > > instrumentation_begin() merely allows one to call !noinstr functions. > > Right, but the handler calls into various other functions. I actually > started to annotate them all with noinstr, but that was a can of worms > when calling into generic kernel functions. And the only problem with > intrumentation in the #VC handler is the #VC-for-#DB exit-code, so I > decided to only handle this one with instrumentation forbidden and allow > it for the rest of the handler.OK, then maybe change the comment to something like: /* * Handle #DB before calling any !noinstr code to avoid * recursive #DB. */ ?