On 24.09.2024 20:36, Andrew Cooper wrote: > On 23/09/2024 2:03 pm, Jan Beulich wrote: >> On 23.09.2024 12:14, Alejandro Vallejo wrote: >>> On Fri Sep 20, 2024 at 3:12 PM BST, Roger Pau Monné wrote: >>>> On Wed, Sep 18, 2024 at 02:05:54PM +0100, Alejandro Vallejo wrote: >>>>> Moves sti directly after the cr2 read and immediately after the #PF >>>>> handler. >>>> I think you need to add some context about why this is needed, iow: >>>> avoid corrupting %cr2 if a nested 3PF happens. >>> I can send a v3 with: >>> >>> ``` >>> Hitting a page fault clobbers %cr2, so if a page fault is handled while >>> handling a previous page fault then %cr2 will hold the address of the >>> latter >>> fault rather than the former. This patch makes the page fault path delay >>> re-enabling IRQs until %cr2 has been read in order to ensure it stays >>> consistent. >> And under what conditions would we experience #PF while already processing >> an earlier #PF? If an interrupt kicks in, that's not supposed to by raising >> any #PF itself. Which isn't to say that the change isn't worthwhile to make, >> but it would be nice if it was explicit whether there are active issues, or >> whether this is merely to be on the safe side going forward. > > My understanding is that this came from code inspection, not an active > issue. > > The same is true for %dr6 and #DB, and MSR_XFD_ERR and #NM. > > I think we can safely agree to veto the use of AMX in the #NM handler, > and IST exceptions don't re-enable interrupts[1], so #PF is the only > problem case. > > Debug keys happen off the back of plain IRQs, and we can get #PF when > interrogating guest stacks.
Hmm, yes, this looks like a case that is actively being fixed here. Wants mentioning, likely wants a respective Fixes: tag, and then also wants backporting. > Also, I'm far from certain we're safe to > spurious #PF's from updating Xen mappings, so I think there are a bunch > of risky corner cases that we might be exposed to. Spurious #PF are possible, but __page_fault_type() explicitly excludes the in_irq() case. Jan
