On Fri, Dec 02, 2022 at 04:35:43PM -0800, Rick Edgecombe wrote: > From: Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx> > > The CPU performs "shadow stack accesses" when it expects to encounter > shadow stack mappings. These accesses can be implicit (via CALL/RET > instructions) or explicit (instructions like WRSS). > > Shadow stacks accesses to shadow-stack mappings can see faults in normal, > valid operation just like regular accesses to regular mappings. Shadow > stacks need some of the same features like delayed allocation, swap and > copy-on-write. The kernel needs to use faults to implement those features. > > The architecture has concepts of both shadow stack reads and shadow stack > writes. Any shadow stack access to non-shadow stack memory will generate > a fault with the shadow stack error code bit set. You lost me here: by "shadow stack access to non-shadow stack memory" you mean the explicit one using WRU*SS? > This means that, unlike normal write protection, the fault handler needs > to create a type of memory that can be written to (with instructions that > generate shadow stack writes), even to fulfill a read access. So in the > case of COW memory, the COW needs to take place even with a shadow stack > read. I guess I'm missing an example here: are we talking here about a user process getting its shadow stack pages allocated and them being COW first and on the first shstk operation, it would generate that fault? > @@ -1331,6 +1345,30 @@ void do_user_addr_fault(struct pt_regs *regs, > > perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); > > + /* > + * When a page becomes COW it changes from a shadow stack permissioned Unknown word [permissioned] in comment. > + * page (Write=0,Dirty=1) to (Write=0,Dirty=0,CoW=1), which is simply > + * read-only to the CPU. When shadow stack is enabled, a RET would > + * normally pop the shadow stack by reading it with a "shadow stack > + * read" access. However, in the COW case the shadow stack memory does > + * not have shadow stack permissions, it is read-only. So it will > + * generate a fault. > + * > + * For conventionally writable pages, a read can be serviced with a > + * read only PTE, and COW would not have to happen. But for shadow > + * stack, there isn't the concept of read-only shadow stack memory. > + * If it is shadow stack permissioned, it can be modified via CALL and Ditto. > + * RET instructions. So COW needs to happen before any memory can be > + * mapped with shadow stack permissions. > + * > + * Shadow stack accesses (read or write) need to be serviced with > + * shadow stack permissioned memory, so in the case of a shadow stack Is this some new formulation I haven't heard about yet? "Permissioned <something>"? > + * read access, treat it as a WRITE fault so both COW will happen and > + * the write fault path will tickle maybe_mkwrite() and map the memory > + * shadow stack. > + */ > + if (error_code & X86_PF_SHSTK) > + flags |= FAULT_FLAG_WRITE; > if (error_code & X86_PF_WRITE) > flags |= FAULT_FLAG_WRITE; > if (error_code & X86_PF_INSTR) > -- > 2.17.1 > -- Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette