From 3386bc8aed825e9f1f65ce38df4b109b2019b71a Mon Sep 17 00:00:00 2001 From: Andy Lutomirski Date: Mon, 4 Dec 2017 15:07:25 +0100 Subject: x86/entry/64: Create a per-CPU SYSCALL entry trampoline Handling SYSCALL is tricky: the SYSCALL handler is entered with every single register (except FLAGS), including RSP, live. It somehow needs to set RSP to point to a valid stack, which means it needs to save the user RSP somewhere and find its own stack pointer. The canonical way to do this is with SWAPGS, which lets us access percpu data using the %gs prefix. With PAGE_TABLE_ISOLATION-like pagetable switching, this is problematic. Without a scratch register, switching CR3 is impossible, so %gs-based percpu memory would need to be mapped in the user pagetables. Doing that without information leaks is difficult or impossible. Instead, use a different sneaky trick. Map a copy of the first part of the SYSCALL asm at a different address for each CPU. Now RIP varies depending on the CPU, so we can use RIP-relative memory access to access percpu memory. By putting the relevant information (one scratch slot and the stack address) at a constant offset relative to RIP, we can make SYSCALL work without relying on %gs. A nice thing about this approach is that we can easily switch it on and off if we want pagetable switching to be configurable. The compat variant of SYSCALL doesn't have this problem in the first place -- there are plenty of scratch registers, since we don't care about preserving r8-r15. This patch therefore doesn't touch SYSCALL32 at all. This patch actually seems to be a small speedup. With this patch, SYSCALL touches an extra cache line and an extra virtual page, but the pipeline no longer stalls waiting for SWAPGS. It seems that, at least in a tight loop, the latter outweights the former. Thanks to David Laight for an optimization tip. Signed-off-by: Andy Lutomirski Signed-off-by: Thomas Gleixner Reviewed-by: Thomas Gleixner Reviewed-by: Borislav Petkov Cc: Boris Ostrovsky Cc: Borislav Petkov Cc: Brian Gerst Cc: Dave Hansen Cc: Dave Hansen Cc: David Laight Cc: Denys Vlasenko Cc: Eduardo Valentin Cc: Greg KH Cc: H. Peter Anvin Cc: Josh Poimboeuf Cc: Juergen Gross Cc: Linus Torvalds Cc: Peter Zijlstra Cc: Rik van Riel Cc: Will Deacon Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: hughd@google.com Cc: keescook@google.com Link: https://lkml.kernel.org/r/20171204150606.403607157@linutronix.de Signed-off-by: Ingo Molnar --- arch/x86/entry/entry_64.S | 58 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 58 insertions(+) (limited to 'arch/x86/entry') diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S index 42a9379f7acb..2582984ffb4b 100644 --- a/arch/x86/entry/entry_64.S +++ b/arch/x86/entry/entry_64.S @@ -136,6 +136,64 @@ END(native_usergs_sysret64) * with them due to bugs in both AMD and Intel CPUs. */ + .pushsection .entry_trampoline, "ax" + +/* + * The code in here gets remapped into cpu_entry_area's trampoline. This means + * that the assembler and linker have the wrong idea as to where this code + * lives (and, in fact, it's mapped more than once, so it's not even at a + * fixed address). So we can't reference any symbols outside the entry + * trampoline and expect it to work. + * + * Instead, we carefully abuse %rip-relative addressing. + * _entry_trampoline(%rip) refers to the start of the remapped) entry + * trampoline. We can thus find cpu_entry_area with this macro: + */ + +#define CPU_ENTRY_AREA \ + _entry_trampoline - CPU_ENTRY_AREA_entry_trampoline(%rip) + +/* The top word of the SYSENTER stack is hot and is usable as scratch space. */ +#define RSP_SCRATCH CPU_ENTRY_AREA_tss + CPU_TSS_SYSENTER_stack + \ + SIZEOF_SYSENTER_stack - 8 + CPU_ENTRY_AREA + +ENTRY(entry_SYSCALL_64_trampoline) + UNWIND_HINT_EMPTY + swapgs + + /* Stash the user RSP. */ + movq %rsp, RSP_SCRATCH + + /* Load the top of the task stack into RSP */ + movq CPU_ENTRY_AREA_tss + TSS_sp1 + CPU_ENTRY_AREA, %rsp + + /* Start building the simulated IRET frame. */ + pushq $__USER_DS /* pt_regs->ss */ + pushq RSP_SCRATCH /* pt_regs->sp */ + pushq %r11 /* pt_regs->flags */ + pushq $__USER_CS /* pt_regs->cs */ + pushq %rcx /* pt_regs->ip */ + + /* + * x86 lacks a near absolute jump, and we can't jump to the real + * entry text with a relative jump. We could push the target + * address and then use retq, but this destroys the pipeline on + * many CPUs (wasting over 20 cycles on Sandy Bridge). Instead, + * spill RDI and restore it in a second-stage trampoline. + */ + pushq %rdi + movq $entry_SYSCALL_64_stage2, %rdi + jmp *%rdi +END(entry_SYSCALL_64_trampoline) + + .popsection + +ENTRY(entry_SYSCALL_64_stage2) + UNWIND_HINT_EMPTY + popq %rdi + jmp entry_SYSCALL_64_after_hwframe +END(entry_SYSCALL_64_stage2) + ENTRY(entry_SYSCALL_64) UNWIND_HINT_EMPTY /* -- cgit v1.2.1