linux/arch/arm64/kernel/syscall.c
Mark Rutland 23529049c6 arm64: entry: fix non-NMI user<->kernel transitions
When built with PROVE_LOCKING, NO_HZ_FULL, and CONTEXT_TRACKING_FORCE
will WARN() at boot time that interrupts are enabled when we call
context_tracking_user_enter(), despite the DAIF flags indicating that
IRQs are masked.

The problem is that we're not tracking IRQ flag changes accurately, and
so lockdep believes interrupts are enabled when they are not (and
vice-versa). We can shuffle things so to make this more accurate. For
kernel->user transitions there are a number of constraints we need to
consider:

1) When we call __context_tracking_user_enter() HW IRQs must be disabled
   and lockdep must be up-to-date with this.

2) Userspace should be treated as having IRQs enabled from the PoV of
   both lockdep and tracing.

3) As context_tracking_user_enter() stops RCU from watching, we cannot
   use RCU after calling it.

4) IRQ flag tracing and lockdep have state that must be manipulated
   before RCU is disabled.

... with similar constraints applying for user->kernel transitions, with
the ordering reversed.

The generic entry code has enter_from_user_mode() and
exit_to_user_mode() helpers to handle this. We can't use those directly,
so we add arm64 copies for now (without the instrumentation markers
which aren't used on arm64). These replace the existing user_exit() and
user_exit_irqoff() calls spread throughout handlers, and the exception
unmasking is left as-is.

Note that:

* The accounting for debug exceptions from userspace now happens in
  el0_dbg() and ret_to_user(), so this is removed from
  debug_exception_enter() and debug_exception_exit(). As
  user_exit_irqoff() wakes RCU, the userspace-specific check is removed.

* The accounting for syscalls now happens in el0_svc(),
  el0_svc_compat(), and ret_to_user(), so this is removed from
  el0_svc_common(). This does not adversely affect the workaround for
  erratum 1463225, as this does not depend on any of the state tracking.

* In ret_to_user() we mask interrupts with local_daif_mask(), and so we
  need to inform lockdep and tracing. Here a trace_hardirqs_off() is
  sufficient and safe as we have not yet exited kernel context and RCU
  is usable.

* As PROVE_LOCKING selects TRACE_IRQFLAGS, the ifdeferry in entry.S only
  needs to check for the latter.

* EL0 SError handling will be dealt with in a subsequent patch, as this
  needs to be treated as an NMI.

Prior to this patch, booting an appropriately-configured kernel would
result in spats as below:

| DEBUG_LOCKS_WARN_ON(lockdep_hardirqs_enabled())
| WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5280 check_flags.part.54+0x1dc/0x1f0
| Modules linked in:
| CPU: 2 PID: 1 Comm: init Not tainted 5.10.0-rc3 #3
| Hardware name: linux,dummy-virt (DT)
| pstate: 804003c5 (Nzcv DAIF +PAN -UAO -TCO BTYPE=--)
| pc : check_flags.part.54+0x1dc/0x1f0
| lr : check_flags.part.54+0x1dc/0x1f0
| sp : ffff80001003bd80
| x29: ffff80001003bd80 x28: ffff66ce801e0000
| x27: 00000000ffffffff x26: 00000000000003c0
| x25: 0000000000000000 x24: ffffc31842527258
| x23: ffffc31842491368 x22: ffffc3184282d000
| x21: 0000000000000000 x20: 0000000000000001
| x19: ffffc318432ce000 x18: 0080000000000000
| x17: 0000000000000000 x16: ffffc31840f18a78
| x15: 0000000000000001 x14: ffffc3184285c810
| x13: 0000000000000001 x12: 0000000000000000
| x11: ffffc318415857a0 x10: ffffc318406614c0
| x9 : ffffc318415857a0 x8 : ffffc31841f1d000
| x7 : 647261685f706564 x6 : ffffc3183ff7c66c
| x5 : ffff66ce801e0000 x4 : 0000000000000000
| x3 : ffffc3183fe00000 x2 : ffffc31841500000
| x1 : e956dc24146b3500 x0 : 0000000000000000
| Call trace:
|  check_flags.part.54+0x1dc/0x1f0
|  lock_is_held_type+0x10c/0x188
|  rcu_read_lock_sched_held+0x70/0x98
|  __context_tracking_enter+0x310/0x350
|  context_tracking_enter.part.3+0x5c/0xc8
|  context_tracking_user_enter+0x6c/0x80
|  finish_ret_to_user+0x2c/0x13cr

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: James Morse <james.morse@arm.com>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20201130115950.22492-8-mark.rutland@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
2020-11-30 12:11:38 +00:00

214 lines
5.7 KiB
C

// SPDX-License-Identifier: GPL-2.0
#include <linux/compiler.h>
#include <linux/context_tracking.h>
#include <linux/errno.h>
#include <linux/nospec.h>
#include <linux/ptrace.h>
#include <linux/syscalls.h>
#include <asm/daifflags.h>
#include <asm/debug-monitors.h>
#include <asm/fpsimd.h>
#include <asm/syscall.h>
#include <asm/thread_info.h>
#include <asm/unistd.h>
long compat_arm_syscall(struct pt_regs *regs, int scno);
long sys_ni_syscall(void);
static long do_ni_syscall(struct pt_regs *regs, int scno)
{
#ifdef CONFIG_COMPAT
long ret;
if (is_compat_task()) {
ret = compat_arm_syscall(regs, scno);
if (ret != -ENOSYS)
return ret;
}
#endif
return sys_ni_syscall();
}
static long __invoke_syscall(struct pt_regs *regs, syscall_fn_t syscall_fn)
{
return syscall_fn(regs);
}
static void invoke_syscall(struct pt_regs *regs, unsigned int scno,
unsigned int sc_nr,
const syscall_fn_t syscall_table[])
{
long ret;
if (scno < sc_nr) {
syscall_fn_t syscall_fn;
syscall_fn = syscall_table[array_index_nospec(scno, sc_nr)];
ret = __invoke_syscall(regs, syscall_fn);
} else {
ret = do_ni_syscall(regs, scno);
}
if (is_compat_task())
ret = lower_32_bits(ret);
regs->regs[0] = ret;
}
static inline bool has_syscall_work(unsigned long flags)
{
return unlikely(flags & _TIF_SYSCALL_WORK);
}
int syscall_trace_enter(struct pt_regs *regs);
void syscall_trace_exit(struct pt_regs *regs);
#ifdef CONFIG_ARM64_ERRATUM_1463225
DECLARE_PER_CPU(int, __in_cortex_a76_erratum_1463225_wa);
static void cortex_a76_erratum_1463225_svc_handler(void)
{
u32 reg, val;
if (!unlikely(test_thread_flag(TIF_SINGLESTEP)))
return;
if (!unlikely(this_cpu_has_cap(ARM64_WORKAROUND_1463225)))
return;
__this_cpu_write(__in_cortex_a76_erratum_1463225_wa, 1);
reg = read_sysreg(mdscr_el1);
val = reg | DBG_MDSCR_SS | DBG_MDSCR_KDE;
write_sysreg(val, mdscr_el1);
asm volatile("msr daifclr, #8");
isb();
/* We will have taken a single-step exception by this point */
write_sysreg(reg, mdscr_el1);
__this_cpu_write(__in_cortex_a76_erratum_1463225_wa, 0);
}
#else
static void cortex_a76_erratum_1463225_svc_handler(void) { }
#endif /* CONFIG_ARM64_ERRATUM_1463225 */
static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
const syscall_fn_t syscall_table[])
{
unsigned long flags = current_thread_info()->flags;
regs->orig_x0 = regs->regs[0];
regs->syscallno = scno;
/*
* BTI note:
* The architecture does not guarantee that SPSR.BTYPE is zero
* on taking an SVC, so we could return to userspace with a
* non-zero BTYPE after the syscall.
*
* This shouldn't matter except when userspace is explicitly
* doing something stupid, such as setting PROT_BTI on a page
* that lacks conforming BTI/PACIxSP instructions, falling
* through from one executable page to another with differing
* PROT_BTI, or messing with BTYPE via ptrace: in such cases,
* userspace should not be surprised if a SIGILL occurs on
* syscall return.
*
* So, don't touch regs->pstate & PSR_BTYPE_MASK here.
* (Similarly for HVC and SMC elsewhere.)
*/
cortex_a76_erratum_1463225_svc_handler();
local_daif_restore(DAIF_PROCCTX);
if (system_supports_mte() && (flags & _TIF_MTE_ASYNC_FAULT)) {
/*
* Process the asynchronous tag check fault before the actual
* syscall. do_notify_resume() will send a signal to userspace
* before the syscall is restarted.
*/
regs->regs[0] = -ERESTARTNOINTR;
return;
}
if (has_syscall_work(flags)) {
/*
* The de-facto standard way to skip a system call using ptrace
* is to set the system call to -1 (NO_SYSCALL) and set x0 to a
* suitable error code for consumption by userspace. However,
* this cannot be distinguished from a user-issued syscall(-1)
* and so we must set x0 to -ENOSYS here in case the tracer doesn't
* issue the skip and we fall into trace_exit with x0 preserved.
*
* This is slightly odd because it also means that if a tracer
* sets the system call number to -1 but does not initialise x0,
* then x0 will be preserved for all system calls apart from a
* user-issued syscall(-1). However, requesting a skip and not
* setting the return value is unlikely to do anything sensible
* anyway.
*/
if (scno == NO_SYSCALL)
regs->regs[0] = -ENOSYS;
scno = syscall_trace_enter(regs);
if (scno == NO_SYSCALL)
goto trace_exit;
}
invoke_syscall(regs, scno, sc_nr, syscall_table);
/*
* The tracing status may have changed under our feet, so we have to
* check again. However, if we were tracing entry, then we always trace
* exit regardless, as the old entry assembly did.
*/
if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
local_daif_mask();
flags = current_thread_info()->flags;
if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP)) {
/*
* We're off to userspace, where interrupts are
* always enabled after we restore the flags from
* the SPSR.
*/
trace_hardirqs_on();
return;
}
local_daif_restore(DAIF_PROCCTX);
}
trace_exit:
syscall_trace_exit(regs);
}
static inline void sve_user_discard(void)
{
if (!system_supports_sve())
return;
clear_thread_flag(TIF_SVE);
/*
* task_fpsimd_load() won't be called to update CPACR_EL1 in
* ret_to_user unless TIF_FOREIGN_FPSTATE is still set, which only
* happens if a context switch or kernel_neon_begin() or context
* modification (sigreturn, ptrace) intervenes.
* So, ensure that CPACR_EL1 is already correct for the fast-path case.
*/
sve_user_disable();
}
void do_el0_svc(struct pt_regs *regs)
{
sve_user_discard();
el0_svc_common(regs, regs->regs[8], __NR_syscalls, sys_call_table);
}
#ifdef CONFIG_COMPAT
void do_el0_svc_compat(struct pt_regs *regs)
{
el0_svc_common(regs, regs->regs[7], __NR_compat_syscalls,
compat_sys_call_table);
}
#endif