ARM: 7748/1: oabi: handle faults when loading swi instruction from userspace

Running an OABI_COMPAT kernel on an SMP platform can lead to fun and
games with page aging.

If one CPU issues a swi instruction immediately before another CPU
decides to mkold the page containing the swi instruction, then we will
fault attempting to load the instruction during the vector_swi handler
in order to retrieve its immediate field. Since this fault is not
currently dealt with by our exception tables, this results in a panic:

  Unable to handle kernel paging request at virtual address 4020841c
  pgd = c490c000
  [4020841c] *pgd=84451831, *pte=bf05859d, *ppte=00000000
  Internal error: Oops: 17 [#1] PREEMPT SMP ARM
  Modules linked in: hid_sony(O)
  CPU: 1    Tainted: G        W  O  (3.4.0-perf-gf496dca-01162-gcbcc62b #1)
  PC is at vector_swi+0x28/0x88
  LR is at 0x40208420

This patch wraps all of the swi instruction loads with the USER macro
and provides a shared exception table entry which simply rewinds the
saved user PC and returns from the system call (without setting tbl, so
there's no worries with tracing or syscall restarting). Returning to
userspace will re-enter the page fault handler, from where we will
probably send SIGSEGV to the current task.

Reported-by: Wang, Yalin <yalin.wang@sonymobile.com>
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

diff --git a/arch/arm/kernel/entry-common.S b/arch/arm/kernel/entry-common.S
index bc5bc0a..4bc816a 100644 (file)
--- a/arch/arm/kernel/entry-common.S
+++ b/arch/arm/kernel/entry-common.S
@@ -362,6 +362,16 @@ ENTRY(vector_swi)
        str     r0, [sp, #S_OLD_R0]             @ Save OLD_R0
        zero_fp
 
+#ifdef CONFIG_ALIGNMENT_TRAP
+       ldr     ip, __cr_alignment
+       ldr     ip, [ip]
+       mcr     p15, 0, ip, c1, c0              @ update control register
+#endif
+
+       enable_irq
+       ct_user_exit
+       get_thread_info tsk
+
        /*
         * Get the system call number.
         */
@@ -375,9 +385,9 @@ ENTRY(vector_swi)
 #ifdef CONFIG_ARM_THUMB
        tst     r8, #PSR_T_BIT
        movne   r10, #0                         @ no thumb OABI emulation
-       ldreq   r10, [lr, #-4]                  @ get SWI instruction
+ USER( ldreq   r10, [lr, #-4]          )       @ get SWI instruction
 #else
-       ldr     r10, [lr, #-4]                  @ get SWI instruction
+ USER( ldr     r10, [lr, #-4]          )       @ get SWI instruction
 #endif
 #ifdef CONFIG_CPU_ENDIAN_BE8
        rev     r10, r10                        @ little endian instruction
@@ -392,22 +402,13 @@ ENTRY(vector_swi)
        /* Legacy ABI only, possibly thumb mode. */
        tst     r8, #PSR_T_BIT                  @ this is SPSR from save_user_regs
        addne   scno, r7, #__NR_SYSCALL_BASE    @ put OS number in
-       ldreq   scno, [lr, #-4]
+ USER( ldreq   scno, [lr, #-4]         )
 
 #else
        /* Legacy ABI only. */
-       ldr     scno, [lr, #-4]                 @ get SWI instruction
+ USER( ldr     scno, [lr, #-4]         )       @ get SWI instruction
 #endif
 
-#ifdef CONFIG_ALIGNMENT_TRAP
-       ldr     ip, __cr_alignment
-       ldr     ip, [ip]
-       mcr     p15, 0, ip, c1, c0              @ update control register
-#endif
-       enable_irq
-       ct_user_exit
-
-       get_thread_info tsk
        adr     tbl, sys_call_table             @ load syscall table pointer
 
 #if defined(CONFIG_OABI_COMPAT)
@@ -442,6 +443,21 @@ local_restart:
        eor     r0, scno, #__NR_SYSCALL_BASE    @ put OS number back
        bcs     arm_syscall     
        b       sys_ni_syscall                  @ not private func
+
+#if defined(CONFIG_OABI_COMPAT) || !defined(CONFIG_AEABI)
+       /*
+        * We failed to handle a fault trying to access the page
+        * containing the swi instruction, but we're not really in a
+        * position to return -EFAULT. Instead, return back to the
+        * instruction and re-enter the user fault handling path trying
+        * to page it in. This will likely result in sending SEGV to the
+        * current task.
+        */
+9001:
+       sub     lr, lr, #4
+       str     lr, [sp, #S_PC]
+       b       ret_fast_syscall
+#endif
 ENDPROC(vector_swi)
 
        /*