xtensa: make fast_unaligned store restartable

fast_unaligned may encounter DTLB miss or SEGFAULT during the store
emulation. Don't update epc1 and lcount until after the store emulation
is complete, so that the faulting store instruction could be replayed.
Remove duplicate code handling zero overhead loops and calculate new
epc1 and lcount in one place.

Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>

diff --git a/arch/xtensa/kernel/align.S b/arch/xtensa/kernel/align.S
index 904f32f..2c7c13d 100644 (file)
--- a/arch/xtensa/kernel/align.S
+++ b/arch/xtensa/kernel/align.S
@@ -277,18 +277,6 @@ ENTRY(fast_unaligned)
        /* Set target register. */
 
 1:
-
-#if XCHAL_HAVE_LOOPS
-       rsr     a5, lend                # check if we reached LEND
-       bne     a7, a5, 1f
-       rsr     a5, lcount              # and LCOUNT != 0
-       beqz    a5, 1f
-       addi    a5, a5, -1              # decrement LCOUNT and set
-       rsr     a7, lbeg                # set PC to LBEGIN
-       wsr     a5, lcount
-#endif
-
-1:     wsr     a7, epc1                # skip load instruction
        extui   a4, a4, INSN_T, 4       # extract target register
        movi    a5, .Lload_table
        addx8   a4, a4, a5
@@ -358,17 +346,6 @@ ENTRY(fast_unaligned)
        /* Get memory address */
 
 1:
-#if XCHAL_HAVE_LOOPS
-       rsr     a4, lend                # check if we reached LEND
-       bne     a7, a4, 1f
-       rsr     a4, lcount              # and LCOUNT != 0
-       beqz    a4, 1f
-       addi    a4, a4, -1              # decrement LCOUNT and set
-       rsr     a7, lbeg                # set PC to LBEGIN
-       wsr     a4, lcount
-#endif
-
-1:     wsr     a7, epc1                # skip store instruction
        movi    a4, ~3
        and     a4, a4, a8              # align memory address
 
@@ -380,25 +357,25 @@ ENTRY(fast_unaligned)
 #endif
 
        __ssa8r a8
-       __src_b a7, a5, a6              # lo-mask  F..F0..0 (BE) 0..0F..F (LE)
+       __src_b a8, a5, a6              # lo-mask  F..F0..0 (BE) 0..0F..F (LE)
        __src_b a6, a6, a5              # hi-mask  0..0F..F (BE) F..F0..0 (LE)
 #ifdef UNALIGNED_USER_EXCEPTION
        l32e    a5, a4, -8
 #else
        l32i    a5, a4, 0               # load lower address word
 #endif
-       and     a5, a5, a7              # mask
-       __sh    a7, a3                  # shift value
-       or      a5, a5, a7              # or with original value
+       and     a5, a5, a8              # mask
+       __sh    a8, a3                  # shift value
+       or      a5, a5, a8              # or with original value
 #ifdef UNALIGNED_USER_EXCEPTION
        s32e    a5, a4, -8
-       l32e    a7, a4, -4
+       l32e    a8, a4, -4
 #else
        s32i    a5, a4, 0               # store
-       l32i    a7, a4, 4               # same for upper address word
+       l32i    a8, a4, 4               # same for upper address word
 #endif
        __sl    a5, a3
-       and     a6, a7, a6
+       and     a6, a8, a6
        or      a6, a6, a5
 #ifdef UNALIGNED_USER_EXCEPTION
        s32e    a6, a4, -4
@@ -406,9 +383,19 @@ ENTRY(fast_unaligned)
        s32i    a6, a4, 4
 #endif
 
-       /* Done. restore stack and return */
-
 .Lexit:
+#if XCHAL_HAVE_LOOPS
+       rsr     a4, lend                # check if we reached LEND
+       bne     a7, a4, 1f
+       rsr     a4, lcount              # and LCOUNT != 0
+       beqz    a4, 1f
+       addi    a4, a4, -1              # decrement LCOUNT and set
+       rsr     a7, lbeg                # set PC to LBEGIN
+       wsr     a4, lcount
+#endif
+
+1:     wsr     a7, epc1                # skip emulated instruction
+
        movi    a4, 0
        rsr     a3, excsave1
        s32i    a4, a3, EXC_TABLE_FIXUP