target/linux/generic/patches-3.3/001-ARM-7668-1-fix-memset-related-crashes-caused-by-rece.patch

   1 From 455bd4c430b0c0a361f38e8658a0d6cb469942b5 Mon Sep 17 00:00:00 2001
   2 From: Ivan Djelic <ivan.djelic@parrot.com>
   3 Date: Wed, 6 Mar 2013 20:09:27 +0100
   4 Subject: [PATCH] ARM: 7668/1: fix memset-related crashes caused by recent GCC
   5  (4.7.2) optimizations
   6
   7 Recent GCC versions (e.g. GCC-4.7.2) perform optimizations based on
   8 assumptions about the implementation of memset and similar functions.
   9 The current ARM optimized memset code does not return the value of
  10 its first argument, as is usually expected from standard implementations.
  11
  12 For instance in the following function:
  13
  14 void debug_mutex_lock_common(struct mutex *lock, struct mutex_waiter *waiter)
  15 {
  16         memset(waiter, MUTEX_DEBUG_INIT, sizeof(*waiter));
  17         waiter->magic = waiter;
  18         INIT_LIST_HEAD(&waiter->list);
  19 }
  20
  21 compiled as:
  22
  23 800554d0 <debug_mutex_lock_common>:
  24 800554d0:       e92d4008        push    {r3, lr}
  25 800554d4:       e1a00001        mov     r0, r1
  26 800554d8:       e3a02010        mov     r2, #16 ; 0x10
  27 800554dc:       e3a01011        mov     r1, #17 ; 0x11
  28 800554e0:       eb04426e        bl      80165ea0 <memset>
  29 800554e4:       e1a03000        mov     r3, r0
  30 800554e8:       e583000c        str     r0, [r3, #12]
  31 800554ec:       e5830000        str     r0, [r3]
  32 800554f0:       e5830004        str     r0, [r3, #4]
  33 800554f4:       e8bd8008        pop     {r3, pc}
  34
  35 GCC assumes memset returns the value of pointer 'waiter' in register r0; causing
  36 register/memory corruptions.
  37
  38 This patch fixes the return value of the assembly version of memset.
  39 It adds a 'mov' instruction and merges an additional load+store into
  40 existing load/store instructions.
  41 For ease of review, here is a breakdown of the patch into 4 simple steps:
  42
  43 Step 1
  44 ======
  45 Perform the following substitutions:
  46 ip -> r8, then
  47 r0 -> ip,
  48 and insert 'mov ip, r0' as the first statement of the function.
  49 At this point, we have a memset() implementation returning the proper result,
  50 but corrupting r8 on some paths (the ones that were using ip).
  51
  52 Step 2
  53 ======
  54 Make sure r8 is saved and restored when (! CALGN(1)+0) == 1:
  55
  56 save r8:
  57 -       str     lr, [sp, #-4]!
  58 +       stmfd   sp!, {r8, lr}
  59
  60 and restore r8 on both exit paths:
  61 -       ldmeqfd sp!, {pc}               @ Now <64 bytes to go.
  62 +       ldmeqfd sp!, {r8, pc}           @ Now <64 bytes to go.
  63 (...)
  64         tst     r2, #16
  65         stmneia ip!, {r1, r3, r8, lr}
  66 -       ldr     lr, [sp], #4
  67 +       ldmfd   sp!, {r8, lr}
  68
  69 Step 3
  70 ======
  71 Make sure r8 is saved and restored when (! CALGN(1)+0) == 0:
  72
  73 save r8:
  74 -       stmfd   sp!, {r4-r7, lr}
  75 +       stmfd   sp!, {r4-r8, lr}
  76
  77 and restore r8 on both exit paths:
  78         bgt     3b
  79 -       ldmeqfd sp!, {r4-r7, pc}
  80 +       ldmeqfd sp!, {r4-r8, pc}
  81 (...)
  82         tst     r2, #16
  83         stmneia ip!, {r4-r7}
  84 -       ldmfd   sp!, {r4-r7, lr}
  85 +       ldmfd   sp!, {r4-r8, lr}
  86
  87 Step 4
  88 ======
  89 Rewrite register list "r4-r7, r8" as "r4-r8".
  90
  91 Signed-off-by: Ivan Djelic <ivan.djelic@parrot.com>
  92 Reviewed-by: Nicolas Pitre <nico@linaro.org>
  93 Signed-off-by: Dirk Behme <dirk.behme@gmail.com>
  94 Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
  95 ---
  96  arch/arm/lib/memset.S | 85 ++++++++++++++++++++++++++-------------------------
  97  1 file changed, 44 insertions(+), 41 deletions(-)
  98
  99 diff --git a/arch/arm/lib/memset.S b/arch/arm/lib/memset.S
 100 index 650d5923ab83..d912e7397ecc 100644
 101 --- a/arch/arm/lib/memset.S
 102 +++ b/arch/arm/lib/memset.S
 103 @@ -19,9 +19,9 @@
 104  1:     subs    r2, r2, #4              @ 1 do we have enough
 105         blt     5f                      @ 1 bytes to align with?
 106         cmp     r3, #2                  @ 1
 107 -       strltb  r1, [r0], #1            @ 1
 108 -       strleb  r1, [r0], #1            @ 1
 109 -       strb    r1, [r0], #1            @ 1
 110 +       strltb  r1, [ip], #1            @ 1
 111 +       strleb  r1, [ip], #1            @ 1
 112 +       strb    r1, [ip], #1            @ 1
 113         add     r2, r2, r3              @ 1 (r2 = r2 - (4 - r3))
 114  /*
 115   * The pointer is now aligned and the length is adjusted.  Try doing the
 116 @@ -29,10 +29,14 @@
 117   */
 118
 119  ENTRY(memset)
 120 -       ands    r3, r0, #3              @ 1 unaligned?
 121 +/*
 122 + * Preserve the contents of r0 for the return value.
 123 + */
 124 +       mov     ip, r0
 125 +       ands    r3, ip, #3              @ 1 unaligned?
 126         bne     1b                      @ 1
 127  /*
 128 - * we know that the pointer in r0 is aligned to a word boundary.
 129 + * we know that the pointer in ip is aligned to a word boundary.
 130   */
 131         orr     r1, r1, r1, lsl #8
 132         orr     r1, r1, r1, lsl #16
 133 @@ -43,29 +47,28 @@ ENTRY(memset)
 134  #if ! CALGN(1)+0
 135
 136  /*
 137 - * We need an extra register for this loop - save the return address and
 138 - * use the LR
 139 + * We need 2 extra registers for this loop - use r8 and the LR
 140   */
 141 -       str     lr, [sp, #-4]!
 142 -       mov     ip, r1
 143 +       stmfd   sp!, {r8, lr}
 144 +       mov     r8, r1
 145         mov     lr, r1
 146
 147  2:     subs    r2, r2, #64
 148 -       stmgeia r0!, {r1, r3, ip, lr}   @ 64 bytes at a time.
 149 -       stmgeia r0!, {r1, r3, ip, lr}
 150 -       stmgeia r0!, {r1, r3, ip, lr}
 151 -       stmgeia r0!, {r1, r3, ip, lr}
 152 +       stmgeia ip!, {r1, r3, r8, lr}   @ 64 bytes at a time.
 153 +       stmgeia ip!, {r1, r3, r8, lr}
 154 +       stmgeia ip!, {r1, r3, r8, lr}
 155 +       stmgeia ip!, {r1, r3, r8, lr}
 156         bgt     2b
 157 -       ldmeqfd sp!, {pc}               @ Now <64 bytes to go.
 158 +       ldmeqfd sp!, {r8, pc}           @ Now <64 bytes to go.
 159  /*
 160   * No need to correct the count; we're only testing bits from now on
 161   */
 162         tst     r2, #32
 163 -       stmneia r0!, {r1, r3, ip, lr}
 164 -       stmneia r0!, {r1, r3, ip, lr}
 165 +       stmneia ip!, {r1, r3, r8, lr}
 166 +       stmneia ip!, {r1, r3, r8, lr}
 167         tst     r2, #16
 168 -       stmneia r0!, {r1, r3, ip, lr}
 169 -       ldr     lr, [sp], #4
 170 +       stmneia ip!, {r1, r3, r8, lr}
 171 +       ldmfd   sp!, {r8, lr}
 172
 173  #else
 174
 175 @@ -74,54 +77,54 @@ ENTRY(memset)
 176   * whole cache lines at once.
 177   */
 178
 179 -       stmfd   sp!, {r4-r7, lr}
 180 +       stmfd   sp!, {r4-r8, lr}
 181         mov     r4, r1
 182         mov     r5, r1
 183         mov     r6, r1
 184         mov     r7, r1
 185 -       mov     ip, r1
 186 +       mov     r8, r1
 187         mov     lr, r1
 188
 189         cmp     r2, #96
 190 -       tstgt   r0, #31
 191 +       tstgt   ip, #31
 192         ble     3f
 193
 194 -       and     ip, r0, #31
 195 -       rsb     ip, ip, #32
 196 -       sub     r2, r2, ip
 197 -       movs    ip, ip, lsl #(32 - 4)
 198 -       stmcsia r0!, {r4, r5, r6, r7}
 199 -       stmmiia r0!, {r4, r5}
 200 -       tst     ip, #(1 << 30)
 201 -       mov     ip, r1
 202 -       strne   r1, [r0], #4
 203 +       and     r8, ip, #31
 204 +       rsb     r8, r8, #32
 205 +       sub     r2, r2, r8
 206 +       movs    r8, r8, lsl #(32 - 4)
 207 +       stmcsia ip!, {r4, r5, r6, r7}
 208 +       stmmiia ip!, {r4, r5}
 209 +       tst     r8, #(1 << 30)
 210 +       mov     r8, r1
 211 +       strne   r1, [ip], #4
 212
 213  3:     subs    r2, r2, #64
 214 -       stmgeia r0!, {r1, r3-r7, ip, lr}
 215 -       stmgeia r0!, {r1, r3-r7, ip, lr}
 216 +       stmgeia ip!, {r1, r3-r8, lr}
 217 +       stmgeia ip!, {r1, r3-r8, lr}
 218         bgt     3b
 219 -       ldmeqfd sp!, {r4-r7, pc}
 220 +       ldmeqfd sp!, {r4-r8, pc}
 221
 222         tst     r2, #32
 223 -       stmneia r0!, {r1, r3-r7, ip, lr}
 224 +       stmneia ip!, {r1, r3-r8, lr}
 225         tst     r2, #16
 226 -       stmneia r0!, {r4-r7}
 227 -       ldmfd   sp!, {r4-r7, lr}
 228 +       stmneia ip!, {r4-r7}
 229 +       ldmfd   sp!, {r4-r8, lr}
 230
 231  #endif
 232
 233  4:     tst     r2, #8
 234 -       stmneia r0!, {r1, r3}
 235 +       stmneia ip!, {r1, r3}
 236         tst     r2, #4
 237 -       strne   r1, [r0], #4
 238 +       strne   r1, [ip], #4
 239  /*
 240   * When we get here, we've got less than 4 bytes to zero.  We
 241   * may have an unaligned pointer as well.
 242   */
 243  5:     tst     r2, #2
 244 -       strneb  r1, [r0], #1
 245 -       strneb  r1, [r0], #1
 246 +       strneb  r1, [ip], #1
 247 +       strneb  r1, [ip], #1
 248         tst     r2, #1
 249 -       strneb  r1, [r0], #1
 250 +       strneb  r1, [ip], #1
 251         mov     pc, lr
 252  ENDPROC(memset)
 253 --
 254 1.8.3.2
 255