The branch optimization fixes in 2.6.21 introduced a bug in
atomic_sub_if_positive that causes it to return even when the sc
instruction fails. The result is that e.g. down_trylock becomes unreliable
as the semaphore counter is not always decremented.
Original MUA-shredded patch from Morten Larsen <mlarsen@broadcom.com>.
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
                "       beqz    %0, 2f                                  \n"
                "        subu   %0, %1, %3                              \n"
                "       .set    reorder                                 \n"
-               "1:                                                     \n"
                "       .subsection 2                                   \n"
                "2:     b       1b                                      \n"
                "       .previous                                       \n"
+               "1:                                                     \n"
                "       .set    mips0                                   \n"
                : "=&r" (result), "=&r" (temp), "=m" (v->counter)
                : "Ir" (i), "m" (v->counter)
                "       beqz    %0, 2f                                  \n"
                "        dsubu  %0, %1, %3                              \n"
                "       .set    reorder                                 \n"
-               "1:                                                     \n"
                "       .subsection 2                                   \n"
                "2:     b       1b                                      \n"
                "       .previous                                       \n"
+               "1:                                                     \n"
                "       .set    mips0                                   \n"
                : "=&r" (result), "=&r" (temp), "=m" (v->counter)
                : "Ir" (i), "m" (v->counter)