summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorDavid S. Miller <davem@davemloft.net>2005-09-08 14:37:53 -0700
committerDavid S. Miller <davem@davemloft.net>2005-09-08 14:37:53 -0700
commit4d803fcdcd97dd346d4b39c3b76e5879cead8a31 (patch)
treef30c8caa998261fc0983121021184f0f6cc555b4
parent1b11d78cf87a7014f96e5b7fa2e1233cc8081a00 (diff)
downloadlinux-4d803fcdcd97dd346d4b39c3b76e5879cead8a31.tar.bz2
[SPARC64]: Inline membar()'s again.
Since GCC has to emit a call and a delay slot to the out-of-line "membar" routines in arch/sparc64/lib/mb.S it is much better to just do the necessary predicted branch inline instead as: ba,pt %xcc, 1f membar #whatever 1: instead of the current: call membar_foo dslot because this way GCC is not required to allocate a stack frame if the function can be a leaf function. This also makes this bug fix easier to backport to 2.4.x Signed-off-by: David S. Miller <davem@davemloft.net>
-rw-r--r--arch/sparc64/kernel/sparc64_ksyms.c9
-rw-r--r--arch/sparc64/lib/Makefile2
-rw-r--r--arch/sparc64/lib/mb.S73
-rw-r--r--include/asm-sparc64/system.h49
4 files changed, 43 insertions, 90 deletions
diff --git a/arch/sparc64/kernel/sparc64_ksyms.c b/arch/sparc64/kernel/sparc64_ksyms.c
index d89fc24808d3..7d9a0f6c437d 100644
--- a/arch/sparc64/kernel/sparc64_ksyms.c
+++ b/arch/sparc64/kernel/sparc64_ksyms.c
@@ -403,12 +403,3 @@ EXPORT_SYMBOL(xor_vis_4);
EXPORT_SYMBOL(xor_vis_5);
EXPORT_SYMBOL(prom_palette);
-
-/* memory barriers */
-EXPORT_SYMBOL(mb);
-EXPORT_SYMBOL(rmb);
-EXPORT_SYMBOL(wmb);
-EXPORT_SYMBOL(membar_storeload);
-EXPORT_SYMBOL(membar_storeload_storestore);
-EXPORT_SYMBOL(membar_storeload_loadload);
-EXPORT_SYMBOL(membar_storestore_loadstore);
diff --git a/arch/sparc64/lib/Makefile b/arch/sparc64/lib/Makefile
index 6201f1040982..40dbeec7e5d6 100644
--- a/arch/sparc64/lib/Makefile
+++ b/arch/sparc64/lib/Makefile
@@ -12,7 +12,7 @@ lib-y := PeeCeeI.o copy_page.o clear_page.o strlen.o strncmp.o \
U1memcpy.o U1copy_from_user.o U1copy_to_user.o \
U3memcpy.o U3copy_from_user.o U3copy_to_user.o U3patch.o \
copy_in_user.o user_fixup.o memmove.o \
- mcount.o ipcsum.o rwsem.o xor.o find_bit.o delay.o mb.o
+ mcount.o ipcsum.o rwsem.o xor.o find_bit.o delay.o
lib-$(CONFIG_DEBUG_SPINLOCK) += debuglocks.o
lib-$(CONFIG_HAVE_DEC_LOCK) += dec_and_lock.o
diff --git a/arch/sparc64/lib/mb.S b/arch/sparc64/lib/mb.S
deleted file mode 100644
index 4004f748619f..000000000000
--- a/arch/sparc64/lib/mb.S
+++ /dev/null
@@ -1,73 +0,0 @@
-/* mb.S: Out of line memory barriers.
- *
- * Copyright (C) 2005 David S. Miller (davem@davemloft.net)
- */
-
- /* These are here in an effort to more fully work around
- * Spitfire Errata #51. Essentially, if a memory barrier
- * occurs soon after a mispredicted branch, the chip can stop
- * executing instructions until a trap occurs. Therefore, if
- * interrupts are disabled, the chip can hang forever.
- *
- * It used to be believed that the memory barrier had to be
- * right in the delay slot, but a case has been traced
- * recently wherein the memory barrier was one instruction
- * after the branch delay slot and the chip still hung. The
- * offending sequence was the following in sym_wakeup_done()
- * of the sym53c8xx_2 driver:
- *
- * call sym_ccb_from_dsa, 0
- * movge %icc, 0, %l0
- * brz,pn %o0, .LL1303
- * mov %o0, %l2
- * membar #LoadLoad
- *
- * The branch has to be mispredicted for the bug to occur.
- * Therefore, we put the memory barrier explicitly into a
- * "branch always, predicted taken" delay slot to avoid the
- * problem case.
- */
-
- .text
-
-99: retl
- nop
-
- .globl mb
-mb: ba,pt %xcc, 99b
- membar #LoadLoad | #LoadStore | #StoreStore | #StoreLoad
- .size mb, .-mb
-
- .globl rmb
-rmb: ba,pt %xcc, 99b
- membar #LoadLoad
- .size rmb, .-rmb
-
- .globl wmb
-wmb: ba,pt %xcc, 99b
- membar #StoreStore
- .size wmb, .-wmb
-
- .globl membar_storeload
-membar_storeload:
- ba,pt %xcc, 99b
- membar #StoreLoad
- .size membar_storeload, .-membar_storeload
-
- .globl membar_storeload_storestore
-membar_storeload_storestore:
- ba,pt %xcc, 99b
- membar #StoreLoad | #StoreStore
- .size membar_storeload_storestore, .-membar_storeload_storestore
-
- .globl membar_storeload_loadload
-membar_storeload_loadload:
- ba,pt %xcc, 99b
- membar #StoreLoad | #LoadLoad
- .size membar_storeload_loadload, .-membar_storeload_loadload
-
- .globl membar_storestore_loadstore
-membar_storestore_loadstore:
- ba,pt %xcc, 99b
- membar #StoreStore | #LoadStore
- .size membar_storestore_loadstore, .-membar_storestore_loadstore
diff --git a/include/asm-sparc64/system.h b/include/asm-sparc64/system.h
index 5e94c05dc2fc..b5417529f6f1 100644
--- a/include/asm-sparc64/system.h
+++ b/include/asm-sparc64/system.h
@@ -28,13 +28,48 @@ enum sparc_cpu {
#define ARCH_SUN4C_SUN4 0
#define ARCH_SUN4 0
-extern void mb(void);
-extern void rmb(void);
-extern void wmb(void);
-extern void membar_storeload(void);
-extern void membar_storeload_storestore(void);
-extern void membar_storeload_loadload(void);
-extern void membar_storestore_loadstore(void);
+/* These are here in an effort to more fully work around Spitfire Errata
+ * #51. Essentially, if a memory barrier occurs soon after a mispredicted
+ * branch, the chip can stop executing instructions until a trap occurs.
+ * Therefore, if interrupts are disabled, the chip can hang forever.
+ *
+ * It used to be believed that the memory barrier had to be right in the
+ * delay slot, but a case has been traced recently wherein the memory barrier
+ * was one instruction after the branch delay slot and the chip still hung.
+ * The offending sequence was the following in sym_wakeup_done() of the
+ * sym53c8xx_2 driver:
+ *
+ * call sym_ccb_from_dsa, 0
+ * movge %icc, 0, %l0
+ * brz,pn %o0, .LL1303
+ * mov %o0, %l2
+ * membar #LoadLoad
+ *
+ * The branch has to be mispredicted for the bug to occur. Therefore, we put
+ * the memory barrier explicitly into a "branch always, predicted taken"
+ * delay slot to avoid the problem case.
+ */
+#define membar_safe(type) \
+do { __asm__ __volatile__("ba,pt %%xcc, 1f\n\t" \
+ " membar " type "\n" \
+ "1:\n" \
+ : : : "memory"); \
+} while (0)
+
+#define mb() \
+ membar_safe("#LoadLoad | #LoadStore | #StoreStore | #StoreLoad")
+#define rmb() \
+ membar_safe("#LoadLoad")
+#define wmb() \
+ membar_safe("#StoreStore")
+#define membar_storeload() \
+ membar_safe("#StoreLoad")
+#define membar_storeload_storestore() \
+ membar_safe("#StoreLoad | #StoreStore")
+#define membar_storeload_loadload() \
+ membar_safe("#StoreLoad | #LoadLoad")
+#define membar_storestore_loadstore() \
+ membar_safe("#StoreStore | #LoadStore")
#endif