parisc: Restore __ldcw_align for PA-RISC 2.0 processors

commit 914988e099fc658436fbd7b8f240160c352b6552 upstream.

Back in 2005, Kyle McMartin removed the 16-byte alignment for
ldcw semaphores on PA 2.0 machines (CONFIG_PA20). This broke
spinlocks on pre PA8800 processors. The main symptom was random
faults in mmap'd memory (e.g., gcc compilations, etc).

Unfortunately, the errata for this ldcw change is lost.

The issue is the 16-byte alignment required for ldcw semaphore
instructions can only be reduced to natural alignment when the
ldcw operation can be handled coherently in cache. Only PA8800
and PA8900 processors actually support doing the operation in
cache.

Aligning the spinlock dynamically adds two integer instructions
to each spinlock.

Tested on rp3440, c8000 and a500.

Signed-off-by: John David Anglin <dave.anglin@bell.net>
Link: https://lore.kernel.org/linux-parisc/6b332788-2227-127f-ba6d-55e99ecf4ed8@bell.net/T/#t
Link: https://lore.kernel.org/linux-parisc/20050609050702.GB4641@roadwarrior.mcmartin.ca/
Cc: stable@vger.kernel.org
Signed-off-by: Helge Deller <deller@gmx.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

arch/parisc/include/asm/ldcw.h

@@ -2,7 +2,6 @@
 #ifndef __PARISC_LDCW_H
 #define __PARISC_LDCW_H
 
-#ifndef CONFIG_PA20
 /* Because kmalloc only guarantees 8-byte alignment for kmalloc'd data,
    and GCC only guarantees 8-byte alignment for stack locals, we can't
    be assured of 16-byte alignment for atomic lock data even if we
@@ -11,6 +10,21 @@
    type and dynamically select the 16-byte aligned int from the array
    for the semaphore. */
 
+/* From: "Jim Hull" <jim.hull of hp.com>
+   I've attached a summary of the change, but basically, for PA 2.0, as
+   long as the ",CO" (coherent operation) completer is implemented, then the
+   16-byte alignment requirement for ldcw and ldcd is relaxed, and instead
+   they only require "natural" alignment (4-byte for ldcw, 8-byte for
+   ldcd).
+
+   Although the cache control hint is accepted by all PA 2.0 processors,
+   it is only implemented on PA8800/PA8900 CPUs. Prior PA8X00 CPUs still
+   require 16-byte alignment. If the address is unaligned, the operation
+   of the instruction is undefined. The ldcw instruction does not generate
+   unaligned data reference traps so misaligned accesses are not detected.
+   This hid the problem for years. So, restore the 16-byte alignment dropped
+   by Kyle McMartin in "Remove __ldcw_align for PA-RISC 2.0 processors". */
+
 #define __PA_LDCW_ALIGNMENT	16
 #define __PA_LDCW_ALIGN_ORDER	4
 #define __ldcw_align(a) ({					\
@@ -19,22 +33,12 @@
 		& ~(__PA_LDCW_ALIGNMENT - 1);			\
 	(volatile unsigned int *) __ret;			\
 })
-#define __LDCW	"ldcw"
 
-#else /*CONFIG_PA20*/
+#ifdef CONFIG_PA20
-/* From: "Jim Hull" <jim.hull of hp.com>
-   I've attached a summary of the change, but basically, for PA 2.0, as
-   long as the ",CO" (coherent operation) completer is specified, then the
-   16-byte alignment requirement for ldcw and ldcd is relaxed, and instead
-   they only require "natural" alignment (4-byte for ldcw, 8-byte for
-   ldcd). */
-
-#define __PA_LDCW_ALIGNMENT	4
-#define __PA_LDCW_ALIGN_ORDER	2
-#define __ldcw_align(a) (&(a)->slock)
 #define __LDCW	"ldcw,co"
-
+#else
-#endif /*!CONFIG_PA20*/
+#define __LDCW	"ldcw"
+#endif
 
 /* LDCW, the only atomic read-write operation PA-RISC has. *sigh*.
    We don't explicitly expose that "*a" may be written as reload

arch/parisc/include/asm/spinlock_types.h

@@ -3,13 +3,8 @@
 #define __ASM_SPINLOCK_TYPES_H
 
 typedef struct {
-#ifdef CONFIG_PA20
-	volatile unsigned int slock;
-# define __ARCH_SPIN_LOCK_UNLOCKED { 1 }
-#else
 	volatile unsigned int lock[4];
 # define __ARCH_SPIN_LOCK_UNLOCKED	{ { 1, 1, 1, 1 } }
-#endif
 } arch_spinlock_t;
 
 typedef struct {