[PATCH v2 1/1] arm64: spinlocks: Fix write starvation with rwlock

[Jeremy Linton]

The ARM64 rwlock is unfair in that readers can perpetually
block the writer.

This patch changes the rwlock behavior so that the writer
unconditionally flags the lock structure (given that its
not already flagged by another writer). This blocks further
readers that aren't in interrupt context from acquiring the
lock. Once all the readers have drained, the writer that
successfully flagged the lock can progress.

With this change, the lock still has a fairness issue caused
by an open race for ownership following a write unlock. If
certain cores/clusters are favored to win these races it
means a small set of writers could starve other users
(including writers). This should not be a common problem
given rwlock users should be read heavy with the occasional
writer. Further, the queued rwlock should also help to
alleviate this problem.

Signed-off-by: Jeremy Linton <jeremy.linton at arm.com>
---
 arch/arm64/include/asm/spinlock.h | 67 ++++++++++++++++++++++++++-------------
 1 file changed, 45 insertions(+), 22 deletions(-)

diff --git a/arch/arm64/include/asm/spinlock.h b/arch/arm64/include/asm/spinlock.h
index 95ad7102b63c..e4f6a1b545ca 100644
--- a/arch/arm64/include/asm/spinlock.h
+++ b/arch/arm64/include/asm/spinlock.h
@@ -143,7 +143,9 @@ static inline int arch_spin_is_contended(arch_spinlock_t *lock)
  * Write lock implementation.
  *
  * Write locks set bit 31. Unlocking, is done by writing 0 since the lock is
- * exclusively held.
+ * exclusively held. Setting the write bit (31) is used as a flag to drain the
+ * readers. The lock is considered taken for the writer only once all the
+ * readers have exited.
  *
  * The memory barriers are implicit with the load-acquire and store-release
  * instructions.
@@ -151,29 +153,41 @@ static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 
 static inline void arch_write_lock(arch_rwlock_t *rw)
 {
-	unsigned int tmp;
+	unsigned int tmp, tmp2, status;
 
 	asm volatile(ARM64_LSE_ATOMIC_INSN(
 	/* LL/SC */
 	"	sevl\n"
 	"1:	wfe\n"
-	"2:	ldaxr	%w0, %1\n"
-	"	cbnz	%w0, 1b\n"
-	"	stxr	%w0, %w2, %1\n"
-	"	cbnz	%w0, 2b\n"
-	__nops(1),
+	"2:	ldaxr	%w0, %3\n"
+	"	tbnz	%w0, #31, 1b\n" /* must be another writer */
+	"	orr	%w1, %w0, %w4\n"
+	"	stxr	%w2, %w1, %3\n"
+	"	cbnz	%w2, 2b\n"  /* failed to store, try again */
+	"	cbz	%w0, 5f\n"  /* if there aren't any readers we're done */
+	"	sevl\n"
+	"3:	wfe\n"		    /* spin waiting for the readers to exit */
+	"4:	ldaxr	%w0, %3\n"
+	"	cmp	%w0, %w4\n"
+	"	b.ne	3b\n"
+	"5:",
 	/* LSE atomics */
-	"1:	mov	%w0, wzr\n"
-	"2:	casa	%w0, %w2, %1\n"
-	"	cbz	%w0, 3f\n"
-	"	ldxr	%w0, %1\n"
-	"	cbz	%w0, 2b\n"
+	"1:	ldseta	%w4, %w0, %3\n"
+	"	cbz	%w0, 5f\n"	/* lock was clear, we are done */
+	"	tbz	%w0, #31, 4f\n" /* we own the lock, wait for readers */
+	"2:	ldxr	%w0, %3\n"	/* spin waiting for writer to exit */
+	"	tbz	%w0, #31, 1b\n"
 	"	wfe\n"
-	"	b	1b\n"
-	"3:")
-	: "=&r" (tmp), "+Q" (rw->lock)
+	"	b	2b\n"
+	__nops(2)
+	"3:	wfe\n"		     /* spin waiting for the readers to exit*/
+	"4:	ldaxr	%w0, %3\n"
+	"	cmp	%w0, %w4\n"
+	"	b.ne	3b\n"
+	"5:")
+	: "=&r" (tmp), "=&r" (tmp2), "=&r" (status), "+Q" (rw->lock)
 	: "r" (0x80000000)
-	: "memory");
+	: "cc", "memory");
 }
 
 static inline int arch_write_trylock(arch_rwlock_t *rw)
@@ -214,7 +228,8 @@ static inline void arch_write_unlock(arch_rwlock_t *rw)
  *
  * It exclusively loads the lock value, increments it and stores the new value
  * back if positive and the CPU still exclusively owns the location. If the
- * value is negative, the lock is already held.
+ * value is negative, a writer is pending. Since the rwlock is rentrant in
+ * interrupt context, ignore the write block in that case.
  *
  * During unlocking there may be multiple active read locks but no write lock.
  *
@@ -228,6 +243,7 @@ static inline void arch_write_unlock(arch_rwlock_t *rw)
 static inline void arch_read_lock(arch_rwlock_t *rw)
 {
 	unsigned int tmp, tmp2;
+	int allow_write_bypass = in_interrupt();
 
 	asm volatile(
 	"	sevl\n"
@@ -235,21 +251,28 @@ static inline void arch_read_lock(arch_rwlock_t *rw)
 	/* LL/SC */
 	"1:	wfe\n"
 	"2:	ldaxr	%w0, %2\n"
-	"	add	%w0, %w0, #1\n"
+	"	cmp	%w0, %w4\n"
+	"       b.eq    1b\n"      /* writer active */
+	"       add     %w0, %w0, #1\n"
+	"       cbnz    %w3, 3f\n" /* in interrupt, skip writer check */
 	"	tbnz	%w0, #31, 1b\n"
-	"	stxr	%w1, %w0, %2\n"
+	"3:	stxr	%w1, %w0, %2\n"
 	"	cbnz	%w1, 2b\n"
 	__nops(1),
 	/* LSE atomics */
 	"1:	wfe\n"
 	"2:	ldxr	%w0, %2\n"
+	"       cmp     %w0, %w4\n"
+	"       b.eq    1b\n"      /* writer active, go wait */
 	"	adds	%w1, %w0, #1\n"
+	"       cbnz    %w3, 3f\n" /* in interrupt, skip writer check */
 	"	tbnz	%w1, #31, 1b\n"
-	"	casa	%w0, %w1, %2\n"
+	"3:	casa	%w0, %w1, %2\n"
 	"	sbc	%w0, %w1, %w0\n"
-	"	cbnz	%w0, 2b")
+	"	cbnz	%w0, 2b"
+	)
 	: "=&r" (tmp), "=&r" (tmp2), "+Q" (rw->lock)
-	:
+	: "r" (allow_write_bypass), "r" (0x80000000)
 	: "cc", "memory");
 }
 
-- 
2.13.5