On 18.03.24 16:39, Jan Beulich wrote:
On 14.03.2024 08:20, Juergen Gross wrote:
The rwlock handling is limiting the number of cpus to 4095 today. The
main reason is the use of the atomic_t data type for the main lock
handling, which needs 2 bits for the locking state (writer waiting or
write locked), 12 bits for the id of a possible writer, and a 12 bit
counter for readers. The limit isn't 4096 due to an off by one sanity
check.
The atomic_t data type is 32 bits wide, so in theory 15 bits for the
writer's cpu id and 15 bits for the reader count seem to be fine, but
via read_trylock() more readers than cpus are possible.
As a result, afaict you choose to use just 14 bits for the CPU, but
still 15 bits (with the 16th to deal with overflow) for the reader count.
That could do with making explicit here, as a question is whether we
deem as sufficient that there is just one extra bit for the reader
count.
Okay, I'll add a sentence to the commit message.
--- a/xen/include/xen/rwlock.h
+++ b/xen/include/xen/rwlock.h
@@ -23,12 +23,12 @@ typedef struct {
#define rwlock_init(l) (*(l) = (rwlock_t)RW_LOCK_UNLOCKED)
/* Writer states & reader shift and bias. */
-#define _QW_CPUMASK 0xfffU /* Writer CPU mask */
-#define _QW_SHIFT 12 /* Writer flags shift */
-#define _QW_WAITING (1U << _QW_SHIFT) /* A writer is waiting */
-#define _QW_LOCKED (3U << _QW_SHIFT) /* A writer holds the lock */
-#define _QW_WMASK (3U << _QW_SHIFT) /* Writer mask */
-#define _QR_SHIFT 14 /* Reader count shift */
+#define _QW_SHIFT 14 /* Writer flags shift */
+#define _QW_CPUMASK ((1U << _QW_SHIFT) - 1) /* Writer CPU mask */
+#define _QW_WAITING (1U << _QW_SHIFT) /* A writer is waiting */
+#define _QW_LOCKED (3U << _QW_SHIFT) /* A writer holds the lock */
+#define _QW_WMASK (3U << _QW_SHIFT) /* Writer mask */
+#define _QR_SHIFT (_QW_SHIFT + 2) /* Reader count shift */
#define _QR_BIAS (1U << _QR_SHIFT)
Btw, seeing all the uppercase U suffixes here, I think you had some
lowercase ones earlier in the series. While Misra doesn't demand
uppercase for U, it does for L and iirc we decided to use all
uppercase suffixes as a result. Would be nice if what goes in could
have this correct right away.
I'll rescan all the patches and change them accordingly.
@@ -36,14 +36,16 @@ void queue_write_lock_slowpath(rwlock_t *lock);
static inline bool _is_write_locked_by_me(unsigned int cnts)
{
- BUILD_BUG_ON(_QW_CPUMASK < NR_CPUS);
+ BUILD_BUG_ON((_QW_CPUMASK + 1) < NR_CPUS);
+ BUILD_BUG_ON(NR_CPUS * _QR_BIAS > INT_MAX);
return (cnts & _QW_WMASK) == _QW_LOCKED &&
(cnts & _QW_CPUMASK) == smp_processor_id();
}
static inline bool _can_read_lock(unsigned int cnts)
{
- return !(cnts & _QW_WMASK) || _is_write_locked_by_me(cnts);
+ return cnts <= INT_MAX &&
+ (!(cnts & _QW_WMASK) || _is_write_locked_by_me(cnts));
}
I view this as problematic: Code knowing that a write lock is being held
may invoke a function using read_trylock() and expect the lock to be
available there.
So you expect it to be fine that someone is using read_trylock() 32768 times
recursively while holding a lock as a writer? Sure, I can change the condition,
but OTOH ...
Juergen
