On 9/7/23 17:06, Waiman Long wrote:
On 9/7/23 15:33, Matthew Wilcox wrote:
On Thu, Sep 07, 2023 at 02:05:54PM -0400, Waiman Long wrote:
On 9/7/23 13:47, Matthew Wilcox (Oracle) wrote:
+static inline int rwsem_is_write_locked(struct rw_semaphore *sem)
+{
+ return atomic_long_read(&sem->count) & 1 /*
RWSEM_WRITER_LOCKED */;
+}
I would prefer you move the various RWSEM_* count bit macros from
kernel/locking/rwsem.c to under the !PREEMPT_RT block and directly use
RWSEM_WRITER_LOCKED instead of hardcoding a value of 1.
Just to be clear, you want the ~50 lines from:
/*
* On 64-bit architectures, the bit definitions of the count are:
...
#define RWSEM_READ_FAILED_MASK (RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
moved from rwsem.c to rwsem.h?
Or just these four lines:
#define RWSEM_WRITER_LOCKED (1UL << 0)
#define RWSEM_FLAG_WAITERS (1UL << 1)
#define RWSEM_FLAG_HANDOFF (1UL << 2)
#define RWSEM_FLAG_READFAIL (1UL << (BITS_PER_LONG - 1))
I think just the first 3 lines will be enough. Maybe a bit of comment
about these bit flags in the count atomic_long value.
Actually, the old rwsem implementation won't allow you to reliably
determine if a rwsem is write locked because the xadd instruction is
used for write locking and the code had to back out the WRITER_BIAS if
the attempt failed. Maybe that is why XFS has its own code to check if a
rwsem is write locked which is needed with the old rwsem implementation.
The new implementation makes this check reliable. Still it is not easy
to check if a rwsem is read locked as the check will be rather
complicated and probably racy.
Cheers,
Longman
