Lines Matching +full:lock +full:- +full:less

81  * Locking strategy for high-resolution timing services
88  * To minimize lock contention and cache thrashing we employ the
90  * on an acquisition-counting mutex, described below; readers (common)
91  * execute in parallel with no synchronization at all -- they don't
93  * readers just examine the writer lock's value before and after loading
98  * The writer lock, hres_lock, is a 32-bit integer consisting of an
99  * 8-bit lock and a 24-bit acquisition count.  To acquire the lock we
100  * set the lock field with ldstub, which sets the low-order 8 bits to
101  * 0xff; to clear the lock, we increment it, which simultaneously clears
102  * the lock field (0xff --> 0x00) and increments the acquisition count
107  * and after loading the time components they need; if either lock value
108  * contains 0xff in the low-order bits (lock held), or if the lock values
109  * are not equal (lock was acquired and released), a writer intervened
110  * and the reader must try again.  If the lock values are equal and the
111  * low-order 8 bits are clear, the timestamp must be valid.  We can check
114  * following table of all possible lock states:
117  *	------------	-----		-----------------
132  * This is essential *even on a CPU that does in-order loads* because
135  * lower bit (andn with 1) of the old lock value right away, since this
137  * anyway to perform the lock comparison described above.
139  * (2) Out-of-order loads.
141  * On a CPU that does out-of-order loads we must ensure that the loads
150  * CBE_HIGH_PIL or higher to prevent single-CPU deadlock.
152  * (4) Cross-calls.
154  * If a cross-call happens while one CPU has hres_lock and another is
157  * waiting to be released from the cross-call, and the cross-call can't
159  * disabled.  Thus cross-calls must be blocked while holding hres_lock.
178  * NATIVE_TIME_TO_NSEC_SCALE is called with NSEC_SHIFT to convert hi-res
180  * hi-res timestamps are in %stick units. On systems that do not have a
181  * %stick register, hi-res timestamps are in %tick units.
195  * Convert hi-res native time (V9's %tick in our case) into nanoseconds.
198  * without using floating point and without overflowing 64-bit integers.
203  * To avoid floating-point we precompute (R * 2^32) during boot and
207  * To avoid 64-bit overflow when multiplying (tick << 4) by nsec_scale,
208  * we split (tick << 4) into its high and low 32-bit pieces, H and L,
218  * floating point and all intermediate results fit in 64-bit registers.
220  * Note that we require that tick is less than (1 << (64 - NSEC_SHIFT));
222  * is a serious problem; (1 << (64 - NSEC_SHIFT)) nanoseconds is over
223  * thirty-six years).  Nonetheless, clients may wish to be aware of this
226  * We provide two versions of this macro: a "full-service" version that
227  * just converts ticks to nanoseconds and a higher-performance version that
230  * we take a fast path if we determine the ticks to be less than 32 bits
234  * Note that in the 32-bit path we don't even bother clearing NPT.
247 	brz,a,pt scr2, 6f;		/* if clear, 32-bit fast path */\
248 	mulx	out, scr1, out;		/* delay: 32-bit fast path */	\
249 	sllx	out, shift, out;	/* clear NPT and pre-scale */	\
255 	ba	7f;			/* branch over 32-bit path */	\
258 	srlx	out, 32 - shift, out;					\
267 	mov	-1, out;						\