1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 *  Extend a 32-bit counter to 63 bits
4 *
5 *  Author:	Nicolas Pitre
6 *  Created:	December 3, 2006
7 *  Copyright:	MontaVista Software, Inc.
8 */
9
10#ifndef __LINUX_CNT32_TO_63_H__
11#define __LINUX_CNT32_TO_63_H__
12
13#include <linux/compiler.h>
14#include <linux/types.h>
15#include <asm/byteorder.h>
16
17/* this is used only to give gcc a clue about good code generation */
18union cnt32_to_63 {
19	struct {
20#if defined(__LITTLE_ENDIAN)
21		u32 lo, hi;
22#elif defined(__BIG_ENDIAN)
23		u32 hi, lo;
24#endif
25	};
26	u64 val;
27};
28
29
30/**
31 * cnt32_to_63 - Expand a 32-bit counter to a 63-bit counter
32 * @cnt_lo: The low part of the counter
33 *
34 * Many hardware clock counters are only 32 bits wide and therefore have
35 * a relatively short period making wrap-arounds rather frequent.  This
36 * is a problem when implementing sched_clock() for example, where a 64-bit
37 * non-wrapping monotonic value is expected to be returned.
38 *
39 * To overcome that limitation, let's extend a 32-bit counter to 63 bits
40 * in a completely lock free fashion. Bits 0 to 31 of the clock are provided
41 * by the hardware while bits 32 to 62 are stored in memory.  The top bit in
42 * memory is used to synchronize with the hardware clock half-period.  When
43 * the top bit of both counters (hardware and in memory) differ then the
44 * memory is updated with a new value, incrementing it when the hardware
45 * counter wraps around.
46 *
47 * Because a word store in memory is atomic then the incremented value will
48 * always be in synch with the top bit indicating to any potential concurrent
49 * reader if the value in memory is up to date or not with regards to the
50 * needed increment.  And any race in updating the value in memory is harmless
51 * as the same value would simply be stored more than once.
52 *
53 * The restrictions for the algorithm to work properly are:
54 *
55 * 1) this code must be called at least once per each half period of the
56 *    32-bit counter;
57 *
58 * 2) this code must not be preempted for a duration longer than the
59 *    32-bit counter half period minus the longest period between two
60 *    calls to this code;
61 *
62 * Those requirements ensure proper update to the state bit in memory.
63 * This is usually not a problem in practice, but if it is then a kernel
64 * timer should be scheduled to manage for this code to be executed often
65 * enough.
66 *
67 * And finally:
68 *
69 * 3) the cnt_lo argument must be seen as a globally incrementing value,
70 *    meaning that it should be a direct reference to the counter data which
71 *    can be evaluated according to a specific ordering within the macro,
72 *    and not the result of a previous evaluation stored in a variable.
73 *
74 * For example, this is wrong:
75 *
76 *	u32 partial = get_hw_count();
77 *	u64 full = cnt32_to_63(partial);
78 *	return full;
79 *
80 * This is fine:
81 *
82 *	u64 full = cnt32_to_63(get_hw_count());
83 *	return full;
84 *
85 * Note that the top bit (bit 63) in the returned value should be considered
86 * as garbage.  It is not cleared here because callers are likely to use a
87 * multiplier on the returned value which can get rid of the top bit
88 * implicitly by making the multiplier even, therefore saving on a runtime
89 * clear-bit instruction. Otherwise caller must remember to clear the top
90 * bit explicitly.
91 */
92#define cnt32_to_63(cnt_lo) \
93({ \
94	static u32 __m_cnt_hi; \
95	union cnt32_to_63 __x; \
96	__x.hi = __m_cnt_hi; \
97 	smp_rmb(); \
98	__x.lo = (cnt_lo); \
99	if (unlikely((s32)(__x.hi ^ __x.lo) < 0)) \
100		__m_cnt_hi = __x.hi = (__x.hi ^ 0x80000000) + (__x.hi >> 31); \
101	__x.val; \
102})
103
104#endif
105