| 1 |
/* |
| 2 |
* Copyright (c) 2015-2020, Arm Limited and Contributors. All rights reserved. |
| 3 |
* Copyright (c) 2020, NVIDIA Corporation. All rights reserved. |
| 4 |
* |
| 5 |
* SPDX-License-Identifier: BSD-3-Clause |
| 6 |
*/ |
| 7 |
|
| 8 |
#include <assert.h> |
| 9 |
#include <string.h> |
| 10 |
|
| 11 |
#include <arch_helpers.h> |
| 12 |
#include <lib/bakery_lock.h> |
| 13 |
#include <lib/el3_runtime/cpu_data.h> |
| 14 |
#include <lib/utils_def.h> |
| 15 |
#include <plat/common/platform.h> |
| 16 |
|
| 17 |
/* |
| 18 |
* Functions in this file implement Bakery Algorithm for mutual exclusion with the |
| 19 |
* bakery lock data structures in cacheable and Normal memory. |
| 20 |
* |
| 21 |
* ARM architecture offers a family of exclusive access instructions to |
| 22 |
* efficiently implement mutual exclusion with hardware support. However, as |
| 23 |
* well as depending on external hardware, these instructions have defined |
| 24 |
* behavior only on certain memory types (cacheable and Normal memory in |
| 25 |
* particular; see ARMv8 Architecture Reference Manual section B2.10). Use cases |
| 26 |
* in trusted firmware are such that mutual exclusion implementation cannot |
| 27 |
* expect that accesses to the lock have the specific type required by the |
| 28 |
* architecture for these primitives to function (for example, not all |
| 29 |
* contenders may have address translation enabled). |
| 30 |
* |
| 31 |
* This implementation does not use mutual exclusion primitives. It expects |
| 32 |
* memory regions where the locks reside to be cacheable and Normal. |
| 33 |
* |
| 34 |
* Note that the ARM architecture guarantees single-copy atomicity for aligned |
| 35 |
* accesses regardless of status of address translation. |
| 36 |
*/ |
| 37 |
|
| 38 |
#ifdef PLAT_PERCPU_BAKERY_LOCK_SIZE |
| 39 |
/* |
| 40 |
* Verify that the platform defined value for the per-cpu space for bakery locks is |
| 41 |
* a multiple of the cache line size, to prevent multiple CPUs writing to the same |
| 42 |
* bakery lock cache line |
| 43 |
* |
| 44 |
* Using this value, if provided, rather than the linker generated value results in |
| 45 |
* more efficient code |
| 46 |
*/ |
| 47 |
CASSERT((PLAT_PERCPU_BAKERY_LOCK_SIZE & (CACHE_WRITEBACK_GRANULE - 1)) == 0, |
| 48 |
PLAT_PERCPU_BAKERY_LOCK_SIZE_not_cacheline_multiple); |
| 49 |
#define PERCPU_BAKERY_LOCK_SIZE (PLAT_PERCPU_BAKERY_LOCK_SIZE) |
| 50 |
#else |
| 51 |
/* |
| 52 |
* Use the linker defined symbol which has evaluated the size reqiurement. |
| 53 |
* This is not as efficient as using a platform defined constant |
| 54 |
*/ |
| 55 |
IMPORT_SYM(uintptr_t, __PERCPU_BAKERY_LOCK_START__, BAKERY_LOCK_START); |
| 56 |
IMPORT_SYM(uintptr_t, __PERCPU_BAKERY_LOCK_END__, BAKERY_LOCK_END); |
| 57 |
#define PERCPU_BAKERY_LOCK_SIZE (BAKERY_LOCK_END - BAKERY_LOCK_START) |
| 58 |
#endif |
| 59 |
|
| 60 |
static inline bakery_lock_t *get_bakery_info(unsigned int cpu_ix, |
| 61 |
bakery_lock_t *lock) |
| 62 |
{ |
| 63 |
return (bakery_info_t *)((uintptr_t)lock + |
| 64 |
cpu_ix * PERCPU_BAKERY_LOCK_SIZE); |
| 65 |
} |
| 66 |
|
| 67 |
static inline void write_cache_op(uintptr_t addr, bool cached) |
| 68 |
{ |
| 69 |
if (cached) |
| 70 |
dccvac(addr); |
| 71 |
else |
| 72 |
dcivac(addr); |
| 73 |
|
| 74 |
dsbish(); |
| 75 |
} |
| 76 |
|
| 77 |
static inline void read_cache_op(uintptr_t addr, bool cached) |
| 78 |
{ |
| 79 |
if (cached) |
| 80 |
dccivac(addr); |
| 81 |
|
| 82 |
dmbish(); |
| 83 |
} |
| 84 |
|
| 85 |
/* Helper function to check if the lock is acquired */ |
| 86 |
static inline __unused bool is_lock_acquired(const bakery_info_t *my_bakery_info, |
| 87 |
bool is_cached) |
| 88 |
{ |
| 89 |
/* |
| 90 |
* Even though lock data is updated only by the owning cpu and |
| 91 |
* appropriate cache maintenance operations are performed, |
| 92 |
* if the previous update was done when the cpu was not participating |
| 93 |
* in coherency, then there is a chance that cache maintenance |
| 94 |
* operations were not propagated to all the caches in the system. |
| 95 |
* Hence do a `read_cache_op()` prior to read. |
| 96 |
*/ |
| 97 |
read_cache_op((uintptr_t)my_bakery_info, is_cached); |
| 98 |
return bakery_ticket_number(my_bakery_info->lock_data) != 0U; |
| 99 |
} |
| 100 |
|
| 101 |
static unsigned int bakery_get_ticket(bakery_lock_t *lock, |
| 102 |
unsigned int me, bool is_cached) |
| 103 |
{ |
| 104 |
unsigned int my_ticket, their_ticket; |
| 105 |
unsigned int they; |
| 106 |
bakery_info_t *my_bakery_info, *their_bakery_info; |
| 107 |
|
| 108 |
/* |
| 109 |
* Obtain a reference to the bakery information for this cpu and ensure |
| 110 |
* it is not NULL. |
| 111 |
*/ |
| 112 |
my_bakery_info = get_bakery_info(me, lock); |
| 113 |
assert(my_bakery_info != NULL); |
| 114 |
|
| 115 |
/* Prevent recursive acquisition.*/ |
| 116 |
assert(!is_lock_acquired(my_bakery_info, is_cached)); |
| 117 |
|
| 118 |
/* |
| 119 |
* Tell other contenders that we are through the bakery doorway i.e. |
| 120 |
* going to allocate a ticket for this cpu. |
| 121 |
*/ |
| 122 |
my_ticket = 0U; |
| 123 |
my_bakery_info->lock_data = make_bakery_data(CHOOSING_TICKET, my_ticket); |
| 124 |
|
| 125 |
write_cache_op((uintptr_t)my_bakery_info, is_cached); |
| 126 |
|
| 127 |
/* |
| 128 |
* Iterate through the bakery information of each contender to allocate |
| 129 |
* the highest ticket number for this cpu. |
| 130 |
*/ |
| 131 |
for (they = 0U; they < BAKERY_LOCK_MAX_CPUS; they++) { |
| 132 |
if (me == they) |
| 133 |
continue; |
| 134 |
|
| 135 |
/* |
| 136 |
* Get a reference to the other contender's bakery info and |
| 137 |
* ensure that a stale copy is not read. |
| 138 |
*/ |
| 139 |
their_bakery_info = get_bakery_info(they, lock); |
| 140 |
assert(their_bakery_info != NULL); |
| 141 |
|
| 142 |
read_cache_op((uintptr_t)their_bakery_info, is_cached); |
| 143 |
|
| 144 |
/* |
| 145 |
* Update this cpu's ticket number if a higher ticket number is |
| 146 |
* seen |
| 147 |
*/ |
| 148 |
their_ticket = bakery_ticket_number(their_bakery_info->lock_data); |
| 149 |
if (their_ticket > my_ticket) |
| 150 |
my_ticket = their_ticket; |
| 151 |
} |
| 152 |
|
| 153 |
/* |
| 154 |
* Compute ticket; then signal to other contenders waiting for us to |
| 155 |
* finish calculating our ticket value that we're done |
| 156 |
*/ |
| 157 |
++my_ticket; |
| 158 |
my_bakery_info->lock_data = make_bakery_data(CHOSEN_TICKET, my_ticket); |
| 159 |
|
| 160 |
write_cache_op((uintptr_t)my_bakery_info, is_cached); |
| 161 |
|
| 162 |
return my_ticket; |
| 163 |
} |
| 164 |
|
| 165 |
void bakery_lock_get(bakery_lock_t *lock) |
| 166 |
{ |
| 167 |
unsigned int they, me; |
| 168 |
unsigned int my_ticket, my_prio, their_ticket; |
| 169 |
bakery_info_t *their_bakery_info; |
| 170 |
unsigned int their_bakery_data; |
| 171 |
bool is_cached; |
| 172 |
|
| 173 |
me = plat_my_core_pos(); |
| 174 |
is_cached = is_dcache_enabled(); |
| 175 |
|
| 176 |
/* Get a ticket */ |
| 177 |
my_ticket = bakery_get_ticket(lock, me, is_cached); |
| 178 |
|
| 179 |
/* |
| 180 |
* Now that we got our ticket, compute our priority value, then compare |
| 181 |
* with that of others, and proceed to acquire the lock |
| 182 |
*/ |
| 183 |
my_prio = bakery_get_priority(my_ticket, me); |
| 184 |
for (they = 0U; they < BAKERY_LOCK_MAX_CPUS; they++) { |
| 185 |
if (me == they) |
| 186 |
continue; |
| 187 |
|
| 188 |
/* |
| 189 |
* Get a reference to the other contender's bakery info and |
| 190 |
* ensure that a stale copy is not read. |
| 191 |
*/ |
| 192 |
their_bakery_info = get_bakery_info(they, lock); |
| 193 |
assert(their_bakery_info != NULL); |
| 194 |
|
| 195 |
/* Wait for the contender to get their ticket */ |
| 196 |
do { |
| 197 |
read_cache_op((uintptr_t)their_bakery_info, is_cached); |
| 198 |
their_bakery_data = their_bakery_info->lock_data; |
| 199 |
} while (bakery_is_choosing(their_bakery_data)); |
| 200 |
|
| 201 |
/* |
| 202 |
* If the other party is a contender, they'll have non-zero |
| 203 |
* (valid) ticket value. If they do, compare priorities |
| 204 |
*/ |
| 205 |
their_ticket = bakery_ticket_number(their_bakery_data); |
| 206 |
if (their_ticket && (bakery_get_priority(their_ticket, they) < my_prio)) { |
| 207 |
/* |
| 208 |
* They have higher priority (lower value). Wait for |
| 209 |
* their ticket value to change (either release the lock |
| 210 |
* to have it dropped to 0; or drop and probably content |
| 211 |
* again for the same lock to have an even higher value) |
| 212 |
*/ |
| 213 |
do { |
| 214 |
wfe(); |
| 215 |
read_cache_op((uintptr_t)their_bakery_info, is_cached); |
| 216 |
} while (their_ticket |
| 217 |
== bakery_ticket_number(their_bakery_info->lock_data)); |
| 218 |
} |
| 219 |
} |
| 220 |
|
| 221 |
/* |
| 222 |
* Lock acquired. Ensure that any reads and writes from a shared |
| 223 |
* resource in the critical section read/write values after the lock is |
| 224 |
* acquired. |
| 225 |
*/ |
| 226 |
dmbish(); |
| 227 |
} |
| 228 |
|
| 229 |
void bakery_lock_release(bakery_lock_t *lock) |
| 230 |
{ |
| 231 |
bakery_info_t *my_bakery_info; |
| 232 |
bool is_cached = is_dcache_enabled(); |
| 233 |
|
| 234 |
my_bakery_info = get_bakery_info(plat_my_core_pos(), lock); |
| 235 |
|
| 236 |
assert(is_lock_acquired(my_bakery_info, is_cached)); |
| 237 |
|
| 238 |
/* |
| 239 |
* Ensure that other observers see any stores in the critical section |
| 240 |
* before releasing the lock. Also ensure all loads in the critical |
| 241 |
* section are complete before releasing the lock. Release the lock by |
| 242 |
* resetting ticket. Then signal other waiting contenders. |
| 243 |
*/ |
| 244 |
dmbish(); |
| 245 |
my_bakery_info->lock_data = 0U; |
| 246 |
write_cache_op((uintptr_t)my_bakery_info, is_cached); |
| 247 |
|
| 248 |
/* This sev is ordered by the dsbish in write_cahce_op */ |
| 249 |
sev(); |
| 250 |
} |