| b.liu | e958203 | 2025-04-17 19:18:16 +0800 | [diff] [blame^] | 1 | /* |
| 2 | * SPDX-License-Identifier: MIT |
| 3 | * |
| 4 | * Copyright © 2019 Intel Corporation |
| 5 | */ |
| 6 | |
| 7 | #include <linux/slab.h> |
| 8 | #include <linux/workqueue.h> |
| 9 | |
| 10 | #include "i915_active.h" |
| 11 | #include "gem/i915_gem_context.h" |
| 12 | #include "gem/i915_gem_object.h" |
| 13 | #include "i915_globals.h" |
| 14 | #include "i915_request.h" |
| 15 | #include "i915_scheduler.h" |
| 16 | #include "i915_vma.h" |
| 17 | |
| 18 | static LIST_HEAD(globals); |
| 19 | |
| 20 | static atomic_t active; |
| 21 | static atomic_t epoch; |
| 22 | static struct park_work { |
| 23 | struct rcu_work work; |
| 24 | int epoch; |
| 25 | } park; |
| 26 | |
| 27 | static void i915_globals_shrink(void) |
| 28 | { |
| 29 | struct i915_global *global; |
| 30 | |
| 31 | /* |
| 32 | * kmem_cache_shrink() discards empty slabs and reorders partially |
| 33 | * filled slabs to prioritise allocating from the mostly full slabs, |
| 34 | * with the aim of reducing fragmentation. |
| 35 | */ |
| 36 | list_for_each_entry(global, &globals, link) |
| 37 | global->shrink(); |
| 38 | } |
| 39 | |
| 40 | static void __i915_globals_park(struct work_struct *work) |
| 41 | { |
| 42 | /* Confirm nothing woke up in the last grace period */ |
| 43 | if (park.epoch == atomic_read(&epoch)) |
| 44 | i915_globals_shrink(); |
| 45 | } |
| 46 | |
| 47 | void __init i915_global_register(struct i915_global *global) |
| 48 | { |
| 49 | GEM_BUG_ON(!global->shrink); |
| 50 | GEM_BUG_ON(!global->exit); |
| 51 | |
| 52 | list_add_tail(&global->link, &globals); |
| 53 | } |
| 54 | |
| 55 | static void __i915_globals_cleanup(void) |
| 56 | { |
| 57 | struct i915_global *global, *next; |
| 58 | |
| 59 | list_for_each_entry_safe_reverse(global, next, &globals, link) |
| 60 | global->exit(); |
| 61 | } |
| 62 | |
| 63 | static __initconst int (* const initfn[])(void) = { |
| 64 | i915_global_active_init, |
| 65 | i915_global_buddy_init, |
| 66 | i915_global_context_init, |
| 67 | i915_global_gem_context_init, |
| 68 | i915_global_objects_init, |
| 69 | i915_global_request_init, |
| 70 | i915_global_scheduler_init, |
| 71 | i915_global_vma_init, |
| 72 | }; |
| 73 | |
| 74 | int __init i915_globals_init(void) |
| 75 | { |
| 76 | int i; |
| 77 | |
| 78 | for (i = 0; i < ARRAY_SIZE(initfn); i++) { |
| 79 | int err; |
| 80 | |
| 81 | err = initfn[i](); |
| 82 | if (err) { |
| 83 | __i915_globals_cleanup(); |
| 84 | return err; |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | INIT_RCU_WORK(&park.work, __i915_globals_park); |
| 89 | return 0; |
| 90 | } |
| 91 | |
| 92 | void i915_globals_park(void) |
| 93 | { |
| 94 | /* |
| 95 | * Defer shrinking the global slab caches (and other work) until |
| 96 | * after a RCU grace period has completed with no activity. This |
| 97 | * is to try and reduce the latency impact on the consumers caused |
| 98 | * by us shrinking the caches the same time as they are trying to |
| 99 | * allocate, with the assumption being that if we idle long enough |
| 100 | * for an RCU grace period to elapse since the last use, it is likely |
| 101 | * to be longer until we need the caches again. |
| 102 | */ |
| 103 | if (!atomic_dec_and_test(&active)) |
| 104 | return; |
| 105 | |
| 106 | park.epoch = atomic_inc_return(&epoch); |
| 107 | queue_rcu_work(system_wq, &park.work); |
| 108 | } |
| 109 | |
| 110 | void i915_globals_unpark(void) |
| 111 | { |
| 112 | atomic_inc(&epoch); |
| 113 | atomic_inc(&active); |
| 114 | } |
| 115 | |
| 116 | void __exit i915_globals_exit(void) |
| 117 | { |
| 118 | /* Flush any residual park_work */ |
| 119 | atomic_inc(&epoch); |
| 120 | flush_rcu_work(&park.work); |
| 121 | |
| 122 | __i915_globals_cleanup(); |
| 123 | |
| 124 | /* And ensure that our DESTROY_BY_RCU slabs are truly destroyed */ |
| 125 | rcu_barrier(); |
| 126 | } |