| 1 | #include <acpi/lapic.h> |
| 2 | #include <mem/alloc.h> |
| 3 | #include <mem/page.h> |
| 4 | #include <mem/tlb.h> |
| 5 | #include <sch/sched.h> |
| 6 | #include <stdatomic.h> |
| 7 | #include <stdint.h> |
| 8 | #include <thread/dpc.h> |
| 9 | |
| 10 | struct spinlock tlb_shootdown_lock = SPINLOCK_INIT; |
| 11 | |
| 12 | static void tlb_shootdown_internal(void) { |
| 13 | size_t cpu = smp_core_id(); |
| 14 | struct tlb_shootdown_cpu *c = &global.shootdown_data[cpu]; |
| 15 | |
| 16 | uint64_t done = atomic_load_explicit(&c->done_gen, memory_order_relaxed); |
| 17 | |
| 18 | while (true) { |
| 19 | uint64_t req = atomic_load_explicit(&c->req_gen, memory_order_acquire); |
| 20 | |
| 21 | if (done >= req) |
| 22 | break; |
| 23 | |
| 24 | for (;;) { |
| 25 | uint32_t tail = |
| 26 | atomic_load_explicit(&c->tail, memory_order_relaxed); |
| 27 | uint32_t head = |
| 28 | atomic_load_explicit(&c->head, memory_order_acquire); |
| 29 | |
| 30 | if (tail == head) |
| 31 | break; |
| 32 | |
| 33 | while (tail != head) { |
| 34 | uintptr_t addr = |
| 35 | atomic_load_explicit(&c->queue[tail & (TLB_QUEUE_SIZE - 1)], |
| 36 | memory_order_acquire); |
| 37 | |
| 38 | if (addr) |
| 39 | invlpg(virt: addr); |
| 40 | |
| 41 | tail++; |
| 42 | } |
| 43 | |
| 44 | atomic_store_explicit(&c->tail, tail, memory_order_release); |
| 45 | } |
| 46 | |
| 47 | if (atomic_exchange_explicit(&c->flush_all, false, |
| 48 | memory_order_acq_rel)) { |
| 49 | tlb_flush(); |
| 50 | uint32_t h = atomic_load_explicit(&c->head, memory_order_acquire); |
| 51 | atomic_store_explicit(&c->tail, h, memory_order_release); |
| 52 | } |
| 53 | |
| 54 | done++; |
| 55 | atomic_store_explicit(&c->done_gen, done, memory_order_release); |
| 56 | } |
| 57 | } |
| 58 | |
| 59 | enum irq_result tlb_shootdown_isr(void *ctx, irq_t irq, |
| 60 | struct irq_context *rsp) { |
| 61 | (void) ctx; |
| 62 | (void) irq; |
| 63 | (void) rsp; |
| 64 | |
| 65 | tlb_shootdown_internal(); |
| 66 | return IRQ_HANDLED; |
| 67 | } |
| 68 | |
| 69 | void tlb_shootdown(uintptr_t addr, bool synchronous) { |
| 70 | if (global.current_bootstage < BOOTSTAGE_MID_MP) |
| 71 | return; |
| 72 | |
| 73 | enum irql lirql = spin_lock(lock: &tlb_shootdown_lock); |
| 74 | |
| 75 | uint64_t gen = atomic_fetch_add_explicit(&global.next_tlb_gen, 1, |
| 76 | memory_order_relaxed) + |
| 77 | 1; |
| 78 | |
| 79 | size_t this_cpu = smp_core_id(); |
| 80 | |
| 81 | size_t i; |
| 82 | for_each_cpu_id(i) { |
| 83 | if (i == this_cpu) { |
| 84 | invlpg(virt: addr); |
| 85 | continue; |
| 86 | } |
| 87 | |
| 88 | struct tlb_shootdown_cpu *t = &global.shootdown_data[i]; |
| 89 | |
| 90 | uint32_t head = atomic_load_explicit(&t->head, memory_order_relaxed); |
| 91 | uint32_t tail = atomic_load_explicit(&t->tail, memory_order_acquire); |
| 92 | |
| 93 | if ((head - tail) >= TLB_QUEUE_SIZE) { |
| 94 | atomic_store_explicit(&t->flush_all, true, memory_order_release); |
| 95 | } else { |
| 96 | atomic_store_explicit(&t->queue[head & (TLB_QUEUE_SIZE - 1)], addr, |
| 97 | memory_order_release); |
| 98 | atomic_store_explicit(&t->head, head + 1, memory_order_release); |
| 99 | } |
| 100 | |
| 101 | atomic_store_explicit(&t->req_gen, gen, memory_order_release); |
| 102 | ipi_send(apic_id: i, IRQ_TLB_SHOOTDOWN); |
| 103 | } |
| 104 | |
| 105 | if (synchronous) { |
| 106 | for_each_cpu_id(i) { |
| 107 | if (i == this_cpu) |
| 108 | continue; |
| 109 | |
| 110 | struct tlb_shootdown_cpu *o = &global.shootdown_data[i]; |
| 111 | |
| 112 | int spins = 0; |
| 113 | |
| 114 | while (atomic_load_explicit(&o->done_gen, memory_order_acquire) < |
| 115 | gen) { |
| 116 | if (spins < 100) { |
| 117 | cpu_relax(); |
| 118 | spins++; |
| 119 | continue; |
| 120 | } |
| 121 | |
| 122 | spins = 0; |
| 123 | ipi_send(apic_id: i, IRQ_TLB_SHOOTDOWN); |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | atomic_fetch_add_explicit(&global.pt_epoch, 1, memory_order_release); |
| 128 | } |
| 129 | |
| 130 | spin_unlock(lock: &tlb_shootdown_lock, old: lirql); |
| 131 | } |
| 132 | |