Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2017 ARM Ltd. |
| 3 | * Author: Marc Zyngier <marc.zyngier@arm.com> |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or modify |
| 6 | * it under the terms of the GNU General Public License version 2 as |
| 7 | * published by the Free Software Foundation. |
| 8 | * |
| 9 | * This program is distributed in the hope that it will be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
| 13 | * |
| 14 | * You should have received a copy of the GNU General Public License |
| 15 | * along with this program. If not, see <http://www.gnu.org/licenses/>. |
| 16 | */ |
| 17 | |
| 18 | #include <linux/interrupt.h> |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 19 | #include <linux/irq.h> |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 20 | #include <linux/irqdomain.h> |
| 21 | #include <linux/kvm_host.h> |
Marc Zyngier | 196b136 | 2017-10-27 15:28:39 +0100 | [diff] [blame] | 22 | #include <linux/irqchip/arm-gic-v3.h> |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 23 | |
| 24 | #include "vgic.h" |
| 25 | |
Marc Zyngier | ed8703a | 2017-10-27 15:28:55 +0100 | [diff] [blame] | 26 | /* |
| 27 | * How KVM uses GICv4 (insert rude comments here): |
| 28 | * |
| 29 | * The vgic-v4 layer acts as a bridge between several entities: |
| 30 | * - The GICv4 ITS representation offered by the ITS driver |
| 31 | * - VFIO, which is in charge of the PCI endpoint |
| 32 | * - The virtual ITS, which is the only thing the guest sees |
| 33 | * |
| 34 | * The configuration of VLPIs is triggered by a callback from VFIO, |
| 35 | * instructing KVM that a PCI device has been configured to deliver |
| 36 | * MSIs to a vITS. |
| 37 | * |
| 38 | * kvm_vgic_v4_set_forwarding() is thus called with the routing entry, |
| 39 | * and this is used to find the corresponding vITS data structures |
| 40 | * (ITS instance, device, event and irq) using a process that is |
| 41 | * extremely similar to the injection of an MSI. |
| 42 | * |
| 43 | * At this stage, we can link the guest's view of an LPI (uniquely |
| 44 | * identified by the routing entry) and the host irq, using the GICv4 |
| 45 | * driver mapping operation. Should the mapping succeed, we've then |
| 46 | * successfully upgraded the guest's LPI to a VLPI. We can then start |
| 47 | * with updating GICv4's view of the property table and generating an |
| 48 | * INValidation in order to kickstart the delivery of this VLPI to the |
| 49 | * guest directly, without software intervention. Well, almost. |
| 50 | * |
| 51 | * When the PCI endpoint is deconfigured, this operation is reversed |
| 52 | * with VFIO calling kvm_vgic_v4_unset_forwarding(). |
| 53 | * |
| 54 | * Once the VLPI has been mapped, it needs to follow any change the |
| 55 | * guest performs on its LPI through the vITS. For that, a number of |
| 56 | * command handlers have hooks to communicate these changes to the HW: |
| 57 | * - Any invalidation triggers a call to its_prop_update_vlpi() |
| 58 | * - The INT command results in a irq_set_irqchip_state(), which |
| 59 | * generates an INT on the corresponding VLPI. |
| 60 | * - The CLEAR command results in a irq_set_irqchip_state(), which |
| 61 | * generates an CLEAR on the corresponding VLPI. |
| 62 | * - DISCARD translates into an unmap, similar to a call to |
| 63 | * kvm_vgic_v4_unset_forwarding(). |
| 64 | * - MOVI is translated by an update of the existing mapping, changing |
| 65 | * the target vcpu, resulting in a VMOVI being generated. |
| 66 | * - MOVALL is translated by a string of mapping updates (similar to |
| 67 | * the handling of MOVI). MOVALL is horrible. |
| 68 | * |
| 69 | * Note that a DISCARD/MAPTI sequence emitted from the guest without |
| 70 | * reprogramming the PCI endpoint after MAPTI does not result in a |
| 71 | * VLPI being mapped, as there is no callback from VFIO (the guest |
| 72 | * will get the interrupt via the normal SW injection). Fixing this is |
| 73 | * not trivial, and requires some horrible messing with the VFIO |
| 74 | * internals. Not fun. Don't do that. |
| 75 | * |
| 76 | * Then there is the scheduling. Each time a vcpu is about to run on a |
| 77 | * physical CPU, KVM must tell the corresponding redistributor about |
| 78 | * it. And if we've migrated our vcpu from one CPU to another, we must |
| 79 | * tell the ITS (so that the messages reach the right redistributor). |
| 80 | * This is done in two steps: first issue a irq_set_affinity() on the |
| 81 | * irq corresponding to the vcpu, then call its_schedule_vpe(). You |
| 82 | * must be in a non-preemptible context. On exit, another call to |
| 83 | * its_schedule_vpe() tells the redistributor that we're done with the |
| 84 | * vcpu. |
| 85 | * |
| 86 | * Finally, the doorbell handling: Each vcpu is allocated an interrupt |
| 87 | * which will fire each time a VLPI is made pending whilst the vcpu is |
| 88 | * not running. Each time the vcpu gets blocked, the doorbell |
| 89 | * interrupt gets enabled. When the vcpu is unblocked (for whatever |
| 90 | * reason), the doorbell interrupt is disabled. |
| 91 | */ |
| 92 | |
Marc Zyngier | d3d83f7 | 2017-10-27 15:28:53 +0100 | [diff] [blame] | 93 | #define DB_IRQ_FLAGS (IRQ_NOAUTOEN | IRQ_DISABLE_UNLAZY | IRQ_NO_BALANCING) |
| 94 | |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 95 | static irqreturn_t vgic_v4_doorbell_handler(int irq, void *info) |
| 96 | { |
| 97 | struct kvm_vcpu *vcpu = info; |
| 98 | |
| 99 | vcpu->arch.vgic_cpu.vgic_v3.its_vpe.pending_last = true; |
| 100 | kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); |
| 101 | kvm_vcpu_kick(vcpu); |
| 102 | |
| 103 | return IRQ_HANDLED; |
| 104 | } |
| 105 | |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 106 | /** |
| 107 | * vgic_v4_init - Initialize the GICv4 data structures |
| 108 | * @kvm: Pointer to the VM being initialized |
| 109 | * |
| 110 | * We may be called each time a vITS is created, or when the |
| 111 | * vgic is initialized. This relies on kvm->lock to be |
| 112 | * held. In both cases, the number of vcpus should now be |
| 113 | * fixed. |
| 114 | */ |
| 115 | int vgic_v4_init(struct kvm *kvm) |
| 116 | { |
| 117 | struct vgic_dist *dist = &kvm->arch.vgic; |
| 118 | struct kvm_vcpu *vcpu; |
| 119 | int i, nr_vcpus, ret; |
| 120 | |
Christoffer Dall | 3d1ad64 | 2017-11-10 09:16:23 +0100 | [diff] [blame] | 121 | if (!vgic_supports_direct_msis(kvm)) |
| 122 | return 0; /* Nothing to see here... move along. */ |
| 123 | |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 124 | if (dist->its_vm.vpes) |
| 125 | return 0; |
| 126 | |
| 127 | nr_vcpus = atomic_read(&kvm->online_vcpus); |
| 128 | |
| 129 | dist->its_vm.vpes = kzalloc(sizeof(*dist->its_vm.vpes) * nr_vcpus, |
| 130 | GFP_KERNEL); |
| 131 | if (!dist->its_vm.vpes) |
| 132 | return -ENOMEM; |
| 133 | |
| 134 | dist->its_vm.nr_vpes = nr_vcpus; |
| 135 | |
| 136 | kvm_for_each_vcpu(i, vcpu, kvm) |
| 137 | dist->its_vm.vpes[i] = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe; |
| 138 | |
| 139 | ret = its_alloc_vcpu_irqs(&dist->its_vm); |
| 140 | if (ret < 0) { |
| 141 | kvm_err("VPE IRQ allocation failure\n"); |
| 142 | kfree(dist->its_vm.vpes); |
| 143 | dist->its_vm.nr_vpes = 0; |
| 144 | dist->its_vm.vpes = NULL; |
| 145 | return ret; |
| 146 | } |
| 147 | |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 148 | kvm_for_each_vcpu(i, vcpu, kvm) { |
| 149 | int irq = dist->its_vm.vpes[i]->irq; |
| 150 | |
| 151 | /* |
| 152 | * Don't automatically enable the doorbell, as we're |
| 153 | * flipping it back and forth when the vcpu gets |
| 154 | * blocked. Also disable the lazy disabling, as the |
| 155 | * doorbell could kick us out of the guest too |
| 156 | * early... |
| 157 | */ |
Marc Zyngier | d3d83f7 | 2017-10-27 15:28:53 +0100 | [diff] [blame] | 158 | irq_set_status_flags(irq, DB_IRQ_FLAGS); |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 159 | ret = request_irq(irq, vgic_v4_doorbell_handler, |
| 160 | 0, "vcpu", vcpu); |
| 161 | if (ret) { |
| 162 | kvm_err("failed to allocate vcpu IRQ%d\n", irq); |
| 163 | /* |
| 164 | * Trick: adjust the number of vpes so we know |
| 165 | * how many to nuke on teardown... |
| 166 | */ |
| 167 | dist->its_vm.nr_vpes = i; |
| 168 | break; |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | if (ret) |
| 173 | vgic_v4_teardown(kvm); |
| 174 | |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 175 | return ret; |
| 176 | } |
| 177 | |
| 178 | /** |
| 179 | * vgic_v4_teardown - Free the GICv4 data structures |
| 180 | * @kvm: Pointer to the VM being destroyed |
| 181 | * |
| 182 | * Relies on kvm->lock to be held. |
| 183 | */ |
| 184 | void vgic_v4_teardown(struct kvm *kvm) |
| 185 | { |
| 186 | struct its_vm *its_vm = &kvm->arch.vgic.its_vm; |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 187 | int i; |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 188 | |
| 189 | if (!its_vm->vpes) |
| 190 | return; |
| 191 | |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 192 | for (i = 0; i < its_vm->nr_vpes; i++) { |
| 193 | struct kvm_vcpu *vcpu = kvm_get_vcpu(kvm, i); |
| 194 | int irq = its_vm->vpes[i]->irq; |
| 195 | |
Marc Zyngier | d3d83f7 | 2017-10-27 15:28:53 +0100 | [diff] [blame] | 196 | irq_clear_status_flags(irq, DB_IRQ_FLAGS); |
Marc Zyngier | bdb2d2c | 2017-10-27 15:28:48 +0100 | [diff] [blame] | 197 | free_irq(irq, vcpu); |
| 198 | } |
| 199 | |
Marc Zyngier | 74fe55d | 2017-10-27 15:28:38 +0100 | [diff] [blame] | 200 | its_free_vcpu_irqs(its_vm); |
| 201 | kfree(its_vm->vpes); |
| 202 | its_vm->nr_vpes = 0; |
| 203 | its_vm->vpes = NULL; |
| 204 | } |
Marc Zyngier | 196b136 | 2017-10-27 15:28:39 +0100 | [diff] [blame] | 205 | |
Marc Zyngier | 6277579 | 2017-10-27 15:28:50 +0100 | [diff] [blame] | 206 | int vgic_v4_sync_hwstate(struct kvm_vcpu *vcpu) |
| 207 | { |
| 208 | if (!vgic_supports_direct_msis(vcpu->kvm)) |
| 209 | return 0; |
| 210 | |
| 211 | return its_schedule_vpe(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe, false); |
| 212 | } |
| 213 | |
| 214 | int vgic_v4_flush_hwstate(struct kvm_vcpu *vcpu) |
| 215 | { |
| 216 | int irq = vcpu->arch.vgic_cpu.vgic_v3.its_vpe.irq; |
| 217 | int err; |
| 218 | |
| 219 | if (!vgic_supports_direct_msis(vcpu->kvm)) |
| 220 | return 0; |
| 221 | |
| 222 | /* |
| 223 | * Before making the VPE resident, make sure the redistributor |
| 224 | * corresponding to our current CPU expects us here. See the |
| 225 | * doc in drivers/irqchip/irq-gic-v4.c to understand how this |
| 226 | * turns into a VMOVP command at the ITS level. |
| 227 | */ |
| 228 | err = irq_set_affinity(irq, cpumask_of(smp_processor_id())); |
| 229 | if (err) |
| 230 | return err; |
| 231 | |
| 232 | err = its_schedule_vpe(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe, true); |
| 233 | if (err) |
| 234 | return err; |
| 235 | |
| 236 | /* |
| 237 | * Now that the VPE is resident, let's get rid of a potential |
| 238 | * doorbell interrupt that would still be pending. |
| 239 | */ |
| 240 | err = irq_set_irqchip_state(irq, IRQCHIP_STATE_PENDING, false); |
| 241 | |
| 242 | return err; |
| 243 | } |
| 244 | |
Marc Zyngier | 196b136 | 2017-10-27 15:28:39 +0100 | [diff] [blame] | 245 | static struct vgic_its *vgic_get_its(struct kvm *kvm, |
| 246 | struct kvm_kernel_irq_routing_entry *irq_entry) |
| 247 | { |
| 248 | struct kvm_msi msi = (struct kvm_msi) { |
| 249 | .address_lo = irq_entry->msi.address_lo, |
| 250 | .address_hi = irq_entry->msi.address_hi, |
| 251 | .data = irq_entry->msi.data, |
| 252 | .flags = irq_entry->msi.flags, |
| 253 | .devid = irq_entry->msi.devid, |
| 254 | }; |
| 255 | |
| 256 | return vgic_msi_to_its(kvm, &msi); |
| 257 | } |
| 258 | |
| 259 | int kvm_vgic_v4_set_forwarding(struct kvm *kvm, int virq, |
| 260 | struct kvm_kernel_irq_routing_entry *irq_entry) |
| 261 | { |
| 262 | struct vgic_its *its; |
| 263 | struct vgic_irq *irq; |
| 264 | struct its_vlpi_map map; |
| 265 | int ret; |
| 266 | |
| 267 | if (!vgic_supports_direct_msis(kvm)) |
| 268 | return 0; |
| 269 | |
| 270 | /* |
| 271 | * Get the ITS, and escape early on error (not a valid |
| 272 | * doorbell for any of our vITSs). |
| 273 | */ |
| 274 | its = vgic_get_its(kvm, irq_entry); |
| 275 | if (IS_ERR(its)) |
| 276 | return 0; |
| 277 | |
| 278 | mutex_lock(&its->its_lock); |
| 279 | |
| 280 | /* Perform then actual DevID/EventID -> LPI translation. */ |
| 281 | ret = vgic_its_resolve_lpi(kvm, its, irq_entry->msi.devid, |
| 282 | irq_entry->msi.data, &irq); |
| 283 | if (ret) |
| 284 | goto out; |
| 285 | |
| 286 | /* |
| 287 | * Emit the mapping request. If it fails, the ITS probably |
| 288 | * isn't v4 compatible, so let's silently bail out. Holding |
| 289 | * the ITS lock should ensure that nothing can modify the |
| 290 | * target vcpu. |
| 291 | */ |
| 292 | map = (struct its_vlpi_map) { |
| 293 | .vm = &kvm->arch.vgic.its_vm, |
| 294 | .vpe = &irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe, |
| 295 | .vintid = irq->intid, |
| 296 | .properties = ((irq->priority & 0xfc) | |
| 297 | (irq->enabled ? LPI_PROP_ENABLED : 0) | |
| 298 | LPI_PROP_GROUP1), |
| 299 | .db_enabled = true, |
| 300 | }; |
| 301 | |
| 302 | ret = its_map_vlpi(virq, &map); |
| 303 | if (ret) |
| 304 | goto out; |
| 305 | |
| 306 | irq->hw = true; |
| 307 | irq->host_irq = virq; |
| 308 | |
| 309 | out: |
| 310 | mutex_unlock(&its->its_lock); |
| 311 | return ret; |
| 312 | } |
| 313 | |
| 314 | int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int virq, |
| 315 | struct kvm_kernel_irq_routing_entry *irq_entry) |
| 316 | { |
| 317 | struct vgic_its *its; |
| 318 | struct vgic_irq *irq; |
| 319 | int ret; |
| 320 | |
| 321 | if (!vgic_supports_direct_msis(kvm)) |
| 322 | return 0; |
| 323 | |
| 324 | /* |
| 325 | * Get the ITS, and escape early on error (not a valid |
| 326 | * doorbell for any of our vITSs). |
| 327 | */ |
| 328 | its = vgic_get_its(kvm, irq_entry); |
| 329 | if (IS_ERR(its)) |
| 330 | return 0; |
| 331 | |
| 332 | mutex_lock(&its->its_lock); |
| 333 | |
| 334 | ret = vgic_its_resolve_lpi(kvm, its, irq_entry->msi.devid, |
| 335 | irq_entry->msi.data, &irq); |
| 336 | if (ret) |
| 337 | goto out; |
| 338 | |
| 339 | WARN_ON(!(irq->hw && irq->host_irq == virq)); |
Marc Zyngier | a05d1c0 | 2017-11-16 17:58:19 +0000 | [diff] [blame] | 340 | if (irq->hw) { |
| 341 | irq->hw = false; |
| 342 | ret = its_unmap_vlpi(virq); |
| 343 | } |
Marc Zyngier | 196b136 | 2017-10-27 15:28:39 +0100 | [diff] [blame] | 344 | |
| 345 | out: |
| 346 | mutex_unlock(&its->its_lock); |
| 347 | return ret; |
| 348 | } |
Marc Zyngier | df9ba95 | 2017-10-27 15:28:49 +0100 | [diff] [blame] | 349 | |
| 350 | void kvm_vgic_v4_enable_doorbell(struct kvm_vcpu *vcpu) |
| 351 | { |
| 352 | if (vgic_supports_direct_msis(vcpu->kvm)) { |
| 353 | int irq = vcpu->arch.vgic_cpu.vgic_v3.its_vpe.irq; |
| 354 | if (irq) |
| 355 | enable_irq(irq); |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | void kvm_vgic_v4_disable_doorbell(struct kvm_vcpu *vcpu) |
| 360 | { |
| 361 | if (vgic_supports_direct_msis(vcpu->kvm)) { |
| 362 | int irq = vcpu->arch.vgic_cpu.vgic_v3.its_vpe.irq; |
| 363 | if (irq) |
| 364 | disable_irq(irq); |
| 365 | } |
| 366 | } |