drivers/iommu/irq_remapping.c

   1 #include <linux/seq_file.h>
   2 #include <linux/cpumask.h>
   3 #include <linux/kernel.h>
   4 #include <linux/string.h>
   5 #include <linux/errno.h>
   6 #include <linux/msi.h>
   7 #include <linux/irq.h>
   8 #include <linux/pci.h>
   9
  10 #include <asm/hw_irq.h>
  11 #include <asm/irq_remapping.h>
  12 #include <asm/processor.h>
  13 #include <asm/x86_init.h>
  14 #include <asm/apic.h>
  15
  16 #include "irq_remapping.h"
  17
  18 int irq_remapping_enabled;
  19
  20 int disable_irq_remap;
  21 int irq_remap_broken;
  22 int disable_sourceid_checking;
  23 int no_x2apic_optout;
  24
  25 static struct irq_remap_ops *remap_ops;
  26
  27 static int msi_alloc_remapped_irq(struct pci_dev *pdev, int irq, int nvec);
  28 static int msi_setup_remapped_irq(struct pci_dev *pdev, unsigned int irq,
  29                                   int index, int sub_handle);
  30 static int set_remapped_irq_affinity(struct irq_data *data,
  31                                      const struct cpumask *mask,
  32                                      bool force);
  33
  34 static bool irq_remapped(struct irq_cfg *cfg)
  35 {
  36         return (cfg->remapped == 1);
  37 }
  38
  39 static void irq_remapping_disable_io_apic(void)
  40 {
  41         /*
  42          * With interrupt-remapping, for now we will use virtual wire A
  43          * mode, as virtual wire B is little complex (need to configure
  44          * both IOAPIC RTE as well as interrupt-remapping table entry).
  45          * As this gets called during crash dump, keep this simple for
  46          * now.
  47          */
  48         if (cpu_has_apic || apic_from_smp_config())
  49                 disconnect_bsp_APIC(0);
  50 }
  51
  52 static int do_setup_msi_irqs(struct pci_dev *dev, int nvec)
  53 {
  54         int ret, sub_handle, nvec_pow2, index = 0;
  55         unsigned int irq;
  56         struct msi_desc *msidesc;
  57
  58         WARN_ON(!list_is_singular(&dev->msi_list));
  59         msidesc = list_entry(dev->msi_list.next, struct msi_desc, list);
  60         WARN_ON(msidesc->irq);
  61         WARN_ON(msidesc->msi_attrib.multiple);
  62         WARN_ON(msidesc->nvec_used);
  63
  64         irq = irq_alloc_hwirqs(nvec, dev_to_node(&dev->dev));
  65         if (irq == 0)
  66                 return -ENOSPC;
  67
  68         nvec_pow2 = __roundup_pow_of_two(nvec);
  69         msidesc->nvec_used = nvec;
  70         msidesc->msi_attrib.multiple = ilog2(nvec_pow2);
  71         for (sub_handle = 0; sub_handle < nvec; sub_handle++) {
  72                 if (!sub_handle) {
  73                         index = msi_alloc_remapped_irq(dev, irq, nvec_pow2);
  74                         if (index < 0) {
  75                                 ret = index;
  76                                 goto error;
  77                         }
  78                 } else {
  79                         ret = msi_setup_remapped_irq(dev, irq + sub_handle,
  80                                                      index, sub_handle);
  81                         if (ret < 0)
  82                                 goto error;
  83                 }
  84                 ret = setup_msi_irq(dev, msidesc, irq, sub_handle);
  85                 if (ret < 0)
  86                         goto error;
  87         }
  88         return 0;
  89
  90 error:
  91         irq_free_hwirqs(irq, nvec);
  92
  93         /*
  94          * Restore altered MSI descriptor fields and prevent just destroyed
  95          * IRQs from tearing down again in default_teardown_msi_irqs()
  96          */
  97         msidesc->irq = 0;
  98         msidesc->nvec_used = 0;
  99         msidesc->msi_attrib.multiple = 0;
 100
 101         return ret;
 102 }
 103
 104 static int do_setup_msix_irqs(struct pci_dev *dev, int nvec)
 105 {
 106         int node, ret, sub_handle, index = 0;
 107         struct msi_desc *msidesc;
 108         unsigned int irq;
 109
 110         node            = dev_to_node(&dev->dev);
 111         sub_handle      = 0;
 112
 113         list_for_each_entry(msidesc, &dev->msi_list, list) {
 114
 115                 irq = irq_alloc_hwirq(node);
 116                 if (irq == 0)
 117                         return -1;
 118
 119                 if (sub_handle == 0)
 120                         ret = index = msi_alloc_remapped_irq(dev, irq, nvec);
 121                 else
 122                         ret = msi_setup_remapped_irq(dev, irq, index, sub_handle);
 123
 124                 if (ret < 0)
 125                         goto error;
 126
 127                 ret = setup_msi_irq(dev, msidesc, irq, 0);
 128                 if (ret < 0)
 129                         goto error;
 130
 131                 sub_handle += 1;
 132                 irq        += 1;
 133         }
 134
 135         return 0;
 136
 137 error:
 138         irq_free_hwirq(irq);
 139         return ret;
 140 }
 141
 142 static int irq_remapping_setup_msi_irqs(struct pci_dev *dev,
 143                                         int nvec, int type)
 144 {
 145         if (type == PCI_CAP_ID_MSI)
 146                 return do_setup_msi_irqs(dev, nvec);
 147         else
 148                 return do_setup_msix_irqs(dev, nvec);
 149 }
 150
 151 static void eoi_ioapic_pin_remapped(int apic, int pin, int vector)
 152 {
 153         /*
 154          * Intr-remapping uses pin number as the virtual vector
 155          * in the RTE. Actual vector is programmed in
 156          * intr-remapping table entry. Hence for the io-apic
 157          * EOI we use the pin number.
 158          */
 159         io_apic_eoi(apic, pin);
 160 }
 161
 162 static void __init irq_remapping_modify_x86_ops(void)
 163 {
 164         x86_io_apic_ops.disable         = irq_remapping_disable_io_apic;
 165         x86_io_apic_ops.set_affinity    = set_remapped_irq_affinity;
 166         x86_io_apic_ops.setup_entry     = setup_ioapic_remapped_entry;
 167         x86_io_apic_ops.eoi_ioapic_pin  = eoi_ioapic_pin_remapped;
 168         x86_msi.setup_msi_irqs          = irq_remapping_setup_msi_irqs;
 169         x86_msi.setup_hpet_msi          = setup_hpet_msi_remapped;
 170         x86_msi.compose_msi_msg         = compose_remapped_msi_msg;
 171 }
 172
 173 static __init int setup_nointremap(char *str)
 174 {
 175         disable_irq_remap = 1;
 176         return 0;
 177 }
 178 early_param("nointremap", setup_nointremap);
 179
 180 static __init int setup_irqremap(char *str)
 181 {
 182         if (!str)
 183                 return -EINVAL;
 184
 185         while (*str) {
 186                 if (!strncmp(str, "on", 2))
 187                         disable_irq_remap = 0;
 188                 else if (!strncmp(str, "off", 3))
 189                         disable_irq_remap = 1;
 190                 else if (!strncmp(str, "nosid", 5))
 191                         disable_sourceid_checking = 1;
 192                 else if (!strncmp(str, "no_x2apic_optout", 16))
 193                         no_x2apic_optout = 1;
 194
 195                 str += strcspn(str, ",");
 196                 while (*str == ',')
 197                         str++;
 198         }
 199
 200         return 0;
 201 }
 202 early_param("intremap", setup_irqremap);
 203
 204 void __init setup_irq_remapping_ops(void)
 205 {
 206         remap_ops = &intel_irq_remap_ops;
 207
 208 #ifdef CONFIG_AMD_IOMMU
 209         if (amd_iommu_irq_ops.prepare() == 0)
 210                 remap_ops = &amd_iommu_irq_ops;
 211 #endif
 212 }
 213
 214 void set_irq_remapping_broken(void)
 215 {
 216         irq_remap_broken = 1;
 217 }
 218
 219 int irq_remapping_supported(void)
 220 {
 221         if (disable_irq_remap)
 222                 return 0;
 223
 224         if (!remap_ops || !remap_ops->supported)
 225                 return 0;
 226
 227         return remap_ops->supported();
 228 }
 229
 230 int __init irq_remapping_prepare(void)
 231 {
 232         if (!remap_ops || !remap_ops->prepare)
 233                 return -ENODEV;
 234
 235         return remap_ops->prepare();
 236 }
 237
 238 int __init irq_remapping_enable(void)
 239 {
 240         int ret;
 241
 242         if (!remap_ops || !remap_ops->enable)
 243                 return -ENODEV;
 244
 245         ret = remap_ops->enable();
 246
 247         if (irq_remapping_enabled)
 248                 irq_remapping_modify_x86_ops();
 249
 250         return ret;
 251 }
 252
 253 void irq_remapping_disable(void)
 254 {
 255         if (!irq_remapping_enabled ||
 256             !remap_ops ||
 257             !remap_ops->disable)
 258                 return;
 259
 260         remap_ops->disable();
 261 }
 262
 263 int irq_remapping_reenable(int mode)
 264 {
 265         if (!irq_remapping_enabled ||
 266             !remap_ops ||
 267             !remap_ops->reenable)
 268                 return 0;
 269
 270         return remap_ops->reenable(mode);
 271 }
 272
 273 int __init irq_remap_enable_fault_handling(void)
 274 {
 275         if (!irq_remapping_enabled)
 276                 return 0;
 277
 278         if (!remap_ops || !remap_ops->enable_faulting)
 279                 return -ENODEV;
 280
 281         return remap_ops->enable_faulting();
 282 }
 283
 284 int setup_ioapic_remapped_entry(int irq,
 285                                 struct IO_APIC_route_entry *entry,
 286                                 unsigned int destination, int vector,
 287                                 struct io_apic_irq_attr *attr)
 288 {
 289         if (!remap_ops || !remap_ops->setup_ioapic_entry)
 290                 return -ENODEV;
 291
 292         return remap_ops->setup_ioapic_entry(irq, entry, destination,
 293                                              vector, attr);
 294 }
 295
 296 static int set_remapped_irq_affinity(struct irq_data *data,
 297                                      const struct cpumask *mask, bool force)
 298 {
 299         if (!config_enabled(CONFIG_SMP) || !remap_ops ||
 300             !remap_ops->set_affinity)
 301                 return 0;
 302
 303         return remap_ops->set_affinity(data, mask, force);
 304 }
 305
 306 void free_remapped_irq(int irq)
 307 {
 308         struct irq_cfg *cfg = irq_get_chip_data(irq);
 309
 310         if (!remap_ops || !remap_ops->free_irq)
 311                 return;
 312
 313         if (irq_remapped(cfg))
 314                 remap_ops->free_irq(irq);
 315 }
 316
 317 void compose_remapped_msi_msg(struct pci_dev *pdev,
 318                               unsigned int irq, unsigned int dest,
 319                               struct msi_msg *msg, u8 hpet_id)
 320 {
 321         struct irq_cfg *cfg = irq_get_chip_data(irq);
 322
 323         if (!irq_remapped(cfg))
 324                 native_compose_msi_msg(pdev, irq, dest, msg, hpet_id);
 325         else if (remap_ops && remap_ops->compose_msi_msg)
 326                 remap_ops->compose_msi_msg(pdev, irq, dest, msg, hpet_id);
 327 }
 328
 329 static int msi_alloc_remapped_irq(struct pci_dev *pdev, int irq, int nvec)
 330 {
 331         if (!remap_ops || !remap_ops->msi_alloc_irq)
 332                 return -ENODEV;
 333
 334         return remap_ops->msi_alloc_irq(pdev, irq, nvec);
 335 }
 336
 337 static int msi_setup_remapped_irq(struct pci_dev *pdev, unsigned int irq,
 338                                   int index, int sub_handle)
 339 {
 340         if (!remap_ops || !remap_ops->msi_setup_irq)
 341                 return -ENODEV;
 342
 343         return remap_ops->msi_setup_irq(pdev, irq, index, sub_handle);
 344 }
 345
 346 int setup_hpet_msi_remapped(unsigned int irq, unsigned int id)
 347 {
 348         if (!remap_ops || !remap_ops->setup_hpet_msi)
 349                 return -ENODEV;
 350
 351         return remap_ops->setup_hpet_msi(irq, id);
 352 }
 353
 354 void panic_if_irq_remap(const char *msg)
 355 {
 356         if (irq_remapping_enabled)
 357                 panic(msg);
 358 }
 359
 360 static void ir_ack_apic_edge(struct irq_data *data)
 361 {
 362         ack_APIC_irq();
 363 }
 364
 365 static void ir_ack_apic_level(struct irq_data *data)
 366 {
 367         ack_APIC_irq();
 368         eoi_ioapic_irq(data->irq, data->chip_data);
 369 }
 370
 371 static void ir_print_prefix(struct irq_data *data, struct seq_file *p)
 372 {
 373         seq_printf(p, " IR-%s", data->chip->name);
 374 }
 375
 376 void irq_remap_modify_chip_defaults(struct irq_chip *chip)
 377 {
 378         chip->irq_print_chip = ir_print_prefix;
 379         chip->irq_ack = ir_ack_apic_edge;
 380         chip->irq_eoi = ir_ack_apic_level;
 381         chip->irq_set_affinity = x86_io_apic_ops.set_affinity;
 382 }
 383
 384 bool setup_remapped_irq(int irq, struct irq_cfg *cfg, struct irq_chip *chip)
 385 {
 386         if (!irq_remapped(cfg))
 387                 return false;
 388         irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
 389         irq_remap_modify_chip_defaults(chip);
 390         return true;
 391 }