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