arch/powerpc/platforms/pseries/msi.c

   1 /*
   2  * Copyright 2006 Jake Moilanen <moilanen@austin.ibm.com>, IBM Corp.
   3  * Copyright 2006-2007 Michael Ellerman, IBM Corp.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License
   7  * as published by the Free Software Foundation; version 2 of the
   8  * License.
   9  *
  10  */
  11
  12 #include <linux/device.h>
  13 #include <linux/irq.h>
  14 #include <linux/msi.h>
  15
  16 #include <asm/rtas.h>
  17 #include <asm/hw_irq.h>
  18 #include <asm/ppc-pci.h>
  19
  20 static int query_token, change_token;
  21
  22 #define RTAS_QUERY_FN           0
  23 #define RTAS_CHANGE_FN          1
  24 #define RTAS_RESET_FN           2
  25 #define RTAS_CHANGE_MSI_FN      3
  26 #define RTAS_CHANGE_MSIX_FN     4
  27 #define RTAS_CHANGE_32MSI_FN    5
  28
  29 /* RTAS Helpers */
  30
  31 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
  32 {
  33         u32 addr, seq_num, rtas_ret[3];
  34         unsigned long buid;
  35         int rc;
  36
  37         addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
  38         buid = pdn->phb->buid;
  39
  40         seq_num = 1;
  41         do {
  42                 if (func == RTAS_CHANGE_MSI_FN || func == RTAS_CHANGE_MSIX_FN ||
  43                     func == RTAS_CHANGE_32MSI_FN)
  44                         rc = rtas_call(change_token, 6, 4, rtas_ret, addr,
  45                                         BUID_HI(buid), BUID_LO(buid),
  46                                         func, num_irqs, seq_num);
  47                 else
  48                         rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
  49                                         BUID_HI(buid), BUID_LO(buid),
  50                                         func, num_irqs, seq_num);
  51
  52                 seq_num = rtas_ret[1];
  53         } while (rtas_busy_delay(rc));
  54
  55         /*
  56          * If the RTAS call succeeded, return the number of irqs allocated.
  57          * If not, make sure we return a negative error code.
  58          */
  59         if (rc == 0)
  60                 rc = rtas_ret[0];
  61         else if (rc > 0)
  62                 rc = -rc;
  63
  64         pr_debug("rtas_msi: ibm,change_msi(func=%d,num=%d), got %d rc = %d\n",
  65                  func, num_irqs, rtas_ret[0], rc);
  66
  67         return rc;
  68 }
  69
  70 static void rtas_disable_msi(struct pci_dev *pdev)
  71 {
  72         struct pci_dn *pdn;
  73
  74         pdn = pci_get_pdn(pdev);
  75         if (!pdn)
  76                 return;
  77
  78         /*
  79          * disabling MSI with the explicit interface also disables MSI-X
  80          */
  81         if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
  82                 /*
  83                  * may have failed because explicit interface is not
  84                  * present
  85                  */
  86                 if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
  87                         pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
  88                 }
  89         }
  90 }
  91
  92 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
  93 {
  94         u32 addr, rtas_ret[2];
  95         unsigned long buid;
  96         int rc;
  97
  98         addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
  99         buid = pdn->phb->buid;
 100
 101         do {
 102                 rc = rtas_call(query_token, 4, 3, rtas_ret, addr,
 103                                BUID_HI(buid), BUID_LO(buid), offset);
 104         } while (rtas_busy_delay(rc));
 105
 106         if (rc) {
 107                 pr_debug("rtas_msi: error (%d) querying source number\n", rc);
 108                 return rc;
 109         }
 110
 111         return rtas_ret[0];
 112 }
 113
 114 static void rtas_teardown_msi_irqs(struct pci_dev *pdev)
 115 {
 116         struct msi_desc *entry;
 117
 118         list_for_each_entry(entry, &pdev->msi_list, list) {
 119                 if (entry->irq == NO_IRQ)
 120                         continue;
 121
 122                 irq_set_msi_desc(entry->irq, NULL);
 123                 irq_dispose_mapping(entry->irq);
 124         }
 125
 126         rtas_disable_msi(pdev);
 127 }
 128
 129 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
 130 {
 131         struct device_node *dn;
 132         struct pci_dn *pdn;
 133         const u32 *req_msi;
 134
 135         pdn = pci_get_pdn(pdev);
 136         if (!pdn)
 137                 return -ENODEV;
 138
 139         dn = pdn->node;
 140
 141         req_msi = of_get_property(dn, prop_name, NULL);
 142         if (!req_msi) {
 143                 pr_debug("rtas_msi: No %s on %s\n", prop_name, dn->full_name);
 144                 return -ENOENT;
 145         }
 146
 147         if (*req_msi < nvec) {
 148                 pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
 149
 150                 if (*req_msi == 0) /* Be paranoid */
 151                         return -ENOSPC;
 152
 153                 return *req_msi;
 154         }
 155
 156         return 0;
 157 }
 158
 159 static int check_req_msi(struct pci_dev *pdev, int nvec)
 160 {
 161         return check_req(pdev, nvec, "ibm,req#msi");
 162 }
 163
 164 static int check_req_msix(struct pci_dev *pdev, int nvec)
 165 {
 166         return check_req(pdev, nvec, "ibm,req#msi-x");
 167 }
 168
 169 /* Quota calculation */
 170
 171 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
 172 {
 173         struct device_node *dn;
 174         const u32 *p;
 175
 176         dn = of_node_get(pci_device_to_OF_node(dev));
 177         while (dn) {
 178                 p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
 179                 if (p) {
 180                         pr_debug("rtas_msi: found prop on dn %s\n",
 181                                 dn->full_name);
 182                         *total = *p;
 183                         return dn;
 184                 }
 185
 186                 dn = of_get_next_parent(dn);
 187         }
 188
 189         return NULL;
 190 }
 191
 192 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
 193 {
 194         struct device_node *dn;
 195         struct eeh_dev *edev;
 196
 197         /* Found our PE and assume 8 at that point. */
 198
 199         dn = pci_device_to_OF_node(dev);
 200         if (!dn)
 201                 return NULL;
 202
 203         /* Get the top level device in the PE */
 204         edev = of_node_to_eeh_dev(dn);
 205         if (edev->pe)
 206                 edev = list_first_entry(&edev->pe->edevs, struct eeh_dev, list);
 207         dn = eeh_dev_to_of_node(edev);
 208         if (!dn)
 209                 return NULL;
 210
 211         /* We actually want the parent */
 212         dn = of_get_parent(dn);
 213         if (!dn)
 214                 return NULL;
 215
 216         /* Hardcode of 8 for old firmwares */
 217         *total = 8;
 218         pr_debug("rtas_msi: using PE dn %s\n", dn->full_name);
 219
 220         return dn;
 221 }
 222
 223 struct msi_counts {
 224         struct device_node *requestor;
 225         int num_devices;
 226         int request;
 227         int quota;
 228         int spare;
 229         int over_quota;
 230 };
 231
 232 static void *count_non_bridge_devices(struct device_node *dn, void *data)
 233 {
 234         struct msi_counts *counts = data;
 235         const u32 *p;
 236         u32 class;
 237
 238         pr_debug("rtas_msi: counting %s\n", dn->full_name);
 239
 240         p = of_get_property(dn, "class-code", NULL);
 241         class = p ? *p : 0;
 242
 243         if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
 244                 counts->num_devices++;
 245
 246         return NULL;
 247 }
 248
 249 static void *count_spare_msis(struct device_node *dn, void *data)
 250 {
 251         struct msi_counts *counts = data;
 252         const u32 *p;
 253         int req;
 254
 255         if (dn == counts->requestor)
 256                 req = counts->request;
 257         else {
 258                 /* We don't know if a driver will try to use MSI or MSI-X,
 259                  * so we just have to punt and use the larger of the two. */
 260                 req = 0;
 261                 p = of_get_property(dn, "ibm,req#msi", NULL);
 262                 if (p)
 263                         req = *p;
 264
 265                 p = of_get_property(dn, "ibm,req#msi-x", NULL);
 266                 if (p)
 267                         req = max(req, (int)*p);
 268         }
 269
 270         if (req < counts->quota)
 271                 counts->spare += counts->quota - req;
 272         else if (req > counts->quota)
 273                 counts->over_quota++;
 274
 275         return NULL;
 276 }
 277
 278 static int msi_quota_for_device(struct pci_dev *dev, int request)
 279 {
 280         struct device_node *pe_dn;
 281         struct msi_counts counts;
 282         int total;
 283
 284         pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
 285                   request);
 286
 287         pe_dn = find_pe_total_msi(dev, &total);
 288         if (!pe_dn)
 289                 pe_dn = find_pe_dn(dev, &total);
 290
 291         if (!pe_dn) {
 292                 pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
 293                 goto out;
 294         }
 295
 296         pr_debug("rtas_msi: found PE %s\n", pe_dn->full_name);
 297
 298         memset(&counts, 0, sizeof(struct msi_counts));
 299
 300         /* Work out how many devices we have below this PE */
 301         traverse_pci_devices(pe_dn, count_non_bridge_devices, &counts);
 302
 303         if (counts.num_devices == 0) {
 304                 pr_err("rtas_msi: found 0 devices under PE for %s\n",
 305                         pci_name(dev));
 306                 goto out;
 307         }
 308
 309         counts.quota = total / counts.num_devices;
 310         if (request <= counts.quota)
 311                 goto out;
 312
 313         /* else, we have some more calculating to do */
 314         counts.requestor = pci_device_to_OF_node(dev);
 315         counts.request = request;
 316         traverse_pci_devices(pe_dn, count_spare_msis, &counts);
 317
 318         /* If the quota isn't an integer multiple of the total, we can
 319          * use the remainder as spare MSIs for anyone that wants them. */
 320         counts.spare += total % counts.num_devices;
 321
 322         /* Divide any spare by the number of over-quota requestors */
 323         if (counts.over_quota)
 324                 counts.quota += counts.spare / counts.over_quota;
 325
 326         /* And finally clamp the request to the possibly adjusted quota */
 327         request = min(counts.quota, request);
 328
 329         pr_debug("rtas_msi: request clamped to quota %d\n", request);
 330 out:
 331         of_node_put(pe_dn);
 332
 333         return request;
 334 }
 335
 336 static int rtas_msi_check_device(struct pci_dev *pdev, int nvec, int type)
 337 {
 338         int quota, rc;
 339
 340         if (type == PCI_CAP_ID_MSIX)
 341                 rc = check_req_msix(pdev, nvec);
 342         else
 343                 rc = check_req_msi(pdev, nvec);
 344
 345         if (rc)
 346                 return rc;
 347
 348         quota = msi_quota_for_device(pdev, nvec);
 349
 350         if (quota && quota < nvec)
 351                 return quota;
 352
 353         return 0;
 354 }
 355
 356 static int check_msix_entries(struct pci_dev *pdev)
 357 {
 358         struct msi_desc *entry;
 359         int expected;
 360
 361         /* There's no way for us to express to firmware that we want
 362          * a discontiguous, or non-zero based, range of MSI-X entries.
 363          * So we must reject such requests. */
 364
 365         expected = 0;
 366         list_for_each_entry(entry, &pdev->msi_list, list) {
 367                 if (entry->msi_attrib.entry_nr != expected) {
 368                         pr_debug("rtas_msi: bad MSI-X entries.\n");
 369                         return -EINVAL;
 370                 }
 371                 expected++;
 372         }
 373
 374         return 0;
 375 }
 376
 377 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
 378 {
 379         u32 addr_hi, addr_lo;
 380
 381         /*
 382          * We should only get in here for IODA1 configs. This is based on the
 383          * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
 384          * support, and we are in a PCIe Gen2 slot.
 385          */
 386         dev_info(&pdev->dev,
 387                  "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
 388         pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
 389         addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
 390         pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
 391         pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
 392 }
 393
 394 static int rtas_setup_msi_irqs(struct pci_dev *pdev, int nvec_in, int type)
 395 {
 396         struct pci_dn *pdn;
 397         int hwirq, virq, i, rc;
 398         struct msi_desc *entry;
 399         struct msi_msg msg;
 400         int nvec = nvec_in;
 401         int use_32bit_msi_hack = 0;
 402
 403         pdn = pci_get_pdn(pdev);
 404         if (!pdn)
 405                 return -ENODEV;
 406
 407         if (type == PCI_CAP_ID_MSIX && check_msix_entries(pdev))
 408                 return -EINVAL;
 409
 410         /*
 411          * Firmware currently refuse any non power of two allocation
 412          * so we round up if the quota will allow it.
 413          */
 414         if (type == PCI_CAP_ID_MSIX) {
 415                 int m = roundup_pow_of_two(nvec);
 416                 int quota = msi_quota_for_device(pdev, m);
 417
 418                 if (quota >= m)
 419                         nvec = m;
 420         }
 421
 422         /*
 423          * Try the new more explicit firmware interface, if that fails fall
 424          * back to the old interface. The old interface is known to never
 425          * return MSI-Xs.
 426          */
 427 again:
 428         if (type == PCI_CAP_ID_MSI) {
 429                 if (pdev->no_64bit_msi) {
 430                         rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
 431                         if (rc < 0) {
 432                                 /*
 433                                  * We only want to run the 32 bit MSI hack below if
 434                                  * the max bus speed is Gen2 speed
 435                                  */
 436                                 if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
 437                                         return rc;
 438
 439                                 use_32bit_msi_hack = 1;
 440                         }
 441                 } else
 442                         rc = -1;
 443
 444                 if (rc < 0)
 445                         rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
 446
 447                 if (rc < 0) {
 448                         pr_debug("rtas_msi: trying the old firmware call.\n");
 449                         rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
 450                 }
 451
 452                 if (use_32bit_msi_hack && rc > 0)
 453                         rtas_hack_32bit_msi_gen2(pdev);
 454         } else
 455                 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
 456
 457         if (rc != nvec) {
 458                 if (nvec != nvec_in) {
 459                         nvec = nvec_in;
 460                         goto again;
 461                 }
 462                 pr_debug("rtas_msi: rtas_change_msi() failed\n");
 463                 return rc;
 464         }
 465
 466         i = 0;
 467         list_for_each_entry(entry, &pdev->msi_list, list) {
 468                 hwirq = rtas_query_irq_number(pdn, i++);
 469                 if (hwirq < 0) {
 470                         pr_debug("rtas_msi: error (%d) getting hwirq\n", rc);
 471                         return hwirq;
 472                 }
 473
 474                 virq = irq_create_mapping(NULL, hwirq);
 475
 476                 if (virq == NO_IRQ) {
 477                         pr_debug("rtas_msi: Failed mapping hwirq %d\n", hwirq);
 478                         return -ENOSPC;
 479                 }
 480
 481                 dev_dbg(&pdev->dev, "rtas_msi: allocated virq %d\n", virq);
 482                 irq_set_msi_desc(virq, entry);
 483
 484                 /* Read config space back so we can restore after reset */
 485                 read_msi_msg(virq, &msg);
 486                 entry->msg = msg;
 487         }
 488
 489         return 0;
 490 }
 491
 492 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
 493 {
 494         /* No LSI -> leave MSIs (if any) configured */
 495         if (pdev->irq == NO_IRQ) {
 496                 dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
 497                 return;
 498         }
 499
 500         /* No MSI -> MSIs can't have been assigned by fw, leave LSI */
 501         if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
 502                 dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
 503                 return;
 504         }
 505
 506         dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
 507         rtas_disable_msi(pdev);
 508 }
 509
 510 static int rtas_msi_init(void)
 511 {
 512         query_token  = rtas_token("ibm,query-interrupt-source-number");
 513         change_token = rtas_token("ibm,change-msi");
 514
 515         if ((query_token == RTAS_UNKNOWN_SERVICE) ||
 516                         (change_token == RTAS_UNKNOWN_SERVICE)) {
 517                 pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
 518                 return -1;
 519         }
 520
 521         pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
 522
 523         WARN_ON(ppc_md.setup_msi_irqs);
 524         ppc_md.setup_msi_irqs = rtas_setup_msi_irqs;
 525         ppc_md.teardown_msi_irqs = rtas_teardown_msi_irqs;
 526         ppc_md.msi_check_device = rtas_msi_check_device;
 527
 528         WARN_ON(ppc_md.pci_irq_fixup);
 529         ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
 530
 531         return 0;
 532 }
 533 arch_initcall(rtas_msi_init);
 534