arch/powerpc/platforms/pseries/papr_scm.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #define pr_fmt(fmt)     "papr-scm: " fmt
   4
   5 #include <linux/of.h>
   6 #include <linux/kernel.h>
   7 #include <linux/module.h>
   8 #include <linux/ioport.h>
   9 #include <linux/slab.h>
  10 #include <linux/ndctl.h>
  11 #include <linux/sched.h>
  12 #include <linux/libnvdimm.h>
  13 #include <linux/platform_device.h>
  14 #include <linux/delay.h>
  15
  16 #include <asm/plpar_wrappers.h>
  17
  18 #define BIND_ANY_ADDR (~0ul)
  19
  20 #define PAPR_SCM_DIMM_CMD_MASK \
  21         ((1ul << ND_CMD_GET_CONFIG_SIZE) | \
  22          (1ul << ND_CMD_GET_CONFIG_DATA) | \
  23          (1ul << ND_CMD_SET_CONFIG_DATA))
  24
  25 struct papr_scm_priv {
  26         struct platform_device *pdev;
  27         struct device_node *dn;
  28         uint32_t drc_index;
  29         uint64_t blocks;
  30         uint64_t block_size;
  31         int metadata_size;
  32         bool is_volatile;
  33
  34         uint64_t bound_addr;
  35
  36         struct nvdimm_bus_descriptor bus_desc;
  37         struct nvdimm_bus *bus;
  38         struct nvdimm *nvdimm;
  39         struct resource res;
  40         struct nd_region *region;
  41         struct nd_interleave_set nd_set;
  42 };
  43
  44 static int drc_pmem_bind(struct papr_scm_priv *p)
  45 {
  46         unsigned long ret[PLPAR_HCALL_BUFSIZE];
  47         uint64_t saved = 0;
  48         uint64_t token;
  49         int64_t rc;
  50
  51         /*
  52          * When the hypervisor cannot map all the requested memory in a single
  53          * hcall it returns H_BUSY and we call again with the token until
  54          * we get H_SUCCESS. Aborting the retry loop before getting H_SUCCESS
  55          * leave the system in an undefined state, so we wait.
  56          */
  57         token = 0;
  58
  59         do {
  60                 rc = plpar_hcall(H_SCM_BIND_MEM, ret, p->drc_index, 0,
  61                                 p->blocks, BIND_ANY_ADDR, token);
  62                 token = ret[0];
  63                 if (!saved)
  64                         saved = ret[1];
  65                 cond_resched();
  66         } while (rc == H_BUSY);
  67
  68         if (rc)
  69                 return rc;
  70
  71         p->bound_addr = saved;
  72         dev_dbg(&p->pdev->dev, "bound drc 0x%x to 0x%lx\n",
  73                 p->drc_index, (unsigned long)saved);
  74         return rc;
  75 }
  76
  77 static void drc_pmem_unbind(struct papr_scm_priv *p)
  78 {
  79         unsigned long ret[PLPAR_HCALL_BUFSIZE];
  80         uint64_t token = 0;
  81         int64_t rc;
  82
  83         dev_dbg(&p->pdev->dev, "unbind drc 0x%x\n", p->drc_index);
  84
  85         /* NB: unbind has the same retry requirements as drc_pmem_bind() */
  86         do {
  87
  88                 /* Unbind of all SCM resources associated with drcIndex */
  89                 rc = plpar_hcall(H_SCM_UNBIND_ALL, ret, H_UNBIND_SCOPE_DRC,
  90                                  p->drc_index, token);
  91                 token = ret[0];
  92
  93                 /* Check if we are stalled for some time */
  94                 if (H_IS_LONG_BUSY(rc)) {
  95                         msleep(get_longbusy_msecs(rc));
  96                         rc = H_BUSY;
  97                 } else if (rc == H_BUSY) {
  98                         cond_resched();
  99                 }
 100
 101         } while (rc == H_BUSY);
 102
 103         if (rc)
 104                 dev_err(&p->pdev->dev, "unbind error: %lld\n", rc);
 105         else
 106                 dev_dbg(&p->pdev->dev, "unbind drc 0x%x complete\n",
 107                         p->drc_index);
 108
 109         return;
 110 }
 111
 112 static int drc_pmem_query_n_bind(struct papr_scm_priv *p)
 113 {
 114         unsigned long start_addr;
 115         unsigned long end_addr;
 116         unsigned long ret[PLPAR_HCALL_BUFSIZE];
 117         int64_t rc;
 118
 119
 120         rc = plpar_hcall(H_SCM_QUERY_BLOCK_MEM_BINDING, ret,
 121                          p->drc_index, 0);
 122         if (rc)
 123                 goto err_out;
 124         start_addr = ret[0];
 125
 126         /* Make sure the full region is bound. */
 127         rc = plpar_hcall(H_SCM_QUERY_BLOCK_MEM_BINDING, ret,
 128                          p->drc_index, p->blocks - 1);
 129         if (rc)
 130                 goto err_out;
 131         end_addr = ret[0];
 132
 133         if ((end_addr - start_addr) != ((p->blocks - 1) * p->block_size))
 134                 goto err_out;
 135
 136         p->bound_addr = start_addr;
 137         dev_dbg(&p->pdev->dev, "bound drc 0x%x to 0x%lx\n", p->drc_index, start_addr);
 138         return rc;
 139
 140 err_out:
 141         dev_info(&p->pdev->dev,
 142                  "Failed to query, trying an unbind followed by bind");
 143         drc_pmem_unbind(p);
 144         return drc_pmem_bind(p);
 145 }
 146
 147
 148 static int papr_scm_meta_get(struct papr_scm_priv *p,
 149                              struct nd_cmd_get_config_data_hdr *hdr)
 150 {
 151         unsigned long data[PLPAR_HCALL_BUFSIZE];
 152         unsigned long offset, data_offset;
 153         int len, read;
 154         int64_t ret;
 155
 156         if ((hdr->in_offset + hdr->in_length) > p->metadata_size)
 157                 return -EINVAL;
 158
 159         for (len = hdr->in_length; len; len -= read) {
 160
 161                 data_offset = hdr->in_length - len;
 162                 offset = hdr->in_offset + data_offset;
 163
 164                 if (len >= 8)
 165                         read = 8;
 166                 else if (len >= 4)
 167                         read = 4;
 168                 else if (len >= 2)
 169                         read = 2;
 170                 else
 171                         read = 1;
 172
 173                 ret = plpar_hcall(H_SCM_READ_METADATA, data, p->drc_index,
 174                                   offset, read);
 175
 176                 if (ret == H_PARAMETER) /* bad DRC index */
 177                         return -ENODEV;
 178                 if (ret)
 179                         return -EINVAL; /* other invalid parameter */
 180
 181                 switch (read) {
 182                 case 8:
 183                         *(uint64_t *)(hdr->out_buf + data_offset) = be64_to_cpu(data[0]);
 184                         break;
 185                 case 4:
 186                         *(uint32_t *)(hdr->out_buf + data_offset) = be32_to_cpu(data[0] & 0xffffffff);
 187                         break;
 188
 189                 case 2:
 190                         *(uint16_t *)(hdr->out_buf + data_offset) = be16_to_cpu(data[0] & 0xffff);
 191                         break;
 192
 193                 case 1:
 194                         *(uint8_t *)(hdr->out_buf + data_offset) = (data[0] & 0xff);
 195                         break;
 196                 }
 197         }
 198         return 0;
 199 }
 200
 201 static int papr_scm_meta_set(struct papr_scm_priv *p,
 202                              struct nd_cmd_set_config_hdr *hdr)
 203 {
 204         unsigned long offset, data_offset;
 205         int len, wrote;
 206         unsigned long data;
 207         __be64 data_be;
 208         int64_t ret;
 209
 210         if ((hdr->in_offset + hdr->in_length) > p->metadata_size)
 211                 return -EINVAL;
 212
 213         for (len = hdr->in_length; len; len -= wrote) {
 214
 215                 data_offset = hdr->in_length - len;
 216                 offset = hdr->in_offset + data_offset;
 217
 218                 if (len >= 8) {
 219                         data = *(uint64_t *)(hdr->in_buf + data_offset);
 220                         data_be = cpu_to_be64(data);
 221                         wrote = 8;
 222                 } else if (len >= 4) {
 223                         data = *(uint32_t *)(hdr->in_buf + data_offset);
 224                         data &= 0xffffffff;
 225                         data_be = cpu_to_be32(data);
 226                         wrote = 4;
 227                 } else if (len >= 2) {
 228                         data = *(uint16_t *)(hdr->in_buf + data_offset);
 229                         data &= 0xffff;
 230                         data_be = cpu_to_be16(data);
 231                         wrote = 2;
 232                 } else {
 233                         data_be = *(uint8_t *)(hdr->in_buf + data_offset);
 234                         data_be &= 0xff;
 235                         wrote = 1;
 236                 }
 237
 238                 ret = plpar_hcall_norets(H_SCM_WRITE_METADATA, p->drc_index,
 239                                          offset, data_be, wrote);
 240                 if (ret == H_PARAMETER) /* bad DRC index */
 241                         return -ENODEV;
 242                 if (ret)
 243                         return -EINVAL; /* other invalid parameter */
 244         }
 245
 246         return 0;
 247 }
 248
 249 int papr_scm_ndctl(struct nvdimm_bus_descriptor *nd_desc, struct nvdimm *nvdimm,
 250                 unsigned int cmd, void *buf, unsigned int buf_len, int *cmd_rc)
 251 {
 252         struct nd_cmd_get_config_size *get_size_hdr;
 253         struct papr_scm_priv *p;
 254
 255         /* Only dimm-specific calls are supported atm */
 256         if (!nvdimm)
 257                 return -EINVAL;
 258
 259         p = nvdimm_provider_data(nvdimm);
 260
 261         switch (cmd) {
 262         case ND_CMD_GET_CONFIG_SIZE:
 263                 get_size_hdr = buf;
 264
 265                 get_size_hdr->status = 0;
 266                 get_size_hdr->max_xfer = 8;
 267                 get_size_hdr->config_size = p->metadata_size;
 268                 *cmd_rc = 0;
 269                 break;
 270
 271         case ND_CMD_GET_CONFIG_DATA:
 272                 *cmd_rc = papr_scm_meta_get(p, buf);
 273                 break;
 274
 275         case ND_CMD_SET_CONFIG_DATA:
 276                 *cmd_rc = papr_scm_meta_set(p, buf);
 277                 break;
 278
 279         default:
 280                 return -EINVAL;
 281         }
 282
 283         dev_dbg(&p->pdev->dev, "returned with cmd_rc = %d\n", *cmd_rc);
 284
 285         return 0;
 286 }
 287
 288 static inline int papr_scm_node(int node)
 289 {
 290         int min_dist = INT_MAX, dist;
 291         int nid, min_node;
 292
 293         if ((node == NUMA_NO_NODE) || node_online(node))
 294                 return node;
 295
 296         min_node = first_online_node;
 297         for_each_online_node(nid) {
 298                 dist = node_distance(node, nid);
 299                 if (dist < min_dist) {
 300                         min_dist = dist;
 301                         min_node = nid;
 302                 }
 303         }
 304         return min_node;
 305 }
 306
 307 static int papr_scm_nvdimm_init(struct papr_scm_priv *p)
 308 {
 309         struct device *dev = &p->pdev->dev;
 310         struct nd_mapping_desc mapping;
 311         struct nd_region_desc ndr_desc;
 312         unsigned long dimm_flags;
 313         int target_nid, online_nid;
 314
 315         p->bus_desc.ndctl = papr_scm_ndctl;
 316         p->bus_desc.module = THIS_MODULE;
 317         p->bus_desc.of_node = p->pdev->dev.of_node;
 318         p->bus_desc.provider_name = kstrdup(p->pdev->name, GFP_KERNEL);
 319
 320         if (!p->bus_desc.provider_name)
 321                 return -ENOMEM;
 322
 323         p->bus = nvdimm_bus_register(NULL, &p->bus_desc);
 324         if (!p->bus) {
 325                 dev_err(dev, "Error creating nvdimm bus %pOF\n", p->dn);
 326                 kfree(p->bus_desc.provider_name);
 327                 return -ENXIO;
 328         }
 329
 330         dimm_flags = 0;
 331         set_bit(NDD_ALIASING, &dimm_flags);
 332
 333         p->nvdimm = nvdimm_create(p->bus, p, NULL, dimm_flags,
 334                                   PAPR_SCM_DIMM_CMD_MASK, 0, NULL);
 335         if (!p->nvdimm) {
 336                 dev_err(dev, "Error creating DIMM object for %pOF\n", p->dn);
 337                 goto err;
 338         }
 339
 340         if (nvdimm_bus_check_dimm_count(p->bus, 1))
 341                 goto err;
 342
 343         /* now add the region */
 344
 345         memset(&mapping, 0, sizeof(mapping));
 346         mapping.nvdimm = p->nvdimm;
 347         mapping.start = 0;
 348         mapping.size = p->blocks * p->block_size; // XXX: potential overflow?
 349
 350         memset(&ndr_desc, 0, sizeof(ndr_desc));
 351         target_nid = dev_to_node(&p->pdev->dev);
 352         online_nid = papr_scm_node(target_nid);
 353         ndr_desc.numa_node = online_nid;
 354         ndr_desc.target_node = target_nid;
 355         ndr_desc.res = &p->res;
 356         ndr_desc.of_node = p->dn;
 357         ndr_desc.provider_data = p;
 358         ndr_desc.mapping = &mapping;
 359         ndr_desc.num_mappings = 1;
 360         ndr_desc.nd_set = &p->nd_set;
 361
 362         if (p->is_volatile)
 363                 p->region = nvdimm_volatile_region_create(p->bus, &ndr_desc);
 364         else
 365                 p->region = nvdimm_pmem_region_create(p->bus, &ndr_desc);
 366         if (!p->region) {
 367                 dev_err(dev, "Error registering region %pR from %pOF\n",
 368                                 ndr_desc.res, p->dn);
 369                 goto err;
 370         }
 371         if (target_nid != online_nid)
 372                 dev_info(dev, "Region registered with target node %d and online node %d",
 373                          target_nid, online_nid);
 374
 375         return 0;
 376
 377 err:    nvdimm_bus_unregister(p->bus);
 378         kfree(p->bus_desc.provider_name);
 379         return -ENXIO;
 380 }
 381
 382 static int papr_scm_probe(struct platform_device *pdev)
 383 {
 384         struct device_node *dn = pdev->dev.of_node;
 385         u32 drc_index, metadata_size;
 386         u64 blocks, block_size;
 387         struct papr_scm_priv *p;
 388         const char *uuid_str;
 389         u64 uuid[2];
 390         int rc;
 391
 392         /* check we have all the required DT properties */
 393         if (of_property_read_u32(dn, "ibm,my-drc-index", &drc_index)) {
 394                 dev_err(&pdev->dev, "%pOF: missing drc-index!\n", dn);
 395                 return -ENODEV;
 396         }
 397
 398         if (of_property_read_u64(dn, "ibm,block-size", &block_size)) {
 399                 dev_err(&pdev->dev, "%pOF: missing block-size!\n", dn);
 400                 return -ENODEV;
 401         }
 402
 403         if (of_property_read_u64(dn, "ibm,number-of-blocks", &blocks)) {
 404                 dev_err(&pdev->dev, "%pOF: missing number-of-blocks!\n", dn);
 405                 return -ENODEV;
 406         }
 407
 408         if (of_property_read_string(dn, "ibm,unit-guid", &uuid_str)) {
 409                 dev_err(&pdev->dev, "%pOF: missing unit-guid!\n", dn);
 410                 return -ENODEV;
 411         }
 412
 413
 414         p = kzalloc(sizeof(*p), GFP_KERNEL);
 415         if (!p)
 416                 return -ENOMEM;
 417
 418         /* optional DT properties */
 419         of_property_read_u32(dn, "ibm,metadata-size", &metadata_size);
 420
 421         p->dn = dn;
 422         p->drc_index = drc_index;
 423         p->block_size = block_size;
 424         p->blocks = blocks;
 425         p->is_volatile = !of_property_read_bool(dn, "ibm,cache-flush-required");
 426
 427         /* We just need to ensure that set cookies are unique across */
 428         uuid_parse(uuid_str, (uuid_t *) uuid);
 429         /*
 430          * cookie1 and cookie2 are not really little endian
 431          * we store a little endian representation of the
 432          * uuid str so that we can compare this with the label
 433          * area cookie irrespective of the endian config with which
 434          * the kernel is built.
 435          */
 436         p->nd_set.cookie1 = cpu_to_le64(uuid[0]);
 437         p->nd_set.cookie2 = cpu_to_le64(uuid[1]);
 438
 439         /* might be zero */
 440         p->metadata_size = metadata_size;
 441         p->pdev = pdev;
 442
 443         /* request the hypervisor to bind this region to somewhere in memory */
 444         rc = drc_pmem_bind(p);
 445
 446         /* If phyp says drc memory still bound then force unbound and retry */
 447         if (rc == H_OVERLAP)
 448                 rc = drc_pmem_query_n_bind(p);
 449
 450         if (rc != H_SUCCESS) {
 451                 dev_err(&p->pdev->dev, "bind err: %d\n", rc);
 452                 rc = -ENXIO;
 453                 goto err;
 454         }
 455
 456         /* setup the resource for the newly bound range */
 457         p->res.start = p->bound_addr;
 458         p->res.end   = p->bound_addr + p->blocks * p->block_size - 1;
 459         p->res.name  = pdev->name;
 460         p->res.flags = IORESOURCE_MEM;
 461
 462         rc = papr_scm_nvdimm_init(p);
 463         if (rc)
 464                 goto err2;
 465
 466         platform_set_drvdata(pdev, p);
 467
 468         return 0;
 469
 470 err2:   drc_pmem_unbind(p);
 471 err:    kfree(p);
 472         return rc;
 473 }
 474
 475 static int papr_scm_remove(struct platform_device *pdev)
 476 {
 477         struct papr_scm_priv *p = platform_get_drvdata(pdev);
 478
 479         nvdimm_bus_unregister(p->bus);
 480         drc_pmem_unbind(p);
 481         kfree(p->bus_desc.provider_name);
 482         kfree(p);
 483
 484         return 0;
 485 }
 486
 487 static const struct of_device_id papr_scm_match[] = {
 488         { .compatible = "ibm,pmemory" },
 489         { },
 490 };
 491
 492 static struct platform_driver papr_scm_driver = {
 493         .probe = papr_scm_probe,
 494         .remove = papr_scm_remove,
 495         .driver = {
 496                 .name = "papr_scm",
 497                 .of_match_table = papr_scm_match,
 498         },
 499 };
 500
 501 module_platform_driver(papr_scm_driver);
 502 MODULE_DEVICE_TABLE(of, papr_scm_match);
 503 MODULE_LICENSE("GPL");
 504 MODULE_AUTHOR("IBM Corporation");