drivers/misc/cxl/native.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright 2014 IBM Corp.
   4  */
   5
   6 #include <linux/spinlock.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/clock.h>
   9 #include <linux/slab.h>
  10 #include <linux/mutex.h>
  11 #include <linux/mm.h>
  12 #include <linux/uaccess.h>
  13 #include <linux/delay.h>
  14 #include <asm/synch.h>
  15 #include <asm/switch_to.h>
  16 #include <misc/cxl-base.h>
  17
  18 #include "cxl.h"
  19 #include "trace.h"
  20
  21 static int afu_control(struct cxl_afu *afu, u64 command, u64 clear,
  22                        u64 result, u64 mask, bool enabled)
  23 {
  24         u64 AFU_Cntl;
  25         unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
  26         int rc = 0;
  27
  28         spin_lock(&afu->afu_cntl_lock);
  29         pr_devel("AFU command starting: %llx\n", command);
  30
  31         trace_cxl_afu_ctrl(afu, command);
  32
  33         AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
  34         cxl_p2n_write(afu, CXL_AFU_Cntl_An, (AFU_Cntl & ~clear) | command);
  35
  36         AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
  37         while ((AFU_Cntl & mask) != result) {
  38                 if (time_after_eq(jiffies, timeout)) {
  39                         dev_warn(&afu->dev, "WARNING: AFU control timed out!\n");
  40                         rc = -EBUSY;
  41                         goto out;
  42                 }
  43
  44                 if (!cxl_ops->link_ok(afu->adapter, afu)) {
  45                         afu->enabled = enabled;
  46                         rc = -EIO;
  47                         goto out;
  48                 }
  49
  50                 pr_devel_ratelimited("AFU control... (0x%016llx)\n",
  51                                      AFU_Cntl | command);
  52                 cpu_relax();
  53                 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
  54         }
  55
  56         if (AFU_Cntl & CXL_AFU_Cntl_An_RA) {
  57                 /*
  58                  * Workaround for a bug in the XSL used in the Mellanox CX4
  59                  * that fails to clear the RA bit after an AFU reset,
  60                  * preventing subsequent AFU resets from working.
  61                  */
  62                 cxl_p2n_write(afu, CXL_AFU_Cntl_An, AFU_Cntl & ~CXL_AFU_Cntl_An_RA);
  63         }
  64
  65         pr_devel("AFU command complete: %llx\n", command);
  66         afu->enabled = enabled;
  67 out:
  68         trace_cxl_afu_ctrl_done(afu, command, rc);
  69         spin_unlock(&afu->afu_cntl_lock);
  70
  71         return rc;
  72 }
  73
  74 static int afu_enable(struct cxl_afu *afu)
  75 {
  76         pr_devel("AFU enable request\n");
  77
  78         return afu_control(afu, CXL_AFU_Cntl_An_E, 0,
  79                            CXL_AFU_Cntl_An_ES_Enabled,
  80                            CXL_AFU_Cntl_An_ES_MASK, true);
  81 }
  82
  83 int cxl_afu_disable(struct cxl_afu *afu)
  84 {
  85         pr_devel("AFU disable request\n");
  86
  87         return afu_control(afu, 0, CXL_AFU_Cntl_An_E,
  88                            CXL_AFU_Cntl_An_ES_Disabled,
  89                            CXL_AFU_Cntl_An_ES_MASK, false);
  90 }
  91
  92 /* This will disable as well as reset */
  93 static int native_afu_reset(struct cxl_afu *afu)
  94 {
  95         int rc;
  96         u64 serr;
  97
  98         pr_devel("AFU reset request\n");
  99
 100         rc = afu_control(afu, CXL_AFU_Cntl_An_RA, 0,
 101                            CXL_AFU_Cntl_An_RS_Complete | CXL_AFU_Cntl_An_ES_Disabled,
 102                            CXL_AFU_Cntl_An_RS_MASK | CXL_AFU_Cntl_An_ES_MASK,
 103                            false);
 104
 105         /*
 106          * Re-enable any masked interrupts when the AFU is not
 107          * activated to avoid side effects after attaching a process
 108          * in dedicated mode.
 109          */
 110         if (afu->current_mode == 0) {
 111                 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
 112                 serr &= ~CXL_PSL_SERR_An_IRQ_MASKS;
 113                 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
 114         }
 115
 116         return rc;
 117 }
 118
 119 static int native_afu_check_and_enable(struct cxl_afu *afu)
 120 {
 121         if (!cxl_ops->link_ok(afu->adapter, afu)) {
 122                 WARN(1, "Refusing to enable afu while link down!\n");
 123                 return -EIO;
 124         }
 125         if (afu->enabled)
 126                 return 0;
 127         return afu_enable(afu);
 128 }
 129
 130 int cxl_psl_purge(struct cxl_afu *afu)
 131 {
 132         u64 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An);
 133         u64 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
 134         u64 dsisr, dar;
 135         u64 start, end;
 136         u64 trans_fault = 0x0ULL;
 137         unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
 138         int rc = 0;
 139
 140         trace_cxl_psl_ctrl(afu, CXL_PSL_SCNTL_An_Pc);
 141
 142         pr_devel("PSL purge request\n");
 143
 144         if (cxl_is_power8())
 145                 trans_fault = CXL_PSL_DSISR_TRANS;
 146         if (cxl_is_power9())
 147                 trans_fault = CXL_PSL9_DSISR_An_TF;
 148
 149         if (!cxl_ops->link_ok(afu->adapter, afu)) {
 150                 dev_warn(&afu->dev, "PSL Purge called with link down, ignoring\n");
 151                 rc = -EIO;
 152                 goto out;
 153         }
 154
 155         if ((AFU_Cntl & CXL_AFU_Cntl_An_ES_MASK) != CXL_AFU_Cntl_An_ES_Disabled) {
 156                 WARN(1, "psl_purge request while AFU not disabled!\n");
 157                 cxl_afu_disable(afu);
 158         }
 159
 160         cxl_p1n_write(afu, CXL_PSL_SCNTL_An,
 161                        PSL_CNTL | CXL_PSL_SCNTL_An_Pc);
 162         start = local_clock();
 163         PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An);
 164         while ((PSL_CNTL &  CXL_PSL_SCNTL_An_Ps_MASK)
 165                         == CXL_PSL_SCNTL_An_Ps_Pending) {
 166                 if (time_after_eq(jiffies, timeout)) {
 167                         dev_warn(&afu->dev, "WARNING: PSL Purge timed out!\n");
 168                         rc = -EBUSY;
 169                         goto out;
 170                 }
 171                 if (!cxl_ops->link_ok(afu->adapter, afu)) {
 172                         rc = -EIO;
 173                         goto out;
 174                 }
 175
 176                 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
 177                 pr_devel_ratelimited("PSL purging... PSL_CNTL: 0x%016llx  PSL_DSISR: 0x%016llx\n",
 178                                      PSL_CNTL, dsisr);
 179
 180                 if (dsisr & trans_fault) {
 181                         dar = cxl_p2n_read(afu, CXL_PSL_DAR_An);
 182                         dev_notice(&afu->dev, "PSL purge terminating pending translation, DSISR: 0x%016llx, DAR: 0x%016llx\n",
 183                                    dsisr, dar);
 184                         cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
 185                 } else if (dsisr) {
 186                         dev_notice(&afu->dev, "PSL purge acknowledging pending non-translation fault, DSISR: 0x%016llx\n",
 187                                    dsisr);
 188                         cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
 189                 } else {
 190                         cpu_relax();
 191                 }
 192                 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An);
 193         }
 194         end = local_clock();
 195         pr_devel("PSL purged in %lld ns\n", end - start);
 196
 197         cxl_p1n_write(afu, CXL_PSL_SCNTL_An,
 198                        PSL_CNTL & ~CXL_PSL_SCNTL_An_Pc);
 199 out:
 200         trace_cxl_psl_ctrl_done(afu, CXL_PSL_SCNTL_An_Pc, rc);
 201         return rc;
 202 }
 203
 204 static int spa_max_procs(int spa_size)
 205 {
 206         /*
 207          * From the CAIA:
 208          *    end_of_SPA_area = SPA_Base + ((n+4) * 128) + (( ((n*8) + 127) >> 7) * 128) + 255
 209          * Most of that junk is really just an overly-complicated way of saying
 210          * the last 256 bytes are __aligned(128), so it's really:
 211          *    end_of_SPA_area = end_of_PSL_queue_area + __aligned(128) 255
 212          * and
 213          *    end_of_PSL_queue_area = SPA_Base + ((n+4) * 128) + (n*8) - 1
 214          * so
 215          *    sizeof(SPA) = ((n+4) * 128) + (n*8) + __aligned(128) 256
 216          * Ignore the alignment (which is safe in this case as long as we are
 217          * careful with our rounding) and solve for n:
 218          */
 219         return ((spa_size / 8) - 96) / 17;
 220 }
 221
 222 static int cxl_alloc_spa(struct cxl_afu *afu, int mode)
 223 {
 224         unsigned spa_size;
 225
 226         /* Work out how many pages to allocate */
 227         afu->native->spa_order = -1;
 228         do {
 229                 afu->native->spa_order++;
 230                 spa_size = (1 << afu->native->spa_order) * PAGE_SIZE;
 231
 232                 if (spa_size > 0x100000) {
 233                         dev_warn(&afu->dev, "num_of_processes too large for the SPA, limiting to %i (0x%x)\n",
 234                                         afu->native->spa_max_procs, afu->native->spa_size);
 235                         if (mode != CXL_MODE_DEDICATED)
 236                                 afu->num_procs = afu->native->spa_max_procs;
 237                         break;
 238                 }
 239
 240                 afu->native->spa_size = spa_size;
 241                 afu->native->spa_max_procs = spa_max_procs(afu->native->spa_size);
 242         } while (afu->native->spa_max_procs < afu->num_procs);
 243
 244         if (!(afu->native->spa = (struct cxl_process_element *)
 245               __get_free_pages(GFP_KERNEL | __GFP_ZERO, afu->native->spa_order))) {
 246                 pr_err("cxl_alloc_spa: Unable to allocate scheduled process area\n");
 247                 return -ENOMEM;
 248         }
 249         pr_devel("spa pages: %i afu->spa_max_procs: %i   afu->num_procs: %i\n",
 250                  1<<afu->native->spa_order, afu->native->spa_max_procs, afu->num_procs);
 251
 252         return 0;
 253 }
 254
 255 static void attach_spa(struct cxl_afu *afu)
 256 {
 257         u64 spap;
 258
 259         afu->native->sw_command_status = (__be64 *)((char *)afu->native->spa +
 260                                             ((afu->native->spa_max_procs + 3) * 128));
 261
 262         spap = virt_to_phys(afu->native->spa) & CXL_PSL_SPAP_Addr;
 263         spap |= ((afu->native->spa_size >> (12 - CXL_PSL_SPAP_Size_Shift)) - 1) & CXL_PSL_SPAP_Size;
 264         spap |= CXL_PSL_SPAP_V;
 265         pr_devel("cxl: SPA allocated at 0x%p. Max processes: %i, sw_command_status: 0x%p CXL_PSL_SPAP_An=0x%016llx\n",
 266                 afu->native->spa, afu->native->spa_max_procs,
 267                 afu->native->sw_command_status, spap);
 268         cxl_p1n_write(afu, CXL_PSL_SPAP_An, spap);
 269 }
 270
 271 static inline void detach_spa(struct cxl_afu *afu)
 272 {
 273         cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0);
 274 }
 275
 276 void cxl_release_spa(struct cxl_afu *afu)
 277 {
 278         if (afu->native->spa) {
 279                 free_pages((unsigned long) afu->native->spa,
 280                         afu->native->spa_order);
 281                 afu->native->spa = NULL;
 282         }
 283 }
 284
 285 /*
 286  * Invalidation of all ERAT entries is no longer required by CAIA2. Use
 287  * only for debug.
 288  */
 289 int cxl_invalidate_all_psl9(struct cxl *adapter)
 290 {
 291         unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
 292         u64 ierat;
 293
 294         pr_devel("CXL adapter - invalidation of all ERAT entries\n");
 295
 296         /* Invalidates all ERAT entries for Radix or HPT */
 297         ierat = CXL_XSL9_IERAT_IALL;
 298         if (radix_enabled())
 299                 ierat |= CXL_XSL9_IERAT_INVR;
 300         cxl_p1_write(adapter, CXL_XSL9_IERAT, ierat);
 301
 302         while (cxl_p1_read(adapter, CXL_XSL9_IERAT) & CXL_XSL9_IERAT_IINPROG) {
 303                 if (time_after_eq(jiffies, timeout)) {
 304                         dev_warn(&adapter->dev,
 305                         "WARNING: CXL adapter invalidation of all ERAT entries timed out!\n");
 306                         return -EBUSY;
 307                 }
 308                 if (!cxl_ops->link_ok(adapter, NULL))
 309                         return -EIO;
 310                 cpu_relax();
 311         }
 312         return 0;
 313 }
 314
 315 int cxl_invalidate_all_psl8(struct cxl *adapter)
 316 {
 317         unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
 318
 319         pr_devel("CXL adapter wide TLBIA & SLBIA\n");
 320
 321         cxl_p1_write(adapter, CXL_PSL_AFUSEL, CXL_PSL_AFUSEL_A);
 322
 323         cxl_p1_write(adapter, CXL_PSL_TLBIA, CXL_TLB_SLB_IQ_ALL);
 324         while (cxl_p1_read(adapter, CXL_PSL_TLBIA) & CXL_TLB_SLB_P) {
 325                 if (time_after_eq(jiffies, timeout)) {
 326                         dev_warn(&adapter->dev, "WARNING: CXL adapter wide TLBIA timed out!\n");
 327                         return -EBUSY;
 328                 }
 329                 if (!cxl_ops->link_ok(adapter, NULL))
 330                         return -EIO;
 331                 cpu_relax();
 332         }
 333
 334         cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_ALL);
 335         while (cxl_p1_read(adapter, CXL_PSL_SLBIA) & CXL_TLB_SLB_P) {
 336                 if (time_after_eq(jiffies, timeout)) {
 337                         dev_warn(&adapter->dev, "WARNING: CXL adapter wide SLBIA timed out!\n");
 338                         return -EBUSY;
 339                 }
 340                 if (!cxl_ops->link_ok(adapter, NULL))
 341                         return -EIO;
 342                 cpu_relax();
 343         }
 344         return 0;
 345 }
 346
 347 int cxl_data_cache_flush(struct cxl *adapter)
 348 {
 349         u64 reg;
 350         unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
 351
 352         /*
 353          * Do a datacache flush only if datacache is available.
 354          * In case of PSL9D datacache absent hence flush operation.
 355          * would timeout.
 356          */
 357         if (adapter->native->no_data_cache) {
 358                 pr_devel("No PSL data cache. Ignoring cache flush req.\n");
 359                 return 0;
 360         }
 361
 362         pr_devel("Flushing data cache\n");
 363         reg = cxl_p1_read(adapter, CXL_PSL_Control);
 364         reg |= CXL_PSL_Control_Fr;
 365         cxl_p1_write(adapter, CXL_PSL_Control, reg);
 366
 367         reg = cxl_p1_read(adapter, CXL_PSL_Control);
 368         while ((reg & CXL_PSL_Control_Fs_MASK) != CXL_PSL_Control_Fs_Complete) {
 369                 if (time_after_eq(jiffies, timeout)) {
 370                         dev_warn(&adapter->dev, "WARNING: cache flush timed out!\n");
 371                         return -EBUSY;
 372                 }
 373
 374                 if (!cxl_ops->link_ok(adapter, NULL)) {
 375                         dev_warn(&adapter->dev, "WARNING: link down when flushing cache\n");
 376                         return -EIO;
 377                 }
 378                 cpu_relax();
 379                 reg = cxl_p1_read(adapter, CXL_PSL_Control);
 380         }
 381
 382         reg &= ~CXL_PSL_Control_Fr;
 383         cxl_p1_write(adapter, CXL_PSL_Control, reg);
 384         return 0;
 385 }
 386
 387 static int cxl_write_sstp(struct cxl_afu *afu, u64 sstp0, u64 sstp1)
 388 {
 389         int rc;
 390
 391         /* 1. Disable SSTP by writing 0 to SSTP1[V] */
 392         cxl_p2n_write(afu, CXL_SSTP1_An, 0);
 393
 394         /* 2. Invalidate all SLB entries */
 395         if ((rc = cxl_afu_slbia(afu)))
 396                 return rc;
 397
 398         /* 3. Set SSTP0_An */
 399         cxl_p2n_write(afu, CXL_SSTP0_An, sstp0);
 400
 401         /* 4. Set SSTP1_An */
 402         cxl_p2n_write(afu, CXL_SSTP1_An, sstp1);
 403
 404         return 0;
 405 }
 406
 407 /* Using per slice version may improve performance here. (ie. SLBIA_An) */
 408 static void slb_invalid(struct cxl_context *ctx)
 409 {
 410         struct cxl *adapter = ctx->afu->adapter;
 411         u64 slbia;
 412
 413         WARN_ON(!mutex_is_locked(&ctx->afu->native->spa_mutex));
 414
 415         cxl_p1_write(adapter, CXL_PSL_LBISEL,
 416                         ((u64)be32_to_cpu(ctx->elem->common.pid) << 32) |
 417                         be32_to_cpu(ctx->elem->lpid));
 418         cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_LPIDPID);
 419
 420         while (1) {
 421                 if (!cxl_ops->link_ok(adapter, NULL))
 422                         break;
 423                 slbia = cxl_p1_read(adapter, CXL_PSL_SLBIA);
 424                 if (!(slbia & CXL_TLB_SLB_P))
 425                         break;
 426                 cpu_relax();
 427         }
 428 }
 429
 430 static int do_process_element_cmd(struct cxl_context *ctx,
 431                                   u64 cmd, u64 pe_state)
 432 {
 433         u64 state;
 434         unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
 435         int rc = 0;
 436
 437         trace_cxl_llcmd(ctx, cmd);
 438
 439         WARN_ON(!ctx->afu->enabled);
 440
 441         ctx->elem->software_state = cpu_to_be32(pe_state);
 442         smp_wmb();
 443         *(ctx->afu->native->sw_command_status) = cpu_to_be64(cmd | 0 | ctx->pe);
 444         smp_mb();
 445         cxl_p1n_write(ctx->afu, CXL_PSL_LLCMD_An, cmd | ctx->pe);
 446         while (1) {
 447                 if (time_after_eq(jiffies, timeout)) {
 448                         dev_warn(&ctx->afu->dev, "WARNING: Process Element Command timed out!\n");
 449                         rc = -EBUSY;
 450                         goto out;
 451                 }
 452                 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
 453                         dev_warn(&ctx->afu->dev, "WARNING: Device link down, aborting Process Element Command!\n");
 454                         rc = -EIO;
 455                         goto out;
 456                 }
 457                 state = be64_to_cpup(ctx->afu->native->sw_command_status);
 458                 if (state == ~0ULL) {
 459                         pr_err("cxl: Error adding process element to AFU\n");
 460                         rc = -1;
 461                         goto out;
 462                 }
 463                 if ((state & (CXL_SPA_SW_CMD_MASK | CXL_SPA_SW_STATE_MASK  | CXL_SPA_SW_LINK_MASK)) ==
 464                     (cmd | (cmd >> 16) | ctx->pe))
 465                         break;
 466                 /*
 467                  * The command won't finish in the PSL if there are
 468                  * outstanding DSIs.  Hence we need to yield here in
 469                  * case there are outstanding DSIs that we need to
 470                  * service.  Tuning possiblity: we could wait for a
 471                  * while before sched
 472                  */
 473                 schedule();
 474
 475         }
 476 out:
 477         trace_cxl_llcmd_done(ctx, cmd, rc);
 478         return rc;
 479 }
 480
 481 static int add_process_element(struct cxl_context *ctx)
 482 {
 483         int rc = 0;
 484
 485         mutex_lock(&ctx->afu->native->spa_mutex);
 486         pr_devel("%s Adding pe: %i started\n", __func__, ctx->pe);
 487         if (!(rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_ADD, CXL_PE_SOFTWARE_STATE_V)))
 488                 ctx->pe_inserted = true;
 489         pr_devel("%s Adding pe: %i finished\n", __func__, ctx->pe);
 490         mutex_unlock(&ctx->afu->native->spa_mutex);
 491         return rc;
 492 }
 493
 494 static int terminate_process_element(struct cxl_context *ctx)
 495 {
 496         int rc = 0;
 497
 498         /* fast path terminate if it's already invalid */
 499         if (!(ctx->elem->software_state & cpu_to_be32(CXL_PE_SOFTWARE_STATE_V)))
 500                 return rc;
 501
 502         mutex_lock(&ctx->afu->native->spa_mutex);
 503         pr_devel("%s Terminate pe: %i started\n", __func__, ctx->pe);
 504         /* We could be asked to terminate when the hw is down. That
 505          * should always succeed: it's not running if the hw has gone
 506          * away and is being reset.
 507          */
 508         if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
 509                 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_TERMINATE,
 510                                             CXL_PE_SOFTWARE_STATE_V | CXL_PE_SOFTWARE_STATE_T);
 511         ctx->elem->software_state = 0;  /* Remove Valid bit */
 512         pr_devel("%s Terminate pe: %i finished\n", __func__, ctx->pe);
 513         mutex_unlock(&ctx->afu->native->spa_mutex);
 514         return rc;
 515 }
 516
 517 static int remove_process_element(struct cxl_context *ctx)
 518 {
 519         int rc = 0;
 520
 521         mutex_lock(&ctx->afu->native->spa_mutex);
 522         pr_devel("%s Remove pe: %i started\n", __func__, ctx->pe);
 523
 524         /* We could be asked to remove when the hw is down. Again, if
 525          * the hw is down, the PE is gone, so we succeed.
 526          */
 527         if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
 528                 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_REMOVE, 0);
 529
 530         if (!rc)
 531                 ctx->pe_inserted = false;
 532         if (cxl_is_power8())
 533                 slb_invalid(ctx);
 534         pr_devel("%s Remove pe: %i finished\n", __func__, ctx->pe);
 535         mutex_unlock(&ctx->afu->native->spa_mutex);
 536
 537         return rc;
 538 }
 539
 540 void cxl_assign_psn_space(struct cxl_context *ctx)
 541 {
 542         if (!ctx->afu->pp_size || ctx->master) {
 543                 ctx->psn_phys = ctx->afu->psn_phys;
 544                 ctx->psn_size = ctx->afu->adapter->ps_size;
 545         } else {
 546                 ctx->psn_phys = ctx->afu->psn_phys +
 547                         (ctx->afu->native->pp_offset + ctx->afu->pp_size * ctx->pe);
 548                 ctx->psn_size = ctx->afu->pp_size;
 549         }
 550 }
 551
 552 static int activate_afu_directed(struct cxl_afu *afu)
 553 {
 554         int rc;
 555
 556         dev_info(&afu->dev, "Activating AFU directed mode\n");
 557
 558         afu->num_procs = afu->max_procs_virtualised;
 559         if (afu->native->spa == NULL) {
 560                 if (cxl_alloc_spa(afu, CXL_MODE_DIRECTED))
 561                         return -ENOMEM;
 562         }
 563         attach_spa(afu);
 564
 565         cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_AFU);
 566         if (cxl_is_power8())
 567                 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL);
 568         cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L);
 569
 570         afu->current_mode = CXL_MODE_DIRECTED;
 571
 572         if ((rc = cxl_chardev_m_afu_add(afu)))
 573                 return rc;
 574
 575         if ((rc = cxl_sysfs_afu_m_add(afu)))
 576                 goto err;
 577
 578         if ((rc = cxl_chardev_s_afu_add(afu)))
 579                 goto err1;
 580
 581         return 0;
 582 err1:
 583         cxl_sysfs_afu_m_remove(afu);
 584 err:
 585         cxl_chardev_afu_remove(afu);
 586         return rc;
 587 }
 588
 589 #ifdef CONFIG_CPU_LITTLE_ENDIAN
 590 #define set_endian(sr) ((sr) |= CXL_PSL_SR_An_LE)
 591 #else
 592 #define set_endian(sr) ((sr) &= ~(CXL_PSL_SR_An_LE))
 593 #endif
 594
 595 u64 cxl_calculate_sr(bool master, bool kernel, bool real_mode, bool p9)
 596 {
 597         u64 sr = 0;
 598
 599         set_endian(sr);
 600         if (master)
 601                 sr |= CXL_PSL_SR_An_MP;
 602         if (mfspr(SPRN_LPCR) & LPCR_TC)
 603                 sr |= CXL_PSL_SR_An_TC;
 604
 605         if (kernel) {
 606                 if (!real_mode)
 607                         sr |= CXL_PSL_SR_An_R;
 608                 sr |= (mfmsr() & MSR_SF) | CXL_PSL_SR_An_HV;
 609         } else {
 610                 sr |= CXL_PSL_SR_An_PR | CXL_PSL_SR_An_R;
 611                 if (radix_enabled())
 612                         sr |= CXL_PSL_SR_An_HV;
 613                 else
 614                         sr &= ~(CXL_PSL_SR_An_HV);
 615                 if (!test_tsk_thread_flag(current, TIF_32BIT))
 616                         sr |= CXL_PSL_SR_An_SF;
 617         }
 618         if (p9) {
 619                 if (radix_enabled())
 620                         sr |= CXL_PSL_SR_An_XLAT_ror;
 621                 else
 622                         sr |= CXL_PSL_SR_An_XLAT_hpt;
 623         }
 624         return sr;
 625 }
 626
 627 static u64 calculate_sr(struct cxl_context *ctx)
 628 {
 629         return cxl_calculate_sr(ctx->master, ctx->kernel, false,
 630                                 cxl_is_power9());
 631 }
 632
 633 static void update_ivtes_directed(struct cxl_context *ctx)
 634 {
 635         bool need_update = (ctx->status == STARTED);
 636         int r;
 637
 638         if (need_update) {
 639                 WARN_ON(terminate_process_element(ctx));
 640                 WARN_ON(remove_process_element(ctx));
 641         }
 642
 643         for (r = 0; r < CXL_IRQ_RANGES; r++) {
 644                 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]);
 645                 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]);
 646         }
 647
 648         /*
 649          * Theoretically we could use the update llcmd, instead of a
 650          * terminate/remove/add (or if an atomic update was required we could
 651          * do a suspend/update/resume), however it seems there might be issues
 652          * with the update llcmd on some cards (including those using an XSL on
 653          * an ASIC) so for now it's safest to go with the commands that are
 654          * known to work. In the future if we come across a situation where the
 655          * card may be performing transactions using the same PE while we are
 656          * doing this update we might need to revisit this.
 657          */
 658         if (need_update)
 659                 WARN_ON(add_process_element(ctx));
 660 }
 661
 662 static int process_element_entry_psl9(struct cxl_context *ctx, u64 wed, u64 amr)
 663 {
 664         u32 pid;
 665         int rc;
 666
 667         cxl_assign_psn_space(ctx);
 668
 669         ctx->elem->ctxtime = 0; /* disable */
 670         ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID));
 671         ctx->elem->haurp = 0; /* disable */
 672
 673         if (ctx->kernel)
 674                 pid = 0;
 675         else {
 676                 if (ctx->mm == NULL) {
 677                         pr_devel("%s: unable to get mm for pe=%d pid=%i\n",
 678                                 __func__, ctx->pe, pid_nr(ctx->pid));
 679                         return -EINVAL;
 680                 }
 681                 pid = ctx->mm->context.id;
 682         }
 683
 684         /* Assign a unique TIDR (thread id) for the current thread */
 685         if (!(ctx->tidr) && (ctx->assign_tidr)) {
 686                 rc = set_thread_tidr(current);
 687                 if (rc)
 688                         return -ENODEV;
 689                 ctx->tidr = current->thread.tidr;
 690                 pr_devel("%s: current tidr: %d\n", __func__, ctx->tidr);
 691         }
 692
 693         ctx->elem->common.tid = cpu_to_be32(ctx->tidr);
 694         ctx->elem->common.pid = cpu_to_be32(pid);
 695
 696         ctx->elem->sr = cpu_to_be64(calculate_sr(ctx));
 697
 698         ctx->elem->common.csrp = 0; /* disable */
 699
 700         cxl_prefault(ctx, wed);
 701
 702         /*
 703          * Ensure we have the multiplexed PSL interrupt set up to take faults
 704          * for kernel contexts that may not have allocated any AFU IRQs at all:
 705          */
 706         if (ctx->irqs.range[0] == 0) {
 707                 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq;
 708                 ctx->irqs.range[0] = 1;
 709         }
 710
 711         ctx->elem->common.amr = cpu_to_be64(amr);
 712         ctx->elem->common.wed = cpu_to_be64(wed);
 713
 714         return 0;
 715 }
 716
 717 int cxl_attach_afu_directed_psl9(struct cxl_context *ctx, u64 wed, u64 amr)
 718 {
 719         int result;
 720
 721         /* fill the process element entry */
 722         result = process_element_entry_psl9(ctx, wed, amr);
 723         if (result)
 724                 return result;
 725
 726         update_ivtes_directed(ctx);
 727
 728         /* first guy needs to enable */
 729         result = cxl_ops->afu_check_and_enable(ctx->afu);
 730         if (result)
 731                 return result;
 732
 733         return add_process_element(ctx);
 734 }
 735
 736 int cxl_attach_afu_directed_psl8(struct cxl_context *ctx, u64 wed, u64 amr)
 737 {
 738         u32 pid;
 739         int result;
 740
 741         cxl_assign_psn_space(ctx);
 742
 743         ctx->elem->ctxtime = 0; /* disable */
 744         ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID));
 745         ctx->elem->haurp = 0; /* disable */
 746         ctx->elem->u.sdr = cpu_to_be64(mfspr(SPRN_SDR1));
 747
 748         pid = current->pid;
 749         if (ctx->kernel)
 750                 pid = 0;
 751         ctx->elem->common.tid = 0;
 752         ctx->elem->common.pid = cpu_to_be32(pid);
 753
 754         ctx->elem->sr = cpu_to_be64(calculate_sr(ctx));
 755
 756         ctx->elem->common.csrp = 0; /* disable */
 757         ctx->elem->common.u.psl8.aurp0 = 0; /* disable */
 758         ctx->elem->common.u.psl8.aurp1 = 0; /* disable */
 759
 760         cxl_prefault(ctx, wed);
 761
 762         ctx->elem->common.u.psl8.sstp0 = cpu_to_be64(ctx->sstp0);
 763         ctx->elem->common.u.psl8.sstp1 = cpu_to_be64(ctx->sstp1);
 764
 765         /*
 766          * Ensure we have the multiplexed PSL interrupt set up to take faults
 767          * for kernel contexts that may not have allocated any AFU IRQs at all:
 768          */
 769         if (ctx->irqs.range[0] == 0) {
 770                 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq;
 771                 ctx->irqs.range[0] = 1;
 772         }
 773
 774         update_ivtes_directed(ctx);
 775
 776         ctx->elem->common.amr = cpu_to_be64(amr);
 777         ctx->elem->common.wed = cpu_to_be64(wed);
 778
 779         /* first guy needs to enable */
 780         if ((result = cxl_ops->afu_check_and_enable(ctx->afu)))
 781                 return result;
 782
 783         return add_process_element(ctx);
 784 }
 785
 786 static int deactivate_afu_directed(struct cxl_afu *afu)
 787 {
 788         dev_info(&afu->dev, "Deactivating AFU directed mode\n");
 789
 790         afu->current_mode = 0;
 791         afu->num_procs = 0;
 792
 793         cxl_sysfs_afu_m_remove(afu);
 794         cxl_chardev_afu_remove(afu);
 795
 796         /*
 797          * The CAIA section 2.2.1 indicates that the procedure for starting and
 798          * stopping an AFU in AFU directed mode is AFU specific, which is not
 799          * ideal since this code is generic and with one exception has no
 800          * knowledge of the AFU. This is in contrast to the procedure for
 801          * disabling a dedicated process AFU, which is documented to just
 802          * require a reset. The architecture does indicate that both an AFU
 803          * reset and an AFU disable should result in the AFU being disabled and
 804          * we do both followed by a PSL purge for safety.
 805          *
 806          * Notably we used to have some issues with the disable sequence on PSL
 807          * cards, which is why we ended up using this heavy weight procedure in
 808          * the first place, however a bug was discovered that had rendered the
 809          * disable operation ineffective, so it is conceivable that was the
 810          * sole explanation for those difficulties. Careful regression testing
 811          * is recommended if anyone attempts to remove or reorder these
 812          * operations.
 813          *
 814          * The XSL on the Mellanox CX4 behaves a little differently from the
 815          * PSL based cards and will time out an AFU reset if the AFU is still
 816          * enabled. That card is special in that we do have a means to identify
 817          * it from this code, so in that case we skip the reset and just use a
 818          * disable/purge to avoid the timeout and corresponding noise in the
 819          * kernel log.
 820          */
 821         if (afu->adapter->native->sl_ops->needs_reset_before_disable)
 822                 cxl_ops->afu_reset(afu);
 823         cxl_afu_disable(afu);
 824         cxl_psl_purge(afu);
 825
 826         return 0;
 827 }
 828
 829 int cxl_activate_dedicated_process_psl9(struct cxl_afu *afu)
 830 {
 831         dev_info(&afu->dev, "Activating dedicated process mode\n");
 832
 833         /*
 834          * If XSL is set to dedicated mode (Set in PSL_SCNTL reg), the
 835          * XSL and AFU are programmed to work with a single context.
 836          * The context information should be configured in the SPA area
 837          * index 0 (so PSL_SPAP must be configured before enabling the
 838          * AFU).
 839          */
 840         afu->num_procs = 1;
 841         if (afu->native->spa == NULL) {
 842                 if (cxl_alloc_spa(afu, CXL_MODE_DEDICATED))
 843                         return -ENOMEM;
 844         }
 845         attach_spa(afu);
 846
 847         cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process);
 848         cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L);
 849
 850         afu->current_mode = CXL_MODE_DEDICATED;
 851
 852         return cxl_chardev_d_afu_add(afu);
 853 }
 854
 855 int cxl_activate_dedicated_process_psl8(struct cxl_afu *afu)
 856 {
 857         dev_info(&afu->dev, "Activating dedicated process mode\n");
 858
 859         cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process);
 860
 861         cxl_p1n_write(afu, CXL_PSL_CtxTime_An, 0); /* disable */
 862         cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0);    /* disable */
 863         cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL);
 864         cxl_p1n_write(afu, CXL_PSL_LPID_An, mfspr(SPRN_LPID));
 865         cxl_p1n_write(afu, CXL_HAURP_An, 0);       /* disable */
 866         cxl_p1n_write(afu, CXL_PSL_SDR_An, mfspr(SPRN_SDR1));
 867
 868         cxl_p2n_write(afu, CXL_CSRP_An, 0);        /* disable */
 869         cxl_p2n_write(afu, CXL_AURP0_An, 0);       /* disable */
 870         cxl_p2n_write(afu, CXL_AURP1_An, 0);       /* disable */
 871
 872         afu->current_mode = CXL_MODE_DEDICATED;
 873         afu->num_procs = 1;
 874
 875         return cxl_chardev_d_afu_add(afu);
 876 }
 877
 878 void cxl_update_dedicated_ivtes_psl9(struct cxl_context *ctx)
 879 {
 880         int r;
 881
 882         for (r = 0; r < CXL_IRQ_RANGES; r++) {
 883                 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]);
 884                 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]);
 885         }
 886 }
 887
 888 void cxl_update_dedicated_ivtes_psl8(struct cxl_context *ctx)
 889 {
 890         struct cxl_afu *afu = ctx->afu;
 891
 892         cxl_p1n_write(afu, CXL_PSL_IVTE_Offset_An,
 893                        (((u64)ctx->irqs.offset[0] & 0xffff) << 48) |
 894                        (((u64)ctx->irqs.offset[1] & 0xffff) << 32) |
 895                        (((u64)ctx->irqs.offset[2] & 0xffff) << 16) |
 896                         ((u64)ctx->irqs.offset[3] & 0xffff));
 897         cxl_p1n_write(afu, CXL_PSL_IVTE_Limit_An, (u64)
 898                        (((u64)ctx->irqs.range[0] & 0xffff) << 48) |
 899                        (((u64)ctx->irqs.range[1] & 0xffff) << 32) |
 900                        (((u64)ctx->irqs.range[2] & 0xffff) << 16) |
 901                         ((u64)ctx->irqs.range[3] & 0xffff));
 902 }
 903
 904 int cxl_attach_dedicated_process_psl9(struct cxl_context *ctx, u64 wed, u64 amr)
 905 {
 906         struct cxl_afu *afu = ctx->afu;
 907         int result;
 908
 909         /* fill the process element entry */
 910         result = process_element_entry_psl9(ctx, wed, amr);
 911         if (result)
 912                 return result;
 913
 914         if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes)
 915                 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx);
 916
 917         ctx->elem->software_state = cpu_to_be32(CXL_PE_SOFTWARE_STATE_V);
 918         /*
 919          * Ideally we should do a wmb() here to make sure the changes to the
 920          * PE are visible to the card before we call afu_enable.
 921          * On ppc64 though all mmios are preceded by a 'sync' instruction hence
 922          * we dont dont need one here.
 923          */
 924
 925         result = cxl_ops->afu_reset(afu);
 926         if (result)
 927                 return result;
 928
 929         return afu_enable(afu);
 930 }
 931
 932 int cxl_attach_dedicated_process_psl8(struct cxl_context *ctx, u64 wed, u64 amr)
 933 {
 934         struct cxl_afu *afu = ctx->afu;
 935         u64 pid;
 936         int rc;
 937
 938         pid = (u64)current->pid << 32;
 939         if (ctx->kernel)
 940                 pid = 0;
 941         cxl_p2n_write(afu, CXL_PSL_PID_TID_An, pid);
 942
 943         cxl_p1n_write(afu, CXL_PSL_SR_An, calculate_sr(ctx));
 944
 945         if ((rc = cxl_write_sstp(afu, ctx->sstp0, ctx->sstp1)))
 946                 return rc;
 947
 948         cxl_prefault(ctx, wed);
 949
 950         if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes)
 951                 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx);
 952
 953         cxl_p2n_write(afu, CXL_PSL_AMR_An, amr);
 954
 955         /* master only context for dedicated */
 956         cxl_assign_psn_space(ctx);
 957
 958         if ((rc = cxl_ops->afu_reset(afu)))
 959                 return rc;
 960
 961         cxl_p2n_write(afu, CXL_PSL_WED_An, wed);
 962
 963         return afu_enable(afu);
 964 }
 965
 966 static int deactivate_dedicated_process(struct cxl_afu *afu)
 967 {
 968         dev_info(&afu->dev, "Deactivating dedicated process mode\n");
 969
 970         afu->current_mode = 0;
 971         afu->num_procs = 0;
 972
 973         cxl_chardev_afu_remove(afu);
 974
 975         return 0;
 976 }
 977
 978 static int native_afu_deactivate_mode(struct cxl_afu *afu, int mode)
 979 {
 980         if (mode == CXL_MODE_DIRECTED)
 981                 return deactivate_afu_directed(afu);
 982         if (mode == CXL_MODE_DEDICATED)
 983                 return deactivate_dedicated_process(afu);
 984         return 0;
 985 }
 986
 987 static int native_afu_activate_mode(struct cxl_afu *afu, int mode)
 988 {
 989         if (!mode)
 990                 return 0;
 991         if (!(mode & afu->modes_supported))
 992                 return -EINVAL;
 993
 994         if (!cxl_ops->link_ok(afu->adapter, afu)) {
 995                 WARN(1, "Device link is down, refusing to activate!\n");
 996                 return -EIO;
 997         }
 998
 999         if (mode == CXL_MODE_DIRECTED)
1000                 return activate_afu_directed(afu);
1001         if ((mode == CXL_MODE_DEDICATED) &&
1002             (afu->adapter->native->sl_ops->activate_dedicated_process))
1003                 return afu->adapter->native->sl_ops->activate_dedicated_process(afu);
1004
1005         return -EINVAL;
1006 }
1007
1008 static int native_attach_process(struct cxl_context *ctx, bool kernel,
1009                                 u64 wed, u64 amr)
1010 {
1011         if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
1012                 WARN(1, "Device link is down, refusing to attach process!\n");
1013                 return -EIO;
1014         }
1015
1016         ctx->kernel = kernel;
1017         if ((ctx->afu->current_mode == CXL_MODE_DIRECTED) &&
1018             (ctx->afu->adapter->native->sl_ops->attach_afu_directed))
1019                 return ctx->afu->adapter->native->sl_ops->attach_afu_directed(ctx, wed, amr);
1020
1021         if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) &&
1022             (ctx->afu->adapter->native->sl_ops->attach_dedicated_process))
1023                 return ctx->afu->adapter->native->sl_ops->attach_dedicated_process(ctx, wed, amr);
1024
1025         return -EINVAL;
1026 }
1027
1028 static inline int detach_process_native_dedicated(struct cxl_context *ctx)
1029 {
1030         /*
1031          * The CAIA section 2.1.1 indicates that we need to do an AFU reset to
1032          * stop the AFU in dedicated mode (we therefore do not make that
1033          * optional like we do in the afu directed path). It does not indicate
1034          * that we need to do an explicit disable (which should occur
1035          * implicitly as part of the reset) or purge, but we do these as well
1036          * to be on the safe side.
1037          *
1038          * Notably we used to have some issues with the disable sequence
1039          * (before the sequence was spelled out in the architecture) which is
1040          * why we were so heavy weight in the first place, however a bug was
1041          * discovered that had rendered the disable operation ineffective, so
1042          * it is conceivable that was the sole explanation for those
1043          * difficulties. Point is, we should be careful and do some regression
1044          * testing if we ever attempt to remove any part of this procedure.
1045          */
1046         cxl_ops->afu_reset(ctx->afu);
1047         cxl_afu_disable(ctx->afu);
1048         cxl_psl_purge(ctx->afu);
1049         return 0;
1050 }
1051
1052 static void native_update_ivtes(struct cxl_context *ctx)
1053 {
1054         if (ctx->afu->current_mode == CXL_MODE_DIRECTED)
1055                 return update_ivtes_directed(ctx);
1056         if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) &&
1057             (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes))
1058                 return ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx);
1059         WARN(1, "native_update_ivtes: Bad mode\n");
1060 }
1061
1062 static inline int detach_process_native_afu_directed(struct cxl_context *ctx)
1063 {
1064         if (!ctx->pe_inserted)
1065                 return 0;
1066         if (terminate_process_element(ctx))
1067                 return -1;
1068         if (remove_process_element(ctx))
1069                 return -1;
1070
1071         return 0;
1072 }
1073
1074 static int native_detach_process(struct cxl_context *ctx)
1075 {
1076         trace_cxl_detach(ctx);
1077
1078         if (ctx->afu->current_mode == CXL_MODE_DEDICATED)
1079                 return detach_process_native_dedicated(ctx);
1080
1081         return detach_process_native_afu_directed(ctx);
1082 }
1083
1084 static int native_get_irq_info(struct cxl_afu *afu, struct cxl_irq_info *info)
1085 {
1086         /* If the adapter has gone away, we can't get any meaningful
1087          * information.
1088          */
1089         if (!cxl_ops->link_ok(afu->adapter, afu))
1090                 return -EIO;
1091
1092         info->dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
1093         info->dar = cxl_p2n_read(afu, CXL_PSL_DAR_An);
1094         if (cxl_is_power8())
1095                 info->dsr = cxl_p2n_read(afu, CXL_PSL_DSR_An);
1096         info->afu_err = cxl_p2n_read(afu, CXL_AFU_ERR_An);
1097         info->errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
1098         info->proc_handle = 0;
1099
1100         return 0;
1101 }
1102
1103 void cxl_native_irq_dump_regs_psl9(struct cxl_context *ctx)
1104 {
1105         u64 fir1, serr;
1106
1107         fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL9_FIR1);
1108
1109         dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1);
1110         if (ctx->afu->adapter->native->sl_ops->register_serr_irq) {
1111                 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An);
1112                 cxl_afu_decode_psl_serr(ctx->afu, serr);
1113         }
1114 }
1115
1116 void cxl_native_irq_dump_regs_psl8(struct cxl_context *ctx)
1117 {
1118         u64 fir1, fir2, fir_slice, serr, afu_debug;
1119
1120         fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR1);
1121         fir2 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR2);
1122         fir_slice = cxl_p1n_read(ctx->afu, CXL_PSL_FIR_SLICE_An);
1123         afu_debug = cxl_p1n_read(ctx->afu, CXL_AFU_DEBUG_An);
1124
1125         dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1);
1126         dev_crit(&ctx->afu->dev, "PSL_FIR2: 0x%016llx\n", fir2);
1127         if (ctx->afu->adapter->native->sl_ops->register_serr_irq) {
1128                 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An);
1129                 cxl_afu_decode_psl_serr(ctx->afu, serr);
1130         }
1131         dev_crit(&ctx->afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice);
1132         dev_crit(&ctx->afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug);
1133 }
1134
1135 static irqreturn_t native_handle_psl_slice_error(struct cxl_context *ctx,
1136                                                 u64 dsisr, u64 errstat)
1137 {
1138
1139         dev_crit(&ctx->afu->dev, "PSL ERROR STATUS: 0x%016llx\n", errstat);
1140
1141         if (ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers)
1142                 ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers(ctx);
1143
1144         if (ctx->afu->adapter->native->sl_ops->debugfs_stop_trace) {
1145                 dev_crit(&ctx->afu->dev, "STOPPING CXL TRACE\n");
1146                 ctx->afu->adapter->native->sl_ops->debugfs_stop_trace(ctx->afu->adapter);
1147         }
1148
1149         return cxl_ops->ack_irq(ctx, 0, errstat);
1150 }
1151
1152 static bool cxl_is_translation_fault(struct cxl_afu *afu, u64 dsisr)
1153 {
1154         if ((cxl_is_power8()) && (dsisr & CXL_PSL_DSISR_TRANS))
1155                 return true;
1156
1157         if ((cxl_is_power9()) && (dsisr & CXL_PSL9_DSISR_An_TF))
1158                 return true;
1159
1160         return false;
1161 }
1162
1163 irqreturn_t cxl_fail_irq_psl(struct cxl_afu *afu, struct cxl_irq_info *irq_info)
1164 {
1165         if (cxl_is_translation_fault(afu, irq_info->dsisr))
1166                 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
1167         else
1168                 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
1169
1170         return IRQ_HANDLED;
1171 }
1172
1173 static irqreturn_t native_irq_multiplexed(int irq, void *data)
1174 {
1175         struct cxl_afu *afu = data;
1176         struct cxl_context *ctx;
1177         struct cxl_irq_info irq_info;
1178         u64 phreg = cxl_p2n_read(afu, CXL_PSL_PEHandle_An);
1179         int ph, ret = IRQ_HANDLED, res;
1180
1181         /* check if eeh kicked in while the interrupt was in flight */
1182         if (unlikely(phreg == ~0ULL)) {
1183                 dev_warn(&afu->dev,
1184                          "Ignoring slice interrupt(%d) due to fenced card",
1185                          irq);
1186                 return IRQ_HANDLED;
1187         }
1188         /* Mask the pe-handle from register value */
1189         ph = phreg & 0xffff;
1190         if ((res = native_get_irq_info(afu, &irq_info))) {
1191                 WARN(1, "Unable to get CXL IRQ Info: %i\n", res);
1192                 if (afu->adapter->native->sl_ops->fail_irq)
1193                         return afu->adapter->native->sl_ops->fail_irq(afu, &irq_info);
1194                 return ret;
1195         }
1196
1197         rcu_read_lock();
1198         ctx = idr_find(&afu->contexts_idr, ph);
1199         if (ctx) {
1200                 if (afu->adapter->native->sl_ops->handle_interrupt)
1201                         ret = afu->adapter->native->sl_ops->handle_interrupt(irq, ctx, &irq_info);
1202                 rcu_read_unlock();
1203                 return ret;
1204         }
1205         rcu_read_unlock();
1206
1207         WARN(1, "Unable to demultiplex CXL PSL IRQ for PE %i DSISR %016llx DAR"
1208                 " %016llx\n(Possible AFU HW issue - was a term/remove acked"
1209                 " with outstanding transactions?)\n", ph, irq_info.dsisr,
1210                 irq_info.dar);
1211         if (afu->adapter->native->sl_ops->fail_irq)
1212                 ret = afu->adapter->native->sl_ops->fail_irq(afu, &irq_info);
1213         return ret;
1214 }
1215
1216 static void native_irq_wait(struct cxl_context *ctx)
1217 {
1218         u64 dsisr;
1219         int timeout = 1000;
1220         int ph;
1221
1222         /*
1223          * Wait until no further interrupts are presented by the PSL
1224          * for this context.
1225          */
1226         while (timeout--) {
1227                 ph = cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) & 0xffff;
1228                 if (ph != ctx->pe)
1229                         return;
1230                 dsisr = cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An);
1231                 if (cxl_is_power8() &&
1232                    ((dsisr & CXL_PSL_DSISR_PENDING) == 0))
1233                         return;
1234                 if (cxl_is_power9() &&
1235                    ((dsisr & CXL_PSL9_DSISR_PENDING) == 0))
1236                         return;
1237                 /*
1238                  * We are waiting for the workqueue to process our
1239                  * irq, so need to let that run here.
1240                  */
1241                 msleep(1);
1242         }
1243
1244         dev_warn(&ctx->afu->dev, "WARNING: waiting on DSI for PE %i"
1245                  " DSISR %016llx!\n", ph, dsisr);
1246         return;
1247 }
1248
1249 static irqreturn_t native_slice_irq_err(int irq, void *data)
1250 {
1251         struct cxl_afu *afu = data;
1252         u64 errstat, serr, afu_error, dsisr;
1253         u64 fir_slice, afu_debug, irq_mask;
1254
1255         /*
1256          * slice err interrupt is only used with full PSL (no XSL)
1257          */
1258         serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
1259         errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
1260         afu_error = cxl_p2n_read(afu, CXL_AFU_ERR_An);
1261         dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
1262         cxl_afu_decode_psl_serr(afu, serr);
1263
1264         if (cxl_is_power8()) {
1265                 fir_slice = cxl_p1n_read(afu, CXL_PSL_FIR_SLICE_An);
1266                 afu_debug = cxl_p1n_read(afu, CXL_AFU_DEBUG_An);
1267                 dev_crit(&afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice);
1268                 dev_crit(&afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug);
1269         }
1270         dev_crit(&afu->dev, "CXL_PSL_ErrStat_An: 0x%016llx\n", errstat);
1271         dev_crit(&afu->dev, "AFU_ERR_An: 0x%.16llx\n", afu_error);
1272         dev_crit(&afu->dev, "PSL_DSISR_An: 0x%.16llx\n", dsisr);
1273
1274         /* mask off the IRQ so it won't retrigger until the AFU is reset */
1275         irq_mask = (serr & CXL_PSL_SERR_An_IRQS) >> 32;
1276         serr |= irq_mask;
1277         cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
1278         dev_info(&afu->dev, "Further such interrupts will be masked until the AFU is reset\n");
1279
1280         return IRQ_HANDLED;
1281 }
1282
1283 void cxl_native_err_irq_dump_regs_psl9(struct cxl *adapter)
1284 {
1285         u64 fir1;
1286
1287         fir1 = cxl_p1_read(adapter, CXL_PSL9_FIR1);
1288         dev_crit(&adapter->dev, "PSL_FIR: 0x%016llx\n", fir1);
1289 }
1290
1291 void cxl_native_err_irq_dump_regs_psl8(struct cxl *adapter)
1292 {
1293         u64 fir1, fir2;
1294
1295         fir1 = cxl_p1_read(adapter, CXL_PSL_FIR1);
1296         fir2 = cxl_p1_read(adapter, CXL_PSL_FIR2);
1297         dev_crit(&adapter->dev,
1298                  "PSL_FIR1: 0x%016llx\nPSL_FIR2: 0x%016llx\n",
1299                  fir1, fir2);
1300 }
1301
1302 static irqreturn_t native_irq_err(int irq, void *data)
1303 {
1304         struct cxl *adapter = data;
1305         u64 err_ivte;
1306
1307         WARN(1, "CXL ERROR interrupt %i\n", irq);
1308
1309         err_ivte = cxl_p1_read(adapter, CXL_PSL_ErrIVTE);
1310         dev_crit(&adapter->dev, "PSL_ErrIVTE: 0x%016llx\n", err_ivte);
1311
1312         if (adapter->native->sl_ops->debugfs_stop_trace) {
1313                 dev_crit(&adapter->dev, "STOPPING CXL TRACE\n");
1314                 adapter->native->sl_ops->debugfs_stop_trace(adapter);
1315         }
1316
1317         if (adapter->native->sl_ops->err_irq_dump_registers)
1318                 adapter->native->sl_ops->err_irq_dump_registers(adapter);
1319
1320         return IRQ_HANDLED;
1321 }
1322
1323 int cxl_native_register_psl_err_irq(struct cxl *adapter)
1324 {
1325         int rc;
1326
1327         adapter->irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err",
1328                                       dev_name(&adapter->dev));
1329         if (!adapter->irq_name)
1330                 return -ENOMEM;
1331
1332         if ((rc = cxl_register_one_irq(adapter, native_irq_err, adapter,
1333                                        &adapter->native->err_hwirq,
1334                                        &adapter->native->err_virq,
1335                                        adapter->irq_name))) {
1336                 kfree(adapter->irq_name);
1337                 adapter->irq_name = NULL;
1338                 return rc;
1339         }
1340
1341         cxl_p1_write(adapter, CXL_PSL_ErrIVTE, adapter->native->err_hwirq & 0xffff);
1342
1343         return 0;
1344 }
1345
1346 void cxl_native_release_psl_err_irq(struct cxl *adapter)
1347 {
1348         if (adapter->native->err_virq == 0 ||
1349             adapter->native->err_virq !=
1350             irq_find_mapping(NULL, adapter->native->err_hwirq))
1351                 return;
1352
1353         cxl_p1_write(adapter, CXL_PSL_ErrIVTE, 0x0000000000000000);
1354         cxl_unmap_irq(adapter->native->err_virq, adapter);
1355         cxl_ops->release_one_irq(adapter, adapter->native->err_hwirq);
1356         kfree(adapter->irq_name);
1357         adapter->native->err_virq = 0;
1358 }
1359
1360 int cxl_native_register_serr_irq(struct cxl_afu *afu)
1361 {
1362         u64 serr;
1363         int rc;
1364
1365         afu->err_irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err",
1366                                       dev_name(&afu->dev));
1367         if (!afu->err_irq_name)
1368                 return -ENOMEM;
1369
1370         if ((rc = cxl_register_one_irq(afu->adapter, native_slice_irq_err, afu,
1371                                        &afu->serr_hwirq,
1372                                        &afu->serr_virq, afu->err_irq_name))) {
1373                 kfree(afu->err_irq_name);
1374                 afu->err_irq_name = NULL;
1375                 return rc;
1376         }
1377
1378         serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
1379         if (cxl_is_power8())
1380                 serr = (serr & 0x00ffffffffff0000ULL) | (afu->serr_hwirq & 0xffff);
1381         if (cxl_is_power9()) {
1382                 /*
1383                  * By default, all errors are masked. So don't set all masks.
1384                  * Slice errors will be transfered.
1385                  */
1386                 serr = (serr & ~0xff0000007fffffffULL) | (afu->serr_hwirq & 0xffff);
1387         }
1388         cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
1389
1390         return 0;
1391 }
1392
1393 void cxl_native_release_serr_irq(struct cxl_afu *afu)
1394 {
1395         if (afu->serr_virq == 0 ||
1396             afu->serr_virq != irq_find_mapping(NULL, afu->serr_hwirq))
1397                 return;
1398
1399         cxl_p1n_write(afu, CXL_PSL_SERR_An, 0x0000000000000000);
1400         cxl_unmap_irq(afu->serr_virq, afu);
1401         cxl_ops->release_one_irq(afu->adapter, afu->serr_hwirq);
1402         kfree(afu->err_irq_name);
1403         afu->serr_virq = 0;
1404 }
1405
1406 int cxl_native_register_psl_irq(struct cxl_afu *afu)
1407 {
1408         int rc;
1409
1410         afu->psl_irq_name = kasprintf(GFP_KERNEL, "cxl-%s",
1411                                       dev_name(&afu->dev));
1412         if (!afu->psl_irq_name)
1413                 return -ENOMEM;
1414
1415         if ((rc = cxl_register_one_irq(afu->adapter, native_irq_multiplexed,
1416                                     afu, &afu->native->psl_hwirq, &afu->native->psl_virq,
1417                                     afu->psl_irq_name))) {
1418                 kfree(afu->psl_irq_name);
1419                 afu->psl_irq_name = NULL;
1420         }
1421         return rc;
1422 }
1423
1424 void cxl_native_release_psl_irq(struct cxl_afu *afu)
1425 {
1426         if (afu->native->psl_virq == 0 ||
1427             afu->native->psl_virq !=
1428             irq_find_mapping(NULL, afu->native->psl_hwirq))
1429                 return;
1430
1431         cxl_unmap_irq(afu->native->psl_virq, afu);
1432         cxl_ops->release_one_irq(afu->adapter, afu->native->psl_hwirq);
1433         kfree(afu->psl_irq_name);
1434         afu->native->psl_virq = 0;
1435 }
1436
1437 static void recover_psl_err(struct cxl_afu *afu, u64 errstat)
1438 {
1439         u64 dsisr;
1440
1441         pr_devel("RECOVERING FROM PSL ERROR... (0x%016llx)\n", errstat);
1442
1443         /* Clear PSL_DSISR[PE] */
1444         dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
1445         cxl_p2n_write(afu, CXL_PSL_DSISR_An, dsisr & ~CXL_PSL_DSISR_An_PE);
1446
1447         /* Write 1s to clear error status bits */
1448         cxl_p2n_write(afu, CXL_PSL_ErrStat_An, errstat);
1449 }
1450
1451 static int native_ack_irq(struct cxl_context *ctx, u64 tfc, u64 psl_reset_mask)
1452 {
1453         trace_cxl_psl_irq_ack(ctx, tfc);
1454         if (tfc)
1455                 cxl_p2n_write(ctx->afu, CXL_PSL_TFC_An, tfc);
1456         if (psl_reset_mask)
1457                 recover_psl_err(ctx->afu, psl_reset_mask);
1458
1459         return 0;
1460 }
1461
1462 int cxl_check_error(struct cxl_afu *afu)
1463 {
1464         return (cxl_p1n_read(afu, CXL_PSL_SCNTL_An) == ~0ULL);
1465 }
1466
1467 static bool native_support_attributes(const char *attr_name,
1468                                       enum cxl_attrs type)
1469 {
1470         return true;
1471 }
1472
1473 static int native_afu_cr_read64(struct cxl_afu *afu, int cr, u64 off, u64 *out)
1474 {
1475         if (unlikely(!cxl_ops->link_ok(afu->adapter, afu)))
1476                 return -EIO;
1477         if (unlikely(off >= afu->crs_len))
1478                 return -ERANGE;
1479         *out = in_le64(afu->native->afu_desc_mmio + afu->crs_offset +
1480                 (cr * afu->crs_len) + off);
1481         return 0;
1482 }
1483
1484 static int native_afu_cr_read32(struct cxl_afu *afu, int cr, u64 off, u32 *out)
1485 {
1486         if (unlikely(!cxl_ops->link_ok(afu->adapter, afu)))
1487                 return -EIO;
1488         if (unlikely(off >= afu->crs_len))
1489                 return -ERANGE;
1490         *out = in_le32(afu->native->afu_desc_mmio + afu->crs_offset +
1491                 (cr * afu->crs_len) + off);
1492         return 0;
1493 }
1494
1495 static int native_afu_cr_read16(struct cxl_afu *afu, int cr, u64 off, u16 *out)
1496 {
1497         u64 aligned_off = off & ~0x3L;
1498         u32 val;
1499         int rc;
1500
1501         rc = native_afu_cr_read32(afu, cr, aligned_off, &val);
1502         if (!rc)
1503                 *out = (val >> ((off & 0x3) * 8)) & 0xffff;
1504         return rc;
1505 }
1506
1507 static int native_afu_cr_read8(struct cxl_afu *afu, int cr, u64 off, u8 *out)
1508 {
1509         u64 aligned_off = off & ~0x3L;
1510         u32 val;
1511         int rc;
1512
1513         rc = native_afu_cr_read32(afu, cr, aligned_off, &val);
1514         if (!rc)
1515                 *out = (val >> ((off & 0x3) * 8)) & 0xff;
1516         return rc;
1517 }
1518
1519 static int native_afu_cr_write32(struct cxl_afu *afu, int cr, u64 off, u32 in)
1520 {
1521         if (unlikely(!cxl_ops->link_ok(afu->adapter, afu)))
1522                 return -EIO;
1523         if (unlikely(off >= afu->crs_len))
1524                 return -ERANGE;
1525         out_le32(afu->native->afu_desc_mmio + afu->crs_offset +
1526                 (cr * afu->crs_len) + off, in);
1527         return 0;
1528 }
1529
1530 static int native_afu_cr_write16(struct cxl_afu *afu, int cr, u64 off, u16 in)
1531 {
1532         u64 aligned_off = off & ~0x3L;
1533         u32 val32, mask, shift;
1534         int rc;
1535
1536         rc = native_afu_cr_read32(afu, cr, aligned_off, &val32);
1537         if (rc)
1538                 return rc;
1539         shift = (off & 0x3) * 8;
1540         WARN_ON(shift == 24);
1541         mask = 0xffff << shift;
1542         val32 = (val32 & ~mask) | (in << shift);
1543
1544         rc = native_afu_cr_write32(afu, cr, aligned_off, val32);
1545         return rc;
1546 }
1547
1548 static int native_afu_cr_write8(struct cxl_afu *afu, int cr, u64 off, u8 in)
1549 {
1550         u64 aligned_off = off & ~0x3L;
1551         u32 val32, mask, shift;
1552         int rc;
1553
1554         rc = native_afu_cr_read32(afu, cr, aligned_off, &val32);
1555         if (rc)
1556                 return rc;
1557         shift = (off & 0x3) * 8;
1558         mask = 0xff << shift;
1559         val32 = (val32 & ~mask) | (in << shift);
1560
1561         rc = native_afu_cr_write32(afu, cr, aligned_off, val32);
1562         return rc;
1563 }
1564
1565 const struct cxl_backend_ops cxl_native_ops = {
1566         .module = THIS_MODULE,
1567         .adapter_reset = cxl_pci_reset,
1568         .alloc_one_irq = cxl_pci_alloc_one_irq,
1569         .release_one_irq = cxl_pci_release_one_irq,
1570         .alloc_irq_ranges = cxl_pci_alloc_irq_ranges,
1571         .release_irq_ranges = cxl_pci_release_irq_ranges,
1572         .setup_irq = cxl_pci_setup_irq,
1573         .handle_psl_slice_error = native_handle_psl_slice_error,
1574         .psl_interrupt = NULL,
1575         .ack_irq = native_ack_irq,
1576         .irq_wait = native_irq_wait,
1577         .attach_process = native_attach_process,
1578         .detach_process = native_detach_process,
1579         .update_ivtes = native_update_ivtes,
1580         .support_attributes = native_support_attributes,
1581         .link_ok = cxl_adapter_link_ok,
1582         .release_afu = cxl_pci_release_afu,
1583         .afu_read_err_buffer = cxl_pci_afu_read_err_buffer,
1584         .afu_check_and_enable = native_afu_check_and_enable,
1585         .afu_activate_mode = native_afu_activate_mode,
1586         .afu_deactivate_mode = native_afu_deactivate_mode,
1587         .afu_reset = native_afu_reset,
1588         .afu_cr_read8 = native_afu_cr_read8,
1589         .afu_cr_read16 = native_afu_cr_read16,
1590         .afu_cr_read32 = native_afu_cr_read32,
1591         .afu_cr_read64 = native_afu_cr_read64,
1592         .afu_cr_write8 = native_afu_cr_write8,
1593         .afu_cr_write16 = native_afu_cr_write16,
1594         .afu_cr_write32 = native_afu_cr_write32,
1595         .read_adapter_vpd = cxl_pci_read_adapter_vpd,
1596 };