drivers/misc/cxl/fault.c

   1 /*
   2  * Copyright 2014 IBM Corp.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public License
   6  * as published by the Free Software Foundation; either version
   7  * 2 of the License, or (at your option) any later version.
   8  */
   9
  10 #include <linux/workqueue.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/sched/mm.h>
  13 #include <linux/pid.h>
  14 #include <linux/mm.h>
  15 #include <linux/moduleparam.h>
  16
  17 #undef MODULE_PARAM_PREFIX
  18 #define MODULE_PARAM_PREFIX "cxl" "."
  19 #include <asm/current.h>
  20 #include <asm/copro.h>
  21 #include <asm/mmu.h>
  22
  23 #include "cxl.h"
  24 #include "trace.h"
  25
  26 static bool sste_matches(struct cxl_sste *sste, struct copro_slb *slb)
  27 {
  28         return ((sste->vsid_data == cpu_to_be64(slb->vsid)) &&
  29                 (sste->esid_data == cpu_to_be64(slb->esid)));
  30 }
  31
  32 /*
  33  * This finds a free SSTE for the given SLB, or returns NULL if it's already in
  34  * the segment table.
  35  */
  36 static struct cxl_sste* find_free_sste(struct cxl_context *ctx,
  37                                        struct copro_slb *slb)
  38 {
  39         struct cxl_sste *primary, *sste, *ret = NULL;
  40         unsigned int mask = (ctx->sst_size >> 7) - 1; /* SSTP0[SegTableSize] */
  41         unsigned int entry;
  42         unsigned int hash;
  43
  44         if (slb->vsid & SLB_VSID_B_1T)
  45                 hash = (slb->esid >> SID_SHIFT_1T) & mask;
  46         else /* 256M */
  47                 hash = (slb->esid >> SID_SHIFT) & mask;
  48
  49         primary = ctx->sstp + (hash << 3);
  50
  51         for (entry = 0, sste = primary; entry < 8; entry++, sste++) {
  52                 if (!ret && !(be64_to_cpu(sste->esid_data) & SLB_ESID_V))
  53                         ret = sste;
  54                 if (sste_matches(sste, slb))
  55                         return NULL;
  56         }
  57         if (ret)
  58                 return ret;
  59
  60         /* Nothing free, select an entry to cast out */
  61         ret = primary + ctx->sst_lru;
  62         ctx->sst_lru = (ctx->sst_lru + 1) & 0x7;
  63
  64         return ret;
  65 }
  66
  67 static void cxl_load_segment(struct cxl_context *ctx, struct copro_slb *slb)
  68 {
  69         /* mask is the group index, we search primary and secondary here. */
  70         struct cxl_sste *sste;
  71         unsigned long flags;
  72
  73         spin_lock_irqsave(&ctx->sste_lock, flags);
  74         sste = find_free_sste(ctx, slb);
  75         if (!sste)
  76                 goto out_unlock;
  77
  78         pr_devel("CXL Populating SST[%li]: %#llx %#llx\n",
  79                         sste - ctx->sstp, slb->vsid, slb->esid);
  80         trace_cxl_ste_write(ctx, sste - ctx->sstp, slb->esid, slb->vsid);
  81
  82         sste->vsid_data = cpu_to_be64(slb->vsid);
  83         sste->esid_data = cpu_to_be64(slb->esid);
  84 out_unlock:
  85         spin_unlock_irqrestore(&ctx->sste_lock, flags);
  86 }
  87
  88 static int cxl_fault_segment(struct cxl_context *ctx, struct mm_struct *mm,
  89                              u64 ea)
  90 {
  91         struct copro_slb slb = {0,0};
  92         int rc;
  93
  94         if (!(rc = copro_calculate_slb(mm, ea, &slb))) {
  95                 cxl_load_segment(ctx, &slb);
  96         }
  97
  98         return rc;
  99 }
 100
 101 static void cxl_ack_ae(struct cxl_context *ctx)
 102 {
 103         unsigned long flags;
 104
 105         cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_AE, 0);
 106
 107         spin_lock_irqsave(&ctx->lock, flags);
 108         ctx->pending_fault = true;
 109         ctx->fault_addr = ctx->dar;
 110         ctx->fault_dsisr = ctx->dsisr;
 111         spin_unlock_irqrestore(&ctx->lock, flags);
 112
 113         wake_up_all(&ctx->wq);
 114 }
 115
 116 static int cxl_handle_segment_miss(struct cxl_context *ctx,
 117                                    struct mm_struct *mm, u64 ea)
 118 {
 119         int rc;
 120
 121         pr_devel("CXL interrupt: Segment fault pe: %i ea: %#llx\n", ctx->pe, ea);
 122         trace_cxl_ste_miss(ctx, ea);
 123
 124         if ((rc = cxl_fault_segment(ctx, mm, ea)))
 125                 cxl_ack_ae(ctx);
 126         else {
 127
 128                 mb(); /* Order seg table write to TFC MMIO write */
 129                 cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_R, 0);
 130         }
 131
 132         return IRQ_HANDLED;
 133 }
 134
 135 static void cxl_handle_page_fault(struct cxl_context *ctx,
 136                                   struct mm_struct *mm, u64 dsisr, u64 dar)
 137 {
 138         unsigned flt = 0;
 139         int result;
 140         unsigned long access, flags, inv_flags = 0;
 141
 142         trace_cxl_pte_miss(ctx, dsisr, dar);
 143
 144         if ((result = copro_handle_mm_fault(mm, dar, dsisr, &flt))) {
 145                 pr_devel("copro_handle_mm_fault failed: %#x\n", result);
 146                 return cxl_ack_ae(ctx);
 147         }
 148
 149         /*
 150          * update_mmu_cache() will not have loaded the hash since current->trap
 151          * is not a 0x400 or 0x300, so just call hash_page_mm() here.
 152          */
 153         access = _PAGE_PRESENT | _PAGE_READ;
 154         if (dsisr & CXL_PSL_DSISR_An_S)
 155                 access |= _PAGE_WRITE;
 156
 157         access |= _PAGE_PRIVILEGED;
 158         if ((!ctx->kernel) || (REGION_ID(dar) == USER_REGION_ID))
 159                 access &= ~_PAGE_PRIVILEGED;
 160
 161         if (dsisr & DSISR_NOHPTE)
 162                 inv_flags |= HPTE_NOHPTE_UPDATE;
 163
 164         local_irq_save(flags);
 165         hash_page_mm(mm, dar, access, 0x300, inv_flags);
 166         local_irq_restore(flags);
 167
 168         pr_devel("Page fault successfully handled for pe: %i!\n", ctx->pe);
 169         cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_R, 0);
 170 }
 171
 172 /*
 173  * Returns the mm_struct corresponding to the context ctx via ctx->pid
 174  * In case the task has exited we use the task group leader accessible
 175  * via ctx->glpid to find the next task in the thread group that has a
 176  * valid  mm_struct associated with it. If a task with valid mm_struct
 177  * is found the ctx->pid is updated to use the task struct for subsequent
 178  * translations. In case no valid mm_struct is found in the task group to
 179  * service the fault a NULL is returned.
 180  */
 181 static struct mm_struct *get_mem_context(struct cxl_context *ctx)
 182 {
 183         struct task_struct *task = NULL;
 184         struct mm_struct *mm = NULL;
 185         struct pid *old_pid = ctx->pid;
 186
 187         if (old_pid == NULL) {
 188                 pr_warn("%s: Invalid context for pe=%d\n",
 189                          __func__, ctx->pe);
 190                 return NULL;
 191         }
 192
 193         task = get_pid_task(old_pid, PIDTYPE_PID);
 194
 195         /*
 196          * pid_alive may look racy but this saves us from costly
 197          * get_task_mm when the task is a zombie. In worst case
 198          * we may think a task is alive, which is about to die
 199          * but get_task_mm will return NULL.
 200          */
 201         if (task != NULL && pid_alive(task))
 202                 mm = get_task_mm(task);
 203
 204         /* release the task struct that was taken earlier */
 205         if (task)
 206                 put_task_struct(task);
 207         else
 208                 pr_devel("%s: Context owning pid=%i for pe=%i dead\n",
 209                         __func__, pid_nr(old_pid), ctx->pe);
 210
 211         /*
 212          * If we couldn't find the mm context then use the group
 213          * leader to iterate over the task group and find a task
 214          * that gives us mm_struct.
 215          */
 216         if (unlikely(mm == NULL && ctx->glpid != NULL)) {
 217
 218                 rcu_read_lock();
 219                 task = pid_task(ctx->glpid, PIDTYPE_PID);
 220                 if (task)
 221                         do {
 222                                 mm = get_task_mm(task);
 223                                 if (mm) {
 224                                         ctx->pid = get_task_pid(task,
 225                                                                 PIDTYPE_PID);
 226                                         break;
 227                                 }
 228                                 task = next_thread(task);
 229                         } while (task && !thread_group_leader(task));
 230                 rcu_read_unlock();
 231
 232                 /* check if we switched pid */
 233                 if (ctx->pid != old_pid) {
 234                         if (mm)
 235                                 pr_devel("%s:pe=%i switch pid %i->%i\n",
 236                                          __func__, ctx->pe, pid_nr(old_pid),
 237                                          pid_nr(ctx->pid));
 238                         else
 239                                 pr_devel("%s:Cannot find mm for pid=%i\n",
 240                                          __func__, pid_nr(old_pid));
 241
 242                         /* drop the reference to older pid */
 243                         put_pid(old_pid);
 244                 }
 245         }
 246
 247         return mm;
 248 }
 249
 250
 251
 252 void cxl_handle_fault(struct work_struct *fault_work)
 253 {
 254         struct cxl_context *ctx =
 255                 container_of(fault_work, struct cxl_context, fault_work);
 256         u64 dsisr = ctx->dsisr;
 257         u64 dar = ctx->dar;
 258         struct mm_struct *mm = NULL;
 259
 260         if (cpu_has_feature(CPU_FTR_HVMODE)) {
 261                 if (cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An) != dsisr ||
 262                     cxl_p2n_read(ctx->afu, CXL_PSL_DAR_An) != dar ||
 263                     cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) != ctx->pe) {
 264                         /* Most likely explanation is harmless - a dedicated
 265                          * process has detached and these were cleared by the
 266                          * PSL purge, but warn about it just in case
 267                          */
 268                         dev_notice(&ctx->afu->dev, "cxl_handle_fault: Translation fault regs changed\n");
 269                         return;
 270                 }
 271         }
 272
 273         /* Early return if the context is being / has been detached */
 274         if (ctx->status == CLOSED) {
 275                 cxl_ack_ae(ctx);
 276                 return;
 277         }
 278
 279         pr_devel("CXL BOTTOM HALF handling fault for afu pe: %i. "
 280                 "DSISR: %#llx DAR: %#llx\n", ctx->pe, dsisr, dar);
 281
 282         if (!ctx->kernel) {
 283
 284                 mm = get_mem_context(ctx);
 285                 /* indicates all the thread in task group have exited */
 286                 if (mm == NULL) {
 287                         pr_devel("%s: unable to get mm for pe=%d pid=%i\n",
 288                                  __func__, ctx->pe, pid_nr(ctx->pid));
 289                         cxl_ack_ae(ctx);
 290                         return;
 291                 } else {
 292                         pr_devel("Handling page fault for pe=%d pid=%i\n",
 293                                  ctx->pe, pid_nr(ctx->pid));
 294                 }
 295         }
 296
 297         if (dsisr & CXL_PSL_DSISR_An_DS)
 298                 cxl_handle_segment_miss(ctx, mm, dar);
 299         else if (dsisr & CXL_PSL_DSISR_An_DM)
 300                 cxl_handle_page_fault(ctx, mm, dsisr, dar);
 301         else
 302                 WARN(1, "cxl_handle_fault has nothing to handle\n");
 303
 304         if (mm)
 305                 mmput(mm);
 306 }
 307
 308 static void cxl_prefault_one(struct cxl_context *ctx, u64 ea)
 309 {
 310         struct mm_struct *mm;
 311
 312         mm = get_mem_context(ctx);
 313         if (mm == NULL) {
 314                 pr_devel("cxl_prefault_one unable to get mm %i\n",
 315                          pid_nr(ctx->pid));
 316                 return;
 317         }
 318
 319         cxl_fault_segment(ctx, mm, ea);
 320
 321         mmput(mm);
 322 }
 323
 324 static u64 next_segment(u64 ea, u64 vsid)
 325 {
 326         if (vsid & SLB_VSID_B_1T)
 327                 ea |= (1ULL << 40) - 1;
 328         else
 329                 ea |= (1ULL << 28) - 1;
 330
 331         return ea + 1;
 332 }
 333
 334 static void cxl_prefault_vma(struct cxl_context *ctx)
 335 {
 336         u64 ea, last_esid = 0;
 337         struct copro_slb slb;
 338         struct vm_area_struct *vma;
 339         int rc;
 340         struct mm_struct *mm;
 341
 342         mm = get_mem_context(ctx);
 343         if (mm == NULL) {
 344                 pr_devel("cxl_prefault_vm unable to get mm %i\n",
 345                          pid_nr(ctx->pid));
 346                 return;
 347         }
 348
 349         down_read(&mm->mmap_sem);
 350         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 351                 for (ea = vma->vm_start; ea < vma->vm_end;
 352                                 ea = next_segment(ea, slb.vsid)) {
 353                         rc = copro_calculate_slb(mm, ea, &slb);
 354                         if (rc)
 355                                 continue;
 356
 357                         if (last_esid == slb.esid)
 358                                 continue;
 359
 360                         cxl_load_segment(ctx, &slb);
 361                         last_esid = slb.esid;
 362                 }
 363         }
 364         up_read(&mm->mmap_sem);
 365
 366         mmput(mm);
 367 }
 368
 369 void cxl_prefault(struct cxl_context *ctx, u64 wed)
 370 {
 371         switch (ctx->afu->prefault_mode) {
 372         case CXL_PREFAULT_WED:
 373                 cxl_prefault_one(ctx, wed);
 374                 break;
 375         case CXL_PREFAULT_ALL:
 376                 cxl_prefault_vma(ctx);
 377                 break;
 378         default:
 379                 break;
 380         }
 381 }