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