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 | _PAGE_READ;
 153         if (dsisr & CXL_PSL_DSISR_An_S)
 154                 access |= _PAGE_WRITE;
 155
 156         access |= _PAGE_PRIVILEGED;
 157         if ((!ctx->kernel) || (REGION_ID(dar) == USER_REGION_ID))
 158                 access &= ~_PAGE_PRIVILEGED;
 159
 160         if (dsisr & DSISR_NOHPTE)
 161                 inv_flags |= HPTE_NOHPTE_UPDATE;
 162
 163         local_irq_save(flags);
 164         hash_page_mm(mm, dar, access, 0x300, inv_flags);
 165         local_irq_restore(flags);
 166
 167         pr_devel("Page fault successfully handled for pe: %i!\n", ctx->pe);
 168         cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_R, 0);
 169 }
 170
 171 /*
 172  * Returns the mm_struct corresponding to the context ctx via ctx->pid
 173  * In case the task has exited we use the task group leader accessible
 174  * via ctx->glpid to find the next task in the thread group that has a
 175  * valid  mm_struct associated with it. If a task with valid mm_struct
 176  * is found the ctx->pid is updated to use the task struct for subsequent
 177  * translations. In case no valid mm_struct is found in the task group to
 178  * service the fault a NULL is returned.
 179  */
 180 static struct mm_struct *get_mem_context(struct cxl_context *ctx)
 181 {
 182         struct task_struct *task = NULL;
 183         struct mm_struct *mm = NULL;
 184         struct pid *old_pid = ctx->pid;
 185
 186         if (old_pid == NULL) {
 187                 pr_warn("%s: Invalid context for pe=%d\n",
 188                          __func__, ctx->pe);
 189                 return NULL;
 190         }
 191
 192         task = get_pid_task(old_pid, PIDTYPE_PID);
 193
 194         /*
 195          * pid_alive may look racy but this saves us from costly
 196          * get_task_mm when the task is a zombie. In worst case
 197          * we may think a task is alive, which is about to die
 198          * but get_task_mm will return NULL.
 199          */
 200         if (task != NULL && pid_alive(task))
 201                 mm = get_task_mm(task);
 202
 203         /* release the task struct that was taken earlier */
 204         if (task)
 205                 put_task_struct(task);
 206         else
 207                 pr_devel("%s: Context owning pid=%i for pe=%i dead\n",
 208                         __func__, pid_nr(old_pid), ctx->pe);
 209
 210         /*
 211          * If we couldn't find the mm context then use the group
 212          * leader to iterate over the task group and find a task
 213          * that gives us mm_struct.
 214          */
 215         if (unlikely(mm == NULL && ctx->glpid != NULL)) {
 216
 217                 rcu_read_lock();
 218                 task = pid_task(ctx->glpid, PIDTYPE_PID);
 219                 if (task)
 220                         do {
 221                                 mm = get_task_mm(task);
 222                                 if (mm) {
 223                                         ctx->pid = get_task_pid(task,
 224                                                                 PIDTYPE_PID);
 225                                         break;
 226                                 }
 227                                 task = next_thread(task);
 228                         } while (task && !thread_group_leader(task));
 229                 rcu_read_unlock();
 230
 231                 /* check if we switched pid */
 232                 if (ctx->pid != old_pid) {
 233                         if (mm)
 234                                 pr_devel("%s:pe=%i switch pid %i->%i\n",
 235                                          __func__, ctx->pe, pid_nr(old_pid),
 236                                          pid_nr(ctx->pid));
 237                         else
 238                                 pr_devel("%s:Cannot find mm for pid=%i\n",
 239                                          __func__, pid_nr(old_pid));
 240
 241                         /* drop the reference to older pid */
 242                         put_pid(old_pid);
 243                 }
 244         }
 245
 246         return mm;
 247 }
 248
 249
 250
 251 void cxl_handle_fault(struct work_struct *fault_work)
 252 {
 253         struct cxl_context *ctx =
 254                 container_of(fault_work, struct cxl_context, fault_work);
 255         u64 dsisr = ctx->dsisr;
 256         u64 dar = ctx->dar;
 257         struct mm_struct *mm = NULL;
 258
 259         if (cpu_has_feature(CPU_FTR_HVMODE)) {
 260                 if (cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An) != dsisr ||
 261                     cxl_p2n_read(ctx->afu, CXL_PSL_DAR_An) != dar ||
 262                     cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) != ctx->pe) {
 263                         /* Most likely explanation is harmless - a dedicated
 264                          * process has detached and these were cleared by the
 265                          * PSL purge, but warn about it just in case
 266                          */
 267                         dev_notice(&ctx->afu->dev, "cxl_handle_fault: Translation fault regs changed\n");
 268                         return;
 269                 }
 270         }
 271
 272         /* Early return if the context is being / has been detached */
 273         if (ctx->status == CLOSED) {
 274                 cxl_ack_ae(ctx);
 275                 return;
 276         }
 277
 278         pr_devel("CXL BOTTOM HALF handling fault for afu pe: %i. "
 279                 "DSISR: %#llx DAR: %#llx\n", ctx->pe, dsisr, dar);
 280
 281         if (!ctx->kernel) {
 282
 283                 mm = get_mem_context(ctx);
 284                 /* indicates all the thread in task group have exited */
 285                 if (mm == NULL) {
 286                         pr_devel("%s: unable to get mm for pe=%d pid=%i\n",
 287                                  __func__, ctx->pe, pid_nr(ctx->pid));
 288                         cxl_ack_ae(ctx);
 289                         return;
 290                 } else {
 291                         pr_devel("Handling page fault for pe=%d pid=%i\n",
 292                                  ctx->pe, pid_nr(ctx->pid));
 293                 }
 294         }
 295
 296         if (dsisr & CXL_PSL_DSISR_An_DS)
 297                 cxl_handle_segment_miss(ctx, mm, dar);
 298         else if (dsisr & CXL_PSL_DSISR_An_DM)
 299                 cxl_handle_page_fault(ctx, mm, dsisr, dar);
 300         else
 301                 WARN(1, "cxl_handle_fault has nothing to handle\n");
 302
 303         if (mm)
 304                 mmput(mm);
 305 }
 306
 307 static void cxl_prefault_one(struct cxl_context *ctx, u64 ea)
 308 {
 309         struct mm_struct *mm;
 310
 311         mm = get_mem_context(ctx);
 312         if (mm == NULL) {
 313                 pr_devel("cxl_prefault_one unable to get mm %i\n",
 314                          pid_nr(ctx->pid));
 315                 return;
 316         }
 317
 318         cxl_fault_segment(ctx, mm, ea);
 319
 320         mmput(mm);
 321 }
 322
 323 static u64 next_segment(u64 ea, u64 vsid)
 324 {
 325         if (vsid & SLB_VSID_B_1T)
 326                 ea |= (1ULL << 40) - 1;
 327         else
 328                 ea |= (1ULL << 28) - 1;
 329
 330         return ea + 1;
 331 }
 332
 333 static void cxl_prefault_vma(struct cxl_context *ctx)
 334 {
 335         u64 ea, last_esid = 0;
 336         struct copro_slb slb;
 337         struct vm_area_struct *vma;
 338         int rc;
 339         struct mm_struct *mm;
 340
 341         mm = get_mem_context(ctx);
 342         if (mm == NULL) {
 343                 pr_devel("cxl_prefault_vm unable to get mm %i\n",
 344                          pid_nr(ctx->pid));
 345                 return;
 346         }
 347
 348         down_read(&mm->mmap_sem);
 349         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 350                 for (ea = vma->vm_start; ea < vma->vm_end;
 351                                 ea = next_segment(ea, slb.vsid)) {
 352                         rc = copro_calculate_slb(mm, ea, &slb);
 353                         if (rc)
 354                                 continue;
 355
 356                         if (last_esid == slb.esid)
 357                                 continue;
 358
 359                         cxl_load_segment(ctx, &slb);
 360                         last_esid = slb.esid;
 361                 }
 362         }
 363         up_read(&mm->mmap_sem);
 364
 365         mmput(mm);
 366 }
 367
 368 void cxl_prefault(struct cxl_context *ctx, u64 wed)
 369 {
 370         switch (ctx->afu->prefault_mode) {
 371         case CXL_PREFAULT_WED:
 372                 cxl_prefault_one(ctx, wed);
 373                 break;
 374         case CXL_PREFAULT_ALL:
 375                 cxl_prefault_vma(ctx);
 376                 break;
 377         default:
 378                 break;
 379         }
 380 }