1c942fddfSThomas Gleixner // SPDX-License-Identifier: GPL-2.0-or-later 2133ff0eaSJérôme Glisse /* 3133ff0eaSJérôme Glisse * Copyright 2013 Red Hat Inc. 4133ff0eaSJérôme Glisse * 5f813f219SJérôme Glisse * Authors: Jérôme Glisse <jglisse@redhat.com> 6133ff0eaSJérôme Glisse */ 7133ff0eaSJérôme Glisse /* 8133ff0eaSJérôme Glisse * Refer to include/linux/hmm.h for information about heterogeneous memory 9133ff0eaSJérôme Glisse * management or HMM for short. 10133ff0eaSJérôme Glisse */ 11133ff0eaSJérôme Glisse #include <linux/mm.h> 12133ff0eaSJérôme Glisse #include <linux/hmm.h> 13858b54daSJérôme Glisse #include <linux/init.h> 14da4c3c73SJérôme Glisse #include <linux/rmap.h> 15da4c3c73SJérôme Glisse #include <linux/swap.h> 16133ff0eaSJérôme Glisse #include <linux/slab.h> 17133ff0eaSJérôme Glisse #include <linux/sched.h> 184ef589dcSJérôme Glisse #include <linux/mmzone.h> 194ef589dcSJérôme Glisse #include <linux/pagemap.h> 20da4c3c73SJérôme Glisse #include <linux/swapops.h> 21da4c3c73SJérôme Glisse #include <linux/hugetlb.h> 224ef589dcSJérôme Glisse #include <linux/memremap.h> 237b2d55d2SJérôme Glisse #include <linux/jump_label.h> 2455c0ece8SJérôme Glisse #include <linux/dma-mapping.h> 25c0b12405SJérôme Glisse #include <linux/mmu_notifier.h> 264ef589dcSJérôme Glisse #include <linux/memory_hotplug.h> 274ef589dcSJérôme Glisse 284ef589dcSJérôme Glisse #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT) 29133ff0eaSJérôme Glisse 306b368cd4SJérôme Glisse #if IS_ENABLED(CONFIG_HMM_MIRROR) 31c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops; 32c0b12405SJérôme Glisse 33704f3f2cSJérôme Glisse static inline struct hmm *mm_get_hmm(struct mm_struct *mm) 34133ff0eaSJérôme Glisse { 35c0b12405SJérôme Glisse struct hmm *hmm = READ_ONCE(mm->hmm); 36704f3f2cSJérôme Glisse 37704f3f2cSJérôme Glisse if (hmm && kref_get_unless_zero(&hmm->kref)) 38704f3f2cSJérôme Glisse return hmm; 39704f3f2cSJérôme Glisse 40704f3f2cSJérôme Glisse return NULL; 41704f3f2cSJérôme Glisse } 42704f3f2cSJérôme Glisse 43704f3f2cSJérôme Glisse /** 44704f3f2cSJérôme Glisse * hmm_get_or_create - register HMM against an mm (HMM internal) 45704f3f2cSJérôme Glisse * 46704f3f2cSJérôme Glisse * @mm: mm struct to attach to 47704f3f2cSJérôme Glisse * Returns: returns an HMM object, either by referencing the existing 48704f3f2cSJérôme Glisse * (per-process) object, or by creating a new one. 49704f3f2cSJérôme Glisse * 50704f3f2cSJérôme Glisse * This is not intended to be used directly by device drivers. If mm already 51704f3f2cSJérôme Glisse * has an HMM struct then it get a reference on it and returns it. Otherwise 52704f3f2cSJérôme Glisse * it allocates an HMM struct, initializes it, associate it with the mm and 53704f3f2cSJérôme Glisse * returns it. 54704f3f2cSJérôme Glisse */ 55704f3f2cSJérôme Glisse static struct hmm *hmm_get_or_create(struct mm_struct *mm) 56704f3f2cSJérôme Glisse { 57704f3f2cSJérôme Glisse struct hmm *hmm = mm_get_hmm(mm); 58c0b12405SJérôme Glisse bool cleanup = false; 59133ff0eaSJérôme Glisse 60c0b12405SJérôme Glisse if (hmm) 61c0b12405SJérôme Glisse return hmm; 62c0b12405SJérôme Glisse 63c0b12405SJérôme Glisse hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); 64c0b12405SJérôme Glisse if (!hmm) 65c0b12405SJérôme Glisse return NULL; 66a3e0d41cSJérôme Glisse init_waitqueue_head(&hmm->wq); 67c0b12405SJérôme Glisse INIT_LIST_HEAD(&hmm->mirrors); 68c0b12405SJérôme Glisse init_rwsem(&hmm->mirrors_sem); 69c0b12405SJérôme Glisse hmm->mmu_notifier.ops = NULL; 70da4c3c73SJérôme Glisse INIT_LIST_HEAD(&hmm->ranges); 71a3e0d41cSJérôme Glisse mutex_init(&hmm->lock); 72704f3f2cSJérôme Glisse kref_init(&hmm->kref); 73a3e0d41cSJérôme Glisse hmm->notifiers = 0; 74a3e0d41cSJérôme Glisse hmm->dead = false; 75c0b12405SJérôme Glisse hmm->mm = mm; 76c0b12405SJérôme Glisse 77c0b12405SJérôme Glisse spin_lock(&mm->page_table_lock); 78c0b12405SJérôme Glisse if (!mm->hmm) 79c0b12405SJérôme Glisse mm->hmm = hmm; 80c0b12405SJérôme Glisse else 81c0b12405SJérôme Glisse cleanup = true; 82c0b12405SJérôme Glisse spin_unlock(&mm->page_table_lock); 83c0b12405SJérôme Glisse 8486a2d598SRalph Campbell if (cleanup) 8586a2d598SRalph Campbell goto error; 8686a2d598SRalph Campbell 8786a2d598SRalph Campbell /* 8886a2d598SRalph Campbell * We should only get here if hold the mmap_sem in write mode ie on 8986a2d598SRalph Campbell * registration of first mirror through hmm_mirror_register() 9086a2d598SRalph Campbell */ 9186a2d598SRalph Campbell hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; 9286a2d598SRalph Campbell if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) 9386a2d598SRalph Campbell goto error_mm; 94c0b12405SJérôme Glisse 95704f3f2cSJérôme Glisse return hmm; 9686a2d598SRalph Campbell 9786a2d598SRalph Campbell error_mm: 9886a2d598SRalph Campbell spin_lock(&mm->page_table_lock); 9986a2d598SRalph Campbell if (mm->hmm == hmm) 10086a2d598SRalph Campbell mm->hmm = NULL; 10186a2d598SRalph Campbell spin_unlock(&mm->page_table_lock); 10286a2d598SRalph Campbell error: 10386a2d598SRalph Campbell kfree(hmm); 10486a2d598SRalph Campbell return NULL; 105133ff0eaSJérôme Glisse } 106133ff0eaSJérôme Glisse 107704f3f2cSJérôme Glisse static void hmm_free(struct kref *kref) 108704f3f2cSJérôme Glisse { 109704f3f2cSJérôme Glisse struct hmm *hmm = container_of(kref, struct hmm, kref); 110704f3f2cSJérôme Glisse struct mm_struct *mm = hmm->mm; 111704f3f2cSJérôme Glisse 112704f3f2cSJérôme Glisse mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm); 113704f3f2cSJérôme Glisse 114704f3f2cSJérôme Glisse spin_lock(&mm->page_table_lock); 115704f3f2cSJérôme Glisse if (mm->hmm == hmm) 116704f3f2cSJérôme Glisse mm->hmm = NULL; 117704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 118704f3f2cSJérôme Glisse 119704f3f2cSJérôme Glisse kfree(hmm); 120704f3f2cSJérôme Glisse } 121704f3f2cSJérôme Glisse 122704f3f2cSJérôme Glisse static inline void hmm_put(struct hmm *hmm) 123704f3f2cSJérôme Glisse { 124704f3f2cSJérôme Glisse kref_put(&hmm->kref, hmm_free); 125704f3f2cSJérôme Glisse } 126704f3f2cSJérôme Glisse 127133ff0eaSJérôme Glisse void hmm_mm_destroy(struct mm_struct *mm) 128133ff0eaSJérôme Glisse { 129704f3f2cSJérôme Glisse struct hmm *hmm; 130704f3f2cSJérôme Glisse 131704f3f2cSJérôme Glisse spin_lock(&mm->page_table_lock); 132704f3f2cSJérôme Glisse hmm = mm_get_hmm(mm); 133704f3f2cSJérôme Glisse mm->hmm = NULL; 134704f3f2cSJérôme Glisse if (hmm) { 135704f3f2cSJérôme Glisse hmm->mm = NULL; 136a3e0d41cSJérôme Glisse hmm->dead = true; 137704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 138704f3f2cSJérôme Glisse hmm_put(hmm); 139704f3f2cSJérôme Glisse return; 140704f3f2cSJérôme Glisse } 141704f3f2cSJérôme Glisse 142704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 143133ff0eaSJérôme Glisse } 144c0b12405SJérôme Glisse 145a3e0d41cSJérôme Glisse static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm) 146c0b12405SJérôme Glisse { 147a3e0d41cSJérôme Glisse struct hmm *hmm = mm_get_hmm(mm); 148c0b12405SJérôme Glisse struct hmm_mirror *mirror; 149da4c3c73SJérôme Glisse struct hmm_range *range; 150da4c3c73SJérôme Glisse 151a3e0d41cSJérôme Glisse /* Report this HMM as dying. */ 152a3e0d41cSJérôme Glisse hmm->dead = true; 153da4c3c73SJérôme Glisse 154a3e0d41cSJérôme Glisse /* Wake-up everyone waiting on any range. */ 155a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 156*085ea250SRalph Campbell list_for_each_entry(range, &hmm->ranges, list) 157da4c3c73SJérôme Glisse range->valid = false; 158a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 159a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 160e1401513SRalph Campbell 161e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 162e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror, 163e1401513SRalph Campbell list); 164e1401513SRalph Campbell while (mirror) { 165e1401513SRalph Campbell list_del_init(&mirror->list); 166e1401513SRalph Campbell if (mirror->ops->release) { 167e1401513SRalph Campbell /* 168*085ea250SRalph Campbell * Drop mirrors_sem so the release callback can wait 169*085ea250SRalph Campbell * on any pending work that might itself trigger a 170*085ea250SRalph Campbell * mmu_notifier callback and thus would deadlock with 171*085ea250SRalph Campbell * us. 172e1401513SRalph Campbell */ 173e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 174e1401513SRalph Campbell mirror->ops->release(mirror); 175e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 176e1401513SRalph Campbell } 177e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, 178e1401513SRalph Campbell struct hmm_mirror, list); 179e1401513SRalph Campbell } 180e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 181704f3f2cSJérôme Glisse 182704f3f2cSJérôme Glisse hmm_put(hmm); 183e1401513SRalph Campbell } 184e1401513SRalph Campbell 18593065ac7SMichal Hocko static int hmm_invalidate_range_start(struct mmu_notifier *mn, 186a3e0d41cSJérôme Glisse const struct mmu_notifier_range *nrange) 187c0b12405SJérôme Glisse { 188a3e0d41cSJérôme Glisse struct hmm *hmm = mm_get_hmm(nrange->mm); 189a3e0d41cSJérôme Glisse struct hmm_mirror *mirror; 190ec131b2dSJérôme Glisse struct hmm_update update; 191a3e0d41cSJérôme Glisse struct hmm_range *range; 192a3e0d41cSJérôme Glisse int ret = 0; 193c0b12405SJérôme Glisse 194c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 195c0b12405SJérôme Glisse 196a3e0d41cSJérôme Glisse update.start = nrange->start; 197a3e0d41cSJérôme Glisse update.end = nrange->end; 198ec131b2dSJérôme Glisse update.event = HMM_UPDATE_INVALIDATE; 199dfcd6660SJérôme Glisse update.blockable = mmu_notifier_range_blockable(nrange); 200a3e0d41cSJérôme Glisse 201dfcd6660SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 202a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 203a3e0d41cSJérôme Glisse else if (!mutex_trylock(&hmm->lock)) { 204a3e0d41cSJérôme Glisse ret = -EAGAIN; 205a3e0d41cSJérôme Glisse goto out; 206a3e0d41cSJérôme Glisse } 207a3e0d41cSJérôme Glisse hmm->notifiers++; 208a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 209a3e0d41cSJérôme Glisse if (update.end < range->start || update.start >= range->end) 210a3e0d41cSJérôme Glisse continue; 211a3e0d41cSJérôme Glisse 212a3e0d41cSJérôme Glisse range->valid = false; 213a3e0d41cSJérôme Glisse } 214a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 215a3e0d41cSJérôme Glisse 216dfcd6660SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 217a3e0d41cSJérôme Glisse down_read(&hmm->mirrors_sem); 218a3e0d41cSJérôme Glisse else if (!down_read_trylock(&hmm->mirrors_sem)) { 219a3e0d41cSJérôme Glisse ret = -EAGAIN; 220a3e0d41cSJérôme Glisse goto out; 221a3e0d41cSJérôme Glisse } 222a3e0d41cSJérôme Glisse list_for_each_entry(mirror, &hmm->mirrors, list) { 223a3e0d41cSJérôme Glisse int ret; 224a3e0d41cSJérôme Glisse 225a3e0d41cSJérôme Glisse ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update); 226*085ea250SRalph Campbell if (!update.blockable && ret == -EAGAIN) 227*085ea250SRalph Campbell break; 228a3e0d41cSJérôme Glisse } 229a3e0d41cSJérôme Glisse up_read(&hmm->mirrors_sem); 230a3e0d41cSJérôme Glisse 231a3e0d41cSJérôme Glisse out: 232704f3f2cSJérôme Glisse hmm_put(hmm); 233704f3f2cSJérôme Glisse return ret; 234c0b12405SJérôme Glisse } 235c0b12405SJérôme Glisse 236c0b12405SJérôme Glisse static void hmm_invalidate_range_end(struct mmu_notifier *mn, 237a3e0d41cSJérôme Glisse const struct mmu_notifier_range *nrange) 238c0b12405SJérôme Glisse { 239a3e0d41cSJérôme Glisse struct hmm *hmm = mm_get_hmm(nrange->mm); 240c0b12405SJérôme Glisse 241c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 242c0b12405SJérôme Glisse 243a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 244a3e0d41cSJérôme Glisse hmm->notifiers--; 245a3e0d41cSJérôme Glisse if (!hmm->notifiers) { 246a3e0d41cSJérôme Glisse struct hmm_range *range; 247a3e0d41cSJérôme Glisse 248a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 249a3e0d41cSJérôme Glisse if (range->valid) 250a3e0d41cSJérôme Glisse continue; 251a3e0d41cSJérôme Glisse range->valid = true; 252a3e0d41cSJérôme Glisse } 253a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 254a3e0d41cSJérôme Glisse } 255a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 256a3e0d41cSJérôme Glisse 257704f3f2cSJérôme Glisse hmm_put(hmm); 258c0b12405SJérôme Glisse } 259c0b12405SJérôme Glisse 260c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { 261e1401513SRalph Campbell .release = hmm_release, 262c0b12405SJérôme Glisse .invalidate_range_start = hmm_invalidate_range_start, 263c0b12405SJérôme Glisse .invalidate_range_end = hmm_invalidate_range_end, 264c0b12405SJérôme Glisse }; 265c0b12405SJérôme Glisse 266c0b12405SJérôme Glisse /* 267c0b12405SJérôme Glisse * hmm_mirror_register() - register a mirror against an mm 268c0b12405SJérôme Glisse * 269c0b12405SJérôme Glisse * @mirror: new mirror struct to register 270c0b12405SJérôme Glisse * @mm: mm to register against 271*085ea250SRalph Campbell * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments 272c0b12405SJérôme Glisse * 273c0b12405SJérôme Glisse * To start mirroring a process address space, the device driver must register 274c0b12405SJérôme Glisse * an HMM mirror struct. 275c0b12405SJérôme Glisse * 276c0b12405SJérôme Glisse * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! 277c0b12405SJérôme Glisse */ 278c0b12405SJérôme Glisse int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) 279c0b12405SJérôme Glisse { 280c0b12405SJérôme Glisse /* Sanity check */ 281c0b12405SJérôme Glisse if (!mm || !mirror || !mirror->ops) 282c0b12405SJérôme Glisse return -EINVAL; 283c0b12405SJérôme Glisse 284704f3f2cSJérôme Glisse mirror->hmm = hmm_get_or_create(mm); 285c0b12405SJérôme Glisse if (!mirror->hmm) 286c0b12405SJérôme Glisse return -ENOMEM; 287c0b12405SJérôme Glisse 288c0b12405SJérôme Glisse down_write(&mirror->hmm->mirrors_sem); 289c0b12405SJérôme Glisse list_add(&mirror->list, &mirror->hmm->mirrors); 290c0b12405SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 291c0b12405SJérôme Glisse 292c0b12405SJérôme Glisse return 0; 293c0b12405SJérôme Glisse } 294c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_register); 295c0b12405SJérôme Glisse 296c0b12405SJérôme Glisse /* 297c0b12405SJérôme Glisse * hmm_mirror_unregister() - unregister a mirror 298c0b12405SJérôme Glisse * 299*085ea250SRalph Campbell * @mirror: mirror struct to unregister 300c0b12405SJérôme Glisse * 301c0b12405SJérôme Glisse * Stop mirroring a process address space, and cleanup. 302c0b12405SJérôme Glisse */ 303c0b12405SJérôme Glisse void hmm_mirror_unregister(struct hmm_mirror *mirror) 304c0b12405SJérôme Glisse { 305704f3f2cSJérôme Glisse struct hmm *hmm = READ_ONCE(mirror->hmm); 306c0b12405SJérôme Glisse 307704f3f2cSJérôme Glisse if (hmm == NULL) 308c01cbba2SJérôme Glisse return; 309c01cbba2SJérôme Glisse 310c0b12405SJérôme Glisse down_write(&hmm->mirrors_sem); 311e1401513SRalph Campbell list_del_init(&mirror->list); 312704f3f2cSJérôme Glisse /* To protect us against double unregister ... */ 313c01cbba2SJérôme Glisse mirror->hmm = NULL; 314c0b12405SJérôme Glisse up_write(&hmm->mirrors_sem); 315c01cbba2SJérôme Glisse 316704f3f2cSJérôme Glisse hmm_put(hmm); 317c0b12405SJérôme Glisse } 318c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_unregister); 319da4c3c73SJérôme Glisse 32074eee180SJérôme Glisse struct hmm_vma_walk { 32174eee180SJérôme Glisse struct hmm_range *range; 322992de9a8SJérôme Glisse struct dev_pagemap *pgmap; 32374eee180SJérôme Glisse unsigned long last; 32474eee180SJérôme Glisse bool fault; 32574eee180SJérôme Glisse bool block; 32674eee180SJérôme Glisse }; 32774eee180SJérôme Glisse 3282aee09d8SJérôme Glisse static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr, 3292aee09d8SJérôme Glisse bool write_fault, uint64_t *pfn) 33074eee180SJérôme Glisse { 33174eee180SJérôme Glisse unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE; 33274eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 333f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 33474eee180SJérôme Glisse struct vm_area_struct *vma = walk->vma; 33550a7ca3cSSouptick Joarder vm_fault_t ret; 33674eee180SJérôme Glisse 33774eee180SJérôme Glisse flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; 3382aee09d8SJérôme Glisse flags |= write_fault ? FAULT_FLAG_WRITE : 0; 33950a7ca3cSSouptick Joarder ret = handle_mm_fault(vma, addr, flags); 34050a7ca3cSSouptick Joarder if (ret & VM_FAULT_RETRY) 34173231612SJérôme Glisse return -EAGAIN; 34250a7ca3cSSouptick Joarder if (ret & VM_FAULT_ERROR) { 343f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 34474eee180SJérôme Glisse return -EFAULT; 34574eee180SJérôme Glisse } 34674eee180SJérôme Glisse 34773231612SJérôme Glisse return -EBUSY; 34874eee180SJérôme Glisse } 34974eee180SJérôme Glisse 350da4c3c73SJérôme Glisse static int hmm_pfns_bad(unsigned long addr, 351da4c3c73SJérôme Glisse unsigned long end, 352da4c3c73SJérôme Glisse struct mm_walk *walk) 353da4c3c73SJérôme Glisse { 354c719547fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 355c719547fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 356ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 357da4c3c73SJérôme Glisse unsigned long i; 358da4c3c73SJérôme Glisse 359da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 360da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) 361f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_ERROR]; 362da4c3c73SJérôme Glisse 363da4c3c73SJérôme Glisse return 0; 364da4c3c73SJérôme Glisse } 365da4c3c73SJérôme Glisse 3665504ed29SJérôme Glisse /* 3675504ed29SJérôme Glisse * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s) 3685504ed29SJérôme Glisse * @start: range virtual start address (inclusive) 3695504ed29SJérôme Glisse * @end: range virtual end address (exclusive) 3702aee09d8SJérôme Glisse * @fault: should we fault or not ? 3712aee09d8SJérôme Glisse * @write_fault: write fault ? 3725504ed29SJérôme Glisse * @walk: mm_walk structure 373*085ea250SRalph Campbell * Return: 0 on success, -EBUSY after page fault, or page fault error 3745504ed29SJérôme Glisse * 3755504ed29SJérôme Glisse * This function will be called whenever pmd_none() or pte_none() returns true, 3765504ed29SJérôme Glisse * or whenever there is no page directory covering the virtual address range. 3775504ed29SJérôme Glisse */ 3782aee09d8SJérôme Glisse static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end, 3792aee09d8SJérôme Glisse bool fault, bool write_fault, 380da4c3c73SJérôme Glisse struct mm_walk *walk) 381da4c3c73SJérôme Glisse { 38274eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 38374eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 384ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 38563d5066fSJérôme Glisse unsigned long i, page_size; 386da4c3c73SJérôme Glisse 38774eee180SJérôme Glisse hmm_vma_walk->last = addr; 38863d5066fSJérôme Glisse page_size = hmm_range_page_size(range); 38963d5066fSJérôme Glisse i = (addr - range->start) >> range->page_shift; 39063d5066fSJérôme Glisse 39163d5066fSJérôme Glisse for (; addr < end; addr += page_size, i++) { 392f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_NONE]; 3932aee09d8SJérôme Glisse if (fault || write_fault) { 39474eee180SJérôme Glisse int ret; 395da4c3c73SJérôme Glisse 3962aee09d8SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, write_fault, 3972aee09d8SJérôme Glisse &pfns[i]); 39873231612SJérôme Glisse if (ret != -EBUSY) 39974eee180SJérôme Glisse return ret; 40074eee180SJérôme Glisse } 40174eee180SJérôme Glisse } 40274eee180SJérôme Glisse 40373231612SJérôme Glisse return (fault || write_fault) ? -EBUSY : 0; 4042aee09d8SJérôme Glisse } 4052aee09d8SJérôme Glisse 4062aee09d8SJérôme Glisse static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4072aee09d8SJérôme Glisse uint64_t pfns, uint64_t cpu_flags, 4082aee09d8SJérôme Glisse bool *fault, bool *write_fault) 4092aee09d8SJérôme Glisse { 410f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 411f88a1e90SJérôme Glisse 4122aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) 4132aee09d8SJérôme Glisse return; 4142aee09d8SJérôme Glisse 415023a019aSJérôme Glisse /* 416023a019aSJérôme Glisse * So we not only consider the individual per page request we also 417023a019aSJérôme Glisse * consider the default flags requested for the range. The API can 418023a019aSJérôme Glisse * be use in 2 fashions. The first one where the HMM user coalesce 419023a019aSJérôme Glisse * multiple page fault into one request and set flags per pfns for 420023a019aSJérôme Glisse * of those faults. The second one where the HMM user want to pre- 421023a019aSJérôme Glisse * fault a range with specific flags. For the latter one it is a 422023a019aSJérôme Glisse * waste to have the user pre-fill the pfn arrays with a default 423023a019aSJérôme Glisse * flags value. 424023a019aSJérôme Glisse */ 425023a019aSJérôme Glisse pfns = (pfns & range->pfn_flags_mask) | range->default_flags; 426023a019aSJérôme Glisse 4272aee09d8SJérôme Glisse /* We aren't ask to do anything ... */ 428f88a1e90SJérôme Glisse if (!(pfns & range->flags[HMM_PFN_VALID])) 4292aee09d8SJérôme Glisse return; 430f88a1e90SJérôme Glisse /* If this is device memory than only fault if explicitly requested */ 431f88a1e90SJérôme Glisse if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) { 432f88a1e90SJérôme Glisse /* Do we fault on device memory ? */ 433f88a1e90SJérôme Glisse if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) { 434f88a1e90SJérôme Glisse *write_fault = pfns & range->flags[HMM_PFN_WRITE]; 435f88a1e90SJérôme Glisse *fault = true; 436f88a1e90SJérôme Glisse } 4372aee09d8SJérôme Glisse return; 4382aee09d8SJérôme Glisse } 439f88a1e90SJérôme Glisse 440f88a1e90SJérôme Glisse /* If CPU page table is not valid then we need to fault */ 441f88a1e90SJérôme Glisse *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]); 442f88a1e90SJérôme Glisse /* Need to write fault ? */ 443f88a1e90SJérôme Glisse if ((pfns & range->flags[HMM_PFN_WRITE]) && 444f88a1e90SJérôme Glisse !(cpu_flags & range->flags[HMM_PFN_WRITE])) { 445f88a1e90SJérôme Glisse *write_fault = true; 4462aee09d8SJérôme Glisse *fault = true; 4472aee09d8SJérôme Glisse } 4482aee09d8SJérôme Glisse } 4492aee09d8SJérôme Glisse 4502aee09d8SJérôme Glisse static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4512aee09d8SJérôme Glisse const uint64_t *pfns, unsigned long npages, 4522aee09d8SJérôme Glisse uint64_t cpu_flags, bool *fault, 4532aee09d8SJérôme Glisse bool *write_fault) 4542aee09d8SJérôme Glisse { 4552aee09d8SJérôme Glisse unsigned long i; 4562aee09d8SJérôme Glisse 4572aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) { 4582aee09d8SJérôme Glisse *fault = *write_fault = false; 4592aee09d8SJérôme Glisse return; 4602aee09d8SJérôme Glisse } 4612aee09d8SJérôme Glisse 462a3e0d41cSJérôme Glisse *fault = *write_fault = false; 4632aee09d8SJérôme Glisse for (i = 0; i < npages; ++i) { 4642aee09d8SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags, 4652aee09d8SJérôme Glisse fault, write_fault); 466a3e0d41cSJérôme Glisse if ((*write_fault)) 4672aee09d8SJérôme Glisse return; 4682aee09d8SJérôme Glisse } 4692aee09d8SJérôme Glisse } 4702aee09d8SJérôme Glisse 4712aee09d8SJérôme Glisse static int hmm_vma_walk_hole(unsigned long addr, unsigned long end, 4722aee09d8SJérôme Glisse struct mm_walk *walk) 4732aee09d8SJérôme Glisse { 4742aee09d8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 4752aee09d8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4762aee09d8SJérôme Glisse bool fault, write_fault; 4772aee09d8SJérôme Glisse unsigned long i, npages; 4782aee09d8SJérôme Glisse uint64_t *pfns; 4792aee09d8SJérôme Glisse 4802aee09d8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 4812aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 4822aee09d8SJérôme Glisse pfns = &range->pfns[i]; 4832aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 4842aee09d8SJérôme Glisse 0, &fault, &write_fault); 4852aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 4862aee09d8SJérôme Glisse } 4872aee09d8SJérôme Glisse 488f88a1e90SJérôme Glisse static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd) 4892aee09d8SJérôme Glisse { 4902aee09d8SJérôme Glisse if (pmd_protnone(pmd)) 4912aee09d8SJérôme Glisse return 0; 492f88a1e90SJérôme Glisse return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] | 493f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 494f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 495da4c3c73SJérôme Glisse } 496da4c3c73SJérôme Glisse 497992de9a8SJérôme Glisse static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud) 498992de9a8SJérôme Glisse { 499992de9a8SJérôme Glisse if (!pud_present(pud)) 500992de9a8SJérôme Glisse return 0; 501992de9a8SJérôme Glisse return pud_write(pud) ? range->flags[HMM_PFN_VALID] | 502992de9a8SJérôme Glisse range->flags[HMM_PFN_WRITE] : 503992de9a8SJérôme Glisse range->flags[HMM_PFN_VALID]; 504992de9a8SJérôme Glisse } 505992de9a8SJérôme Glisse 50653f5c3f4SJérôme Glisse static int hmm_vma_handle_pmd(struct mm_walk *walk, 50753f5c3f4SJérôme Glisse unsigned long addr, 50853f5c3f4SJérôme Glisse unsigned long end, 50953f5c3f4SJérôme Glisse uint64_t *pfns, 51053f5c3f4SJérôme Glisse pmd_t pmd) 51153f5c3f4SJérôme Glisse { 512992de9a8SJérôme Glisse #ifdef CONFIG_TRANSPARENT_HUGEPAGE 51353f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 514f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 5152aee09d8SJérôme Glisse unsigned long pfn, npages, i; 5162aee09d8SJérôme Glisse bool fault, write_fault; 517f88a1e90SJérôme Glisse uint64_t cpu_flags; 51853f5c3f4SJérôme Glisse 5192aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 520f88a1e90SJérôme Glisse cpu_flags = pmd_to_hmm_pfn_flags(range, pmd); 5212aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags, 5222aee09d8SJérôme Glisse &fault, &write_fault); 52353f5c3f4SJérôme Glisse 5242aee09d8SJérôme Glisse if (pmd_protnone(pmd) || fault || write_fault) 5252aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 52653f5c3f4SJérôme Glisse 52753f5c3f4SJérôme Glisse pfn = pmd_pfn(pmd) + pte_index(addr); 528992de9a8SJérôme Glisse for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) { 529992de9a8SJérôme Glisse if (pmd_devmap(pmd)) { 530992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 531992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 532992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 533992de9a8SJérôme Glisse return -EBUSY; 534992de9a8SJérôme Glisse } 535391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags; 536992de9a8SJérôme Glisse } 537992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 538992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 539992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 540992de9a8SJérôme Glisse } 54153f5c3f4SJérôme Glisse hmm_vma_walk->last = end; 54253f5c3f4SJérôme Glisse return 0; 543992de9a8SJérôme Glisse #else 544992de9a8SJérôme Glisse /* If THP is not enabled then we should never reach that code ! */ 545992de9a8SJérôme Glisse return -EINVAL; 546992de9a8SJérôme Glisse #endif 54753f5c3f4SJérôme Glisse } 54853f5c3f4SJérôme Glisse 549f88a1e90SJérôme Glisse static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte) 5502aee09d8SJérôme Glisse { 5512aee09d8SJérôme Glisse if (pte_none(pte) || !pte_present(pte)) 5522aee09d8SJérôme Glisse return 0; 553f88a1e90SJérôme Glisse return pte_write(pte) ? range->flags[HMM_PFN_VALID] | 554f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 555f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 5562aee09d8SJérôme Glisse } 5572aee09d8SJérôme Glisse 55853f5c3f4SJérôme Glisse static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr, 55953f5c3f4SJérôme Glisse unsigned long end, pmd_t *pmdp, pte_t *ptep, 56053f5c3f4SJérôme Glisse uint64_t *pfn) 56153f5c3f4SJérôme Glisse { 56253f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 563f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 56453f5c3f4SJérôme Glisse struct vm_area_struct *vma = walk->vma; 5652aee09d8SJérôme Glisse bool fault, write_fault; 5662aee09d8SJérôme Glisse uint64_t cpu_flags; 56753f5c3f4SJérôme Glisse pte_t pte = *ptep; 568f88a1e90SJérôme Glisse uint64_t orig_pfn = *pfn; 56953f5c3f4SJérôme Glisse 570f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_NONE]; 57173231612SJérôme Glisse fault = write_fault = false; 57253f5c3f4SJérôme Glisse 57353f5c3f4SJérôme Glisse if (pte_none(pte)) { 57473231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0, 57573231612SJérôme Glisse &fault, &write_fault); 5762aee09d8SJérôme Glisse if (fault || write_fault) 57753f5c3f4SJérôme Glisse goto fault; 57853f5c3f4SJérôme Glisse return 0; 57953f5c3f4SJérôme Glisse } 58053f5c3f4SJérôme Glisse 58153f5c3f4SJérôme Glisse if (!pte_present(pte)) { 58253f5c3f4SJérôme Glisse swp_entry_t entry = pte_to_swp_entry(pte); 58353f5c3f4SJérôme Glisse 58453f5c3f4SJérôme Glisse if (!non_swap_entry(entry)) { 5852aee09d8SJérôme Glisse if (fault || write_fault) 58653f5c3f4SJérôme Glisse goto fault; 58753f5c3f4SJérôme Glisse return 0; 58853f5c3f4SJérôme Glisse } 58953f5c3f4SJérôme Glisse 59053f5c3f4SJérôme Glisse /* 59153f5c3f4SJérôme Glisse * This is a special swap entry, ignore migration, use 59253f5c3f4SJérôme Glisse * device and report anything else as error. 59353f5c3f4SJérôme Glisse */ 59453f5c3f4SJérôme Glisse if (is_device_private_entry(entry)) { 595f88a1e90SJérôme Glisse cpu_flags = range->flags[HMM_PFN_VALID] | 596f88a1e90SJérôme Glisse range->flags[HMM_PFN_DEVICE_PRIVATE]; 5972aee09d8SJérôme Glisse cpu_flags |= is_write_device_private_entry(entry) ? 598f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 0; 599f88a1e90SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 600f88a1e90SJérôme Glisse &fault, &write_fault); 601f88a1e90SJérôme Glisse if (fault || write_fault) 602f88a1e90SJérôme Glisse goto fault; 603391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, 604391aab11SJérôme Glisse swp_offset(entry)); 605f88a1e90SJérôme Glisse *pfn |= cpu_flags; 60653f5c3f4SJérôme Glisse return 0; 60753f5c3f4SJérôme Glisse } 60853f5c3f4SJérôme Glisse 60953f5c3f4SJérôme Glisse if (is_migration_entry(entry)) { 6102aee09d8SJérôme Glisse if (fault || write_fault) { 61153f5c3f4SJérôme Glisse pte_unmap(ptep); 61253f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 61353f5c3f4SJérôme Glisse migration_entry_wait(vma->vm_mm, 61453f5c3f4SJérôme Glisse pmdp, addr); 61573231612SJérôme Glisse return -EBUSY; 61653f5c3f4SJérôme Glisse } 61753f5c3f4SJérôme Glisse return 0; 61853f5c3f4SJérôme Glisse } 61953f5c3f4SJérôme Glisse 62053f5c3f4SJérôme Glisse /* Report error for everything else */ 621f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 62253f5c3f4SJérôme Glisse return -EFAULT; 62373231612SJérôme Glisse } else { 62473231612SJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, pte); 62573231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 62673231612SJérôme Glisse &fault, &write_fault); 62753f5c3f4SJérôme Glisse } 62853f5c3f4SJérôme Glisse 6292aee09d8SJérôme Glisse if (fault || write_fault) 63053f5c3f4SJérôme Glisse goto fault; 63153f5c3f4SJérôme Glisse 632992de9a8SJérôme Glisse if (pte_devmap(pte)) { 633992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte), 634992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 635992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 636992de9a8SJérôme Glisse return -EBUSY; 637992de9a8SJérôme Glisse } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) { 638992de9a8SJérôme Glisse *pfn = range->values[HMM_PFN_SPECIAL]; 639992de9a8SJérôme Glisse return -EFAULT; 640992de9a8SJérôme Glisse } 641992de9a8SJérôme Glisse 642391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags; 64353f5c3f4SJérôme Glisse return 0; 64453f5c3f4SJérôme Glisse 64553f5c3f4SJérôme Glisse fault: 646992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 647992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 648992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 649992de9a8SJérôme Glisse } 65053f5c3f4SJérôme Glisse pte_unmap(ptep); 65153f5c3f4SJérôme Glisse /* Fault any virtual address we were asked to fault */ 6522aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 65353f5c3f4SJérôme Glisse } 65453f5c3f4SJérôme Glisse 655da4c3c73SJérôme Glisse static int hmm_vma_walk_pmd(pmd_t *pmdp, 656da4c3c73SJérôme Glisse unsigned long start, 657da4c3c73SJérôme Glisse unsigned long end, 658da4c3c73SJérôme Glisse struct mm_walk *walk) 659da4c3c73SJérôme Glisse { 66074eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 66174eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 662d08faca0SJérôme Glisse struct vm_area_struct *vma = walk->vma; 663ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 664da4c3c73SJérôme Glisse unsigned long addr = start, i; 665da4c3c73SJérôme Glisse pte_t *ptep; 666da4c3c73SJérôme Glisse pmd_t pmd; 667da4c3c73SJérôme Glisse 668d08faca0SJérôme Glisse 669d08faca0SJérôme Glisse again: 670d08faca0SJérôme Glisse pmd = READ_ONCE(*pmdp); 671d08faca0SJérôme Glisse if (pmd_none(pmd)) 672d08faca0SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 673d08faca0SJérôme Glisse 674d08faca0SJérôme Glisse if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB)) 675d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 676d08faca0SJérôme Glisse 677d08faca0SJérôme Glisse if (thp_migration_supported() && is_pmd_migration_entry(pmd)) { 678d08faca0SJérôme Glisse bool fault, write_fault; 679d08faca0SJérôme Glisse unsigned long npages; 680d08faca0SJérôme Glisse uint64_t *pfns; 681d08faca0SJérôme Glisse 682d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 683d08faca0SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 684d08faca0SJérôme Glisse pfns = &range->pfns[i]; 685d08faca0SJérôme Glisse 686d08faca0SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 687d08faca0SJérôme Glisse 0, &fault, &write_fault); 688d08faca0SJérôme Glisse if (fault || write_fault) { 689d08faca0SJérôme Glisse hmm_vma_walk->last = addr; 690d08faca0SJérôme Glisse pmd_migration_entry_wait(vma->vm_mm, pmdp); 69173231612SJérôme Glisse return -EBUSY; 692d08faca0SJérôme Glisse } 693d08faca0SJérôme Glisse return 0; 694d08faca0SJérôme Glisse } else if (!pmd_present(pmd)) 695d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 696d08faca0SJérôme Glisse 697d08faca0SJérôme Glisse if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) { 698da4c3c73SJérôme Glisse /* 699da4c3c73SJérôme Glisse * No need to take pmd_lock here, even if some other threads 700da4c3c73SJérôme Glisse * is splitting the huge pmd we will get that event through 701da4c3c73SJérôme Glisse * mmu_notifier callback. 702da4c3c73SJérôme Glisse * 703da4c3c73SJérôme Glisse * So just read pmd value and check again its a transparent 704da4c3c73SJérôme Glisse * huge or device mapping one and compute corresponding pfn 705da4c3c73SJérôme Glisse * values. 706da4c3c73SJérôme Glisse */ 707da4c3c73SJérôme Glisse pmd = pmd_read_atomic(pmdp); 708da4c3c73SJérôme Glisse barrier(); 709da4c3c73SJérôme Glisse if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) 710da4c3c73SJérôme Glisse goto again; 711da4c3c73SJérôme Glisse 712d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 71353f5c3f4SJérôme Glisse return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd); 714da4c3c73SJérôme Glisse } 715da4c3c73SJérôme Glisse 716d08faca0SJérôme Glisse /* 717d08faca0SJérôme Glisse * We have handled all the valid case above ie either none, migration, 718d08faca0SJérôme Glisse * huge or transparent huge. At this point either it is a valid pmd 719d08faca0SJérôme Glisse * entry pointing to pte directory or it is a bad pmd that will not 720d08faca0SJérôme Glisse * recover. 721d08faca0SJérôme Glisse */ 722d08faca0SJérôme Glisse if (pmd_bad(pmd)) 723da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 724da4c3c73SJérôme Glisse 725da4c3c73SJérôme Glisse ptep = pte_offset_map(pmdp, addr); 726d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 727da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { 72853f5c3f4SJérôme Glisse int r; 729da4c3c73SJérôme Glisse 73053f5c3f4SJérôme Glisse r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]); 73153f5c3f4SJérôme Glisse if (r) { 73253f5c3f4SJérôme Glisse /* hmm_vma_handle_pte() did unmap pte directory */ 73374eee180SJérôme Glisse hmm_vma_walk->last = addr; 73453f5c3f4SJérôme Glisse return r; 73574eee180SJérôme Glisse } 736da4c3c73SJérôme Glisse } 737992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 738992de9a8SJérôme Glisse /* 739992de9a8SJérôme Glisse * We do put_dev_pagemap() here and not in hmm_vma_handle_pte() 740992de9a8SJérôme Glisse * so that we can leverage get_dev_pagemap() optimization which 741992de9a8SJérôme Glisse * will not re-take a reference on a pgmap if we already have 742992de9a8SJérôme Glisse * one. 743992de9a8SJérôme Glisse */ 744992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 745992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 746992de9a8SJérôme Glisse } 747da4c3c73SJérôme Glisse pte_unmap(ptep - 1); 748da4c3c73SJérôme Glisse 74953f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 750da4c3c73SJérôme Glisse return 0; 751da4c3c73SJérôme Glisse } 752da4c3c73SJérôme Glisse 753992de9a8SJérôme Glisse static int hmm_vma_walk_pud(pud_t *pudp, 754992de9a8SJérôme Glisse unsigned long start, 755992de9a8SJérôme Glisse unsigned long end, 756992de9a8SJérôme Glisse struct mm_walk *walk) 757992de9a8SJérôme Glisse { 758992de9a8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 759992de9a8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 760992de9a8SJérôme Glisse unsigned long addr = start, next; 761992de9a8SJérôme Glisse pmd_t *pmdp; 762992de9a8SJérôme Glisse pud_t pud; 763992de9a8SJérôme Glisse int ret; 764992de9a8SJérôme Glisse 765992de9a8SJérôme Glisse again: 766992de9a8SJérôme Glisse pud = READ_ONCE(*pudp); 767992de9a8SJérôme Glisse if (pud_none(pud)) 768992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 769992de9a8SJérôme Glisse 770992de9a8SJérôme Glisse if (pud_huge(pud) && pud_devmap(pud)) { 771992de9a8SJérôme Glisse unsigned long i, npages, pfn; 772992de9a8SJérôme Glisse uint64_t *pfns, cpu_flags; 773992de9a8SJérôme Glisse bool fault, write_fault; 774992de9a8SJérôme Glisse 775992de9a8SJérôme Glisse if (!pud_present(pud)) 776992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 777992de9a8SJérôme Glisse 778992de9a8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 779992de9a8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 780992de9a8SJérôme Glisse pfns = &range->pfns[i]; 781992de9a8SJérôme Glisse 782992de9a8SJérôme Glisse cpu_flags = pud_to_hmm_pfn_flags(range, pud); 783992de9a8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 784992de9a8SJérôme Glisse cpu_flags, &fault, &write_fault); 785992de9a8SJérôme Glisse if (fault || write_fault) 786992de9a8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, 787992de9a8SJérôme Glisse write_fault, walk); 788992de9a8SJérôme Glisse 789992de9a8SJérôme Glisse pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); 790992de9a8SJérôme Glisse for (i = 0; i < npages; ++i, ++pfn) { 791992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 792992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 793992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 794992de9a8SJérôme Glisse return -EBUSY; 795391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 796391aab11SJérôme Glisse cpu_flags; 797992de9a8SJérôme Glisse } 798992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 799992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 800992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 801992de9a8SJérôme Glisse } 802992de9a8SJérôme Glisse hmm_vma_walk->last = end; 803992de9a8SJérôme Glisse return 0; 804992de9a8SJérôme Glisse } 805992de9a8SJérôme Glisse 806992de9a8SJérôme Glisse split_huge_pud(walk->vma, pudp, addr); 807992de9a8SJérôme Glisse if (pud_none(*pudp)) 808992de9a8SJérôme Glisse goto again; 809992de9a8SJérôme Glisse 810992de9a8SJérôme Glisse pmdp = pmd_offset(pudp, addr); 811992de9a8SJérôme Glisse do { 812992de9a8SJérôme Glisse next = pmd_addr_end(addr, end); 813992de9a8SJérôme Glisse ret = hmm_vma_walk_pmd(pmdp, addr, next, walk); 814992de9a8SJérôme Glisse if (ret) 815992de9a8SJérôme Glisse return ret; 816992de9a8SJérôme Glisse } while (pmdp++, addr = next, addr != end); 817992de9a8SJérôme Glisse 818992de9a8SJérôme Glisse return 0; 819992de9a8SJérôme Glisse } 820992de9a8SJérôme Glisse 82163d5066fSJérôme Glisse static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask, 82263d5066fSJérôme Glisse unsigned long start, unsigned long end, 82363d5066fSJérôme Glisse struct mm_walk *walk) 82463d5066fSJérôme Glisse { 82563d5066fSJérôme Glisse #ifdef CONFIG_HUGETLB_PAGE 82663d5066fSJérôme Glisse unsigned long addr = start, i, pfn, mask, size, pfn_inc; 82763d5066fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 82863d5066fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 82963d5066fSJérôme Glisse struct vm_area_struct *vma = walk->vma; 83063d5066fSJérôme Glisse struct hstate *h = hstate_vma(vma); 83163d5066fSJérôme Glisse uint64_t orig_pfn, cpu_flags; 83263d5066fSJérôme Glisse bool fault, write_fault; 83363d5066fSJérôme Glisse spinlock_t *ptl; 83463d5066fSJérôme Glisse pte_t entry; 83563d5066fSJérôme Glisse int ret = 0; 83663d5066fSJérôme Glisse 83763d5066fSJérôme Glisse size = 1UL << huge_page_shift(h); 83863d5066fSJérôme Glisse mask = size - 1; 83963d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) { 84063d5066fSJérôme Glisse /* Make sure we are looking at full page. */ 84163d5066fSJérôme Glisse if (start & mask) 84263d5066fSJérôme Glisse return -EINVAL; 84363d5066fSJérôme Glisse if (end < (start + size)) 84463d5066fSJérôme Glisse return -EINVAL; 84563d5066fSJérôme Glisse pfn_inc = size >> PAGE_SHIFT; 84663d5066fSJérôme Glisse } else { 84763d5066fSJérôme Glisse pfn_inc = 1; 84863d5066fSJérôme Glisse size = PAGE_SIZE; 84963d5066fSJérôme Glisse } 85063d5066fSJérôme Glisse 85163d5066fSJérôme Glisse 85263d5066fSJérôme Glisse ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); 85363d5066fSJérôme Glisse entry = huge_ptep_get(pte); 85463d5066fSJérôme Glisse 85563d5066fSJérôme Glisse i = (start - range->start) >> range->page_shift; 85663d5066fSJérôme Glisse orig_pfn = range->pfns[i]; 85763d5066fSJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 85863d5066fSJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, entry); 85963d5066fSJérôme Glisse fault = write_fault = false; 86063d5066fSJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 86163d5066fSJérôme Glisse &fault, &write_fault); 86263d5066fSJérôme Glisse if (fault || write_fault) { 86363d5066fSJérôme Glisse ret = -ENOENT; 86463d5066fSJérôme Glisse goto unlock; 86563d5066fSJérôme Glisse } 86663d5066fSJérôme Glisse 86763d5066fSJérôme Glisse pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift); 86863d5066fSJérôme Glisse for (; addr < end; addr += size, i++, pfn += pfn_inc) 869391aab11SJérôme Glisse range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 870391aab11SJérôme Glisse cpu_flags; 87163d5066fSJérôme Glisse hmm_vma_walk->last = end; 87263d5066fSJérôme Glisse 87363d5066fSJérôme Glisse unlock: 87463d5066fSJérôme Glisse spin_unlock(ptl); 87563d5066fSJérôme Glisse 87663d5066fSJérôme Glisse if (ret == -ENOENT) 87763d5066fSJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 87863d5066fSJérôme Glisse 87963d5066fSJérôme Glisse return ret; 88063d5066fSJérôme Glisse #else /* CONFIG_HUGETLB_PAGE */ 88163d5066fSJérôme Glisse return -EINVAL; 88263d5066fSJérôme Glisse #endif 88363d5066fSJérôme Glisse } 88463d5066fSJérôme Glisse 885f88a1e90SJérôme Glisse static void hmm_pfns_clear(struct hmm_range *range, 886f88a1e90SJérôme Glisse uint64_t *pfns, 88733cd47dcSJérôme Glisse unsigned long addr, 88833cd47dcSJérôme Glisse unsigned long end) 88933cd47dcSJérôme Glisse { 89033cd47dcSJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 891f88a1e90SJérôme Glisse *pfns = range->values[HMM_PFN_NONE]; 89233cd47dcSJérôme Glisse } 89333cd47dcSJérôme Glisse 894da4c3c73SJérôme Glisse /* 895a3e0d41cSJérôme Glisse * hmm_range_register() - start tracking change to CPU page table over a range 896a3e0d41cSJérôme Glisse * @range: range 897a3e0d41cSJérôme Glisse * @mm: the mm struct for the range of virtual address 898a3e0d41cSJérôme Glisse * @start: start virtual address (inclusive) 899a3e0d41cSJérôme Glisse * @end: end virtual address (exclusive) 90063d5066fSJérôme Glisse * @page_shift: expect page shift for the range 901a3e0d41cSJérôme Glisse * Returns 0 on success, -EFAULT if the address space is no longer valid 902a3e0d41cSJérôme Glisse * 903a3e0d41cSJérôme Glisse * Track updates to the CPU page table see include/linux/hmm.h 904a3e0d41cSJérôme Glisse */ 905a3e0d41cSJérôme Glisse int hmm_range_register(struct hmm_range *range, 906a3e0d41cSJérôme Glisse struct mm_struct *mm, 907a3e0d41cSJérôme Glisse unsigned long start, 90863d5066fSJérôme Glisse unsigned long end, 90963d5066fSJérôme Glisse unsigned page_shift) 910a3e0d41cSJérôme Glisse { 91163d5066fSJérôme Glisse unsigned long mask = ((1UL << page_shift) - 1UL); 912*085ea250SRalph Campbell struct hmm *hmm; 91363d5066fSJérôme Glisse 914a3e0d41cSJérôme Glisse range->valid = false; 915a3e0d41cSJérôme Glisse range->hmm = NULL; 916a3e0d41cSJérôme Glisse 91763d5066fSJérôme Glisse if ((start & mask) || (end & mask)) 91863d5066fSJérôme Glisse return -EINVAL; 91963d5066fSJérôme Glisse if (start >= end) 920a3e0d41cSJérôme Glisse return -EINVAL; 921a3e0d41cSJérôme Glisse 92263d5066fSJérôme Glisse range->page_shift = page_shift; 923a3e0d41cSJérôme Glisse range->start = start; 924a3e0d41cSJérôme Glisse range->end = end; 925a3e0d41cSJérôme Glisse 926*085ea250SRalph Campbell hmm = hmm_get_or_create(mm); 927*085ea250SRalph Campbell if (!hmm) 928a3e0d41cSJérôme Glisse return -EFAULT; 929a3e0d41cSJérôme Glisse 930a3e0d41cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 931*085ea250SRalph Campbell if (hmm->mm == NULL || hmm->dead) { 932*085ea250SRalph Campbell hmm_put(hmm); 933a3e0d41cSJérôme Glisse return -EFAULT; 934a3e0d41cSJérôme Glisse } 935a3e0d41cSJérôme Glisse 936*085ea250SRalph Campbell /* Initialize range to track CPU page table updates. */ 937*085ea250SRalph Campbell mutex_lock(&hmm->lock); 938a3e0d41cSJérôme Glisse 939*085ea250SRalph Campbell range->hmm = hmm; 940*085ea250SRalph Campbell list_add_rcu(&range->list, &hmm->ranges); 941a3e0d41cSJérôme Glisse 942a3e0d41cSJérôme Glisse /* 943a3e0d41cSJérôme Glisse * If there are any concurrent notifiers we have to wait for them for 944a3e0d41cSJérôme Glisse * the range to be valid (see hmm_range_wait_until_valid()). 945a3e0d41cSJérôme Glisse */ 946*085ea250SRalph Campbell if (!hmm->notifiers) 947a3e0d41cSJérôme Glisse range->valid = true; 948*085ea250SRalph Campbell mutex_unlock(&hmm->lock); 949a3e0d41cSJérôme Glisse 950a3e0d41cSJérôme Glisse return 0; 951a3e0d41cSJérôme Glisse } 952a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_register); 953a3e0d41cSJérôme Glisse 954a3e0d41cSJérôme Glisse /* 955a3e0d41cSJérôme Glisse * hmm_range_unregister() - stop tracking change to CPU page table over a range 956a3e0d41cSJérôme Glisse * @range: range 957a3e0d41cSJérôme Glisse * 958a3e0d41cSJérôme Glisse * Range struct is used to track updates to the CPU page table after a call to 959a3e0d41cSJérôme Glisse * hmm_range_register(). See include/linux/hmm.h for how to use it. 960a3e0d41cSJérôme Glisse */ 961a3e0d41cSJérôme Glisse void hmm_range_unregister(struct hmm_range *range) 962a3e0d41cSJérôme Glisse { 963*085ea250SRalph Campbell struct hmm *hmm = range->hmm; 964*085ea250SRalph Campbell 965a3e0d41cSJérôme Glisse /* Sanity check this really should not happen. */ 966*085ea250SRalph Campbell if (hmm == NULL || range->end <= range->start) 967a3e0d41cSJérôme Glisse return; 968a3e0d41cSJérôme Glisse 969*085ea250SRalph Campbell mutex_lock(&hmm->lock); 970a3e0d41cSJérôme Glisse list_del_rcu(&range->list); 971*085ea250SRalph Campbell mutex_unlock(&hmm->lock); 972a3e0d41cSJérôme Glisse 973a3e0d41cSJérôme Glisse /* Drop reference taken by hmm_range_register() */ 974a3e0d41cSJérôme Glisse range->valid = false; 975*085ea250SRalph Campbell hmm_put(hmm); 976a3e0d41cSJérôme Glisse range->hmm = NULL; 977a3e0d41cSJérôme Glisse } 978a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_unregister); 979a3e0d41cSJérôme Glisse 980a3e0d41cSJérôme Glisse /* 98125f23a0cSJérôme Glisse * hmm_range_snapshot() - snapshot CPU page table for a range 98225f23a0cSJérôme Glisse * @range: range 983*085ea250SRalph Campbell * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid 984a3e0d41cSJérôme Glisse * permission (for instance asking for write and range is read only), 985a3e0d41cSJérôme Glisse * -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid 986a3e0d41cSJérôme Glisse * vma or it is illegal to access that range), number of valid pages 987a3e0d41cSJérôme Glisse * in range->pfns[] (from range start address). 988da4c3c73SJérôme Glisse * 989da4c3c73SJérôme Glisse * This snapshots the CPU page table for a range of virtual addresses. Snapshot 990a3e0d41cSJérôme Glisse * validity is tracked by range struct. See in include/linux/hmm.h for example 991a3e0d41cSJérôme Glisse * on how to use. 992da4c3c73SJérôme Glisse */ 99325f23a0cSJérôme Glisse long hmm_range_snapshot(struct hmm_range *range) 994da4c3c73SJérôme Glisse { 99563d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 996a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 99774eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 998a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 999a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 1000da4c3c73SJérôme Glisse struct mm_walk mm_walk; 1001704f3f2cSJérôme Glisse 1002704f3f2cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 1003a3e0d41cSJérôme Glisse if (hmm->mm == NULL || hmm->dead) 1004a3e0d41cSJérôme Glisse return -EFAULT; 1005da4c3c73SJérôme Glisse 1006a3e0d41cSJérôme Glisse do { 1007a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1008a3e0d41cSJérôme Glisse if (!range->valid) 1009a3e0d41cSJérôme Glisse return -EAGAIN; 1010a3e0d41cSJérôme Glisse 1011a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 101263d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1013a3e0d41cSJérôme Glisse return -EFAULT; 1014a3e0d41cSJérôme Glisse 101563d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 10161c2308f0SJason Gunthorpe if (huge_page_shift(hstate_vma(vma)) != 10171c2308f0SJason Gunthorpe range->page_shift && 101863d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 101963d5066fSJérôme Glisse return -EINVAL; 102063d5066fSJérôme Glisse } else { 102163d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 102263d5066fSJérôme Glisse return -EINVAL; 102363d5066fSJérôme Glisse } 102463d5066fSJérôme Glisse 102586586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 102686586a41SJérôme Glisse /* 1027a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1028a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1029a3e0d41cSJérôme Glisse * support architecture that allow write without read. 103086586a41SJérôme Glisse */ 1031a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1032a3e0d41cSJérôme Glisse range->start, range->end); 103386586a41SJérôme Glisse return -EPERM; 103486586a41SJérôme Glisse } 103586586a41SJérôme Glisse 1036a3e0d41cSJérôme Glisse range->vma = vma; 1037992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1038a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 103974eee180SJérôme Glisse hmm_vma_walk.fault = false; 104074eee180SJérôme Glisse hmm_vma_walk.range = range; 104174eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1042a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 104374eee180SJérôme Glisse 1044da4c3c73SJérôme Glisse mm_walk.vma = vma; 1045da4c3c73SJérôme Glisse mm_walk.mm = vma->vm_mm; 1046da4c3c73SJérôme Glisse mm_walk.pte_entry = NULL; 1047da4c3c73SJérôme Glisse mm_walk.test_walk = NULL; 1048da4c3c73SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1049992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 1050da4c3c73SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 1051da4c3c73SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 105263d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 1053da4c3c73SJérôme Glisse 1054a3e0d41cSJérôme Glisse walk_page_range(start, end, &mm_walk); 1055a3e0d41cSJérôme Glisse start = end; 1056a3e0d41cSJérôme Glisse } while (start < range->end); 1057a3e0d41cSJérôme Glisse 105825f23a0cSJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1059da4c3c73SJérôme Glisse } 106025f23a0cSJérôme Glisse EXPORT_SYMBOL(hmm_range_snapshot); 1061da4c3c73SJérôme Glisse 1062da4c3c73SJérôme Glisse /* 106373231612SJérôme Glisse * hmm_range_fault() - try to fault some address in a virtual address range 106408232a45SJérôme Glisse * @range: range being faulted 106574eee180SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 1066*085ea250SRalph Campbell * Return: number of valid pages in range->pfns[] (from range start 106773231612SJérôme Glisse * address). This may be zero. If the return value is negative, 106873231612SJérôme Glisse * then one of the following values may be returned: 106973231612SJérôme Glisse * 107073231612SJérôme Glisse * -EINVAL invalid arguments or mm or virtual address are in an 107163d5066fSJérôme Glisse * invalid vma (for instance device file vma). 107273231612SJérôme Glisse * -ENOMEM: Out of memory. 107373231612SJérôme Glisse * -EPERM: Invalid permission (for instance asking for write and 107473231612SJérôme Glisse * range is read only). 107573231612SJérôme Glisse * -EAGAIN: If you need to retry and mmap_sem was drop. This can only 107673231612SJérôme Glisse * happens if block argument is false. 107773231612SJérôme Glisse * -EBUSY: If the the range is being invalidated and you should wait 107873231612SJérôme Glisse * for invalidation to finish. 107973231612SJérôme Glisse * -EFAULT: Invalid (ie either no valid vma or it is illegal to access 108073231612SJérôme Glisse * that range), number of valid pages in range->pfns[] (from 108173231612SJérôme Glisse * range start address). 108274eee180SJérôme Glisse * 108374eee180SJérôme Glisse * This is similar to a regular CPU page fault except that it will not trigger 108473231612SJérôme Glisse * any memory migration if the memory being faulted is not accessible by CPUs 108573231612SJérôme Glisse * and caller does not ask for migration. 108674eee180SJérôme Glisse * 1087ff05c0c6SJérôme Glisse * On error, for one virtual address in the range, the function will mark the 1088ff05c0c6SJérôme Glisse * corresponding HMM pfn entry with an error flag. 108974eee180SJérôme Glisse */ 109073231612SJérôme Glisse long hmm_range_fault(struct hmm_range *range, bool block) 109174eee180SJérôme Glisse { 109263d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 1093a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 109474eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 1095a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 1096a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 109774eee180SJérôme Glisse struct mm_walk mm_walk; 109874eee180SJérôme Glisse int ret; 109974eee180SJérôme Glisse 1100704f3f2cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 1101a3e0d41cSJérôme Glisse if (hmm->mm == NULL || hmm->dead) 1102a3e0d41cSJérôme Glisse return -EFAULT; 1103a3e0d41cSJérôme Glisse 1104a3e0d41cSJérôme Glisse do { 1105a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1106a3e0d41cSJérôme Glisse if (!range->valid) { 1107a3e0d41cSJérôme Glisse up_read(&hmm->mm->mmap_sem); 1108a3e0d41cSJérôme Glisse return -EAGAIN; 1109704f3f2cSJérôme Glisse } 111074eee180SJérôme Glisse 1111a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 111263d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1113a3e0d41cSJérôme Glisse return -EFAULT; 1114a3e0d41cSJérôme Glisse 111563d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 111663d5066fSJérôme Glisse if (huge_page_shift(hstate_vma(vma)) != 111763d5066fSJérôme Glisse range->page_shift && 111863d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 111963d5066fSJérôme Glisse return -EINVAL; 112063d5066fSJérôme Glisse } else { 112163d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 112263d5066fSJérôme Glisse return -EINVAL; 112363d5066fSJérôme Glisse } 112463d5066fSJérôme Glisse 112586586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 112686586a41SJérôme Glisse /* 1127a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1128a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1129a3e0d41cSJérôme Glisse * support architecture that allow write without read. 113086586a41SJérôme Glisse */ 1131a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1132a3e0d41cSJérôme Glisse range->start, range->end); 113386586a41SJérôme Glisse return -EPERM; 113486586a41SJérôme Glisse } 113574eee180SJérôme Glisse 1136a3e0d41cSJérôme Glisse range->vma = vma; 1137992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1138a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 113974eee180SJérôme Glisse hmm_vma_walk.fault = true; 114074eee180SJérôme Glisse hmm_vma_walk.block = block; 114174eee180SJérôme Glisse hmm_vma_walk.range = range; 114274eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1143a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 114474eee180SJérôme Glisse 114574eee180SJérôme Glisse mm_walk.vma = vma; 114674eee180SJérôme Glisse mm_walk.mm = vma->vm_mm; 114774eee180SJérôme Glisse mm_walk.pte_entry = NULL; 114874eee180SJérôme Glisse mm_walk.test_walk = NULL; 114974eee180SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1150992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 115174eee180SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 115274eee180SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 115363d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 115474eee180SJérôme Glisse 115574eee180SJérôme Glisse do { 1156a3e0d41cSJérôme Glisse ret = walk_page_range(start, end, &mm_walk); 115774eee180SJérôme Glisse start = hmm_vma_walk.last; 1158a3e0d41cSJérôme Glisse 115973231612SJérôme Glisse /* Keep trying while the range is valid. */ 116073231612SJérôme Glisse } while (ret == -EBUSY && range->valid); 116174eee180SJérôme Glisse 116274eee180SJérôme Glisse if (ret) { 116374eee180SJérôme Glisse unsigned long i; 116474eee180SJérôme Glisse 116574eee180SJérôme Glisse i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1166a3e0d41cSJérôme Glisse hmm_pfns_clear(range, &range->pfns[i], 1167a3e0d41cSJérôme Glisse hmm_vma_walk.last, range->end); 116873231612SJérôme Glisse return ret; 116974eee180SJérôme Glisse } 1170a3e0d41cSJérôme Glisse start = end; 1171a3e0d41cSJérôme Glisse 1172a3e0d41cSJérôme Glisse } while (start < range->end); 1173704f3f2cSJérôme Glisse 117473231612SJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 117574eee180SJérôme Glisse } 117673231612SJérôme Glisse EXPORT_SYMBOL(hmm_range_fault); 117755c0ece8SJérôme Glisse 117855c0ece8SJérôme Glisse /** 117955c0ece8SJérôme Glisse * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one. 118055c0ece8SJérôme Glisse * @range: range being faulted 118155c0ece8SJérôme Glisse * @device: device against to dma map page to 118255c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 118355c0ece8SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 1184*085ea250SRalph Campbell * Return: number of pages mapped on success, -EAGAIN if mmap_sem have been 118555c0ece8SJérôme Glisse * drop and you need to try again, some other error value otherwise 118655c0ece8SJérôme Glisse * 118755c0ece8SJérôme Glisse * Note same usage pattern as hmm_range_fault(). 118855c0ece8SJérôme Glisse */ 118955c0ece8SJérôme Glisse long hmm_range_dma_map(struct hmm_range *range, 119055c0ece8SJérôme Glisse struct device *device, 119155c0ece8SJérôme Glisse dma_addr_t *daddrs, 119255c0ece8SJérôme Glisse bool block) 119355c0ece8SJérôme Glisse { 119455c0ece8SJérôme Glisse unsigned long i, npages, mapped; 119555c0ece8SJérôme Glisse long ret; 119655c0ece8SJérôme Glisse 119755c0ece8SJérôme Glisse ret = hmm_range_fault(range, block); 119855c0ece8SJérôme Glisse if (ret <= 0) 119955c0ece8SJérôme Glisse return ret ? ret : -EBUSY; 120055c0ece8SJérôme Glisse 120155c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 120255c0ece8SJérôme Glisse for (i = 0, mapped = 0; i < npages; ++i) { 120355c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 120455c0ece8SJérôme Glisse struct page *page; 120555c0ece8SJérôme Glisse 120655c0ece8SJérôme Glisse /* 120755c0ece8SJérôme Glisse * FIXME need to update DMA API to provide invalid DMA address 120855c0ece8SJérôme Glisse * value instead of a function to test dma address value. This 120955c0ece8SJérôme Glisse * would remove lot of dumb code duplicated accross many arch. 121055c0ece8SJérôme Glisse * 121155c0ece8SJérôme Glisse * For now setting it to 0 here is good enough as the pfns[] 121255c0ece8SJérôme Glisse * value is what is use to check what is valid and what isn't. 121355c0ece8SJérôme Glisse */ 121455c0ece8SJérôme Glisse daddrs[i] = 0; 121555c0ece8SJérôme Glisse 1216391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 121755c0ece8SJérôme Glisse if (page == NULL) 121855c0ece8SJérôme Glisse continue; 121955c0ece8SJérôme Glisse 122055c0ece8SJérôme Glisse /* Check if range is being invalidated */ 122155c0ece8SJérôme Glisse if (!range->valid) { 122255c0ece8SJérôme Glisse ret = -EBUSY; 122355c0ece8SJérôme Glisse goto unmap; 122455c0ece8SJérôme Glisse } 122555c0ece8SJérôme Glisse 122655c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 122755c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 122855c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 122955c0ece8SJérôme Glisse 123055c0ece8SJérôme Glisse daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir); 123155c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) { 123255c0ece8SJérôme Glisse ret = -EFAULT; 123355c0ece8SJérôme Glisse goto unmap; 123455c0ece8SJérôme Glisse } 123555c0ece8SJérôme Glisse 123655c0ece8SJérôme Glisse mapped++; 123755c0ece8SJérôme Glisse } 123855c0ece8SJérôme Glisse 123955c0ece8SJérôme Glisse return mapped; 124055c0ece8SJérôme Glisse 124155c0ece8SJérôme Glisse unmap: 124255c0ece8SJérôme Glisse for (npages = i, i = 0; (i < npages) && mapped; ++i) { 124355c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 124455c0ece8SJérôme Glisse struct page *page; 124555c0ece8SJérôme Glisse 1246391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 124755c0ece8SJérôme Glisse if (page == NULL) 124855c0ece8SJérôme Glisse continue; 124955c0ece8SJérôme Glisse 125055c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) 125155c0ece8SJérôme Glisse continue; 125255c0ece8SJérôme Glisse 125355c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 125455c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 125555c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 125655c0ece8SJérôme Glisse 125755c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 125855c0ece8SJérôme Glisse mapped--; 125955c0ece8SJérôme Glisse } 126055c0ece8SJérôme Glisse 126155c0ece8SJérôme Glisse return ret; 126255c0ece8SJérôme Glisse } 126355c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_map); 126455c0ece8SJérôme Glisse 126555c0ece8SJérôme Glisse /** 126655c0ece8SJérôme Glisse * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map() 126755c0ece8SJérôme Glisse * @range: range being unmapped 126855c0ece8SJérôme Glisse * @vma: the vma against which the range (optional) 126955c0ece8SJérôme Glisse * @device: device against which dma map was done 127055c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 127155c0ece8SJérôme Glisse * @dirty: dirty page if it had the write flag set 1272*085ea250SRalph Campbell * Return: number of page unmapped on success, -EINVAL otherwise 127355c0ece8SJérôme Glisse * 127455c0ece8SJérôme Glisse * Note that caller MUST abide by mmu notifier or use HMM mirror and abide 127555c0ece8SJérôme Glisse * to the sync_cpu_device_pagetables() callback so that it is safe here to 127655c0ece8SJérôme Glisse * call set_page_dirty(). Caller must also take appropriate locks to avoid 127755c0ece8SJérôme Glisse * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress. 127855c0ece8SJérôme Glisse */ 127955c0ece8SJérôme Glisse long hmm_range_dma_unmap(struct hmm_range *range, 128055c0ece8SJérôme Glisse struct vm_area_struct *vma, 128155c0ece8SJérôme Glisse struct device *device, 128255c0ece8SJérôme Glisse dma_addr_t *daddrs, 128355c0ece8SJérôme Glisse bool dirty) 128455c0ece8SJérôme Glisse { 128555c0ece8SJérôme Glisse unsigned long i, npages; 128655c0ece8SJérôme Glisse long cpages = 0; 128755c0ece8SJérôme Glisse 128855c0ece8SJérôme Glisse /* Sanity check. */ 128955c0ece8SJérôme Glisse if (range->end <= range->start) 129055c0ece8SJérôme Glisse return -EINVAL; 129155c0ece8SJérôme Glisse if (!daddrs) 129255c0ece8SJérôme Glisse return -EINVAL; 129355c0ece8SJérôme Glisse if (!range->pfns) 129455c0ece8SJérôme Glisse return -EINVAL; 129555c0ece8SJérôme Glisse 129655c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 129755c0ece8SJérôme Glisse for (i = 0; i < npages; ++i) { 129855c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 129955c0ece8SJérôme Glisse struct page *page; 130055c0ece8SJérôme Glisse 1301391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 130255c0ece8SJérôme Glisse if (page == NULL) 130355c0ece8SJérôme Glisse continue; 130455c0ece8SJérôme Glisse 130555c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 130655c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) { 130755c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 130855c0ece8SJérôme Glisse 130955c0ece8SJérôme Glisse /* 131055c0ece8SJérôme Glisse * See comments in function description on why it is 131155c0ece8SJérôme Glisse * safe here to call set_page_dirty() 131255c0ece8SJérôme Glisse */ 131355c0ece8SJérôme Glisse if (dirty) 131455c0ece8SJérôme Glisse set_page_dirty(page); 131555c0ece8SJérôme Glisse } 131655c0ece8SJérôme Glisse 131755c0ece8SJérôme Glisse /* Unmap and clear pfns/dma address */ 131855c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 131955c0ece8SJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 132055c0ece8SJérôme Glisse /* FIXME see comments in hmm_vma_dma_map() */ 132155c0ece8SJérôme Glisse daddrs[i] = 0; 132255c0ece8SJérôme Glisse cpages++; 132355c0ece8SJérôme Glisse } 132455c0ece8SJérôme Glisse 132555c0ece8SJérôme Glisse return cpages; 132655c0ece8SJérôme Glisse } 132755c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_unmap); 1328c0b12405SJérôme Glisse #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ 13294ef589dcSJérôme Glisse 13304ef589dcSJérôme Glisse 1331df6ad698SJérôme Glisse #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC) 13324ef589dcSJérôme Glisse struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma, 13334ef589dcSJérôme Glisse unsigned long addr) 13344ef589dcSJérôme Glisse { 13354ef589dcSJérôme Glisse struct page *page; 13364ef589dcSJérôme Glisse 13374ef589dcSJérôme Glisse page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 13384ef589dcSJérôme Glisse if (!page) 13394ef589dcSJérôme Glisse return NULL; 13404ef589dcSJérôme Glisse lock_page(page); 13414ef589dcSJérôme Glisse return page; 13424ef589dcSJérôme Glisse } 13434ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_vma_alloc_locked_page); 13444ef589dcSJérôme Glisse 13454ef589dcSJérôme Glisse 13464ef589dcSJérôme Glisse static void hmm_devmem_ref_release(struct percpu_ref *ref) 13474ef589dcSJérôme Glisse { 13484ef589dcSJérôme Glisse struct hmm_devmem *devmem; 13494ef589dcSJérôme Glisse 13504ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 13514ef589dcSJérôme Glisse complete(&devmem->completion); 13524ef589dcSJérôme Glisse } 13534ef589dcSJérôme Glisse 13544ef589dcSJérôme Glisse static void hmm_devmem_ref_exit(void *data) 13554ef589dcSJérôme Glisse { 13564ef589dcSJérôme Glisse struct percpu_ref *ref = data; 13574ef589dcSJérôme Glisse struct hmm_devmem *devmem; 13584ef589dcSJérôme Glisse 13594ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 1360bbecd94eSDan Williams wait_for_completion(&devmem->completion); 13614ef589dcSJérôme Glisse percpu_ref_exit(ref); 13624ef589dcSJérôme Glisse } 13634ef589dcSJérôme Glisse 1364bbecd94eSDan Williams static void hmm_devmem_ref_kill(struct percpu_ref *ref) 13654ef589dcSJérôme Glisse { 13664ef589dcSJérôme Glisse percpu_ref_kill(ref); 13674ef589dcSJérôme Glisse } 13684ef589dcSJérôme Glisse 1369b57e622eSSouptick Joarder static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma, 13704ef589dcSJérôme Glisse unsigned long addr, 13714ef589dcSJérôme Glisse const struct page *page, 13724ef589dcSJérôme Glisse unsigned int flags, 13734ef589dcSJérôme Glisse pmd_t *pmdp) 13744ef589dcSJérôme Glisse { 13754ef589dcSJérôme Glisse struct hmm_devmem *devmem = page->pgmap->data; 13764ef589dcSJérôme Glisse 13774ef589dcSJérôme Glisse return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp); 13784ef589dcSJérôme Glisse } 13794ef589dcSJérôme Glisse 13804ef589dcSJérôme Glisse static void hmm_devmem_free(struct page *page, void *data) 13814ef589dcSJérôme Glisse { 13824ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 13834ef589dcSJérôme Glisse 13842fa147bdSDan Williams page->mapping = NULL; 13852fa147bdSDan Williams 13864ef589dcSJérôme Glisse devmem->ops->free(devmem, page); 13874ef589dcSJérôme Glisse } 13884ef589dcSJérôme Glisse 13894ef589dcSJérôme Glisse /* 13904ef589dcSJérôme Glisse * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory 13914ef589dcSJérôme Glisse * 13924ef589dcSJérôme Glisse * @ops: memory event device driver callback (see struct hmm_devmem_ops) 13934ef589dcSJérôme Glisse * @device: device struct to bind the resource too 13944ef589dcSJérôme Glisse * @size: size in bytes of the device memory to add 1395*085ea250SRalph Campbell * Return: pointer to new hmm_devmem struct ERR_PTR otherwise 13964ef589dcSJérôme Glisse * 13974ef589dcSJérôme Glisse * This function first finds an empty range of physical address big enough to 13984ef589dcSJérôme Glisse * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which 13994ef589dcSJérôme Glisse * in turn allocates struct pages. It does not do anything beyond that; all 14004ef589dcSJérôme Glisse * events affecting the memory will go through the various callbacks provided 14014ef589dcSJérôme Glisse * by hmm_devmem_ops struct. 14024ef589dcSJérôme Glisse * 14034ef589dcSJérôme Glisse * Device driver should call this function during device initialization and 14044ef589dcSJérôme Glisse * is then responsible of memory management. HMM only provides helpers. 14054ef589dcSJérôme Glisse */ 14064ef589dcSJérôme Glisse struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, 14074ef589dcSJérôme Glisse struct device *device, 14084ef589dcSJérôme Glisse unsigned long size) 14094ef589dcSJérôme Glisse { 14104ef589dcSJérôme Glisse struct hmm_devmem *devmem; 14114ef589dcSJérôme Glisse resource_size_t addr; 1412bbecd94eSDan Williams void *result; 14134ef589dcSJérôme Glisse int ret; 14144ef589dcSJérôme Glisse 1415e7638488SDan Williams dev_pagemap_get_ops(); 14164ef589dcSJérôme Glisse 141758ef15b7SDan Williams devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); 14184ef589dcSJérôme Glisse if (!devmem) 14194ef589dcSJérôme Glisse return ERR_PTR(-ENOMEM); 14204ef589dcSJérôme Glisse 14214ef589dcSJérôme Glisse init_completion(&devmem->completion); 14224ef589dcSJérôme Glisse devmem->pfn_first = -1UL; 14234ef589dcSJérôme Glisse devmem->pfn_last = -1UL; 14244ef589dcSJérôme Glisse devmem->resource = NULL; 14254ef589dcSJérôme Glisse devmem->device = device; 14264ef589dcSJérôme Glisse devmem->ops = ops; 14274ef589dcSJérôme Glisse 14284ef589dcSJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 14294ef589dcSJérôme Glisse 0, GFP_KERNEL); 14304ef589dcSJérôme Glisse if (ret) 143158ef15b7SDan Williams return ERR_PTR(ret); 14324ef589dcSJérôme Glisse 143358ef15b7SDan Williams ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref); 14344ef589dcSJérôme Glisse if (ret) 143558ef15b7SDan Williams return ERR_PTR(ret); 14364ef589dcSJérôme Glisse 14374ef589dcSJérôme Glisse size = ALIGN(size, PA_SECTION_SIZE); 14384ef589dcSJérôme Glisse addr = min((unsigned long)iomem_resource.end, 14394ef589dcSJérôme Glisse (1UL << MAX_PHYSMEM_BITS) - 1); 14404ef589dcSJérôme Glisse addr = addr - size + 1UL; 14414ef589dcSJérôme Glisse 14424ef589dcSJérôme Glisse /* 14434ef589dcSJérôme Glisse * FIXME add a new helper to quickly walk resource tree and find free 14444ef589dcSJérôme Glisse * range 14454ef589dcSJérôme Glisse * 14464ef589dcSJérôme Glisse * FIXME what about ioport_resource resource ? 14474ef589dcSJérôme Glisse */ 14484ef589dcSJérôme Glisse for (; addr > size && addr >= iomem_resource.start; addr -= size) { 14494ef589dcSJérôme Glisse ret = region_intersects(addr, size, 0, IORES_DESC_NONE); 14504ef589dcSJérôme Glisse if (ret != REGION_DISJOINT) 14514ef589dcSJérôme Glisse continue; 14524ef589dcSJérôme Glisse 14534ef589dcSJérôme Glisse devmem->resource = devm_request_mem_region(device, addr, size, 14544ef589dcSJérôme Glisse dev_name(device)); 145558ef15b7SDan Williams if (!devmem->resource) 145658ef15b7SDan Williams return ERR_PTR(-ENOMEM); 14574ef589dcSJérôme Glisse break; 14584ef589dcSJérôme Glisse } 145958ef15b7SDan Williams if (!devmem->resource) 146058ef15b7SDan Williams return ERR_PTR(-ERANGE); 14614ef589dcSJérôme Glisse 14624ef589dcSJérôme Glisse devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY; 14634ef589dcSJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 14644ef589dcSJérôme Glisse devmem->pfn_last = devmem->pfn_first + 14654ef589dcSJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1466063a7d1dSDan Williams devmem->page_fault = hmm_devmem_fault; 14674ef589dcSJérôme Glisse 1468bbecd94eSDan Williams devmem->pagemap.type = MEMORY_DEVICE_PRIVATE; 1469bbecd94eSDan Williams devmem->pagemap.res = *devmem->resource; 1470bbecd94eSDan Williams devmem->pagemap.page_free = hmm_devmem_free; 1471bbecd94eSDan Williams devmem->pagemap.altmap_valid = false; 1472bbecd94eSDan Williams devmem->pagemap.ref = &devmem->ref; 1473bbecd94eSDan Williams devmem->pagemap.data = devmem; 1474bbecd94eSDan Williams devmem->pagemap.kill = hmm_devmem_ref_kill; 147558ef15b7SDan Williams 1476bbecd94eSDan Williams result = devm_memremap_pages(devmem->device, &devmem->pagemap); 1477bbecd94eSDan Williams if (IS_ERR(result)) 1478bbecd94eSDan Williams return result; 14794ef589dcSJérôme Glisse return devmem; 14804ef589dcSJérôme Glisse } 148102917e9fSDan Williams EXPORT_SYMBOL_GPL(hmm_devmem_add); 14824ef589dcSJérôme Glisse 1483d3df0a42SJérôme Glisse struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops, 1484d3df0a42SJérôme Glisse struct device *device, 1485d3df0a42SJérôme Glisse struct resource *res) 1486d3df0a42SJérôme Glisse { 1487d3df0a42SJérôme Glisse struct hmm_devmem *devmem; 1488bbecd94eSDan Williams void *result; 1489d3df0a42SJérôme Glisse int ret; 1490d3df0a42SJérôme Glisse 1491d3df0a42SJérôme Glisse if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY) 1492d3df0a42SJérôme Glisse return ERR_PTR(-EINVAL); 1493d3df0a42SJérôme Glisse 1494e7638488SDan Williams dev_pagemap_get_ops(); 1495d3df0a42SJérôme Glisse 149658ef15b7SDan Williams devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); 1497d3df0a42SJérôme Glisse if (!devmem) 1498d3df0a42SJérôme Glisse return ERR_PTR(-ENOMEM); 1499d3df0a42SJérôme Glisse 1500d3df0a42SJérôme Glisse init_completion(&devmem->completion); 1501d3df0a42SJérôme Glisse devmem->pfn_first = -1UL; 1502d3df0a42SJérôme Glisse devmem->pfn_last = -1UL; 1503d3df0a42SJérôme Glisse devmem->resource = res; 1504d3df0a42SJérôme Glisse devmem->device = device; 1505d3df0a42SJérôme Glisse devmem->ops = ops; 1506d3df0a42SJérôme Glisse 1507d3df0a42SJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 1508d3df0a42SJérôme Glisse 0, GFP_KERNEL); 1509d3df0a42SJérôme Glisse if (ret) 151058ef15b7SDan Williams return ERR_PTR(ret); 1511d3df0a42SJérôme Glisse 151258ef15b7SDan Williams ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, 151358ef15b7SDan Williams &devmem->ref); 1514d3df0a42SJérôme Glisse if (ret) 151558ef15b7SDan Williams return ERR_PTR(ret); 1516d3df0a42SJérôme Glisse 1517d3df0a42SJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 1518d3df0a42SJérôme Glisse devmem->pfn_last = devmem->pfn_first + 1519d3df0a42SJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1520063a7d1dSDan Williams devmem->page_fault = hmm_devmem_fault; 1521d3df0a42SJérôme Glisse 1522bbecd94eSDan Williams devmem->pagemap.type = MEMORY_DEVICE_PUBLIC; 1523bbecd94eSDan Williams devmem->pagemap.res = *devmem->resource; 1524bbecd94eSDan Williams devmem->pagemap.page_free = hmm_devmem_free; 1525bbecd94eSDan Williams devmem->pagemap.altmap_valid = false; 1526bbecd94eSDan Williams devmem->pagemap.ref = &devmem->ref; 1527bbecd94eSDan Williams devmem->pagemap.data = devmem; 1528bbecd94eSDan Williams devmem->pagemap.kill = hmm_devmem_ref_kill; 152958ef15b7SDan Williams 1530bbecd94eSDan Williams result = devm_memremap_pages(devmem->device, &devmem->pagemap); 1531bbecd94eSDan Williams if (IS_ERR(result)) 1532bbecd94eSDan Williams return result; 1533d3df0a42SJérôme Glisse return devmem; 1534d3df0a42SJérôme Glisse } 153502917e9fSDan Williams EXPORT_SYMBOL_GPL(hmm_devmem_add_resource); 1536d3df0a42SJérôme Glisse 15374ef589dcSJérôme Glisse /* 1538858b54daSJérôme Glisse * A device driver that wants to handle multiple devices memory through a 1539858b54daSJérôme Glisse * single fake device can use hmm_device to do so. This is purely a helper 1540858b54daSJérôme Glisse * and it is not needed to make use of any HMM functionality. 1541858b54daSJérôme Glisse */ 1542858b54daSJérôme Glisse #define HMM_DEVICE_MAX 256 1543858b54daSJérôme Glisse 1544858b54daSJérôme Glisse static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX); 1545858b54daSJérôme Glisse static DEFINE_SPINLOCK(hmm_device_lock); 1546858b54daSJérôme Glisse static struct class *hmm_device_class; 1547858b54daSJérôme Glisse static dev_t hmm_device_devt; 1548858b54daSJérôme Glisse 1549858b54daSJérôme Glisse static void hmm_device_release(struct device *device) 1550858b54daSJérôme Glisse { 1551858b54daSJérôme Glisse struct hmm_device *hmm_device; 1552858b54daSJérôme Glisse 1553858b54daSJérôme Glisse hmm_device = container_of(device, struct hmm_device, device); 1554858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1555858b54daSJérôme Glisse clear_bit(hmm_device->minor, hmm_device_mask); 1556858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1557858b54daSJérôme Glisse 1558858b54daSJérôme Glisse kfree(hmm_device); 1559858b54daSJérôme Glisse } 1560858b54daSJérôme Glisse 1561858b54daSJérôme Glisse struct hmm_device *hmm_device_new(void *drvdata) 1562858b54daSJérôme Glisse { 1563858b54daSJérôme Glisse struct hmm_device *hmm_device; 1564858b54daSJérôme Glisse 1565858b54daSJérôme Glisse hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL); 1566858b54daSJérôme Glisse if (!hmm_device) 1567858b54daSJérôme Glisse return ERR_PTR(-ENOMEM); 1568858b54daSJérôme Glisse 1569858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1570858b54daSJérôme Glisse hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX); 1571858b54daSJérôme Glisse if (hmm_device->minor >= HMM_DEVICE_MAX) { 1572858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1573858b54daSJérôme Glisse kfree(hmm_device); 1574858b54daSJérôme Glisse return ERR_PTR(-EBUSY); 1575858b54daSJérôme Glisse } 1576858b54daSJérôme Glisse set_bit(hmm_device->minor, hmm_device_mask); 1577858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1578858b54daSJérôme Glisse 1579858b54daSJérôme Glisse dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor); 1580858b54daSJérôme Glisse hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt), 1581858b54daSJérôme Glisse hmm_device->minor); 1582858b54daSJérôme Glisse hmm_device->device.release = hmm_device_release; 1583858b54daSJérôme Glisse dev_set_drvdata(&hmm_device->device, drvdata); 1584858b54daSJérôme Glisse hmm_device->device.class = hmm_device_class; 1585858b54daSJérôme Glisse device_initialize(&hmm_device->device); 1586858b54daSJérôme Glisse 1587858b54daSJérôme Glisse return hmm_device; 1588858b54daSJérôme Glisse } 1589858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_new); 1590858b54daSJérôme Glisse 1591858b54daSJérôme Glisse void hmm_device_put(struct hmm_device *hmm_device) 1592858b54daSJérôme Glisse { 1593858b54daSJérôme Glisse put_device(&hmm_device->device); 1594858b54daSJérôme Glisse } 1595858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_put); 1596858b54daSJérôme Glisse 1597858b54daSJérôme Glisse static int __init hmm_init(void) 1598858b54daSJérôme Glisse { 1599858b54daSJérôme Glisse int ret; 1600858b54daSJérôme Glisse 1601858b54daSJérôme Glisse ret = alloc_chrdev_region(&hmm_device_devt, 0, 1602858b54daSJérôme Glisse HMM_DEVICE_MAX, 1603858b54daSJérôme Glisse "hmm_device"); 1604858b54daSJérôme Glisse if (ret) 1605858b54daSJérôme Glisse return ret; 1606858b54daSJérôme Glisse 1607858b54daSJérôme Glisse hmm_device_class = class_create(THIS_MODULE, "hmm_device"); 1608858b54daSJérôme Glisse if (IS_ERR(hmm_device_class)) { 1609858b54daSJérôme Glisse unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX); 1610858b54daSJérôme Glisse return PTR_ERR(hmm_device_class); 1611858b54daSJérôme Glisse } 1612858b54daSJérôme Glisse return 0; 1613858b54daSJérôme Glisse } 1614858b54daSJérôme Glisse 1615858b54daSJérôme Glisse device_initcall(hmm_init); 1616df6ad698SJérôme Glisse #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ 1617