1133ff0eaSJérôme Glisse /* 2133ff0eaSJérôme Glisse * Copyright 2013 Red Hat Inc. 3133ff0eaSJérôme Glisse * 4133ff0eaSJérôme Glisse * This program is free software; you can redistribute it and/or modify 5133ff0eaSJérôme Glisse * it under the terms of the GNU General Public License as published by 6133ff0eaSJérôme Glisse * the Free Software Foundation; either version 2 of the License, or 7133ff0eaSJérôme Glisse * (at your option) any later version. 8133ff0eaSJérôme Glisse * 9133ff0eaSJérôme Glisse * This program is distributed in the hope that it will be useful, 10133ff0eaSJérôme Glisse * but WITHOUT ANY WARRANTY; without even the implied warranty of 11133ff0eaSJérôme Glisse * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12133ff0eaSJérôme Glisse * GNU General Public License for more details. 13133ff0eaSJérôme Glisse * 14f813f219SJérôme Glisse * Authors: Jérôme Glisse <jglisse@redhat.com> 15133ff0eaSJérôme Glisse */ 16133ff0eaSJérôme Glisse /* 17133ff0eaSJérôme Glisse * Refer to include/linux/hmm.h for information about heterogeneous memory 18133ff0eaSJérôme Glisse * management or HMM for short. 19133ff0eaSJérôme Glisse */ 20133ff0eaSJérôme Glisse #include <linux/mm.h> 21133ff0eaSJérôme Glisse #include <linux/hmm.h> 22858b54daSJérôme Glisse #include <linux/init.h> 23da4c3c73SJérôme Glisse #include <linux/rmap.h> 24da4c3c73SJérôme Glisse #include <linux/swap.h> 25133ff0eaSJérôme Glisse #include <linux/slab.h> 26133ff0eaSJérôme Glisse #include <linux/sched.h> 274ef589dcSJérôme Glisse #include <linux/mmzone.h> 284ef589dcSJérôme Glisse #include <linux/pagemap.h> 29da4c3c73SJérôme Glisse #include <linux/swapops.h> 30da4c3c73SJérôme Glisse #include <linux/hugetlb.h> 314ef589dcSJérôme Glisse #include <linux/memremap.h> 327b2d55d2SJérôme Glisse #include <linux/jump_label.h> 3355c0ece8SJérôme Glisse #include <linux/dma-mapping.h> 34c0b12405SJérôme Glisse #include <linux/mmu_notifier.h> 354ef589dcSJérôme Glisse #include <linux/memory_hotplug.h> 364ef589dcSJérôme Glisse 374ef589dcSJérôme Glisse #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT) 38133ff0eaSJérôme Glisse 396b368cd4SJérôme Glisse #if IS_ENABLED(CONFIG_HMM_MIRROR) 40c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops; 41c0b12405SJérôme Glisse 42704f3f2cSJérôme Glisse static inline struct hmm *mm_get_hmm(struct mm_struct *mm) 43133ff0eaSJérôme Glisse { 44c0b12405SJérôme Glisse struct hmm *hmm = READ_ONCE(mm->hmm); 45704f3f2cSJérôme Glisse 46704f3f2cSJérôme Glisse if (hmm && kref_get_unless_zero(&hmm->kref)) 47704f3f2cSJérôme Glisse return hmm; 48704f3f2cSJérôme Glisse 49704f3f2cSJérôme Glisse return NULL; 50704f3f2cSJérôme Glisse } 51704f3f2cSJérôme Glisse 52704f3f2cSJérôme Glisse /** 53704f3f2cSJérôme Glisse * hmm_get_or_create - register HMM against an mm (HMM internal) 54704f3f2cSJérôme Glisse * 55704f3f2cSJérôme Glisse * @mm: mm struct to attach to 56704f3f2cSJérôme Glisse * Returns: returns an HMM object, either by referencing the existing 57704f3f2cSJérôme Glisse * (per-process) object, or by creating a new one. 58704f3f2cSJérôme Glisse * 59704f3f2cSJérôme Glisse * This is not intended to be used directly by device drivers. If mm already 60704f3f2cSJérôme Glisse * has an HMM struct then it get a reference on it and returns it. Otherwise 61704f3f2cSJérôme Glisse * it allocates an HMM struct, initializes it, associate it with the mm and 62704f3f2cSJérôme Glisse * returns it. 63704f3f2cSJérôme Glisse */ 64704f3f2cSJérôme Glisse static struct hmm *hmm_get_or_create(struct mm_struct *mm) 65704f3f2cSJérôme Glisse { 66704f3f2cSJérôme Glisse struct hmm *hmm = mm_get_hmm(mm); 67c0b12405SJérôme Glisse bool cleanup = false; 68133ff0eaSJérôme Glisse 69c0b12405SJérôme Glisse if (hmm) 70c0b12405SJérôme Glisse return hmm; 71c0b12405SJérôme Glisse 72c0b12405SJérôme Glisse hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); 73c0b12405SJérôme Glisse if (!hmm) 74c0b12405SJérôme Glisse return NULL; 75a3e0d41cSJérôme Glisse init_waitqueue_head(&hmm->wq); 76c0b12405SJérôme Glisse INIT_LIST_HEAD(&hmm->mirrors); 77c0b12405SJérôme Glisse init_rwsem(&hmm->mirrors_sem); 78c0b12405SJérôme Glisse hmm->mmu_notifier.ops = NULL; 79da4c3c73SJérôme Glisse INIT_LIST_HEAD(&hmm->ranges); 80a3e0d41cSJérôme Glisse mutex_init(&hmm->lock); 81704f3f2cSJérôme Glisse kref_init(&hmm->kref); 82a3e0d41cSJérôme Glisse hmm->notifiers = 0; 83a3e0d41cSJérôme Glisse hmm->dead = false; 84c0b12405SJérôme Glisse hmm->mm = mm; 85c0b12405SJérôme Glisse 86c0b12405SJérôme Glisse spin_lock(&mm->page_table_lock); 87c0b12405SJérôme Glisse if (!mm->hmm) 88c0b12405SJérôme Glisse mm->hmm = hmm; 89c0b12405SJérôme Glisse else 90c0b12405SJérôme Glisse cleanup = true; 91c0b12405SJérôme Glisse spin_unlock(&mm->page_table_lock); 92c0b12405SJérôme Glisse 9386a2d598SRalph Campbell if (cleanup) 9486a2d598SRalph Campbell goto error; 9586a2d598SRalph Campbell 9686a2d598SRalph Campbell /* 9786a2d598SRalph Campbell * We should only get here if hold the mmap_sem in write mode ie on 9886a2d598SRalph Campbell * registration of first mirror through hmm_mirror_register() 9986a2d598SRalph Campbell */ 10086a2d598SRalph Campbell hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; 10186a2d598SRalph Campbell if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) 10286a2d598SRalph Campbell goto error_mm; 103c0b12405SJérôme Glisse 104704f3f2cSJérôme Glisse return hmm; 10586a2d598SRalph Campbell 10686a2d598SRalph Campbell error_mm: 10786a2d598SRalph Campbell spin_lock(&mm->page_table_lock); 10886a2d598SRalph Campbell if (mm->hmm == hmm) 10986a2d598SRalph Campbell mm->hmm = NULL; 11086a2d598SRalph Campbell spin_unlock(&mm->page_table_lock); 11186a2d598SRalph Campbell error: 11286a2d598SRalph Campbell kfree(hmm); 11386a2d598SRalph Campbell return NULL; 114133ff0eaSJérôme Glisse } 115133ff0eaSJérôme Glisse 116704f3f2cSJérôme Glisse static void hmm_free(struct kref *kref) 117704f3f2cSJérôme Glisse { 118704f3f2cSJérôme Glisse struct hmm *hmm = container_of(kref, struct hmm, kref); 119704f3f2cSJérôme Glisse struct mm_struct *mm = hmm->mm; 120704f3f2cSJérôme Glisse 121704f3f2cSJérôme Glisse mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm); 122704f3f2cSJérôme Glisse 123704f3f2cSJérôme Glisse spin_lock(&mm->page_table_lock); 124704f3f2cSJérôme Glisse if (mm->hmm == hmm) 125704f3f2cSJérôme Glisse mm->hmm = NULL; 126704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 127704f3f2cSJérôme Glisse 128704f3f2cSJérôme Glisse kfree(hmm); 129704f3f2cSJérôme Glisse } 130704f3f2cSJérôme Glisse 131704f3f2cSJérôme Glisse static inline void hmm_put(struct hmm *hmm) 132704f3f2cSJérôme Glisse { 133704f3f2cSJérôme Glisse kref_put(&hmm->kref, hmm_free); 134704f3f2cSJérôme Glisse } 135704f3f2cSJérôme Glisse 136133ff0eaSJérôme Glisse void hmm_mm_destroy(struct mm_struct *mm) 137133ff0eaSJérôme Glisse { 138704f3f2cSJérôme Glisse struct hmm *hmm; 139704f3f2cSJérôme Glisse 140704f3f2cSJérôme Glisse spin_lock(&mm->page_table_lock); 141704f3f2cSJérôme Glisse hmm = mm_get_hmm(mm); 142704f3f2cSJérôme Glisse mm->hmm = NULL; 143704f3f2cSJérôme Glisse if (hmm) { 144704f3f2cSJérôme Glisse hmm->mm = NULL; 145a3e0d41cSJérôme Glisse hmm->dead = true; 146704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 147704f3f2cSJérôme Glisse hmm_put(hmm); 148704f3f2cSJérôme Glisse return; 149704f3f2cSJérôme Glisse } 150704f3f2cSJérôme Glisse 151704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 152133ff0eaSJérôme Glisse } 153c0b12405SJérôme Glisse 154a3e0d41cSJérôme Glisse static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm) 155c0b12405SJérôme Glisse { 156a3e0d41cSJérôme Glisse struct hmm *hmm = mm_get_hmm(mm); 157c0b12405SJérôme Glisse struct hmm_mirror *mirror; 158da4c3c73SJérôme Glisse struct hmm_range *range; 159da4c3c73SJérôme Glisse 160a3e0d41cSJérôme Glisse /* Report this HMM as dying. */ 161a3e0d41cSJérôme Glisse hmm->dead = true; 162da4c3c73SJérôme Glisse 163a3e0d41cSJérôme Glisse /* Wake-up everyone waiting on any range. */ 164a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 165a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 166da4c3c73SJérôme Glisse range->valid = false; 167da4c3c73SJérôme Glisse } 168a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 169a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 170e1401513SRalph Campbell 171e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 172e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror, 173e1401513SRalph Campbell list); 174e1401513SRalph Campbell while (mirror) { 175e1401513SRalph Campbell list_del_init(&mirror->list); 176e1401513SRalph Campbell if (mirror->ops->release) { 177e1401513SRalph Campbell /* 178e1401513SRalph Campbell * Drop mirrors_sem so callback can wait on any pending 179e1401513SRalph Campbell * work that might itself trigger mmu_notifier callback 180e1401513SRalph Campbell * and thus would deadlock with us. 181e1401513SRalph Campbell */ 182e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 183e1401513SRalph Campbell mirror->ops->release(mirror); 184e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 185e1401513SRalph Campbell } 186e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, 187e1401513SRalph Campbell struct hmm_mirror, list); 188e1401513SRalph Campbell } 189e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 190704f3f2cSJérôme Glisse 191704f3f2cSJérôme Glisse hmm_put(hmm); 192e1401513SRalph Campbell } 193e1401513SRalph Campbell 19493065ac7SMichal Hocko static int hmm_invalidate_range_start(struct mmu_notifier *mn, 195a3e0d41cSJérôme Glisse const struct mmu_notifier_range *nrange) 196c0b12405SJérôme Glisse { 197a3e0d41cSJérôme Glisse struct hmm *hmm = mm_get_hmm(nrange->mm); 198a3e0d41cSJérôme Glisse struct hmm_mirror *mirror; 199ec131b2dSJérôme Glisse struct hmm_update update; 200a3e0d41cSJérôme Glisse struct hmm_range *range; 201a3e0d41cSJérôme Glisse int ret = 0; 202c0b12405SJérôme Glisse 203c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 204c0b12405SJérôme Glisse 205a3e0d41cSJérôme Glisse update.start = nrange->start; 206a3e0d41cSJérôme Glisse update.end = nrange->end; 207ec131b2dSJérôme Glisse update.event = HMM_UPDATE_INVALIDATE; 208*dfcd6660SJérôme Glisse update.blockable = mmu_notifier_range_blockable(nrange); 209a3e0d41cSJérôme Glisse 210*dfcd6660SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 211a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 212a3e0d41cSJérôme Glisse else if (!mutex_trylock(&hmm->lock)) { 213a3e0d41cSJérôme Glisse ret = -EAGAIN; 214a3e0d41cSJérôme Glisse goto out; 215a3e0d41cSJérôme Glisse } 216a3e0d41cSJérôme Glisse hmm->notifiers++; 217a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 218a3e0d41cSJérôme Glisse if (update.end < range->start || update.start >= range->end) 219a3e0d41cSJérôme Glisse continue; 220a3e0d41cSJérôme Glisse 221a3e0d41cSJérôme Glisse range->valid = false; 222a3e0d41cSJérôme Glisse } 223a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 224a3e0d41cSJérôme Glisse 225*dfcd6660SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 226a3e0d41cSJérôme Glisse down_read(&hmm->mirrors_sem); 227a3e0d41cSJérôme Glisse else if (!down_read_trylock(&hmm->mirrors_sem)) { 228a3e0d41cSJérôme Glisse ret = -EAGAIN; 229a3e0d41cSJérôme Glisse goto out; 230a3e0d41cSJérôme Glisse } 231a3e0d41cSJérôme Glisse list_for_each_entry(mirror, &hmm->mirrors, list) { 232a3e0d41cSJérôme Glisse int ret; 233a3e0d41cSJérôme Glisse 234a3e0d41cSJérôme Glisse ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update); 235a3e0d41cSJérôme Glisse if (!update.blockable && ret == -EAGAIN) { 236a3e0d41cSJérôme Glisse up_read(&hmm->mirrors_sem); 237a3e0d41cSJérôme Glisse ret = -EAGAIN; 238a3e0d41cSJérôme Glisse goto out; 239a3e0d41cSJérôme Glisse } 240a3e0d41cSJérôme Glisse } 241a3e0d41cSJérôme Glisse up_read(&hmm->mirrors_sem); 242a3e0d41cSJérôme Glisse 243a3e0d41cSJérôme Glisse out: 244704f3f2cSJérôme Glisse hmm_put(hmm); 245704f3f2cSJérôme Glisse return ret; 246c0b12405SJérôme Glisse } 247c0b12405SJérôme Glisse 248c0b12405SJérôme Glisse static void hmm_invalidate_range_end(struct mmu_notifier *mn, 249a3e0d41cSJérôme Glisse const struct mmu_notifier_range *nrange) 250c0b12405SJérôme Glisse { 251a3e0d41cSJérôme Glisse struct hmm *hmm = mm_get_hmm(nrange->mm); 252c0b12405SJérôme Glisse 253c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 254c0b12405SJérôme Glisse 255a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 256a3e0d41cSJérôme Glisse hmm->notifiers--; 257a3e0d41cSJérôme Glisse if (!hmm->notifiers) { 258a3e0d41cSJérôme Glisse struct hmm_range *range; 259a3e0d41cSJérôme Glisse 260a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 261a3e0d41cSJérôme Glisse if (range->valid) 262a3e0d41cSJérôme Glisse continue; 263a3e0d41cSJérôme Glisse range->valid = true; 264a3e0d41cSJérôme Glisse } 265a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 266a3e0d41cSJérôme Glisse } 267a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 268a3e0d41cSJérôme Glisse 269704f3f2cSJérôme Glisse hmm_put(hmm); 270c0b12405SJérôme Glisse } 271c0b12405SJérôme Glisse 272c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { 273e1401513SRalph Campbell .release = hmm_release, 274c0b12405SJérôme Glisse .invalidate_range_start = hmm_invalidate_range_start, 275c0b12405SJérôme Glisse .invalidate_range_end = hmm_invalidate_range_end, 276c0b12405SJérôme Glisse }; 277c0b12405SJérôme Glisse 278c0b12405SJérôme Glisse /* 279c0b12405SJérôme Glisse * hmm_mirror_register() - register a mirror against an mm 280c0b12405SJérôme Glisse * 281c0b12405SJérôme Glisse * @mirror: new mirror struct to register 282c0b12405SJérôme Glisse * @mm: mm to register against 283c0b12405SJérôme Glisse * 284c0b12405SJérôme Glisse * To start mirroring a process address space, the device driver must register 285c0b12405SJérôme Glisse * an HMM mirror struct. 286c0b12405SJérôme Glisse * 287c0b12405SJérôme Glisse * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! 288c0b12405SJérôme Glisse */ 289c0b12405SJérôme Glisse int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) 290c0b12405SJérôme Glisse { 291c0b12405SJérôme Glisse /* Sanity check */ 292c0b12405SJérôme Glisse if (!mm || !mirror || !mirror->ops) 293c0b12405SJérôme Glisse return -EINVAL; 294c0b12405SJérôme Glisse 295704f3f2cSJérôme Glisse mirror->hmm = hmm_get_or_create(mm); 296c0b12405SJérôme Glisse if (!mirror->hmm) 297c0b12405SJérôme Glisse return -ENOMEM; 298c0b12405SJérôme Glisse 299c0b12405SJérôme Glisse down_write(&mirror->hmm->mirrors_sem); 300c0b12405SJérôme Glisse list_add(&mirror->list, &mirror->hmm->mirrors); 301c0b12405SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 302c0b12405SJérôme Glisse 303c0b12405SJérôme Glisse return 0; 304c0b12405SJérôme Glisse } 305c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_register); 306c0b12405SJérôme Glisse 307c0b12405SJérôme Glisse /* 308c0b12405SJérôme Glisse * hmm_mirror_unregister() - unregister a mirror 309c0b12405SJérôme Glisse * 310c0b12405SJérôme Glisse * @mirror: new mirror struct to register 311c0b12405SJérôme Glisse * 312c0b12405SJérôme Glisse * Stop mirroring a process address space, and cleanup. 313c0b12405SJérôme Glisse */ 314c0b12405SJérôme Glisse void hmm_mirror_unregister(struct hmm_mirror *mirror) 315c0b12405SJérôme Glisse { 316704f3f2cSJérôme Glisse struct hmm *hmm = READ_ONCE(mirror->hmm); 317c0b12405SJérôme Glisse 318704f3f2cSJérôme Glisse if (hmm == NULL) 319c01cbba2SJérôme Glisse return; 320c01cbba2SJérôme Glisse 321c0b12405SJérôme Glisse down_write(&hmm->mirrors_sem); 322e1401513SRalph Campbell list_del_init(&mirror->list); 323704f3f2cSJérôme Glisse /* To protect us against double unregister ... */ 324c01cbba2SJérôme Glisse mirror->hmm = NULL; 325c0b12405SJérôme Glisse up_write(&hmm->mirrors_sem); 326c01cbba2SJérôme Glisse 327704f3f2cSJérôme Glisse hmm_put(hmm); 328c0b12405SJérôme Glisse } 329c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_unregister); 330da4c3c73SJérôme Glisse 33174eee180SJérôme Glisse struct hmm_vma_walk { 33274eee180SJérôme Glisse struct hmm_range *range; 333992de9a8SJérôme Glisse struct dev_pagemap *pgmap; 33474eee180SJérôme Glisse unsigned long last; 33574eee180SJérôme Glisse bool fault; 33674eee180SJérôme Glisse bool block; 33774eee180SJérôme Glisse }; 33874eee180SJérôme Glisse 3392aee09d8SJérôme Glisse static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr, 3402aee09d8SJérôme Glisse bool write_fault, uint64_t *pfn) 34174eee180SJérôme Glisse { 34274eee180SJérôme Glisse unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE; 34374eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 344f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 34574eee180SJérôme Glisse struct vm_area_struct *vma = walk->vma; 34650a7ca3cSSouptick Joarder vm_fault_t ret; 34774eee180SJérôme Glisse 34874eee180SJérôme Glisse flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; 3492aee09d8SJérôme Glisse flags |= write_fault ? FAULT_FLAG_WRITE : 0; 35050a7ca3cSSouptick Joarder ret = handle_mm_fault(vma, addr, flags); 35150a7ca3cSSouptick Joarder if (ret & VM_FAULT_RETRY) 35273231612SJérôme Glisse return -EAGAIN; 35350a7ca3cSSouptick Joarder if (ret & VM_FAULT_ERROR) { 354f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 35574eee180SJérôme Glisse return -EFAULT; 35674eee180SJérôme Glisse } 35774eee180SJérôme Glisse 35873231612SJérôme Glisse return -EBUSY; 35974eee180SJérôme Glisse } 36074eee180SJérôme Glisse 361da4c3c73SJérôme Glisse static int hmm_pfns_bad(unsigned long addr, 362da4c3c73SJérôme Glisse unsigned long end, 363da4c3c73SJérôme Glisse struct mm_walk *walk) 364da4c3c73SJérôme Glisse { 365c719547fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 366c719547fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 367ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 368da4c3c73SJérôme Glisse unsigned long i; 369da4c3c73SJérôme Glisse 370da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 371da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) 372f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_ERROR]; 373da4c3c73SJérôme Glisse 374da4c3c73SJérôme Glisse return 0; 375da4c3c73SJérôme Glisse } 376da4c3c73SJérôme Glisse 3775504ed29SJérôme Glisse /* 3785504ed29SJérôme Glisse * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s) 3795504ed29SJérôme Glisse * @start: range virtual start address (inclusive) 3805504ed29SJérôme Glisse * @end: range virtual end address (exclusive) 3812aee09d8SJérôme Glisse * @fault: should we fault or not ? 3822aee09d8SJérôme Glisse * @write_fault: write fault ? 3835504ed29SJérôme Glisse * @walk: mm_walk structure 38473231612SJérôme Glisse * Returns: 0 on success, -EBUSY after page fault, or page fault error 3855504ed29SJérôme Glisse * 3865504ed29SJérôme Glisse * This function will be called whenever pmd_none() or pte_none() returns true, 3875504ed29SJérôme Glisse * or whenever there is no page directory covering the virtual address range. 3885504ed29SJérôme Glisse */ 3892aee09d8SJérôme Glisse static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end, 3902aee09d8SJérôme Glisse bool fault, bool write_fault, 391da4c3c73SJérôme Glisse struct mm_walk *walk) 392da4c3c73SJérôme Glisse { 39374eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 39474eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 395ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 39663d5066fSJérôme Glisse unsigned long i, page_size; 397da4c3c73SJérôme Glisse 39874eee180SJérôme Glisse hmm_vma_walk->last = addr; 39963d5066fSJérôme Glisse page_size = hmm_range_page_size(range); 40063d5066fSJérôme Glisse i = (addr - range->start) >> range->page_shift; 40163d5066fSJérôme Glisse 40263d5066fSJérôme Glisse for (; addr < end; addr += page_size, i++) { 403f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_NONE]; 4042aee09d8SJérôme Glisse if (fault || write_fault) { 40574eee180SJérôme Glisse int ret; 406da4c3c73SJérôme Glisse 4072aee09d8SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, write_fault, 4082aee09d8SJérôme Glisse &pfns[i]); 40973231612SJérôme Glisse if (ret != -EBUSY) 41074eee180SJérôme Glisse return ret; 41174eee180SJérôme Glisse } 41274eee180SJérôme Glisse } 41374eee180SJérôme Glisse 41473231612SJérôme Glisse return (fault || write_fault) ? -EBUSY : 0; 4152aee09d8SJérôme Glisse } 4162aee09d8SJérôme Glisse 4172aee09d8SJérôme Glisse static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4182aee09d8SJérôme Glisse uint64_t pfns, uint64_t cpu_flags, 4192aee09d8SJérôme Glisse bool *fault, bool *write_fault) 4202aee09d8SJérôme Glisse { 421f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 422f88a1e90SJérôme Glisse 4232aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) 4242aee09d8SJérôme Glisse return; 4252aee09d8SJérôme Glisse 426023a019aSJérôme Glisse /* 427023a019aSJérôme Glisse * So we not only consider the individual per page request we also 428023a019aSJérôme Glisse * consider the default flags requested for the range. The API can 429023a019aSJérôme Glisse * be use in 2 fashions. The first one where the HMM user coalesce 430023a019aSJérôme Glisse * multiple page fault into one request and set flags per pfns for 431023a019aSJérôme Glisse * of those faults. The second one where the HMM user want to pre- 432023a019aSJérôme Glisse * fault a range with specific flags. For the latter one it is a 433023a019aSJérôme Glisse * waste to have the user pre-fill the pfn arrays with a default 434023a019aSJérôme Glisse * flags value. 435023a019aSJérôme Glisse */ 436023a019aSJérôme Glisse pfns = (pfns & range->pfn_flags_mask) | range->default_flags; 437023a019aSJérôme Glisse 4382aee09d8SJérôme Glisse /* We aren't ask to do anything ... */ 439f88a1e90SJérôme Glisse if (!(pfns & range->flags[HMM_PFN_VALID])) 4402aee09d8SJérôme Glisse return; 441f88a1e90SJérôme Glisse /* If this is device memory than only fault if explicitly requested */ 442f88a1e90SJérôme Glisse if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) { 443f88a1e90SJérôme Glisse /* Do we fault on device memory ? */ 444f88a1e90SJérôme Glisse if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) { 445f88a1e90SJérôme Glisse *write_fault = pfns & range->flags[HMM_PFN_WRITE]; 446f88a1e90SJérôme Glisse *fault = true; 447f88a1e90SJérôme Glisse } 4482aee09d8SJérôme Glisse return; 4492aee09d8SJérôme Glisse } 450f88a1e90SJérôme Glisse 451f88a1e90SJérôme Glisse /* If CPU page table is not valid then we need to fault */ 452f88a1e90SJérôme Glisse *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]); 453f88a1e90SJérôme Glisse /* Need to write fault ? */ 454f88a1e90SJérôme Glisse if ((pfns & range->flags[HMM_PFN_WRITE]) && 455f88a1e90SJérôme Glisse !(cpu_flags & range->flags[HMM_PFN_WRITE])) { 456f88a1e90SJérôme Glisse *write_fault = true; 4572aee09d8SJérôme Glisse *fault = true; 4582aee09d8SJérôme Glisse } 4592aee09d8SJérôme Glisse } 4602aee09d8SJérôme Glisse 4612aee09d8SJérôme Glisse static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4622aee09d8SJérôme Glisse const uint64_t *pfns, unsigned long npages, 4632aee09d8SJérôme Glisse uint64_t cpu_flags, bool *fault, 4642aee09d8SJérôme Glisse bool *write_fault) 4652aee09d8SJérôme Glisse { 4662aee09d8SJérôme Glisse unsigned long i; 4672aee09d8SJérôme Glisse 4682aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) { 4692aee09d8SJérôme Glisse *fault = *write_fault = false; 4702aee09d8SJérôme Glisse return; 4712aee09d8SJérôme Glisse } 4722aee09d8SJérôme Glisse 473a3e0d41cSJérôme Glisse *fault = *write_fault = false; 4742aee09d8SJérôme Glisse for (i = 0; i < npages; ++i) { 4752aee09d8SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags, 4762aee09d8SJérôme Glisse fault, write_fault); 477a3e0d41cSJérôme Glisse if ((*write_fault)) 4782aee09d8SJérôme Glisse return; 4792aee09d8SJérôme Glisse } 4802aee09d8SJérôme Glisse } 4812aee09d8SJérôme Glisse 4822aee09d8SJérôme Glisse static int hmm_vma_walk_hole(unsigned long addr, unsigned long end, 4832aee09d8SJérôme Glisse struct mm_walk *walk) 4842aee09d8SJérôme Glisse { 4852aee09d8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 4862aee09d8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4872aee09d8SJérôme Glisse bool fault, write_fault; 4882aee09d8SJérôme Glisse unsigned long i, npages; 4892aee09d8SJérôme Glisse uint64_t *pfns; 4902aee09d8SJérôme Glisse 4912aee09d8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 4922aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 4932aee09d8SJérôme Glisse pfns = &range->pfns[i]; 4942aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 4952aee09d8SJérôme Glisse 0, &fault, &write_fault); 4962aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 4972aee09d8SJérôme Glisse } 4982aee09d8SJérôme Glisse 499f88a1e90SJérôme Glisse static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd) 5002aee09d8SJérôme Glisse { 5012aee09d8SJérôme Glisse if (pmd_protnone(pmd)) 5022aee09d8SJérôme Glisse return 0; 503f88a1e90SJérôme Glisse return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] | 504f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 505f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 506da4c3c73SJérôme Glisse } 507da4c3c73SJérôme Glisse 508992de9a8SJérôme Glisse static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud) 509992de9a8SJérôme Glisse { 510992de9a8SJérôme Glisse if (!pud_present(pud)) 511992de9a8SJérôme Glisse return 0; 512992de9a8SJérôme Glisse return pud_write(pud) ? range->flags[HMM_PFN_VALID] | 513992de9a8SJérôme Glisse range->flags[HMM_PFN_WRITE] : 514992de9a8SJérôme Glisse range->flags[HMM_PFN_VALID]; 515992de9a8SJérôme Glisse } 516992de9a8SJérôme Glisse 51753f5c3f4SJérôme Glisse static int hmm_vma_handle_pmd(struct mm_walk *walk, 51853f5c3f4SJérôme Glisse unsigned long addr, 51953f5c3f4SJérôme Glisse unsigned long end, 52053f5c3f4SJérôme Glisse uint64_t *pfns, 52153f5c3f4SJérôme Glisse pmd_t pmd) 52253f5c3f4SJérôme Glisse { 523992de9a8SJérôme Glisse #ifdef CONFIG_TRANSPARENT_HUGEPAGE 52453f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 525f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 5262aee09d8SJérôme Glisse unsigned long pfn, npages, i; 5272aee09d8SJérôme Glisse bool fault, write_fault; 528f88a1e90SJérôme Glisse uint64_t cpu_flags; 52953f5c3f4SJérôme Glisse 5302aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 531f88a1e90SJérôme Glisse cpu_flags = pmd_to_hmm_pfn_flags(range, pmd); 5322aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags, 5332aee09d8SJérôme Glisse &fault, &write_fault); 53453f5c3f4SJérôme Glisse 5352aee09d8SJérôme Glisse if (pmd_protnone(pmd) || fault || write_fault) 5362aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 53753f5c3f4SJérôme Glisse 53853f5c3f4SJérôme Glisse pfn = pmd_pfn(pmd) + pte_index(addr); 539992de9a8SJérôme Glisse for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) { 540992de9a8SJérôme Glisse if (pmd_devmap(pmd)) { 541992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 542992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 543992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 544992de9a8SJérôme Glisse return -EBUSY; 545992de9a8SJérôme Glisse } 546391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags; 547992de9a8SJérôme Glisse } 548992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 549992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 550992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 551992de9a8SJérôme Glisse } 55253f5c3f4SJérôme Glisse hmm_vma_walk->last = end; 55353f5c3f4SJérôme Glisse return 0; 554992de9a8SJérôme Glisse #else 555992de9a8SJérôme Glisse /* If THP is not enabled then we should never reach that code ! */ 556992de9a8SJérôme Glisse return -EINVAL; 557992de9a8SJérôme Glisse #endif 55853f5c3f4SJérôme Glisse } 55953f5c3f4SJérôme Glisse 560f88a1e90SJérôme Glisse static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte) 5612aee09d8SJérôme Glisse { 5622aee09d8SJérôme Glisse if (pte_none(pte) || !pte_present(pte)) 5632aee09d8SJérôme Glisse return 0; 564f88a1e90SJérôme Glisse return pte_write(pte) ? range->flags[HMM_PFN_VALID] | 565f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 566f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 5672aee09d8SJérôme Glisse } 5682aee09d8SJérôme Glisse 56953f5c3f4SJérôme Glisse static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr, 57053f5c3f4SJérôme Glisse unsigned long end, pmd_t *pmdp, pte_t *ptep, 57153f5c3f4SJérôme Glisse uint64_t *pfn) 57253f5c3f4SJérôme Glisse { 57353f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 574f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 57553f5c3f4SJérôme Glisse struct vm_area_struct *vma = walk->vma; 5762aee09d8SJérôme Glisse bool fault, write_fault; 5772aee09d8SJérôme Glisse uint64_t cpu_flags; 57853f5c3f4SJérôme Glisse pte_t pte = *ptep; 579f88a1e90SJérôme Glisse uint64_t orig_pfn = *pfn; 58053f5c3f4SJérôme Glisse 581f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_NONE]; 58273231612SJérôme Glisse fault = write_fault = false; 58353f5c3f4SJérôme Glisse 58453f5c3f4SJérôme Glisse if (pte_none(pte)) { 58573231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0, 58673231612SJérôme Glisse &fault, &write_fault); 5872aee09d8SJérôme Glisse if (fault || write_fault) 58853f5c3f4SJérôme Glisse goto fault; 58953f5c3f4SJérôme Glisse return 0; 59053f5c3f4SJérôme Glisse } 59153f5c3f4SJérôme Glisse 59253f5c3f4SJérôme Glisse if (!pte_present(pte)) { 59353f5c3f4SJérôme Glisse swp_entry_t entry = pte_to_swp_entry(pte); 59453f5c3f4SJérôme Glisse 59553f5c3f4SJérôme Glisse if (!non_swap_entry(entry)) { 5962aee09d8SJérôme Glisse if (fault || write_fault) 59753f5c3f4SJérôme Glisse goto fault; 59853f5c3f4SJérôme Glisse return 0; 59953f5c3f4SJérôme Glisse } 60053f5c3f4SJérôme Glisse 60153f5c3f4SJérôme Glisse /* 60253f5c3f4SJérôme Glisse * This is a special swap entry, ignore migration, use 60353f5c3f4SJérôme Glisse * device and report anything else as error. 60453f5c3f4SJérôme Glisse */ 60553f5c3f4SJérôme Glisse if (is_device_private_entry(entry)) { 606f88a1e90SJérôme Glisse cpu_flags = range->flags[HMM_PFN_VALID] | 607f88a1e90SJérôme Glisse range->flags[HMM_PFN_DEVICE_PRIVATE]; 6082aee09d8SJérôme Glisse cpu_flags |= is_write_device_private_entry(entry) ? 609f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 0; 610f88a1e90SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 611f88a1e90SJérôme Glisse &fault, &write_fault); 612f88a1e90SJérôme Glisse if (fault || write_fault) 613f88a1e90SJérôme Glisse goto fault; 614391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, 615391aab11SJérôme Glisse swp_offset(entry)); 616f88a1e90SJérôme Glisse *pfn |= cpu_flags; 61753f5c3f4SJérôme Glisse return 0; 61853f5c3f4SJérôme Glisse } 61953f5c3f4SJérôme Glisse 62053f5c3f4SJérôme Glisse if (is_migration_entry(entry)) { 6212aee09d8SJérôme Glisse if (fault || write_fault) { 62253f5c3f4SJérôme Glisse pte_unmap(ptep); 62353f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 62453f5c3f4SJérôme Glisse migration_entry_wait(vma->vm_mm, 62553f5c3f4SJérôme Glisse pmdp, addr); 62673231612SJérôme Glisse return -EBUSY; 62753f5c3f4SJérôme Glisse } 62853f5c3f4SJérôme Glisse return 0; 62953f5c3f4SJérôme Glisse } 63053f5c3f4SJérôme Glisse 63153f5c3f4SJérôme Glisse /* Report error for everything else */ 632f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 63353f5c3f4SJérôme Glisse return -EFAULT; 63473231612SJérôme Glisse } else { 63573231612SJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, pte); 63673231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 63773231612SJérôme Glisse &fault, &write_fault); 63853f5c3f4SJérôme Glisse } 63953f5c3f4SJérôme Glisse 6402aee09d8SJérôme Glisse if (fault || write_fault) 64153f5c3f4SJérôme Glisse goto fault; 64253f5c3f4SJérôme Glisse 643992de9a8SJérôme Glisse if (pte_devmap(pte)) { 644992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte), 645992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 646992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 647992de9a8SJérôme Glisse return -EBUSY; 648992de9a8SJérôme Glisse } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) { 649992de9a8SJérôme Glisse *pfn = range->values[HMM_PFN_SPECIAL]; 650992de9a8SJérôme Glisse return -EFAULT; 651992de9a8SJérôme Glisse } 652992de9a8SJérôme Glisse 653391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags; 65453f5c3f4SJérôme Glisse return 0; 65553f5c3f4SJérôme Glisse 65653f5c3f4SJérôme Glisse fault: 657992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 658992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 659992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 660992de9a8SJérôme Glisse } 66153f5c3f4SJérôme Glisse pte_unmap(ptep); 66253f5c3f4SJérôme Glisse /* Fault any virtual address we were asked to fault */ 6632aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 66453f5c3f4SJérôme Glisse } 66553f5c3f4SJérôme Glisse 666da4c3c73SJérôme Glisse static int hmm_vma_walk_pmd(pmd_t *pmdp, 667da4c3c73SJérôme Glisse unsigned long start, 668da4c3c73SJérôme Glisse unsigned long end, 669da4c3c73SJérôme Glisse struct mm_walk *walk) 670da4c3c73SJérôme Glisse { 67174eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 67274eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 673d08faca0SJérôme Glisse struct vm_area_struct *vma = walk->vma; 674ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 675da4c3c73SJérôme Glisse unsigned long addr = start, i; 676da4c3c73SJérôme Glisse pte_t *ptep; 677da4c3c73SJérôme Glisse pmd_t pmd; 678da4c3c73SJérôme Glisse 679d08faca0SJérôme Glisse 680d08faca0SJérôme Glisse again: 681d08faca0SJérôme Glisse pmd = READ_ONCE(*pmdp); 682d08faca0SJérôme Glisse if (pmd_none(pmd)) 683d08faca0SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 684d08faca0SJérôme Glisse 685d08faca0SJérôme Glisse if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB)) 686d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 687d08faca0SJérôme Glisse 688d08faca0SJérôme Glisse if (thp_migration_supported() && is_pmd_migration_entry(pmd)) { 689d08faca0SJérôme Glisse bool fault, write_fault; 690d08faca0SJérôme Glisse unsigned long npages; 691d08faca0SJérôme Glisse uint64_t *pfns; 692d08faca0SJérôme Glisse 693d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 694d08faca0SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 695d08faca0SJérôme Glisse pfns = &range->pfns[i]; 696d08faca0SJérôme Glisse 697d08faca0SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 698d08faca0SJérôme Glisse 0, &fault, &write_fault); 699d08faca0SJérôme Glisse if (fault || write_fault) { 700d08faca0SJérôme Glisse hmm_vma_walk->last = addr; 701d08faca0SJérôme Glisse pmd_migration_entry_wait(vma->vm_mm, pmdp); 70273231612SJérôme Glisse return -EBUSY; 703d08faca0SJérôme Glisse } 704d08faca0SJérôme Glisse return 0; 705d08faca0SJérôme Glisse } else if (!pmd_present(pmd)) 706d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 707d08faca0SJérôme Glisse 708d08faca0SJérôme Glisse if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) { 709da4c3c73SJérôme Glisse /* 710da4c3c73SJérôme Glisse * No need to take pmd_lock here, even if some other threads 711da4c3c73SJérôme Glisse * is splitting the huge pmd we will get that event through 712da4c3c73SJérôme Glisse * mmu_notifier callback. 713da4c3c73SJérôme Glisse * 714da4c3c73SJérôme Glisse * So just read pmd value and check again its a transparent 715da4c3c73SJérôme Glisse * huge or device mapping one and compute corresponding pfn 716da4c3c73SJérôme Glisse * values. 717da4c3c73SJérôme Glisse */ 718da4c3c73SJérôme Glisse pmd = pmd_read_atomic(pmdp); 719da4c3c73SJérôme Glisse barrier(); 720da4c3c73SJérôme Glisse if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) 721da4c3c73SJérôme Glisse goto again; 722da4c3c73SJérôme Glisse 723d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 72453f5c3f4SJérôme Glisse return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd); 725da4c3c73SJérôme Glisse } 726da4c3c73SJérôme Glisse 727d08faca0SJérôme Glisse /* 728d08faca0SJérôme Glisse * We have handled all the valid case above ie either none, migration, 729d08faca0SJérôme Glisse * huge or transparent huge. At this point either it is a valid pmd 730d08faca0SJérôme Glisse * entry pointing to pte directory or it is a bad pmd that will not 731d08faca0SJérôme Glisse * recover. 732d08faca0SJérôme Glisse */ 733d08faca0SJérôme Glisse if (pmd_bad(pmd)) 734da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 735da4c3c73SJérôme Glisse 736da4c3c73SJérôme Glisse ptep = pte_offset_map(pmdp, addr); 737d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 738da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { 73953f5c3f4SJérôme Glisse int r; 740da4c3c73SJérôme Glisse 74153f5c3f4SJérôme Glisse r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]); 74253f5c3f4SJérôme Glisse if (r) { 74353f5c3f4SJérôme Glisse /* hmm_vma_handle_pte() did unmap pte directory */ 74474eee180SJérôme Glisse hmm_vma_walk->last = addr; 74553f5c3f4SJérôme Glisse return r; 74674eee180SJérôme Glisse } 747da4c3c73SJérôme Glisse } 748992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 749992de9a8SJérôme Glisse /* 750992de9a8SJérôme Glisse * We do put_dev_pagemap() here and not in hmm_vma_handle_pte() 751992de9a8SJérôme Glisse * so that we can leverage get_dev_pagemap() optimization which 752992de9a8SJérôme Glisse * will not re-take a reference on a pgmap if we already have 753992de9a8SJérôme Glisse * one. 754992de9a8SJérôme Glisse */ 755992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 756992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 757992de9a8SJérôme Glisse } 758da4c3c73SJérôme Glisse pte_unmap(ptep - 1); 759da4c3c73SJérôme Glisse 76053f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 761da4c3c73SJérôme Glisse return 0; 762da4c3c73SJérôme Glisse } 763da4c3c73SJérôme Glisse 764992de9a8SJérôme Glisse static int hmm_vma_walk_pud(pud_t *pudp, 765992de9a8SJérôme Glisse unsigned long start, 766992de9a8SJérôme Glisse unsigned long end, 767992de9a8SJérôme Glisse struct mm_walk *walk) 768992de9a8SJérôme Glisse { 769992de9a8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 770992de9a8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 771992de9a8SJérôme Glisse unsigned long addr = start, next; 772992de9a8SJérôme Glisse pmd_t *pmdp; 773992de9a8SJérôme Glisse pud_t pud; 774992de9a8SJérôme Glisse int ret; 775992de9a8SJérôme Glisse 776992de9a8SJérôme Glisse again: 777992de9a8SJérôme Glisse pud = READ_ONCE(*pudp); 778992de9a8SJérôme Glisse if (pud_none(pud)) 779992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 780992de9a8SJérôme Glisse 781992de9a8SJérôme Glisse if (pud_huge(pud) && pud_devmap(pud)) { 782992de9a8SJérôme Glisse unsigned long i, npages, pfn; 783992de9a8SJérôme Glisse uint64_t *pfns, cpu_flags; 784992de9a8SJérôme Glisse bool fault, write_fault; 785992de9a8SJérôme Glisse 786992de9a8SJérôme Glisse if (!pud_present(pud)) 787992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 788992de9a8SJérôme Glisse 789992de9a8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 790992de9a8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 791992de9a8SJérôme Glisse pfns = &range->pfns[i]; 792992de9a8SJérôme Glisse 793992de9a8SJérôme Glisse cpu_flags = pud_to_hmm_pfn_flags(range, pud); 794992de9a8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 795992de9a8SJérôme Glisse cpu_flags, &fault, &write_fault); 796992de9a8SJérôme Glisse if (fault || write_fault) 797992de9a8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, 798992de9a8SJérôme Glisse write_fault, walk); 799992de9a8SJérôme Glisse 800992de9a8SJérôme Glisse #ifdef CONFIG_HUGETLB_PAGE 801992de9a8SJérôme Glisse pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); 802992de9a8SJérôme Glisse for (i = 0; i < npages; ++i, ++pfn) { 803992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 804992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 805992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 806992de9a8SJérôme Glisse return -EBUSY; 807391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 808391aab11SJérôme Glisse cpu_flags; 809992de9a8SJérôme Glisse } 810992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 811992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 812992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 813992de9a8SJérôme Glisse } 814992de9a8SJérôme Glisse hmm_vma_walk->last = end; 815992de9a8SJérôme Glisse return 0; 816992de9a8SJérôme Glisse #else 817992de9a8SJérôme Glisse return -EINVAL; 818992de9a8SJérôme Glisse #endif 819992de9a8SJérôme Glisse } 820992de9a8SJérôme Glisse 821992de9a8SJérôme Glisse split_huge_pud(walk->vma, pudp, addr); 822992de9a8SJérôme Glisse if (pud_none(*pudp)) 823992de9a8SJérôme Glisse goto again; 824992de9a8SJérôme Glisse 825992de9a8SJérôme Glisse pmdp = pmd_offset(pudp, addr); 826992de9a8SJérôme Glisse do { 827992de9a8SJérôme Glisse next = pmd_addr_end(addr, end); 828992de9a8SJérôme Glisse ret = hmm_vma_walk_pmd(pmdp, addr, next, walk); 829992de9a8SJérôme Glisse if (ret) 830992de9a8SJérôme Glisse return ret; 831992de9a8SJérôme Glisse } while (pmdp++, addr = next, addr != end); 832992de9a8SJérôme Glisse 833992de9a8SJérôme Glisse return 0; 834992de9a8SJérôme Glisse } 835992de9a8SJérôme Glisse 83663d5066fSJérôme Glisse static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask, 83763d5066fSJérôme Glisse unsigned long start, unsigned long end, 83863d5066fSJérôme Glisse struct mm_walk *walk) 83963d5066fSJérôme Glisse { 84063d5066fSJérôme Glisse #ifdef CONFIG_HUGETLB_PAGE 84163d5066fSJérôme Glisse unsigned long addr = start, i, pfn, mask, size, pfn_inc; 84263d5066fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 84363d5066fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 84463d5066fSJérôme Glisse struct vm_area_struct *vma = walk->vma; 84563d5066fSJérôme Glisse struct hstate *h = hstate_vma(vma); 84663d5066fSJérôme Glisse uint64_t orig_pfn, cpu_flags; 84763d5066fSJérôme Glisse bool fault, write_fault; 84863d5066fSJérôme Glisse spinlock_t *ptl; 84963d5066fSJérôme Glisse pte_t entry; 85063d5066fSJérôme Glisse int ret = 0; 85163d5066fSJérôme Glisse 85263d5066fSJérôme Glisse size = 1UL << huge_page_shift(h); 85363d5066fSJérôme Glisse mask = size - 1; 85463d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) { 85563d5066fSJérôme Glisse /* Make sure we are looking at full page. */ 85663d5066fSJérôme Glisse if (start & mask) 85763d5066fSJérôme Glisse return -EINVAL; 85863d5066fSJérôme Glisse if (end < (start + size)) 85963d5066fSJérôme Glisse return -EINVAL; 86063d5066fSJérôme Glisse pfn_inc = size >> PAGE_SHIFT; 86163d5066fSJérôme Glisse } else { 86263d5066fSJérôme Glisse pfn_inc = 1; 86363d5066fSJérôme Glisse size = PAGE_SIZE; 86463d5066fSJérôme Glisse } 86563d5066fSJérôme Glisse 86663d5066fSJérôme Glisse 86763d5066fSJérôme Glisse ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); 86863d5066fSJérôme Glisse entry = huge_ptep_get(pte); 86963d5066fSJérôme Glisse 87063d5066fSJérôme Glisse i = (start - range->start) >> range->page_shift; 87163d5066fSJérôme Glisse orig_pfn = range->pfns[i]; 87263d5066fSJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 87363d5066fSJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, entry); 87463d5066fSJérôme Glisse fault = write_fault = false; 87563d5066fSJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 87663d5066fSJérôme Glisse &fault, &write_fault); 87763d5066fSJérôme Glisse if (fault || write_fault) { 87863d5066fSJérôme Glisse ret = -ENOENT; 87963d5066fSJérôme Glisse goto unlock; 88063d5066fSJérôme Glisse } 88163d5066fSJérôme Glisse 88263d5066fSJérôme Glisse pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift); 88363d5066fSJérôme Glisse for (; addr < end; addr += size, i++, pfn += pfn_inc) 884391aab11SJérôme Glisse range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 885391aab11SJérôme Glisse cpu_flags; 88663d5066fSJérôme Glisse hmm_vma_walk->last = end; 88763d5066fSJérôme Glisse 88863d5066fSJérôme Glisse unlock: 88963d5066fSJérôme Glisse spin_unlock(ptl); 89063d5066fSJérôme Glisse 89163d5066fSJérôme Glisse if (ret == -ENOENT) 89263d5066fSJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 89363d5066fSJérôme Glisse 89463d5066fSJérôme Glisse return ret; 89563d5066fSJérôme Glisse #else /* CONFIG_HUGETLB_PAGE */ 89663d5066fSJérôme Glisse return -EINVAL; 89763d5066fSJérôme Glisse #endif 89863d5066fSJérôme Glisse } 89963d5066fSJérôme Glisse 900f88a1e90SJérôme Glisse static void hmm_pfns_clear(struct hmm_range *range, 901f88a1e90SJérôme Glisse uint64_t *pfns, 90233cd47dcSJérôme Glisse unsigned long addr, 90333cd47dcSJérôme Glisse unsigned long end) 90433cd47dcSJérôme Glisse { 90533cd47dcSJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 906f88a1e90SJérôme Glisse *pfns = range->values[HMM_PFN_NONE]; 90733cd47dcSJérôme Glisse } 90833cd47dcSJérôme Glisse 909da4c3c73SJérôme Glisse /* 910a3e0d41cSJérôme Glisse * hmm_range_register() - start tracking change to CPU page table over a range 911a3e0d41cSJérôme Glisse * @range: range 912a3e0d41cSJérôme Glisse * @mm: the mm struct for the range of virtual address 913a3e0d41cSJérôme Glisse * @start: start virtual address (inclusive) 914a3e0d41cSJérôme Glisse * @end: end virtual address (exclusive) 91563d5066fSJérôme Glisse * @page_shift: expect page shift for the range 916a3e0d41cSJérôme Glisse * Returns 0 on success, -EFAULT if the address space is no longer valid 917a3e0d41cSJérôme Glisse * 918a3e0d41cSJérôme Glisse * Track updates to the CPU page table see include/linux/hmm.h 919a3e0d41cSJérôme Glisse */ 920a3e0d41cSJérôme Glisse int hmm_range_register(struct hmm_range *range, 921a3e0d41cSJérôme Glisse struct mm_struct *mm, 922a3e0d41cSJérôme Glisse unsigned long start, 92363d5066fSJérôme Glisse unsigned long end, 92463d5066fSJérôme Glisse unsigned page_shift) 925a3e0d41cSJérôme Glisse { 92663d5066fSJérôme Glisse unsigned long mask = ((1UL << page_shift) - 1UL); 92763d5066fSJérôme Glisse 928a3e0d41cSJérôme Glisse range->valid = false; 929a3e0d41cSJérôme Glisse range->hmm = NULL; 930a3e0d41cSJérôme Glisse 93163d5066fSJérôme Glisse if ((start & mask) || (end & mask)) 93263d5066fSJérôme Glisse return -EINVAL; 93363d5066fSJérôme Glisse if (start >= end) 934a3e0d41cSJérôme Glisse return -EINVAL; 935a3e0d41cSJérôme Glisse 93663d5066fSJérôme Glisse range->page_shift = page_shift; 937a3e0d41cSJérôme Glisse range->start = start; 938a3e0d41cSJérôme Glisse range->end = end; 939a3e0d41cSJérôme Glisse 940a3e0d41cSJérôme Glisse range->hmm = hmm_get_or_create(mm); 941a3e0d41cSJérôme Glisse if (!range->hmm) 942a3e0d41cSJérôme Glisse return -EFAULT; 943a3e0d41cSJérôme Glisse 944a3e0d41cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 945a3e0d41cSJérôme Glisse if (range->hmm->mm == NULL || range->hmm->dead) { 946a3e0d41cSJérôme Glisse hmm_put(range->hmm); 947a3e0d41cSJérôme Glisse return -EFAULT; 948a3e0d41cSJérôme Glisse } 949a3e0d41cSJérôme Glisse 950a3e0d41cSJérôme Glisse /* Initialize range to track CPU page table update */ 951a3e0d41cSJérôme Glisse mutex_lock(&range->hmm->lock); 952a3e0d41cSJérôme Glisse 953a3e0d41cSJérôme Glisse list_add_rcu(&range->list, &range->hmm->ranges); 954a3e0d41cSJérôme Glisse 955a3e0d41cSJérôme Glisse /* 956a3e0d41cSJérôme Glisse * If there are any concurrent notifiers we have to wait for them for 957a3e0d41cSJérôme Glisse * the range to be valid (see hmm_range_wait_until_valid()). 958a3e0d41cSJérôme Glisse */ 959a3e0d41cSJérôme Glisse if (!range->hmm->notifiers) 960a3e0d41cSJérôme Glisse range->valid = true; 961a3e0d41cSJérôme Glisse mutex_unlock(&range->hmm->lock); 962a3e0d41cSJérôme Glisse 963a3e0d41cSJérôme Glisse return 0; 964a3e0d41cSJérôme Glisse } 965a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_register); 966a3e0d41cSJérôme Glisse 967a3e0d41cSJérôme Glisse /* 968a3e0d41cSJérôme Glisse * hmm_range_unregister() - stop tracking change to CPU page table over a range 969a3e0d41cSJérôme Glisse * @range: range 970a3e0d41cSJérôme Glisse * 971a3e0d41cSJérôme Glisse * Range struct is used to track updates to the CPU page table after a call to 972a3e0d41cSJérôme Glisse * hmm_range_register(). See include/linux/hmm.h for how to use it. 973a3e0d41cSJérôme Glisse */ 974a3e0d41cSJérôme Glisse void hmm_range_unregister(struct hmm_range *range) 975a3e0d41cSJérôme Glisse { 976a3e0d41cSJérôme Glisse /* Sanity check this really should not happen. */ 977a3e0d41cSJérôme Glisse if (range->hmm == NULL || range->end <= range->start) 978a3e0d41cSJérôme Glisse return; 979a3e0d41cSJérôme Glisse 980a3e0d41cSJérôme Glisse mutex_lock(&range->hmm->lock); 981a3e0d41cSJérôme Glisse list_del_rcu(&range->list); 982a3e0d41cSJérôme Glisse mutex_unlock(&range->hmm->lock); 983a3e0d41cSJérôme Glisse 984a3e0d41cSJérôme Glisse /* Drop reference taken by hmm_range_register() */ 985a3e0d41cSJérôme Glisse range->valid = false; 986a3e0d41cSJérôme Glisse hmm_put(range->hmm); 987a3e0d41cSJérôme Glisse range->hmm = NULL; 988a3e0d41cSJérôme Glisse } 989a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_unregister); 990a3e0d41cSJérôme Glisse 991a3e0d41cSJérôme Glisse /* 99225f23a0cSJérôme Glisse * hmm_range_snapshot() - snapshot CPU page table for a range 99325f23a0cSJérôme Glisse * @range: range 994a3e0d41cSJérôme Glisse * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid 995a3e0d41cSJérôme Glisse * permission (for instance asking for write and range is read only), 996a3e0d41cSJérôme Glisse * -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid 997a3e0d41cSJérôme Glisse * vma or it is illegal to access that range), number of valid pages 998a3e0d41cSJérôme Glisse * in range->pfns[] (from range start address). 999da4c3c73SJérôme Glisse * 1000da4c3c73SJérôme Glisse * This snapshots the CPU page table for a range of virtual addresses. Snapshot 1001a3e0d41cSJérôme Glisse * validity is tracked by range struct. See in include/linux/hmm.h for example 1002a3e0d41cSJérôme Glisse * on how to use. 1003da4c3c73SJérôme Glisse */ 100425f23a0cSJérôme Glisse long hmm_range_snapshot(struct hmm_range *range) 1005da4c3c73SJérôme Glisse { 100663d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 1007a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 100874eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 1009a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 1010a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 1011da4c3c73SJérôme Glisse struct mm_walk mm_walk; 1012704f3f2cSJérôme Glisse 1013704f3f2cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 1014a3e0d41cSJérôme Glisse if (hmm->mm == NULL || hmm->dead) 1015a3e0d41cSJérôme Glisse return -EFAULT; 1016da4c3c73SJérôme Glisse 1017a3e0d41cSJérôme Glisse do { 1018a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1019a3e0d41cSJérôme Glisse if (!range->valid) 1020a3e0d41cSJérôme Glisse return -EAGAIN; 1021a3e0d41cSJérôme Glisse 1022a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 102363d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1024a3e0d41cSJérôme Glisse return -EFAULT; 1025a3e0d41cSJérôme Glisse 102663d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 102763d5066fSJérôme Glisse struct hstate *h = hstate_vma(vma); 102863d5066fSJérôme Glisse 102963d5066fSJérôme Glisse if (huge_page_shift(h) != range->page_shift && 103063d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 103163d5066fSJérôme Glisse return -EINVAL; 103263d5066fSJérôme Glisse } else { 103363d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 103463d5066fSJérôme Glisse return -EINVAL; 103563d5066fSJérôme Glisse } 103663d5066fSJérôme Glisse 103786586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 103886586a41SJérôme Glisse /* 1039a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1040a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1041a3e0d41cSJérôme Glisse * support architecture that allow write without read. 104286586a41SJérôme Glisse */ 1043a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1044a3e0d41cSJérôme Glisse range->start, range->end); 104586586a41SJérôme Glisse return -EPERM; 104686586a41SJérôme Glisse } 104786586a41SJérôme Glisse 1048a3e0d41cSJérôme Glisse range->vma = vma; 1049992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1050a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 105174eee180SJérôme Glisse hmm_vma_walk.fault = false; 105274eee180SJérôme Glisse hmm_vma_walk.range = range; 105374eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1054a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 105574eee180SJérôme Glisse 1056da4c3c73SJérôme Glisse mm_walk.vma = vma; 1057da4c3c73SJérôme Glisse mm_walk.mm = vma->vm_mm; 1058da4c3c73SJérôme Glisse mm_walk.pte_entry = NULL; 1059da4c3c73SJérôme Glisse mm_walk.test_walk = NULL; 1060da4c3c73SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1061992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 1062da4c3c73SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 1063da4c3c73SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 106463d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 1065da4c3c73SJérôme Glisse 1066a3e0d41cSJérôme Glisse walk_page_range(start, end, &mm_walk); 1067a3e0d41cSJérôme Glisse start = end; 1068a3e0d41cSJérôme Glisse } while (start < range->end); 1069a3e0d41cSJérôme Glisse 107025f23a0cSJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1071da4c3c73SJérôme Glisse } 107225f23a0cSJérôme Glisse EXPORT_SYMBOL(hmm_range_snapshot); 1073da4c3c73SJérôme Glisse 1074da4c3c73SJérôme Glisse /* 107573231612SJérôme Glisse * hmm_range_fault() - try to fault some address in a virtual address range 107608232a45SJérôme Glisse * @range: range being faulted 107774eee180SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 107873231612SJérôme Glisse * Returns: number of valid pages in range->pfns[] (from range start 107973231612SJérôme Glisse * address). This may be zero. If the return value is negative, 108073231612SJérôme Glisse * then one of the following values may be returned: 108173231612SJérôme Glisse * 108273231612SJérôme Glisse * -EINVAL invalid arguments or mm or virtual address are in an 108363d5066fSJérôme Glisse * invalid vma (for instance device file vma). 108473231612SJérôme Glisse * -ENOMEM: Out of memory. 108573231612SJérôme Glisse * -EPERM: Invalid permission (for instance asking for write and 108673231612SJérôme Glisse * range is read only). 108773231612SJérôme Glisse * -EAGAIN: If you need to retry and mmap_sem was drop. This can only 108873231612SJérôme Glisse * happens if block argument is false. 108973231612SJérôme Glisse * -EBUSY: If the the range is being invalidated and you should wait 109073231612SJérôme Glisse * for invalidation to finish. 109173231612SJérôme Glisse * -EFAULT: Invalid (ie either no valid vma or it is illegal to access 109273231612SJérôme Glisse * that range), number of valid pages in range->pfns[] (from 109373231612SJérôme Glisse * range start address). 109474eee180SJérôme Glisse * 109574eee180SJérôme Glisse * This is similar to a regular CPU page fault except that it will not trigger 109673231612SJérôme Glisse * any memory migration if the memory being faulted is not accessible by CPUs 109773231612SJérôme Glisse * and caller does not ask for migration. 109874eee180SJérôme Glisse * 1099ff05c0c6SJérôme Glisse * On error, for one virtual address in the range, the function will mark the 1100ff05c0c6SJérôme Glisse * corresponding HMM pfn entry with an error flag. 110174eee180SJérôme Glisse */ 110273231612SJérôme Glisse long hmm_range_fault(struct hmm_range *range, bool block) 110374eee180SJérôme Glisse { 110463d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 1105a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 110674eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 1107a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 1108a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 110974eee180SJérôme Glisse struct mm_walk mm_walk; 111074eee180SJérôme Glisse int ret; 111174eee180SJérôme Glisse 1112704f3f2cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 1113a3e0d41cSJérôme Glisse if (hmm->mm == NULL || hmm->dead) 1114a3e0d41cSJérôme Glisse return -EFAULT; 1115a3e0d41cSJérôme Glisse 1116a3e0d41cSJérôme Glisse do { 1117a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1118a3e0d41cSJérôme Glisse if (!range->valid) { 1119a3e0d41cSJérôme Glisse up_read(&hmm->mm->mmap_sem); 1120a3e0d41cSJérôme Glisse return -EAGAIN; 1121704f3f2cSJérôme Glisse } 112274eee180SJérôme Glisse 1123a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 112463d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1125a3e0d41cSJérôme Glisse return -EFAULT; 1126a3e0d41cSJérôme Glisse 112763d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 112863d5066fSJérôme Glisse if (huge_page_shift(hstate_vma(vma)) != 112963d5066fSJérôme Glisse range->page_shift && 113063d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 113163d5066fSJérôme Glisse return -EINVAL; 113263d5066fSJérôme Glisse } else { 113363d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 113463d5066fSJérôme Glisse return -EINVAL; 113563d5066fSJérôme Glisse } 113663d5066fSJérôme Glisse 113786586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 113886586a41SJérôme Glisse /* 1139a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1140a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1141a3e0d41cSJérôme Glisse * support architecture that allow write without read. 114286586a41SJérôme Glisse */ 1143a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1144a3e0d41cSJérôme Glisse range->start, range->end); 114586586a41SJérôme Glisse return -EPERM; 114686586a41SJérôme Glisse } 114774eee180SJérôme Glisse 1148a3e0d41cSJérôme Glisse range->vma = vma; 1149992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1150a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 115174eee180SJérôme Glisse hmm_vma_walk.fault = true; 115274eee180SJérôme Glisse hmm_vma_walk.block = block; 115374eee180SJérôme Glisse hmm_vma_walk.range = range; 115474eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1155a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 115674eee180SJérôme Glisse 115774eee180SJérôme Glisse mm_walk.vma = vma; 115874eee180SJérôme Glisse mm_walk.mm = vma->vm_mm; 115974eee180SJérôme Glisse mm_walk.pte_entry = NULL; 116074eee180SJérôme Glisse mm_walk.test_walk = NULL; 116174eee180SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1162992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 116374eee180SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 116474eee180SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 116563d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 116674eee180SJérôme Glisse 116774eee180SJérôme Glisse do { 1168a3e0d41cSJérôme Glisse ret = walk_page_range(start, end, &mm_walk); 116974eee180SJérôme Glisse start = hmm_vma_walk.last; 1170a3e0d41cSJérôme Glisse 117173231612SJérôme Glisse /* Keep trying while the range is valid. */ 117273231612SJérôme Glisse } while (ret == -EBUSY && range->valid); 117374eee180SJérôme Glisse 117474eee180SJérôme Glisse if (ret) { 117574eee180SJérôme Glisse unsigned long i; 117674eee180SJérôme Glisse 117774eee180SJérôme Glisse i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1178a3e0d41cSJérôme Glisse hmm_pfns_clear(range, &range->pfns[i], 1179a3e0d41cSJérôme Glisse hmm_vma_walk.last, range->end); 118073231612SJérôme Glisse return ret; 118174eee180SJérôme Glisse } 1182a3e0d41cSJérôme Glisse start = end; 1183a3e0d41cSJérôme Glisse 1184a3e0d41cSJérôme Glisse } while (start < range->end); 1185704f3f2cSJérôme Glisse 118673231612SJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 118774eee180SJérôme Glisse } 118873231612SJérôme Glisse EXPORT_SYMBOL(hmm_range_fault); 118955c0ece8SJérôme Glisse 119055c0ece8SJérôme Glisse /** 119155c0ece8SJérôme Glisse * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one. 119255c0ece8SJérôme Glisse * @range: range being faulted 119355c0ece8SJérôme Glisse * @device: device against to dma map page to 119455c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 119555c0ece8SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 119655c0ece8SJérôme Glisse * Returns: number of pages mapped on success, -EAGAIN if mmap_sem have been 119755c0ece8SJérôme Glisse * drop and you need to try again, some other error value otherwise 119855c0ece8SJérôme Glisse * 119955c0ece8SJérôme Glisse * Note same usage pattern as hmm_range_fault(). 120055c0ece8SJérôme Glisse */ 120155c0ece8SJérôme Glisse long hmm_range_dma_map(struct hmm_range *range, 120255c0ece8SJérôme Glisse struct device *device, 120355c0ece8SJérôme Glisse dma_addr_t *daddrs, 120455c0ece8SJérôme Glisse bool block) 120555c0ece8SJérôme Glisse { 120655c0ece8SJérôme Glisse unsigned long i, npages, mapped; 120755c0ece8SJérôme Glisse long ret; 120855c0ece8SJérôme Glisse 120955c0ece8SJérôme Glisse ret = hmm_range_fault(range, block); 121055c0ece8SJérôme Glisse if (ret <= 0) 121155c0ece8SJérôme Glisse return ret ? ret : -EBUSY; 121255c0ece8SJérôme Glisse 121355c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 121455c0ece8SJérôme Glisse for (i = 0, mapped = 0; i < npages; ++i) { 121555c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 121655c0ece8SJérôme Glisse struct page *page; 121755c0ece8SJérôme Glisse 121855c0ece8SJérôme Glisse /* 121955c0ece8SJérôme Glisse * FIXME need to update DMA API to provide invalid DMA address 122055c0ece8SJérôme Glisse * value instead of a function to test dma address value. This 122155c0ece8SJérôme Glisse * would remove lot of dumb code duplicated accross many arch. 122255c0ece8SJérôme Glisse * 122355c0ece8SJérôme Glisse * For now setting it to 0 here is good enough as the pfns[] 122455c0ece8SJérôme Glisse * value is what is use to check what is valid and what isn't. 122555c0ece8SJérôme Glisse */ 122655c0ece8SJérôme Glisse daddrs[i] = 0; 122755c0ece8SJérôme Glisse 1228391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 122955c0ece8SJérôme Glisse if (page == NULL) 123055c0ece8SJérôme Glisse continue; 123155c0ece8SJérôme Glisse 123255c0ece8SJérôme Glisse /* Check if range is being invalidated */ 123355c0ece8SJérôme Glisse if (!range->valid) { 123455c0ece8SJérôme Glisse ret = -EBUSY; 123555c0ece8SJérôme Glisse goto unmap; 123655c0ece8SJérôme Glisse } 123755c0ece8SJérôme Glisse 123855c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 123955c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 124055c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 124155c0ece8SJérôme Glisse 124255c0ece8SJérôme Glisse daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir); 124355c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) { 124455c0ece8SJérôme Glisse ret = -EFAULT; 124555c0ece8SJérôme Glisse goto unmap; 124655c0ece8SJérôme Glisse } 124755c0ece8SJérôme Glisse 124855c0ece8SJérôme Glisse mapped++; 124955c0ece8SJérôme Glisse } 125055c0ece8SJérôme Glisse 125155c0ece8SJérôme Glisse return mapped; 125255c0ece8SJérôme Glisse 125355c0ece8SJérôme Glisse unmap: 125455c0ece8SJérôme Glisse for (npages = i, i = 0; (i < npages) && mapped; ++i) { 125555c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 125655c0ece8SJérôme Glisse struct page *page; 125755c0ece8SJérôme Glisse 1258391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 125955c0ece8SJérôme Glisse if (page == NULL) 126055c0ece8SJérôme Glisse continue; 126155c0ece8SJérôme Glisse 126255c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) 126355c0ece8SJérôme Glisse continue; 126455c0ece8SJérôme Glisse 126555c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 126655c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 126755c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 126855c0ece8SJérôme Glisse 126955c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 127055c0ece8SJérôme Glisse mapped--; 127155c0ece8SJérôme Glisse } 127255c0ece8SJérôme Glisse 127355c0ece8SJérôme Glisse return ret; 127455c0ece8SJérôme Glisse } 127555c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_map); 127655c0ece8SJérôme Glisse 127755c0ece8SJérôme Glisse /** 127855c0ece8SJérôme Glisse * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map() 127955c0ece8SJérôme Glisse * @range: range being unmapped 128055c0ece8SJérôme Glisse * @vma: the vma against which the range (optional) 128155c0ece8SJérôme Glisse * @device: device against which dma map was done 128255c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 128355c0ece8SJérôme Glisse * @dirty: dirty page if it had the write flag set 128455c0ece8SJérôme Glisse * Returns: number of page unmapped on success, -EINVAL otherwise 128555c0ece8SJérôme Glisse * 128655c0ece8SJérôme Glisse * Note that caller MUST abide by mmu notifier or use HMM mirror and abide 128755c0ece8SJérôme Glisse * to the sync_cpu_device_pagetables() callback so that it is safe here to 128855c0ece8SJérôme Glisse * call set_page_dirty(). Caller must also take appropriate locks to avoid 128955c0ece8SJérôme Glisse * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress. 129055c0ece8SJérôme Glisse */ 129155c0ece8SJérôme Glisse long hmm_range_dma_unmap(struct hmm_range *range, 129255c0ece8SJérôme Glisse struct vm_area_struct *vma, 129355c0ece8SJérôme Glisse struct device *device, 129455c0ece8SJérôme Glisse dma_addr_t *daddrs, 129555c0ece8SJérôme Glisse bool dirty) 129655c0ece8SJérôme Glisse { 129755c0ece8SJérôme Glisse unsigned long i, npages; 129855c0ece8SJérôme Glisse long cpages = 0; 129955c0ece8SJérôme Glisse 130055c0ece8SJérôme Glisse /* Sanity check. */ 130155c0ece8SJérôme Glisse if (range->end <= range->start) 130255c0ece8SJérôme Glisse return -EINVAL; 130355c0ece8SJérôme Glisse if (!daddrs) 130455c0ece8SJérôme Glisse return -EINVAL; 130555c0ece8SJérôme Glisse if (!range->pfns) 130655c0ece8SJérôme Glisse return -EINVAL; 130755c0ece8SJérôme Glisse 130855c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 130955c0ece8SJérôme Glisse for (i = 0; i < npages; ++i) { 131055c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 131155c0ece8SJérôme Glisse struct page *page; 131255c0ece8SJérôme Glisse 1313391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 131455c0ece8SJérôme Glisse if (page == NULL) 131555c0ece8SJérôme Glisse continue; 131655c0ece8SJérôme Glisse 131755c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 131855c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) { 131955c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 132055c0ece8SJérôme Glisse 132155c0ece8SJérôme Glisse /* 132255c0ece8SJérôme Glisse * See comments in function description on why it is 132355c0ece8SJérôme Glisse * safe here to call set_page_dirty() 132455c0ece8SJérôme Glisse */ 132555c0ece8SJérôme Glisse if (dirty) 132655c0ece8SJérôme Glisse set_page_dirty(page); 132755c0ece8SJérôme Glisse } 132855c0ece8SJérôme Glisse 132955c0ece8SJérôme Glisse /* Unmap and clear pfns/dma address */ 133055c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 133155c0ece8SJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 133255c0ece8SJérôme Glisse /* FIXME see comments in hmm_vma_dma_map() */ 133355c0ece8SJérôme Glisse daddrs[i] = 0; 133455c0ece8SJérôme Glisse cpages++; 133555c0ece8SJérôme Glisse } 133655c0ece8SJérôme Glisse 133755c0ece8SJérôme Glisse return cpages; 133855c0ece8SJérôme Glisse } 133955c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_unmap); 1340c0b12405SJérôme Glisse #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ 13414ef589dcSJérôme Glisse 13424ef589dcSJérôme Glisse 1343df6ad698SJérôme Glisse #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC) 13444ef589dcSJérôme Glisse struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma, 13454ef589dcSJérôme Glisse unsigned long addr) 13464ef589dcSJérôme Glisse { 13474ef589dcSJérôme Glisse struct page *page; 13484ef589dcSJérôme Glisse 13494ef589dcSJérôme Glisse page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 13504ef589dcSJérôme Glisse if (!page) 13514ef589dcSJérôme Glisse return NULL; 13524ef589dcSJérôme Glisse lock_page(page); 13534ef589dcSJérôme Glisse return page; 13544ef589dcSJérôme Glisse } 13554ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_vma_alloc_locked_page); 13564ef589dcSJérôme Glisse 13574ef589dcSJérôme Glisse 13584ef589dcSJérôme Glisse static void hmm_devmem_ref_release(struct percpu_ref *ref) 13594ef589dcSJérôme Glisse { 13604ef589dcSJérôme Glisse struct hmm_devmem *devmem; 13614ef589dcSJérôme Glisse 13624ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 13634ef589dcSJérôme Glisse complete(&devmem->completion); 13644ef589dcSJérôme Glisse } 13654ef589dcSJérôme Glisse 13664ef589dcSJérôme Glisse static void hmm_devmem_ref_exit(void *data) 13674ef589dcSJérôme Glisse { 13684ef589dcSJérôme Glisse struct percpu_ref *ref = data; 13694ef589dcSJérôme Glisse struct hmm_devmem *devmem; 13704ef589dcSJérôme Glisse 13714ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 1372bbecd94eSDan Williams wait_for_completion(&devmem->completion); 13734ef589dcSJérôme Glisse percpu_ref_exit(ref); 13744ef589dcSJérôme Glisse } 13754ef589dcSJérôme Glisse 1376bbecd94eSDan Williams static void hmm_devmem_ref_kill(struct percpu_ref *ref) 13774ef589dcSJérôme Glisse { 13784ef589dcSJérôme Glisse percpu_ref_kill(ref); 13794ef589dcSJérôme Glisse } 13804ef589dcSJérôme Glisse 1381b57e622eSSouptick Joarder static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma, 13824ef589dcSJérôme Glisse unsigned long addr, 13834ef589dcSJérôme Glisse const struct page *page, 13844ef589dcSJérôme Glisse unsigned int flags, 13854ef589dcSJérôme Glisse pmd_t *pmdp) 13864ef589dcSJérôme Glisse { 13874ef589dcSJérôme Glisse struct hmm_devmem *devmem = page->pgmap->data; 13884ef589dcSJérôme Glisse 13894ef589dcSJérôme Glisse return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp); 13904ef589dcSJérôme Glisse } 13914ef589dcSJérôme Glisse 13924ef589dcSJérôme Glisse static void hmm_devmem_free(struct page *page, void *data) 13934ef589dcSJérôme Glisse { 13944ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 13954ef589dcSJérôme Glisse 13962fa147bdSDan Williams page->mapping = NULL; 13972fa147bdSDan Williams 13984ef589dcSJérôme Glisse devmem->ops->free(devmem, page); 13994ef589dcSJérôme Glisse } 14004ef589dcSJérôme Glisse 14014ef589dcSJérôme Glisse /* 14024ef589dcSJérôme Glisse * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory 14034ef589dcSJérôme Glisse * 14044ef589dcSJérôme Glisse * @ops: memory event device driver callback (see struct hmm_devmem_ops) 14054ef589dcSJérôme Glisse * @device: device struct to bind the resource too 14064ef589dcSJérôme Glisse * @size: size in bytes of the device memory to add 14074ef589dcSJérôme Glisse * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise 14084ef589dcSJérôme Glisse * 14094ef589dcSJérôme Glisse * This function first finds an empty range of physical address big enough to 14104ef589dcSJérôme Glisse * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which 14114ef589dcSJérôme Glisse * in turn allocates struct pages. It does not do anything beyond that; all 14124ef589dcSJérôme Glisse * events affecting the memory will go through the various callbacks provided 14134ef589dcSJérôme Glisse * by hmm_devmem_ops struct. 14144ef589dcSJérôme Glisse * 14154ef589dcSJérôme Glisse * Device driver should call this function during device initialization and 14164ef589dcSJérôme Glisse * is then responsible of memory management. HMM only provides helpers. 14174ef589dcSJérôme Glisse */ 14184ef589dcSJérôme Glisse struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, 14194ef589dcSJérôme Glisse struct device *device, 14204ef589dcSJérôme Glisse unsigned long size) 14214ef589dcSJérôme Glisse { 14224ef589dcSJérôme Glisse struct hmm_devmem *devmem; 14234ef589dcSJérôme Glisse resource_size_t addr; 1424bbecd94eSDan Williams void *result; 14254ef589dcSJérôme Glisse int ret; 14264ef589dcSJérôme Glisse 1427e7638488SDan Williams dev_pagemap_get_ops(); 14284ef589dcSJérôme Glisse 142958ef15b7SDan Williams devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); 14304ef589dcSJérôme Glisse if (!devmem) 14314ef589dcSJérôme Glisse return ERR_PTR(-ENOMEM); 14324ef589dcSJérôme Glisse 14334ef589dcSJérôme Glisse init_completion(&devmem->completion); 14344ef589dcSJérôme Glisse devmem->pfn_first = -1UL; 14354ef589dcSJérôme Glisse devmem->pfn_last = -1UL; 14364ef589dcSJérôme Glisse devmem->resource = NULL; 14374ef589dcSJérôme Glisse devmem->device = device; 14384ef589dcSJérôme Glisse devmem->ops = ops; 14394ef589dcSJérôme Glisse 14404ef589dcSJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 14414ef589dcSJérôme Glisse 0, GFP_KERNEL); 14424ef589dcSJérôme Glisse if (ret) 144358ef15b7SDan Williams return ERR_PTR(ret); 14444ef589dcSJérôme Glisse 144558ef15b7SDan Williams ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref); 14464ef589dcSJérôme Glisse if (ret) 144758ef15b7SDan Williams return ERR_PTR(ret); 14484ef589dcSJérôme Glisse 14494ef589dcSJérôme Glisse size = ALIGN(size, PA_SECTION_SIZE); 14504ef589dcSJérôme Glisse addr = min((unsigned long)iomem_resource.end, 14514ef589dcSJérôme Glisse (1UL << MAX_PHYSMEM_BITS) - 1); 14524ef589dcSJérôme Glisse addr = addr - size + 1UL; 14534ef589dcSJérôme Glisse 14544ef589dcSJérôme Glisse /* 14554ef589dcSJérôme Glisse * FIXME add a new helper to quickly walk resource tree and find free 14564ef589dcSJérôme Glisse * range 14574ef589dcSJérôme Glisse * 14584ef589dcSJérôme Glisse * FIXME what about ioport_resource resource ? 14594ef589dcSJérôme Glisse */ 14604ef589dcSJérôme Glisse for (; addr > size && addr >= iomem_resource.start; addr -= size) { 14614ef589dcSJérôme Glisse ret = region_intersects(addr, size, 0, IORES_DESC_NONE); 14624ef589dcSJérôme Glisse if (ret != REGION_DISJOINT) 14634ef589dcSJérôme Glisse continue; 14644ef589dcSJérôme Glisse 14654ef589dcSJérôme Glisse devmem->resource = devm_request_mem_region(device, addr, size, 14664ef589dcSJérôme Glisse dev_name(device)); 146758ef15b7SDan Williams if (!devmem->resource) 146858ef15b7SDan Williams return ERR_PTR(-ENOMEM); 14694ef589dcSJérôme Glisse break; 14704ef589dcSJérôme Glisse } 147158ef15b7SDan Williams if (!devmem->resource) 147258ef15b7SDan Williams return ERR_PTR(-ERANGE); 14734ef589dcSJérôme Glisse 14744ef589dcSJérôme Glisse devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY; 14754ef589dcSJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 14764ef589dcSJérôme Glisse devmem->pfn_last = devmem->pfn_first + 14774ef589dcSJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1478063a7d1dSDan Williams devmem->page_fault = hmm_devmem_fault; 14794ef589dcSJérôme Glisse 1480bbecd94eSDan Williams devmem->pagemap.type = MEMORY_DEVICE_PRIVATE; 1481bbecd94eSDan Williams devmem->pagemap.res = *devmem->resource; 1482bbecd94eSDan Williams devmem->pagemap.page_free = hmm_devmem_free; 1483bbecd94eSDan Williams devmem->pagemap.altmap_valid = false; 1484bbecd94eSDan Williams devmem->pagemap.ref = &devmem->ref; 1485bbecd94eSDan Williams devmem->pagemap.data = devmem; 1486bbecd94eSDan Williams devmem->pagemap.kill = hmm_devmem_ref_kill; 148758ef15b7SDan Williams 1488bbecd94eSDan Williams result = devm_memremap_pages(devmem->device, &devmem->pagemap); 1489bbecd94eSDan Williams if (IS_ERR(result)) 1490bbecd94eSDan Williams return result; 14914ef589dcSJérôme Glisse return devmem; 14924ef589dcSJérôme Glisse } 149302917e9fSDan Williams EXPORT_SYMBOL_GPL(hmm_devmem_add); 14944ef589dcSJérôme Glisse 1495d3df0a42SJérôme Glisse struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops, 1496d3df0a42SJérôme Glisse struct device *device, 1497d3df0a42SJérôme Glisse struct resource *res) 1498d3df0a42SJérôme Glisse { 1499d3df0a42SJérôme Glisse struct hmm_devmem *devmem; 1500bbecd94eSDan Williams void *result; 1501d3df0a42SJérôme Glisse int ret; 1502d3df0a42SJérôme Glisse 1503d3df0a42SJérôme Glisse if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY) 1504d3df0a42SJérôme Glisse return ERR_PTR(-EINVAL); 1505d3df0a42SJérôme Glisse 1506e7638488SDan Williams dev_pagemap_get_ops(); 1507d3df0a42SJérôme Glisse 150858ef15b7SDan Williams devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); 1509d3df0a42SJérôme Glisse if (!devmem) 1510d3df0a42SJérôme Glisse return ERR_PTR(-ENOMEM); 1511d3df0a42SJérôme Glisse 1512d3df0a42SJérôme Glisse init_completion(&devmem->completion); 1513d3df0a42SJérôme Glisse devmem->pfn_first = -1UL; 1514d3df0a42SJérôme Glisse devmem->pfn_last = -1UL; 1515d3df0a42SJérôme Glisse devmem->resource = res; 1516d3df0a42SJérôme Glisse devmem->device = device; 1517d3df0a42SJérôme Glisse devmem->ops = ops; 1518d3df0a42SJérôme Glisse 1519d3df0a42SJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 1520d3df0a42SJérôme Glisse 0, GFP_KERNEL); 1521d3df0a42SJérôme Glisse if (ret) 152258ef15b7SDan Williams return ERR_PTR(ret); 1523d3df0a42SJérôme Glisse 152458ef15b7SDan Williams ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, 152558ef15b7SDan Williams &devmem->ref); 1526d3df0a42SJérôme Glisse if (ret) 152758ef15b7SDan Williams return ERR_PTR(ret); 1528d3df0a42SJérôme Glisse 1529d3df0a42SJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 1530d3df0a42SJérôme Glisse devmem->pfn_last = devmem->pfn_first + 1531d3df0a42SJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1532063a7d1dSDan Williams devmem->page_fault = hmm_devmem_fault; 1533d3df0a42SJérôme Glisse 1534bbecd94eSDan Williams devmem->pagemap.type = MEMORY_DEVICE_PUBLIC; 1535bbecd94eSDan Williams devmem->pagemap.res = *devmem->resource; 1536bbecd94eSDan Williams devmem->pagemap.page_free = hmm_devmem_free; 1537bbecd94eSDan Williams devmem->pagemap.altmap_valid = false; 1538bbecd94eSDan Williams devmem->pagemap.ref = &devmem->ref; 1539bbecd94eSDan Williams devmem->pagemap.data = devmem; 1540bbecd94eSDan Williams devmem->pagemap.kill = hmm_devmem_ref_kill; 154158ef15b7SDan Williams 1542bbecd94eSDan Williams result = devm_memremap_pages(devmem->device, &devmem->pagemap); 1543bbecd94eSDan Williams if (IS_ERR(result)) 1544bbecd94eSDan Williams return result; 1545d3df0a42SJérôme Glisse return devmem; 1546d3df0a42SJérôme Glisse } 154702917e9fSDan Williams EXPORT_SYMBOL_GPL(hmm_devmem_add_resource); 1548d3df0a42SJérôme Glisse 15494ef589dcSJérôme Glisse /* 1550858b54daSJérôme Glisse * A device driver that wants to handle multiple devices memory through a 1551858b54daSJérôme Glisse * single fake device can use hmm_device to do so. This is purely a helper 1552858b54daSJérôme Glisse * and it is not needed to make use of any HMM functionality. 1553858b54daSJérôme Glisse */ 1554858b54daSJérôme Glisse #define HMM_DEVICE_MAX 256 1555858b54daSJérôme Glisse 1556858b54daSJérôme Glisse static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX); 1557858b54daSJérôme Glisse static DEFINE_SPINLOCK(hmm_device_lock); 1558858b54daSJérôme Glisse static struct class *hmm_device_class; 1559858b54daSJérôme Glisse static dev_t hmm_device_devt; 1560858b54daSJérôme Glisse 1561858b54daSJérôme Glisse static void hmm_device_release(struct device *device) 1562858b54daSJérôme Glisse { 1563858b54daSJérôme Glisse struct hmm_device *hmm_device; 1564858b54daSJérôme Glisse 1565858b54daSJérôme Glisse hmm_device = container_of(device, struct hmm_device, device); 1566858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1567858b54daSJérôme Glisse clear_bit(hmm_device->minor, hmm_device_mask); 1568858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1569858b54daSJérôme Glisse 1570858b54daSJérôme Glisse kfree(hmm_device); 1571858b54daSJérôme Glisse } 1572858b54daSJérôme Glisse 1573858b54daSJérôme Glisse struct hmm_device *hmm_device_new(void *drvdata) 1574858b54daSJérôme Glisse { 1575858b54daSJérôme Glisse struct hmm_device *hmm_device; 1576858b54daSJérôme Glisse 1577858b54daSJérôme Glisse hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL); 1578858b54daSJérôme Glisse if (!hmm_device) 1579858b54daSJérôme Glisse return ERR_PTR(-ENOMEM); 1580858b54daSJérôme Glisse 1581858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1582858b54daSJérôme Glisse hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX); 1583858b54daSJérôme Glisse if (hmm_device->minor >= HMM_DEVICE_MAX) { 1584858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1585858b54daSJérôme Glisse kfree(hmm_device); 1586858b54daSJérôme Glisse return ERR_PTR(-EBUSY); 1587858b54daSJérôme Glisse } 1588858b54daSJérôme Glisse set_bit(hmm_device->minor, hmm_device_mask); 1589858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1590858b54daSJérôme Glisse 1591858b54daSJérôme Glisse dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor); 1592858b54daSJérôme Glisse hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt), 1593858b54daSJérôme Glisse hmm_device->minor); 1594858b54daSJérôme Glisse hmm_device->device.release = hmm_device_release; 1595858b54daSJérôme Glisse dev_set_drvdata(&hmm_device->device, drvdata); 1596858b54daSJérôme Glisse hmm_device->device.class = hmm_device_class; 1597858b54daSJérôme Glisse device_initialize(&hmm_device->device); 1598858b54daSJérôme Glisse 1599858b54daSJérôme Glisse return hmm_device; 1600858b54daSJérôme Glisse } 1601858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_new); 1602858b54daSJérôme Glisse 1603858b54daSJérôme Glisse void hmm_device_put(struct hmm_device *hmm_device) 1604858b54daSJérôme Glisse { 1605858b54daSJérôme Glisse put_device(&hmm_device->device); 1606858b54daSJérôme Glisse } 1607858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_put); 1608858b54daSJérôme Glisse 1609858b54daSJérôme Glisse static int __init hmm_init(void) 1610858b54daSJérôme Glisse { 1611858b54daSJérôme Glisse int ret; 1612858b54daSJérôme Glisse 1613858b54daSJérôme Glisse ret = alloc_chrdev_region(&hmm_device_devt, 0, 1614858b54daSJérôme Glisse HMM_DEVICE_MAX, 1615858b54daSJérôme Glisse "hmm_device"); 1616858b54daSJérôme Glisse if (ret) 1617858b54daSJérôme Glisse return ret; 1618858b54daSJérôme Glisse 1619858b54daSJérôme Glisse hmm_device_class = class_create(THIS_MODULE, "hmm_device"); 1620858b54daSJérôme Glisse if (IS_ERR(hmm_device_class)) { 1621858b54daSJérôme Glisse unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX); 1622858b54daSJérôme Glisse return PTR_ERR(hmm_device_class); 1623858b54daSJérôme Glisse } 1624858b54daSJérôme Glisse return 0; 1625858b54daSJérôme Glisse } 1626858b54daSJérôme Glisse 1627858b54daSJérôme Glisse device_initcall(hmm_init); 1628df6ad698SJérôme Glisse #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ 1629