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> 23c8a53b2dSJason Gunthorpe #include <linux/sched/mm.h> 247b2d55d2SJérôme Glisse #include <linux/jump_label.h> 2555c0ece8SJérôme Glisse #include <linux/dma-mapping.h> 26c0b12405SJérôme Glisse #include <linux/mmu_notifier.h> 274ef589dcSJérôme Glisse #include <linux/memory_hotplug.h> 284ef589dcSJérôme Glisse 29c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops; 30c0b12405SJérôme Glisse 31704f3f2cSJérôme Glisse /** 32704f3f2cSJérôme Glisse * hmm_get_or_create - register HMM against an mm (HMM internal) 33704f3f2cSJérôme Glisse * 34704f3f2cSJérôme Glisse * @mm: mm struct to attach to 35704f3f2cSJérôme Glisse * Returns: returns an HMM object, either by referencing the existing 36704f3f2cSJérôme Glisse * (per-process) object, or by creating a new one. 37704f3f2cSJérôme Glisse * 38704f3f2cSJérôme Glisse * This is not intended to be used directly by device drivers. If mm already 39704f3f2cSJérôme Glisse * has an HMM struct then it get a reference on it and returns it. Otherwise 40704f3f2cSJérôme Glisse * it allocates an HMM struct, initializes it, associate it with the mm and 41704f3f2cSJérôme Glisse * returns it. 42704f3f2cSJérôme Glisse */ 43704f3f2cSJérôme Glisse static struct hmm *hmm_get_or_create(struct mm_struct *mm) 44704f3f2cSJérôme Glisse { 458a9320b7SJason Gunthorpe struct hmm *hmm; 46133ff0eaSJérôme Glisse 47*fec88ab0SLinus Torvalds lockdep_assert_held_write(&mm->mmap_sem); 488a9320b7SJason Gunthorpe 498a9320b7SJason Gunthorpe /* Abuse the page_table_lock to also protect mm->hmm. */ 508a9320b7SJason Gunthorpe spin_lock(&mm->page_table_lock); 518a9320b7SJason Gunthorpe hmm = mm->hmm; 528a9320b7SJason Gunthorpe if (mm->hmm && kref_get_unless_zero(&mm->hmm->kref)) 538a9320b7SJason Gunthorpe goto out_unlock; 548a9320b7SJason Gunthorpe spin_unlock(&mm->page_table_lock); 55c0b12405SJérôme Glisse 56c0b12405SJérôme Glisse hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); 57c0b12405SJérôme Glisse if (!hmm) 58c0b12405SJérôme Glisse return NULL; 59a3e0d41cSJérôme Glisse init_waitqueue_head(&hmm->wq); 60c0b12405SJérôme Glisse INIT_LIST_HEAD(&hmm->mirrors); 61c0b12405SJérôme Glisse init_rwsem(&hmm->mirrors_sem); 62c0b12405SJérôme Glisse hmm->mmu_notifier.ops = NULL; 63da4c3c73SJérôme Glisse INIT_LIST_HEAD(&hmm->ranges); 645a136b4aSJason Gunthorpe spin_lock_init(&hmm->ranges_lock); 65704f3f2cSJérôme Glisse kref_init(&hmm->kref); 66a3e0d41cSJérôme Glisse hmm->notifiers = 0; 67c0b12405SJérôme Glisse hmm->mm = mm; 68c0b12405SJérôme Glisse 6986a2d598SRalph Campbell hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; 708a9320b7SJason Gunthorpe if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) { 7186a2d598SRalph Campbell kfree(hmm); 7286a2d598SRalph Campbell return NULL; 73133ff0eaSJérôme Glisse } 74133ff0eaSJérôme Glisse 758a9320b7SJason Gunthorpe mmgrab(hmm->mm); 768a9320b7SJason Gunthorpe 778a9320b7SJason Gunthorpe /* 788a9320b7SJason Gunthorpe * We hold the exclusive mmap_sem here so we know that mm->hmm is 798a9320b7SJason Gunthorpe * still NULL or 0 kref, and is safe to update. 808a9320b7SJason Gunthorpe */ 818a9320b7SJason Gunthorpe spin_lock(&mm->page_table_lock); 828a9320b7SJason Gunthorpe mm->hmm = hmm; 838a9320b7SJason Gunthorpe 848a9320b7SJason Gunthorpe out_unlock: 858a9320b7SJason Gunthorpe spin_unlock(&mm->page_table_lock); 868a9320b7SJason Gunthorpe return hmm; 878a9320b7SJason Gunthorpe } 888a9320b7SJason Gunthorpe 896d7c3cdeSJason Gunthorpe static void hmm_free_rcu(struct rcu_head *rcu) 906d7c3cdeSJason Gunthorpe { 918a9320b7SJason Gunthorpe struct hmm *hmm = container_of(rcu, struct hmm, rcu); 928a9320b7SJason Gunthorpe 938a9320b7SJason Gunthorpe mmdrop(hmm->mm); 948a9320b7SJason Gunthorpe kfree(hmm); 956d7c3cdeSJason Gunthorpe } 966d7c3cdeSJason Gunthorpe 97704f3f2cSJérôme Glisse static void hmm_free(struct kref *kref) 98704f3f2cSJérôme Glisse { 99704f3f2cSJérôme Glisse struct hmm *hmm = container_of(kref, struct hmm, kref); 100704f3f2cSJérôme Glisse 1018a9320b7SJason Gunthorpe spin_lock(&hmm->mm->page_table_lock); 1028a9320b7SJason Gunthorpe if (hmm->mm->hmm == hmm) 1038a9320b7SJason Gunthorpe hmm->mm->hmm = NULL; 1048a9320b7SJason Gunthorpe spin_unlock(&hmm->mm->page_table_lock); 105704f3f2cSJérôme Glisse 1068a9320b7SJason Gunthorpe mmu_notifier_unregister_no_release(&hmm->mmu_notifier, hmm->mm); 1076d7c3cdeSJason Gunthorpe mmu_notifier_call_srcu(&hmm->rcu, hmm_free_rcu); 108704f3f2cSJérôme Glisse } 109704f3f2cSJérôme Glisse 110704f3f2cSJérôme Glisse static inline void hmm_put(struct hmm *hmm) 111704f3f2cSJérôme Glisse { 112704f3f2cSJérôme Glisse kref_put(&hmm->kref, hmm_free); 113704f3f2cSJérôme Glisse } 114704f3f2cSJérôme Glisse 115a3e0d41cSJérôme Glisse static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm) 116c0b12405SJérôme Glisse { 1176d7c3cdeSJason Gunthorpe struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier); 118c0b12405SJérôme Glisse struct hmm_mirror *mirror; 119da4c3c73SJérôme Glisse 1206d7c3cdeSJason Gunthorpe /* Bail out if hmm is in the process of being freed */ 1216d7c3cdeSJason Gunthorpe if (!kref_get_unless_zero(&hmm->kref)) 1226d7c3cdeSJason Gunthorpe return; 123da4c3c73SJérôme Glisse 124e1401513SRalph Campbell /* 12547f24598SJason Gunthorpe * Since hmm_range_register() holds the mmget() lock hmm_release() is 12647f24598SJason Gunthorpe * prevented as long as a range exists. 127e1401513SRalph Campbell */ 12847f24598SJason Gunthorpe WARN_ON(!list_empty_careful(&hmm->ranges)); 129704f3f2cSJérôme Glisse 130a3e0d41cSJérôme Glisse down_read(&hmm->mirrors_sem); 131a3e0d41cSJérôme Glisse list_for_each_entry(mirror, &hmm->mirrors, list) { 132e1401513SRalph Campbell /* 13314331726SJason Gunthorpe * Note: The driver is not allowed to trigger 13414331726SJason Gunthorpe * hmm_mirror_unregister() from this thread. 135e1401513SRalph Campbell */ 13614331726SJason Gunthorpe if (mirror->ops->release) 137e1401513SRalph Campbell mirror->ops->release(mirror); 138a3e0d41cSJérôme Glisse } 139a3e0d41cSJérôme Glisse up_read(&hmm->mirrors_sem); 140a3e0d41cSJérôme Glisse 141704f3f2cSJérôme Glisse hmm_put(hmm); 142c0b12405SJérôme Glisse } 143c0b12405SJérôme Glisse 1445a136b4aSJason Gunthorpe static void notifiers_decrement(struct hmm *hmm) 145c0b12405SJérôme Glisse { 1465a136b4aSJason Gunthorpe unsigned long flags; 147c0b12405SJérôme Glisse 1485a136b4aSJason Gunthorpe spin_lock_irqsave(&hmm->ranges_lock, flags); 149a3e0d41cSJérôme Glisse hmm->notifiers--; 150a3e0d41cSJérôme Glisse if (!hmm->notifiers) { 151a3e0d41cSJérôme Glisse struct hmm_range *range; 152a3e0d41cSJérôme Glisse 153a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 154a3e0d41cSJérôme Glisse if (range->valid) 155a3e0d41cSJérôme Glisse continue; 156a3e0d41cSJérôme Glisse range->valid = true; 157a3e0d41cSJérôme Glisse } 158a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 159a3e0d41cSJérôme Glisse } 1605a136b4aSJason Gunthorpe spin_unlock_irqrestore(&hmm->ranges_lock, flags); 1615a136b4aSJason Gunthorpe } 162a3e0d41cSJérôme Glisse 163133ff0eaSJérôme Glisse static int hmm_invalidate_range_start(struct mmu_notifier *mn, 164133ff0eaSJérôme Glisse const struct mmu_notifier_range *nrange) 165133ff0eaSJérôme Glisse { 1666d7c3cdeSJason Gunthorpe struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier); 167133ff0eaSJérôme Glisse struct hmm_mirror *mirror; 168133ff0eaSJérôme Glisse struct hmm_update update; 169133ff0eaSJérôme Glisse struct hmm_range *range; 1705a136b4aSJason Gunthorpe unsigned long flags; 171133ff0eaSJérôme Glisse int ret = 0; 172133ff0eaSJérôme Glisse 1736d7c3cdeSJason Gunthorpe if (!kref_get_unless_zero(&hmm->kref)) 1746d7c3cdeSJason Gunthorpe return 0; 175133ff0eaSJérôme Glisse 176133ff0eaSJérôme Glisse update.start = nrange->start; 177133ff0eaSJérôme Glisse update.end = nrange->end; 178133ff0eaSJérôme Glisse update.event = HMM_UPDATE_INVALIDATE; 179133ff0eaSJérôme Glisse update.blockable = mmu_notifier_range_blockable(nrange); 180133ff0eaSJérôme Glisse 1815a136b4aSJason Gunthorpe spin_lock_irqsave(&hmm->ranges_lock, flags); 182133ff0eaSJérôme Glisse hmm->notifiers++; 183133ff0eaSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 184133ff0eaSJérôme Glisse if (update.end < range->start || update.start >= range->end) 185133ff0eaSJérôme Glisse continue; 186133ff0eaSJérôme Glisse 187133ff0eaSJérôme Glisse range->valid = false; 188133ff0eaSJérôme Glisse } 1895a136b4aSJason Gunthorpe spin_unlock_irqrestore(&hmm->ranges_lock, flags); 190c0b12405SJérôme Glisse 191c0b12405SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 192c0b12405SJérôme Glisse down_read(&hmm->mirrors_sem); 193c0b12405SJérôme Glisse else if (!down_read_trylock(&hmm->mirrors_sem)) { 194c0b12405SJérôme Glisse ret = -EAGAIN; 195c0b12405SJérôme Glisse goto out; 196c0b12405SJérôme Glisse } 197c0b12405SJérôme Glisse 1985a136b4aSJason Gunthorpe list_for_each_entry(mirror, &hmm->mirrors, list) { 1995a136b4aSJason Gunthorpe int rc; 2005a136b4aSJason Gunthorpe 2015a136b4aSJason Gunthorpe rc = mirror->ops->sync_cpu_device_pagetables(mirror, &update); 2025a136b4aSJason Gunthorpe if (rc) { 2035a136b4aSJason Gunthorpe if (WARN_ON(update.blockable || rc != -EAGAIN)) 2045a136b4aSJason Gunthorpe continue; 2055a136b4aSJason Gunthorpe ret = -EAGAIN; 206085ea250SRalph Campbell break; 207c0b12405SJérôme Glisse } 2085a136b4aSJason Gunthorpe } 209c0b12405SJérôme Glisse up_read(&hmm->mirrors_sem); 210c0b12405SJérôme Glisse 211c0b12405SJérôme Glisse out: 2125a136b4aSJason Gunthorpe if (ret) 2135a136b4aSJason Gunthorpe notifiers_decrement(hmm); 214c0b12405SJérôme Glisse hmm_put(hmm); 215c0b12405SJérôme Glisse return ret; 216c0b12405SJérôme Glisse } 217c0b12405SJérôme Glisse 218c0b12405SJérôme Glisse static void hmm_invalidate_range_end(struct mmu_notifier *mn, 219c0b12405SJérôme Glisse const struct mmu_notifier_range *nrange) 220c0b12405SJérôme Glisse { 2216d7c3cdeSJason Gunthorpe struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier); 222c0b12405SJérôme Glisse 2236d7c3cdeSJason Gunthorpe if (!kref_get_unless_zero(&hmm->kref)) 2246d7c3cdeSJason Gunthorpe return; 225c0b12405SJérôme Glisse 2265a136b4aSJason Gunthorpe notifiers_decrement(hmm); 227704f3f2cSJérôme Glisse hmm_put(hmm); 228c0b12405SJérôme Glisse } 229c0b12405SJérôme Glisse 230c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { 231e1401513SRalph Campbell .release = hmm_release, 232c0b12405SJérôme Glisse .invalidate_range_start = hmm_invalidate_range_start, 233c0b12405SJérôme Glisse .invalidate_range_end = hmm_invalidate_range_end, 234c0b12405SJérôme Glisse }; 235c0b12405SJérôme Glisse 236c0b12405SJérôme Glisse /* 237c0b12405SJérôme Glisse * hmm_mirror_register() - register a mirror against an mm 238c0b12405SJérôme Glisse * 239c0b12405SJérôme Glisse * @mirror: new mirror struct to register 240c0b12405SJérôme Glisse * @mm: mm to register against 241085ea250SRalph Campbell * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments 242c0b12405SJérôme Glisse * 243c0b12405SJérôme Glisse * To start mirroring a process address space, the device driver must register 244c0b12405SJérôme Glisse * an HMM mirror struct. 245c0b12405SJérôme Glisse */ 246c0b12405SJérôme Glisse int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) 247c0b12405SJérôme Glisse { 248*fec88ab0SLinus Torvalds lockdep_assert_held_write(&mm->mmap_sem); 2498a1a0cd0SJason Gunthorpe 250c0b12405SJérôme Glisse /* Sanity check */ 251c0b12405SJérôme Glisse if (!mm || !mirror || !mirror->ops) 252c0b12405SJérôme Glisse return -EINVAL; 253c0b12405SJérôme Glisse 254704f3f2cSJérôme Glisse mirror->hmm = hmm_get_or_create(mm); 255c0b12405SJérôme Glisse if (!mirror->hmm) 256c0b12405SJérôme Glisse return -ENOMEM; 257c0b12405SJérôme Glisse 258c0b12405SJérôme Glisse down_write(&mirror->hmm->mirrors_sem); 259c0b12405SJérôme Glisse list_add(&mirror->list, &mirror->hmm->mirrors); 260c0b12405SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 261c0b12405SJérôme Glisse 262c0b12405SJérôme Glisse return 0; 263c0b12405SJérôme Glisse } 264c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_register); 265c0b12405SJérôme Glisse 266c0b12405SJérôme Glisse /* 267c0b12405SJérôme Glisse * hmm_mirror_unregister() - unregister a mirror 268c0b12405SJérôme Glisse * 269085ea250SRalph Campbell * @mirror: mirror struct to unregister 270c0b12405SJérôme Glisse * 271c0b12405SJérôme Glisse * Stop mirroring a process address space, and cleanup. 272c0b12405SJérôme Glisse */ 273c0b12405SJérôme Glisse void hmm_mirror_unregister(struct hmm_mirror *mirror) 274c0b12405SJérôme Glisse { 275187229c2SJason Gunthorpe struct hmm *hmm = mirror->hmm; 276c01cbba2SJérôme Glisse 277c0b12405SJérôme Glisse down_write(&hmm->mirrors_sem); 27814331726SJason Gunthorpe list_del(&mirror->list); 279c0b12405SJérôme Glisse up_write(&hmm->mirrors_sem); 280704f3f2cSJérôme Glisse hmm_put(hmm); 281c0b12405SJérôme Glisse } 282c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_unregister); 283da4c3c73SJérôme Glisse 28474eee180SJérôme Glisse struct hmm_vma_walk { 28574eee180SJérôme Glisse struct hmm_range *range; 286992de9a8SJérôme Glisse struct dev_pagemap *pgmap; 28774eee180SJérôme Glisse unsigned long last; 28874eee180SJérôme Glisse bool fault; 28974eee180SJérôme Glisse bool block; 29074eee180SJérôme Glisse }; 29174eee180SJérôme Glisse 2922aee09d8SJérôme Glisse static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr, 2932aee09d8SJérôme Glisse bool write_fault, uint64_t *pfn) 29474eee180SJérôme Glisse { 2959b1ae605SKuehling, Felix unsigned int flags = FAULT_FLAG_REMOTE; 29674eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 297f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 29874eee180SJérôme Glisse struct vm_area_struct *vma = walk->vma; 29950a7ca3cSSouptick Joarder vm_fault_t ret; 30074eee180SJérôme Glisse 30174eee180SJérôme Glisse flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; 3022aee09d8SJérôme Glisse flags |= write_fault ? FAULT_FLAG_WRITE : 0; 30350a7ca3cSSouptick Joarder ret = handle_mm_fault(vma, addr, flags); 30450a7ca3cSSouptick Joarder if (ret & VM_FAULT_RETRY) 30573231612SJérôme Glisse return -EAGAIN; 30650a7ca3cSSouptick Joarder if (ret & VM_FAULT_ERROR) { 307f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 30874eee180SJérôme Glisse return -EFAULT; 30974eee180SJérôme Glisse } 31074eee180SJérôme Glisse 31173231612SJérôme Glisse return -EBUSY; 31274eee180SJérôme Glisse } 31374eee180SJérôme Glisse 314da4c3c73SJérôme Glisse static int hmm_pfns_bad(unsigned long addr, 315da4c3c73SJérôme Glisse unsigned long end, 316da4c3c73SJérôme Glisse struct mm_walk *walk) 317da4c3c73SJérôme Glisse { 318c719547fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 319c719547fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 320ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 321da4c3c73SJérôme Glisse unsigned long i; 322da4c3c73SJérôme Glisse 323da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 324da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) 325f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_ERROR]; 326da4c3c73SJérôme Glisse 327da4c3c73SJérôme Glisse return 0; 328da4c3c73SJérôme Glisse } 329da4c3c73SJérôme Glisse 3305504ed29SJérôme Glisse /* 3315504ed29SJérôme Glisse * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s) 3325504ed29SJérôme Glisse * @start: range virtual start address (inclusive) 3335504ed29SJérôme Glisse * @end: range virtual end address (exclusive) 3342aee09d8SJérôme Glisse * @fault: should we fault or not ? 3352aee09d8SJérôme Glisse * @write_fault: write fault ? 3365504ed29SJérôme Glisse * @walk: mm_walk structure 337085ea250SRalph Campbell * Return: 0 on success, -EBUSY after page fault, or page fault error 3385504ed29SJérôme Glisse * 3395504ed29SJérôme Glisse * This function will be called whenever pmd_none() or pte_none() returns true, 3405504ed29SJérôme Glisse * or whenever there is no page directory covering the virtual address range. 3415504ed29SJérôme Glisse */ 3422aee09d8SJérôme Glisse static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end, 3432aee09d8SJérôme Glisse bool fault, bool write_fault, 344da4c3c73SJérôme Glisse struct mm_walk *walk) 345da4c3c73SJérôme Glisse { 34674eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 34774eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 348ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 34963d5066fSJérôme Glisse unsigned long i, page_size; 350da4c3c73SJérôme Glisse 35174eee180SJérôme Glisse hmm_vma_walk->last = addr; 35263d5066fSJérôme Glisse page_size = hmm_range_page_size(range); 35363d5066fSJérôme Glisse i = (addr - range->start) >> range->page_shift; 35463d5066fSJérôme Glisse 35563d5066fSJérôme Glisse for (; addr < end; addr += page_size, i++) { 356f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_NONE]; 3572aee09d8SJérôme Glisse if (fault || write_fault) { 35874eee180SJérôme Glisse int ret; 359da4c3c73SJérôme Glisse 3602aee09d8SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, write_fault, 3612aee09d8SJérôme Glisse &pfns[i]); 36273231612SJérôme Glisse if (ret != -EBUSY) 36374eee180SJérôme Glisse return ret; 36474eee180SJérôme Glisse } 36574eee180SJérôme Glisse } 36674eee180SJérôme Glisse 36773231612SJérôme Glisse return (fault || write_fault) ? -EBUSY : 0; 3682aee09d8SJérôme Glisse } 3692aee09d8SJérôme Glisse 3702aee09d8SJérôme Glisse static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 3712aee09d8SJérôme Glisse uint64_t pfns, uint64_t cpu_flags, 3722aee09d8SJérôme Glisse bool *fault, bool *write_fault) 3732aee09d8SJérôme Glisse { 374f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 375f88a1e90SJérôme Glisse 3762aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) 3772aee09d8SJérôme Glisse return; 3782aee09d8SJérôme Glisse 379023a019aSJérôme Glisse /* 380023a019aSJérôme Glisse * So we not only consider the individual per page request we also 381023a019aSJérôme Glisse * consider the default flags requested for the range. The API can 382023a019aSJérôme Glisse * be use in 2 fashions. The first one where the HMM user coalesce 383023a019aSJérôme Glisse * multiple page fault into one request and set flags per pfns for 384023a019aSJérôme Glisse * of those faults. The second one where the HMM user want to pre- 385023a019aSJérôme Glisse * fault a range with specific flags. For the latter one it is a 386023a019aSJérôme Glisse * waste to have the user pre-fill the pfn arrays with a default 387023a019aSJérôme Glisse * flags value. 388023a019aSJérôme Glisse */ 389023a019aSJérôme Glisse pfns = (pfns & range->pfn_flags_mask) | range->default_flags; 390023a019aSJérôme Glisse 3912aee09d8SJérôme Glisse /* We aren't ask to do anything ... */ 392f88a1e90SJérôme Glisse if (!(pfns & range->flags[HMM_PFN_VALID])) 3932aee09d8SJérôme Glisse return; 394f88a1e90SJérôme Glisse /* If this is device memory than only fault if explicitly requested */ 395f88a1e90SJérôme Glisse if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) { 396f88a1e90SJérôme Glisse /* Do we fault on device memory ? */ 397f88a1e90SJérôme Glisse if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) { 398f88a1e90SJérôme Glisse *write_fault = pfns & range->flags[HMM_PFN_WRITE]; 399f88a1e90SJérôme Glisse *fault = true; 400f88a1e90SJérôme Glisse } 4012aee09d8SJérôme Glisse return; 4022aee09d8SJérôme Glisse } 403f88a1e90SJérôme Glisse 404f88a1e90SJérôme Glisse /* If CPU page table is not valid then we need to fault */ 405f88a1e90SJérôme Glisse *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]); 406f88a1e90SJérôme Glisse /* Need to write fault ? */ 407f88a1e90SJérôme Glisse if ((pfns & range->flags[HMM_PFN_WRITE]) && 408f88a1e90SJérôme Glisse !(cpu_flags & range->flags[HMM_PFN_WRITE])) { 409f88a1e90SJérôme Glisse *write_fault = true; 4102aee09d8SJérôme Glisse *fault = true; 4112aee09d8SJérôme Glisse } 4122aee09d8SJérôme Glisse } 4132aee09d8SJérôme Glisse 4142aee09d8SJérôme Glisse static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4152aee09d8SJérôme Glisse const uint64_t *pfns, unsigned long npages, 4162aee09d8SJérôme Glisse uint64_t cpu_flags, bool *fault, 4172aee09d8SJérôme Glisse bool *write_fault) 4182aee09d8SJérôme Glisse { 4192aee09d8SJérôme Glisse unsigned long i; 4202aee09d8SJérôme Glisse 4212aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) { 4222aee09d8SJérôme Glisse *fault = *write_fault = false; 4232aee09d8SJérôme Glisse return; 4242aee09d8SJérôme Glisse } 4252aee09d8SJérôme Glisse 426a3e0d41cSJérôme Glisse *fault = *write_fault = false; 4272aee09d8SJérôme Glisse for (i = 0; i < npages; ++i) { 4282aee09d8SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags, 4292aee09d8SJérôme Glisse fault, write_fault); 430a3e0d41cSJérôme Glisse if ((*write_fault)) 4312aee09d8SJérôme Glisse return; 4322aee09d8SJérôme Glisse } 4332aee09d8SJérôme Glisse } 4342aee09d8SJérôme Glisse 4352aee09d8SJérôme Glisse static int hmm_vma_walk_hole(unsigned long addr, unsigned long end, 4362aee09d8SJérôme Glisse struct mm_walk *walk) 4372aee09d8SJérôme Glisse { 4382aee09d8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 4392aee09d8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4402aee09d8SJérôme Glisse bool fault, write_fault; 4412aee09d8SJérôme Glisse unsigned long i, npages; 4422aee09d8SJérôme Glisse uint64_t *pfns; 4432aee09d8SJérôme Glisse 4442aee09d8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 4452aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 4462aee09d8SJérôme Glisse pfns = &range->pfns[i]; 4472aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 4482aee09d8SJérôme Glisse 0, &fault, &write_fault); 4492aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 4502aee09d8SJérôme Glisse } 4512aee09d8SJérôme Glisse 452f88a1e90SJérôme Glisse static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd) 4532aee09d8SJérôme Glisse { 4542aee09d8SJérôme Glisse if (pmd_protnone(pmd)) 4552aee09d8SJérôme Glisse return 0; 456f88a1e90SJérôme Glisse return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] | 457f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 458f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 459da4c3c73SJérôme Glisse } 460da4c3c73SJérôme Glisse 461992de9a8SJérôme Glisse static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud) 462992de9a8SJérôme Glisse { 463992de9a8SJérôme Glisse if (!pud_present(pud)) 464992de9a8SJérôme Glisse return 0; 465992de9a8SJérôme Glisse return pud_write(pud) ? range->flags[HMM_PFN_VALID] | 466992de9a8SJérôme Glisse range->flags[HMM_PFN_WRITE] : 467992de9a8SJérôme Glisse range->flags[HMM_PFN_VALID]; 468992de9a8SJérôme Glisse } 469992de9a8SJérôme Glisse 47053f5c3f4SJérôme Glisse static int hmm_vma_handle_pmd(struct mm_walk *walk, 47153f5c3f4SJérôme Glisse unsigned long addr, 47253f5c3f4SJérôme Glisse unsigned long end, 47353f5c3f4SJérôme Glisse uint64_t *pfns, 47453f5c3f4SJérôme Glisse pmd_t pmd) 47553f5c3f4SJérôme Glisse { 476992de9a8SJérôme Glisse #ifdef CONFIG_TRANSPARENT_HUGEPAGE 47753f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 478f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4792aee09d8SJérôme Glisse unsigned long pfn, npages, i; 4802aee09d8SJérôme Glisse bool fault, write_fault; 481f88a1e90SJérôme Glisse uint64_t cpu_flags; 48253f5c3f4SJérôme Glisse 4832aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 484f88a1e90SJérôme Glisse cpu_flags = pmd_to_hmm_pfn_flags(range, pmd); 4852aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags, 4862aee09d8SJérôme Glisse &fault, &write_fault); 48753f5c3f4SJérôme Glisse 4882aee09d8SJérôme Glisse if (pmd_protnone(pmd) || fault || write_fault) 4892aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 49053f5c3f4SJérôme Glisse 49153f5c3f4SJérôme Glisse pfn = pmd_pfn(pmd) + pte_index(addr); 492992de9a8SJérôme Glisse for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) { 493992de9a8SJérôme Glisse if (pmd_devmap(pmd)) { 494992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 495992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 496992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 497992de9a8SJérôme Glisse return -EBUSY; 498992de9a8SJérôme Glisse } 499391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags; 500992de9a8SJérôme Glisse } 501992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 502992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 503992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 504992de9a8SJérôme Glisse } 50553f5c3f4SJérôme Glisse hmm_vma_walk->last = end; 50653f5c3f4SJérôme Glisse return 0; 507992de9a8SJérôme Glisse #else 508992de9a8SJérôme Glisse /* If THP is not enabled then we should never reach that code ! */ 509992de9a8SJérôme Glisse return -EINVAL; 510992de9a8SJérôme Glisse #endif 51153f5c3f4SJérôme Glisse } 51253f5c3f4SJérôme Glisse 513f88a1e90SJérôme Glisse static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte) 5142aee09d8SJérôme Glisse { 515789c2af8SPhilip Yang if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte)) 5162aee09d8SJérôme Glisse return 0; 517f88a1e90SJérôme Glisse return pte_write(pte) ? range->flags[HMM_PFN_VALID] | 518f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 519f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 5202aee09d8SJérôme Glisse } 5212aee09d8SJérôme Glisse 52253f5c3f4SJérôme Glisse static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr, 52353f5c3f4SJérôme Glisse unsigned long end, pmd_t *pmdp, pte_t *ptep, 52453f5c3f4SJérôme Glisse uint64_t *pfn) 52553f5c3f4SJérôme Glisse { 52653f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 527f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 52853f5c3f4SJérôme Glisse struct vm_area_struct *vma = walk->vma; 5292aee09d8SJérôme Glisse bool fault, write_fault; 5302aee09d8SJérôme Glisse uint64_t cpu_flags; 53153f5c3f4SJérôme Glisse pte_t pte = *ptep; 532f88a1e90SJérôme Glisse uint64_t orig_pfn = *pfn; 53353f5c3f4SJérôme Glisse 534f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_NONE]; 53573231612SJérôme Glisse fault = write_fault = false; 53653f5c3f4SJérôme Glisse 53753f5c3f4SJérôme Glisse if (pte_none(pte)) { 53873231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0, 53973231612SJérôme Glisse &fault, &write_fault); 5402aee09d8SJérôme Glisse if (fault || write_fault) 54153f5c3f4SJérôme Glisse goto fault; 54253f5c3f4SJérôme Glisse return 0; 54353f5c3f4SJérôme Glisse } 54453f5c3f4SJérôme Glisse 54553f5c3f4SJérôme Glisse if (!pte_present(pte)) { 54653f5c3f4SJérôme Glisse swp_entry_t entry = pte_to_swp_entry(pte); 54753f5c3f4SJérôme Glisse 54853f5c3f4SJérôme Glisse if (!non_swap_entry(entry)) { 5492aee09d8SJérôme Glisse if (fault || write_fault) 55053f5c3f4SJérôme Glisse goto fault; 55153f5c3f4SJérôme Glisse return 0; 55253f5c3f4SJérôme Glisse } 55353f5c3f4SJérôme Glisse 55453f5c3f4SJérôme Glisse /* 55553f5c3f4SJérôme Glisse * This is a special swap entry, ignore migration, use 55653f5c3f4SJérôme Glisse * device and report anything else as error. 55753f5c3f4SJérôme Glisse */ 55853f5c3f4SJérôme Glisse if (is_device_private_entry(entry)) { 559f88a1e90SJérôme Glisse cpu_flags = range->flags[HMM_PFN_VALID] | 560f88a1e90SJérôme Glisse range->flags[HMM_PFN_DEVICE_PRIVATE]; 5612aee09d8SJérôme Glisse cpu_flags |= is_write_device_private_entry(entry) ? 562f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 0; 563f88a1e90SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 564f88a1e90SJérôme Glisse &fault, &write_fault); 565f88a1e90SJérôme Glisse if (fault || write_fault) 566f88a1e90SJérôme Glisse goto fault; 567391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, 568391aab11SJérôme Glisse swp_offset(entry)); 569f88a1e90SJérôme Glisse *pfn |= cpu_flags; 57053f5c3f4SJérôme Glisse return 0; 57153f5c3f4SJérôme Glisse } 57253f5c3f4SJérôme Glisse 57353f5c3f4SJérôme Glisse if (is_migration_entry(entry)) { 5742aee09d8SJérôme Glisse if (fault || write_fault) { 57553f5c3f4SJérôme Glisse pte_unmap(ptep); 57653f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 57753f5c3f4SJérôme Glisse migration_entry_wait(vma->vm_mm, 57853f5c3f4SJérôme Glisse pmdp, addr); 57973231612SJérôme Glisse return -EBUSY; 58053f5c3f4SJérôme Glisse } 58153f5c3f4SJérôme Glisse return 0; 58253f5c3f4SJérôme Glisse } 58353f5c3f4SJérôme Glisse 58453f5c3f4SJérôme Glisse /* Report error for everything else */ 585f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 58653f5c3f4SJérôme Glisse return -EFAULT; 58773231612SJérôme Glisse } else { 58873231612SJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, pte); 58973231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 59073231612SJérôme Glisse &fault, &write_fault); 59153f5c3f4SJérôme Glisse } 59253f5c3f4SJérôme Glisse 5932aee09d8SJérôme Glisse if (fault || write_fault) 59453f5c3f4SJérôme Glisse goto fault; 59553f5c3f4SJérôme Glisse 596992de9a8SJérôme Glisse if (pte_devmap(pte)) { 597992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte), 598992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 599992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 600992de9a8SJérôme Glisse return -EBUSY; 601992de9a8SJérôme Glisse } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) { 602992de9a8SJérôme Glisse *pfn = range->values[HMM_PFN_SPECIAL]; 603992de9a8SJérôme Glisse return -EFAULT; 604992de9a8SJérôme Glisse } 605992de9a8SJérôme Glisse 606391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags; 60753f5c3f4SJérôme Glisse return 0; 60853f5c3f4SJérôme Glisse 60953f5c3f4SJérôme Glisse fault: 610992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 611992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 612992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 613992de9a8SJérôme Glisse } 61453f5c3f4SJérôme Glisse pte_unmap(ptep); 61553f5c3f4SJérôme Glisse /* Fault any virtual address we were asked to fault */ 6162aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 61753f5c3f4SJérôme Glisse } 61853f5c3f4SJérôme Glisse 619da4c3c73SJérôme Glisse static int hmm_vma_walk_pmd(pmd_t *pmdp, 620da4c3c73SJérôme Glisse unsigned long start, 621da4c3c73SJérôme Glisse unsigned long end, 622da4c3c73SJérôme Glisse struct mm_walk *walk) 623da4c3c73SJérôme Glisse { 62474eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 62574eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 626d08faca0SJérôme Glisse struct vm_area_struct *vma = walk->vma; 627ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 628da4c3c73SJérôme Glisse unsigned long addr = start, i; 629da4c3c73SJérôme Glisse pte_t *ptep; 630da4c3c73SJérôme Glisse pmd_t pmd; 631da4c3c73SJérôme Glisse 632d08faca0SJérôme Glisse 633d08faca0SJérôme Glisse again: 634d08faca0SJérôme Glisse pmd = READ_ONCE(*pmdp); 635d08faca0SJérôme Glisse if (pmd_none(pmd)) 636d08faca0SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 637d08faca0SJérôme Glisse 638d08faca0SJérôme Glisse if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB)) 639d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 640d08faca0SJérôme Glisse 641d08faca0SJérôme Glisse if (thp_migration_supported() && is_pmd_migration_entry(pmd)) { 642d08faca0SJérôme Glisse bool fault, write_fault; 643d08faca0SJérôme Glisse unsigned long npages; 644d08faca0SJérôme Glisse uint64_t *pfns; 645d08faca0SJérôme Glisse 646d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 647d08faca0SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 648d08faca0SJérôme Glisse pfns = &range->pfns[i]; 649d08faca0SJérôme Glisse 650d08faca0SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 651d08faca0SJérôme Glisse 0, &fault, &write_fault); 652d08faca0SJérôme Glisse if (fault || write_fault) { 653d08faca0SJérôme Glisse hmm_vma_walk->last = addr; 654d08faca0SJérôme Glisse pmd_migration_entry_wait(vma->vm_mm, pmdp); 65573231612SJérôme Glisse return -EBUSY; 656d08faca0SJérôme Glisse } 657d08faca0SJérôme Glisse return 0; 658d08faca0SJérôme Glisse } else if (!pmd_present(pmd)) 659d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 660d08faca0SJérôme Glisse 661d08faca0SJérôme Glisse if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) { 662da4c3c73SJérôme Glisse /* 663da4c3c73SJérôme Glisse * No need to take pmd_lock here, even if some other threads 664da4c3c73SJérôme Glisse * is splitting the huge pmd we will get that event through 665da4c3c73SJérôme Glisse * mmu_notifier callback. 666da4c3c73SJérôme Glisse * 667da4c3c73SJérôme Glisse * So just read pmd value and check again its a transparent 668da4c3c73SJérôme Glisse * huge or device mapping one and compute corresponding pfn 669da4c3c73SJérôme Glisse * values. 670da4c3c73SJérôme Glisse */ 671da4c3c73SJérôme Glisse pmd = pmd_read_atomic(pmdp); 672da4c3c73SJérôme Glisse barrier(); 673da4c3c73SJérôme Glisse if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) 674da4c3c73SJérôme Glisse goto again; 675da4c3c73SJérôme Glisse 676d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 67753f5c3f4SJérôme Glisse return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd); 678da4c3c73SJérôme Glisse } 679da4c3c73SJérôme Glisse 680d08faca0SJérôme Glisse /* 681d08faca0SJérôme Glisse * We have handled all the valid case above ie either none, migration, 682d08faca0SJérôme Glisse * huge or transparent huge. At this point either it is a valid pmd 683d08faca0SJérôme Glisse * entry pointing to pte directory or it is a bad pmd that will not 684d08faca0SJérôme Glisse * recover. 685d08faca0SJérôme Glisse */ 686d08faca0SJérôme Glisse if (pmd_bad(pmd)) 687da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 688da4c3c73SJérôme Glisse 689da4c3c73SJérôme Glisse ptep = pte_offset_map(pmdp, addr); 690d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 691da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { 69253f5c3f4SJérôme Glisse int r; 693da4c3c73SJérôme Glisse 69453f5c3f4SJérôme Glisse r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]); 69553f5c3f4SJérôme Glisse if (r) { 69653f5c3f4SJérôme Glisse /* hmm_vma_handle_pte() did unmap pte directory */ 69774eee180SJérôme Glisse hmm_vma_walk->last = addr; 69853f5c3f4SJérôme Glisse return r; 69974eee180SJérôme Glisse } 700da4c3c73SJérôme Glisse } 701992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 702992de9a8SJérôme Glisse /* 703992de9a8SJérôme Glisse * We do put_dev_pagemap() here and not in hmm_vma_handle_pte() 704992de9a8SJérôme Glisse * so that we can leverage get_dev_pagemap() optimization which 705992de9a8SJérôme Glisse * will not re-take a reference on a pgmap if we already have 706992de9a8SJérôme Glisse * one. 707992de9a8SJérôme Glisse */ 708992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 709992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 710992de9a8SJérôme Glisse } 711da4c3c73SJérôme Glisse pte_unmap(ptep - 1); 712da4c3c73SJérôme Glisse 71353f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 714da4c3c73SJérôme Glisse return 0; 715da4c3c73SJérôme Glisse } 716da4c3c73SJérôme Glisse 717992de9a8SJérôme Glisse static int hmm_vma_walk_pud(pud_t *pudp, 718992de9a8SJérôme Glisse unsigned long start, 719992de9a8SJérôme Glisse unsigned long end, 720992de9a8SJérôme Glisse struct mm_walk *walk) 721992de9a8SJérôme Glisse { 722992de9a8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 723992de9a8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 724992de9a8SJérôme Glisse unsigned long addr = start, next; 725992de9a8SJérôme Glisse pmd_t *pmdp; 726992de9a8SJérôme Glisse pud_t pud; 727992de9a8SJérôme Glisse int ret; 728992de9a8SJérôme Glisse 729992de9a8SJérôme Glisse again: 730992de9a8SJérôme Glisse pud = READ_ONCE(*pudp); 731992de9a8SJérôme Glisse if (pud_none(pud)) 732992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 733992de9a8SJérôme Glisse 734992de9a8SJérôme Glisse if (pud_huge(pud) && pud_devmap(pud)) { 735992de9a8SJérôme Glisse unsigned long i, npages, pfn; 736992de9a8SJérôme Glisse uint64_t *pfns, cpu_flags; 737992de9a8SJérôme Glisse bool fault, write_fault; 738992de9a8SJérôme Glisse 739992de9a8SJérôme Glisse if (!pud_present(pud)) 740992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 741992de9a8SJérôme Glisse 742992de9a8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 743992de9a8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 744992de9a8SJérôme Glisse pfns = &range->pfns[i]; 745992de9a8SJérôme Glisse 746992de9a8SJérôme Glisse cpu_flags = pud_to_hmm_pfn_flags(range, pud); 747992de9a8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 748992de9a8SJérôme Glisse cpu_flags, &fault, &write_fault); 749992de9a8SJérôme Glisse if (fault || write_fault) 750992de9a8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, 751992de9a8SJérôme Glisse write_fault, walk); 752992de9a8SJérôme Glisse 753992de9a8SJérôme Glisse pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); 754992de9a8SJérôme Glisse for (i = 0; i < npages; ++i, ++pfn) { 755992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 756992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 757992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 758992de9a8SJérôme Glisse return -EBUSY; 759391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 760391aab11SJérôme Glisse cpu_flags; 761992de9a8SJérôme Glisse } 762992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 763992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 764992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 765992de9a8SJérôme Glisse } 766992de9a8SJérôme Glisse hmm_vma_walk->last = end; 767992de9a8SJérôme Glisse return 0; 768992de9a8SJérôme Glisse } 769992de9a8SJérôme Glisse 770992de9a8SJérôme Glisse split_huge_pud(walk->vma, pudp, addr); 771992de9a8SJérôme Glisse if (pud_none(*pudp)) 772992de9a8SJérôme Glisse goto again; 773992de9a8SJérôme Glisse 774992de9a8SJérôme Glisse pmdp = pmd_offset(pudp, addr); 775992de9a8SJérôme Glisse do { 776992de9a8SJérôme Glisse next = pmd_addr_end(addr, end); 777992de9a8SJérôme Glisse ret = hmm_vma_walk_pmd(pmdp, addr, next, walk); 778992de9a8SJérôme Glisse if (ret) 779992de9a8SJérôme Glisse return ret; 780992de9a8SJérôme Glisse } while (pmdp++, addr = next, addr != end); 781992de9a8SJérôme Glisse 782992de9a8SJérôme Glisse return 0; 783992de9a8SJérôme Glisse } 784992de9a8SJérôme Glisse 78563d5066fSJérôme Glisse static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask, 78663d5066fSJérôme Glisse unsigned long start, unsigned long end, 78763d5066fSJérôme Glisse struct mm_walk *walk) 78863d5066fSJérôme Glisse { 78963d5066fSJérôme Glisse #ifdef CONFIG_HUGETLB_PAGE 79063d5066fSJérôme Glisse unsigned long addr = start, i, pfn, mask, size, pfn_inc; 79163d5066fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 79263d5066fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 79363d5066fSJérôme Glisse struct vm_area_struct *vma = walk->vma; 79463d5066fSJérôme Glisse struct hstate *h = hstate_vma(vma); 79563d5066fSJérôme Glisse uint64_t orig_pfn, cpu_flags; 79663d5066fSJérôme Glisse bool fault, write_fault; 79763d5066fSJérôme Glisse spinlock_t *ptl; 79863d5066fSJérôme Glisse pte_t entry; 79963d5066fSJérôme Glisse int ret = 0; 80063d5066fSJérôme Glisse 80163d5066fSJérôme Glisse size = 1UL << huge_page_shift(h); 80263d5066fSJérôme Glisse mask = size - 1; 80363d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) { 80463d5066fSJérôme Glisse /* Make sure we are looking at full page. */ 80563d5066fSJérôme Glisse if (start & mask) 80663d5066fSJérôme Glisse return -EINVAL; 80763d5066fSJérôme Glisse if (end < (start + size)) 80863d5066fSJérôme Glisse return -EINVAL; 80963d5066fSJérôme Glisse pfn_inc = size >> PAGE_SHIFT; 81063d5066fSJérôme Glisse } else { 81163d5066fSJérôme Glisse pfn_inc = 1; 81263d5066fSJérôme Glisse size = PAGE_SIZE; 81363d5066fSJérôme Glisse } 81463d5066fSJérôme Glisse 81563d5066fSJérôme Glisse 81663d5066fSJérôme Glisse ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); 81763d5066fSJérôme Glisse entry = huge_ptep_get(pte); 81863d5066fSJérôme Glisse 81963d5066fSJérôme Glisse i = (start - range->start) >> range->page_shift; 82063d5066fSJérôme Glisse orig_pfn = range->pfns[i]; 82163d5066fSJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 82263d5066fSJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, entry); 82363d5066fSJérôme Glisse fault = write_fault = false; 82463d5066fSJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 82563d5066fSJérôme Glisse &fault, &write_fault); 82663d5066fSJérôme Glisse if (fault || write_fault) { 82763d5066fSJérôme Glisse ret = -ENOENT; 82863d5066fSJérôme Glisse goto unlock; 82963d5066fSJérôme Glisse } 83063d5066fSJérôme Glisse 83163d5066fSJérôme Glisse pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift); 83263d5066fSJérôme Glisse for (; addr < end; addr += size, i++, pfn += pfn_inc) 833391aab11SJérôme Glisse range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 834391aab11SJérôme Glisse cpu_flags; 83563d5066fSJérôme Glisse hmm_vma_walk->last = end; 83663d5066fSJérôme Glisse 83763d5066fSJérôme Glisse unlock: 83863d5066fSJérôme Glisse spin_unlock(ptl); 83963d5066fSJérôme Glisse 84063d5066fSJérôme Glisse if (ret == -ENOENT) 84163d5066fSJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 84263d5066fSJérôme Glisse 84363d5066fSJérôme Glisse return ret; 84463d5066fSJérôme Glisse #else /* CONFIG_HUGETLB_PAGE */ 84563d5066fSJérôme Glisse return -EINVAL; 84663d5066fSJérôme Glisse #endif 84763d5066fSJérôme Glisse } 84863d5066fSJérôme Glisse 849f88a1e90SJérôme Glisse static void hmm_pfns_clear(struct hmm_range *range, 850f88a1e90SJérôme Glisse uint64_t *pfns, 85133cd47dcSJérôme Glisse unsigned long addr, 85233cd47dcSJérôme Glisse unsigned long end) 85333cd47dcSJérôme Glisse { 85433cd47dcSJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 855f88a1e90SJérôme Glisse *pfns = range->values[HMM_PFN_NONE]; 85633cd47dcSJérôme Glisse } 85733cd47dcSJérôme Glisse 858da4c3c73SJérôme Glisse /* 859a3e0d41cSJérôme Glisse * hmm_range_register() - start tracking change to CPU page table over a range 860a3e0d41cSJérôme Glisse * @range: range 861a3e0d41cSJérôme Glisse * @mm: the mm struct for the range of virtual address 862a3e0d41cSJérôme Glisse * @start: start virtual address (inclusive) 863a3e0d41cSJérôme Glisse * @end: end virtual address (exclusive) 86463d5066fSJérôme Glisse * @page_shift: expect page shift for the range 865a3e0d41cSJérôme Glisse * Returns 0 on success, -EFAULT if the address space is no longer valid 866a3e0d41cSJérôme Glisse * 867a3e0d41cSJérôme Glisse * Track updates to the CPU page table see include/linux/hmm.h 868a3e0d41cSJérôme Glisse */ 869a3e0d41cSJérôme Glisse int hmm_range_register(struct hmm_range *range, 870e36acfe6SJason Gunthorpe struct hmm_mirror *mirror, 871a3e0d41cSJérôme Glisse unsigned long start, 87263d5066fSJérôme Glisse unsigned long end, 87363d5066fSJérôme Glisse unsigned page_shift) 874a3e0d41cSJérôme Glisse { 87563d5066fSJérôme Glisse unsigned long mask = ((1UL << page_shift) - 1UL); 876e36acfe6SJason Gunthorpe struct hmm *hmm = mirror->hmm; 8775a136b4aSJason Gunthorpe unsigned long flags; 87863d5066fSJérôme Glisse 879a3e0d41cSJérôme Glisse range->valid = false; 880a3e0d41cSJérôme Glisse range->hmm = NULL; 881a3e0d41cSJérôme Glisse 88263d5066fSJérôme Glisse if ((start & mask) || (end & mask)) 88363d5066fSJérôme Glisse return -EINVAL; 88463d5066fSJérôme Glisse if (start >= end) 885a3e0d41cSJérôme Glisse return -EINVAL; 886a3e0d41cSJérôme Glisse 88763d5066fSJérôme Glisse range->page_shift = page_shift; 888a3e0d41cSJérôme Glisse range->start = start; 889a3e0d41cSJérôme Glisse range->end = end; 890a3e0d41cSJérôme Glisse 89147f24598SJason Gunthorpe /* Prevent hmm_release() from running while the range is valid */ 89247f24598SJason Gunthorpe if (!mmget_not_zero(hmm->mm)) 893a3e0d41cSJérôme Glisse return -EFAULT; 894a3e0d41cSJérôme Glisse 895085ea250SRalph Campbell /* Initialize range to track CPU page table updates. */ 8965a136b4aSJason Gunthorpe spin_lock_irqsave(&hmm->ranges_lock, flags); 897a3e0d41cSJérôme Glisse 898085ea250SRalph Campbell range->hmm = hmm; 899e36acfe6SJason Gunthorpe kref_get(&hmm->kref); 900157816f3SJason Gunthorpe list_add(&range->list, &hmm->ranges); 901a3e0d41cSJérôme Glisse 902a3e0d41cSJérôme Glisse /* 903a3e0d41cSJérôme Glisse * If there are any concurrent notifiers we have to wait for them for 904a3e0d41cSJérôme Glisse * the range to be valid (see hmm_range_wait_until_valid()). 905a3e0d41cSJérôme Glisse */ 906085ea250SRalph Campbell if (!hmm->notifiers) 907a3e0d41cSJérôme Glisse range->valid = true; 9085a136b4aSJason Gunthorpe spin_unlock_irqrestore(&hmm->ranges_lock, flags); 909a3e0d41cSJérôme Glisse 910a3e0d41cSJérôme Glisse return 0; 911a3e0d41cSJérôme Glisse } 912a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_register); 913a3e0d41cSJérôme Glisse 914a3e0d41cSJérôme Glisse /* 915a3e0d41cSJérôme Glisse * hmm_range_unregister() - stop tracking change to CPU page table over a range 916a3e0d41cSJérôme Glisse * @range: range 917a3e0d41cSJérôme Glisse * 918a3e0d41cSJérôme Glisse * Range struct is used to track updates to the CPU page table after a call to 919a3e0d41cSJérôme Glisse * hmm_range_register(). See include/linux/hmm.h for how to use it. 920a3e0d41cSJérôme Glisse */ 921a3e0d41cSJérôme Glisse void hmm_range_unregister(struct hmm_range *range) 922a3e0d41cSJérôme Glisse { 923085ea250SRalph Campbell struct hmm *hmm = range->hmm; 9245a136b4aSJason Gunthorpe unsigned long flags; 925a3e0d41cSJérôme Glisse 9265a136b4aSJason Gunthorpe spin_lock_irqsave(&hmm->ranges_lock, flags); 92747f24598SJason Gunthorpe list_del_init(&range->list); 9285a136b4aSJason Gunthorpe spin_unlock_irqrestore(&hmm->ranges_lock, flags); 929a3e0d41cSJérôme Glisse 930a3e0d41cSJérôme Glisse /* Drop reference taken by hmm_range_register() */ 93147f24598SJason Gunthorpe mmput(hmm->mm); 932085ea250SRalph Campbell hmm_put(hmm); 9332dcc3eb8SJason Gunthorpe 9342dcc3eb8SJason Gunthorpe /* 9352dcc3eb8SJason Gunthorpe * The range is now invalid and the ref on the hmm is dropped, so 9362dcc3eb8SJason Gunthorpe * poison the pointer. Leave other fields in place, for the caller's 9372dcc3eb8SJason Gunthorpe * use. 9382dcc3eb8SJason Gunthorpe */ 939a3e0d41cSJérôme Glisse range->valid = false; 9402dcc3eb8SJason Gunthorpe memset(&range->hmm, POISON_INUSE, sizeof(range->hmm)); 941a3e0d41cSJérôme Glisse } 942a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_unregister); 943a3e0d41cSJérôme Glisse 944a3e0d41cSJérôme Glisse /* 94525f23a0cSJérôme Glisse * hmm_range_snapshot() - snapshot CPU page table for a range 94625f23a0cSJérôme Glisse * @range: range 947085ea250SRalph Campbell * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid 948a3e0d41cSJérôme Glisse * permission (for instance asking for write and range is read only), 949a3e0d41cSJérôme Glisse * -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid 950a3e0d41cSJérôme Glisse * vma or it is illegal to access that range), number of valid pages 951a3e0d41cSJérôme Glisse * in range->pfns[] (from range start address). 952da4c3c73SJérôme Glisse * 953da4c3c73SJérôme Glisse * This snapshots the CPU page table for a range of virtual addresses. Snapshot 954a3e0d41cSJérôme Glisse * validity is tracked by range struct. See in include/linux/hmm.h for example 955a3e0d41cSJérôme Glisse * on how to use. 956da4c3c73SJérôme Glisse */ 95725f23a0cSJérôme Glisse long hmm_range_snapshot(struct hmm_range *range) 958da4c3c73SJérôme Glisse { 95963d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 960a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 96174eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 962a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 963a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 964da4c3c73SJérôme Glisse struct mm_walk mm_walk; 965704f3f2cSJérôme Glisse 96647f24598SJason Gunthorpe lockdep_assert_held(&hmm->mm->mmap_sem); 967a3e0d41cSJérôme Glisse do { 968a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 969a3e0d41cSJérôme Glisse if (!range->valid) 970a3e0d41cSJérôme Glisse return -EAGAIN; 971a3e0d41cSJérôme Glisse 972a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 97363d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 974a3e0d41cSJérôme Glisse return -EFAULT; 975a3e0d41cSJérôme Glisse 97663d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 9771c2308f0SJason Gunthorpe if (huge_page_shift(hstate_vma(vma)) != 9781c2308f0SJason Gunthorpe range->page_shift && 97963d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 98063d5066fSJérôme Glisse return -EINVAL; 98163d5066fSJérôme Glisse } else { 98263d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 98363d5066fSJérôme Glisse return -EINVAL; 98463d5066fSJérôme Glisse } 98563d5066fSJérôme Glisse 98686586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 98786586a41SJérôme Glisse /* 988a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 989a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 990a3e0d41cSJérôme Glisse * support architecture that allow write without read. 99186586a41SJérôme Glisse */ 992a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 993a3e0d41cSJérôme Glisse range->start, range->end); 99486586a41SJérôme Glisse return -EPERM; 99586586a41SJérôme Glisse } 99686586a41SJérôme Glisse 997a3e0d41cSJérôme Glisse range->vma = vma; 998992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 999a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 100074eee180SJérôme Glisse hmm_vma_walk.fault = false; 100174eee180SJérôme Glisse hmm_vma_walk.range = range; 100274eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1003a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 100474eee180SJérôme Glisse 1005da4c3c73SJérôme Glisse mm_walk.vma = vma; 1006da4c3c73SJérôme Glisse mm_walk.mm = vma->vm_mm; 1007da4c3c73SJérôme Glisse mm_walk.pte_entry = NULL; 1008da4c3c73SJérôme Glisse mm_walk.test_walk = NULL; 1009da4c3c73SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1010992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 1011da4c3c73SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 1012da4c3c73SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 101363d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 1014da4c3c73SJérôme Glisse 1015a3e0d41cSJérôme Glisse walk_page_range(start, end, &mm_walk); 1016a3e0d41cSJérôme Glisse start = end; 1017a3e0d41cSJérôme Glisse } while (start < range->end); 1018a3e0d41cSJérôme Glisse 101925f23a0cSJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1020da4c3c73SJérôme Glisse } 102125f23a0cSJérôme Glisse EXPORT_SYMBOL(hmm_range_snapshot); 1022da4c3c73SJérôme Glisse 1023da4c3c73SJérôme Glisse /* 102473231612SJérôme Glisse * hmm_range_fault() - try to fault some address in a virtual address range 102508232a45SJérôme Glisse * @range: range being faulted 102674eee180SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 1027085ea250SRalph Campbell * Return: number of valid pages in range->pfns[] (from range start 102873231612SJérôme Glisse * address). This may be zero. If the return value is negative, 102973231612SJérôme Glisse * then one of the following values may be returned: 103073231612SJérôme Glisse * 103173231612SJérôme Glisse * -EINVAL invalid arguments or mm or virtual address are in an 103263d5066fSJérôme Glisse * invalid vma (for instance device file vma). 103373231612SJérôme Glisse * -ENOMEM: Out of memory. 103473231612SJérôme Glisse * -EPERM: Invalid permission (for instance asking for write and 103573231612SJérôme Glisse * range is read only). 103673231612SJérôme Glisse * -EAGAIN: If you need to retry and mmap_sem was drop. This can only 103773231612SJérôme Glisse * happens if block argument is false. 103873231612SJérôme Glisse * -EBUSY: If the the range is being invalidated and you should wait 103973231612SJérôme Glisse * for invalidation to finish. 104073231612SJérôme Glisse * -EFAULT: Invalid (ie either no valid vma or it is illegal to access 104173231612SJérôme Glisse * that range), number of valid pages in range->pfns[] (from 104273231612SJérôme Glisse * range start address). 104374eee180SJérôme Glisse * 104474eee180SJérôme Glisse * This is similar to a regular CPU page fault except that it will not trigger 104573231612SJérôme Glisse * any memory migration if the memory being faulted is not accessible by CPUs 104673231612SJérôme Glisse * and caller does not ask for migration. 104774eee180SJérôme Glisse * 1048ff05c0c6SJérôme Glisse * On error, for one virtual address in the range, the function will mark the 1049ff05c0c6SJérôme Glisse * corresponding HMM pfn entry with an error flag. 105074eee180SJérôme Glisse */ 105173231612SJérôme Glisse long hmm_range_fault(struct hmm_range *range, bool block) 105274eee180SJérôme Glisse { 105363d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 1054a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 105574eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 1056a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 1057a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 105874eee180SJérôme Glisse struct mm_walk mm_walk; 105974eee180SJérôme Glisse int ret; 106074eee180SJérôme Glisse 106147f24598SJason Gunthorpe lockdep_assert_held(&hmm->mm->mmap_sem); 1062a3e0d41cSJérôme Glisse 1063a3e0d41cSJérôme Glisse do { 1064a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1065a3e0d41cSJérôme Glisse if (!range->valid) { 1066a3e0d41cSJérôme Glisse up_read(&hmm->mm->mmap_sem); 1067a3e0d41cSJérôme Glisse return -EAGAIN; 1068704f3f2cSJérôme Glisse } 106974eee180SJérôme Glisse 1070a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 107163d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1072a3e0d41cSJérôme Glisse return -EFAULT; 1073a3e0d41cSJérôme Glisse 107463d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 107563d5066fSJérôme Glisse if (huge_page_shift(hstate_vma(vma)) != 107663d5066fSJérôme Glisse range->page_shift && 107763d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 107863d5066fSJérôme Glisse return -EINVAL; 107963d5066fSJérôme Glisse } else { 108063d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 108163d5066fSJérôme Glisse return -EINVAL; 108263d5066fSJérôme Glisse } 108363d5066fSJérôme Glisse 108486586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 108586586a41SJérôme Glisse /* 1086a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1087a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1088a3e0d41cSJérôme Glisse * support architecture that allow write without read. 108986586a41SJérôme Glisse */ 1090a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1091a3e0d41cSJérôme Glisse range->start, range->end); 109286586a41SJérôme Glisse return -EPERM; 109386586a41SJérôme Glisse } 109474eee180SJérôme Glisse 1095a3e0d41cSJérôme Glisse range->vma = vma; 1096992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1097a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 109874eee180SJérôme Glisse hmm_vma_walk.fault = true; 109974eee180SJérôme Glisse hmm_vma_walk.block = block; 110074eee180SJérôme Glisse hmm_vma_walk.range = range; 110174eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1102a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 110374eee180SJérôme Glisse 110474eee180SJérôme Glisse mm_walk.vma = vma; 110574eee180SJérôme Glisse mm_walk.mm = vma->vm_mm; 110674eee180SJérôme Glisse mm_walk.pte_entry = NULL; 110774eee180SJérôme Glisse mm_walk.test_walk = NULL; 110874eee180SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1109992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 111074eee180SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 111174eee180SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 111263d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 111374eee180SJérôme Glisse 111474eee180SJérôme Glisse do { 1115a3e0d41cSJérôme Glisse ret = walk_page_range(start, end, &mm_walk); 111674eee180SJérôme Glisse start = hmm_vma_walk.last; 1117a3e0d41cSJérôme Glisse 111873231612SJérôme Glisse /* Keep trying while the range is valid. */ 111973231612SJérôme Glisse } while (ret == -EBUSY && range->valid); 112074eee180SJérôme Glisse 112174eee180SJérôme Glisse if (ret) { 112274eee180SJérôme Glisse unsigned long i; 112374eee180SJérôme Glisse 112474eee180SJérôme Glisse i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1125a3e0d41cSJérôme Glisse hmm_pfns_clear(range, &range->pfns[i], 1126a3e0d41cSJérôme Glisse hmm_vma_walk.last, range->end); 112773231612SJérôme Glisse return ret; 112874eee180SJérôme Glisse } 1129a3e0d41cSJérôme Glisse start = end; 1130a3e0d41cSJérôme Glisse 1131a3e0d41cSJérôme Glisse } while (start < range->end); 1132704f3f2cSJérôme Glisse 113373231612SJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 113474eee180SJérôme Glisse } 113573231612SJérôme Glisse EXPORT_SYMBOL(hmm_range_fault); 113655c0ece8SJérôme Glisse 113755c0ece8SJérôme Glisse /** 113855c0ece8SJérôme Glisse * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one. 113955c0ece8SJérôme Glisse * @range: range being faulted 114055c0ece8SJérôme Glisse * @device: device against to dma map page to 114155c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 114255c0ece8SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 1143085ea250SRalph Campbell * Return: number of pages mapped on success, -EAGAIN if mmap_sem have been 114455c0ece8SJérôme Glisse * drop and you need to try again, some other error value otherwise 114555c0ece8SJérôme Glisse * 114655c0ece8SJérôme Glisse * Note same usage pattern as hmm_range_fault(). 114755c0ece8SJérôme Glisse */ 114855c0ece8SJérôme Glisse long hmm_range_dma_map(struct hmm_range *range, 114955c0ece8SJérôme Glisse struct device *device, 115055c0ece8SJérôme Glisse dma_addr_t *daddrs, 115155c0ece8SJérôme Glisse bool block) 115255c0ece8SJérôme Glisse { 115355c0ece8SJérôme Glisse unsigned long i, npages, mapped; 115455c0ece8SJérôme Glisse long ret; 115555c0ece8SJérôme Glisse 115655c0ece8SJérôme Glisse ret = hmm_range_fault(range, block); 115755c0ece8SJérôme Glisse if (ret <= 0) 115855c0ece8SJérôme Glisse return ret ? ret : -EBUSY; 115955c0ece8SJérôme Glisse 116055c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 116155c0ece8SJérôme Glisse for (i = 0, mapped = 0; i < npages; ++i) { 116255c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 116355c0ece8SJérôme Glisse struct page *page; 116455c0ece8SJérôme Glisse 116555c0ece8SJérôme Glisse /* 116655c0ece8SJérôme Glisse * FIXME need to update DMA API to provide invalid DMA address 116755c0ece8SJérôme Glisse * value instead of a function to test dma address value. This 116855c0ece8SJérôme Glisse * would remove lot of dumb code duplicated accross many arch. 116955c0ece8SJérôme Glisse * 117055c0ece8SJérôme Glisse * For now setting it to 0 here is good enough as the pfns[] 117155c0ece8SJérôme Glisse * value is what is use to check what is valid and what isn't. 117255c0ece8SJérôme Glisse */ 117355c0ece8SJérôme Glisse daddrs[i] = 0; 117455c0ece8SJérôme Glisse 1175391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 117655c0ece8SJérôme Glisse if (page == NULL) 117755c0ece8SJérôme Glisse continue; 117855c0ece8SJérôme Glisse 117955c0ece8SJérôme Glisse /* Check if range is being invalidated */ 118055c0ece8SJérôme Glisse if (!range->valid) { 118155c0ece8SJérôme Glisse ret = -EBUSY; 118255c0ece8SJérôme Glisse goto unmap; 118355c0ece8SJérôme Glisse } 118455c0ece8SJérôme Glisse 118555c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 118655c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 118755c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 118855c0ece8SJérôme Glisse 118955c0ece8SJérôme Glisse daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir); 119055c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) { 119155c0ece8SJérôme Glisse ret = -EFAULT; 119255c0ece8SJérôme Glisse goto unmap; 119355c0ece8SJérôme Glisse } 119455c0ece8SJérôme Glisse 119555c0ece8SJérôme Glisse mapped++; 119655c0ece8SJérôme Glisse } 119755c0ece8SJérôme Glisse 119855c0ece8SJérôme Glisse return mapped; 119955c0ece8SJérôme Glisse 120055c0ece8SJérôme Glisse unmap: 120155c0ece8SJérôme Glisse for (npages = i, i = 0; (i < npages) && mapped; ++i) { 120255c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 120355c0ece8SJérôme Glisse struct page *page; 120455c0ece8SJérôme Glisse 1205391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 120655c0ece8SJérôme Glisse if (page == NULL) 120755c0ece8SJérôme Glisse continue; 120855c0ece8SJérôme Glisse 120955c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) 121055c0ece8SJérôme Glisse continue; 121155c0ece8SJérôme Glisse 121255c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 121355c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 121455c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 121555c0ece8SJérôme Glisse 121655c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 121755c0ece8SJérôme Glisse mapped--; 121855c0ece8SJérôme Glisse } 121955c0ece8SJérôme Glisse 122055c0ece8SJérôme Glisse return ret; 122155c0ece8SJérôme Glisse } 122255c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_map); 122355c0ece8SJérôme Glisse 122455c0ece8SJérôme Glisse /** 122555c0ece8SJérôme Glisse * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map() 122655c0ece8SJérôme Glisse * @range: range being unmapped 122755c0ece8SJérôme Glisse * @vma: the vma against which the range (optional) 122855c0ece8SJérôme Glisse * @device: device against which dma map was done 122955c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 123055c0ece8SJérôme Glisse * @dirty: dirty page if it had the write flag set 1231085ea250SRalph Campbell * Return: number of page unmapped on success, -EINVAL otherwise 123255c0ece8SJérôme Glisse * 123355c0ece8SJérôme Glisse * Note that caller MUST abide by mmu notifier or use HMM mirror and abide 123455c0ece8SJérôme Glisse * to the sync_cpu_device_pagetables() callback so that it is safe here to 123555c0ece8SJérôme Glisse * call set_page_dirty(). Caller must also take appropriate locks to avoid 123655c0ece8SJérôme Glisse * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress. 123755c0ece8SJérôme Glisse */ 123855c0ece8SJérôme Glisse long hmm_range_dma_unmap(struct hmm_range *range, 123955c0ece8SJérôme Glisse struct vm_area_struct *vma, 124055c0ece8SJérôme Glisse struct device *device, 124155c0ece8SJérôme Glisse dma_addr_t *daddrs, 124255c0ece8SJérôme Glisse bool dirty) 124355c0ece8SJérôme Glisse { 124455c0ece8SJérôme Glisse unsigned long i, npages; 124555c0ece8SJérôme Glisse long cpages = 0; 124655c0ece8SJérôme Glisse 124755c0ece8SJérôme Glisse /* Sanity check. */ 124855c0ece8SJérôme Glisse if (range->end <= range->start) 124955c0ece8SJérôme Glisse return -EINVAL; 125055c0ece8SJérôme Glisse if (!daddrs) 125155c0ece8SJérôme Glisse return -EINVAL; 125255c0ece8SJérôme Glisse if (!range->pfns) 125355c0ece8SJérôme Glisse return -EINVAL; 125455c0ece8SJérôme Glisse 125555c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 125655c0ece8SJérôme Glisse for (i = 0; i < npages; ++i) { 125755c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 125855c0ece8SJérôme Glisse struct page *page; 125955c0ece8SJérôme Glisse 1260391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 126155c0ece8SJérôme Glisse if (page == NULL) 126255c0ece8SJérôme Glisse continue; 126355c0ece8SJérôme Glisse 126455c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 126555c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) { 126655c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 126755c0ece8SJérôme Glisse 126855c0ece8SJérôme Glisse /* 126955c0ece8SJérôme Glisse * See comments in function description on why it is 127055c0ece8SJérôme Glisse * safe here to call set_page_dirty() 127155c0ece8SJérôme Glisse */ 127255c0ece8SJérôme Glisse if (dirty) 127355c0ece8SJérôme Glisse set_page_dirty(page); 127455c0ece8SJérôme Glisse } 127555c0ece8SJérôme Glisse 127655c0ece8SJérôme Glisse /* Unmap and clear pfns/dma address */ 127755c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 127855c0ece8SJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 127955c0ece8SJérôme Glisse /* FIXME see comments in hmm_vma_dma_map() */ 128055c0ece8SJérôme Glisse daddrs[i] = 0; 128155c0ece8SJérôme Glisse cpages++; 128255c0ece8SJérôme Glisse } 128355c0ece8SJérôme Glisse 128455c0ece8SJérôme Glisse return cpages; 128555c0ece8SJérôme Glisse } 128655c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_unmap); 1287