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 * 14133ff0eaSJérôme 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> 33c0b12405SJérôme Glisse #include <linux/mmu_notifier.h> 344ef589dcSJérôme Glisse #include <linux/memory_hotplug.h> 354ef589dcSJérôme Glisse 364ef589dcSJérôme Glisse #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT) 37133ff0eaSJérôme Glisse 386b368cd4SJérôme Glisse #if IS_ENABLED(CONFIG_HMM_MIRROR) 39c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops; 40c0b12405SJérôme Glisse 41133ff0eaSJérôme Glisse /* 42133ff0eaSJérôme Glisse * struct hmm - HMM per mm struct 43133ff0eaSJérôme Glisse * 44133ff0eaSJérôme Glisse * @mm: mm struct this HMM struct is bound to 45da4c3c73SJérôme Glisse * @lock: lock protecting ranges list 46c0b12405SJérôme Glisse * @sequence: we track updates to the CPU page table with a sequence number 47da4c3c73SJérôme Glisse * @ranges: list of range being snapshotted 48c0b12405SJérôme Glisse * @mirrors: list of mirrors for this mm 49c0b12405SJérôme Glisse * @mmu_notifier: mmu notifier to track updates to CPU page table 50c0b12405SJérôme Glisse * @mirrors_sem: read/write semaphore protecting the mirrors list 51133ff0eaSJérôme Glisse */ 52133ff0eaSJérôme Glisse struct hmm { 53133ff0eaSJérôme Glisse struct mm_struct *mm; 54da4c3c73SJérôme Glisse spinlock_t lock; 55c0b12405SJérôme Glisse atomic_t sequence; 56da4c3c73SJérôme Glisse struct list_head ranges; 57c0b12405SJérôme Glisse struct list_head mirrors; 58c0b12405SJérôme Glisse struct mmu_notifier mmu_notifier; 59c0b12405SJérôme Glisse struct rw_semaphore mirrors_sem; 60133ff0eaSJérôme Glisse }; 61133ff0eaSJérôme Glisse 62133ff0eaSJérôme Glisse /* 63133ff0eaSJérôme Glisse * hmm_register - register HMM against an mm (HMM internal) 64133ff0eaSJérôme Glisse * 65133ff0eaSJérôme Glisse * @mm: mm struct to attach to 66133ff0eaSJérôme Glisse * 67133ff0eaSJérôme Glisse * This is not intended to be used directly by device drivers. It allocates an 68133ff0eaSJérôme Glisse * HMM struct if mm does not have one, and initializes it. 69133ff0eaSJérôme Glisse */ 70133ff0eaSJérôme Glisse static struct hmm *hmm_register(struct mm_struct *mm) 71133ff0eaSJérôme Glisse { 72c0b12405SJérôme Glisse struct hmm *hmm = READ_ONCE(mm->hmm); 73c0b12405SJérôme Glisse bool cleanup = false; 74133ff0eaSJérôme Glisse 75133ff0eaSJérôme Glisse /* 76133ff0eaSJérôme Glisse * The hmm struct can only be freed once the mm_struct goes away, 77133ff0eaSJérôme Glisse * hence we should always have pre-allocated an new hmm struct 78133ff0eaSJérôme Glisse * above. 79133ff0eaSJérôme Glisse */ 80c0b12405SJérôme Glisse if (hmm) 81c0b12405SJérôme Glisse return hmm; 82c0b12405SJérôme Glisse 83c0b12405SJérôme Glisse hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); 84c0b12405SJérôme Glisse if (!hmm) 85c0b12405SJérôme Glisse return NULL; 86c0b12405SJérôme Glisse INIT_LIST_HEAD(&hmm->mirrors); 87c0b12405SJérôme Glisse init_rwsem(&hmm->mirrors_sem); 88c0b12405SJérôme Glisse atomic_set(&hmm->sequence, 0); 89c0b12405SJérôme Glisse hmm->mmu_notifier.ops = NULL; 90da4c3c73SJérôme Glisse INIT_LIST_HEAD(&hmm->ranges); 91da4c3c73SJérôme Glisse spin_lock_init(&hmm->lock); 92c0b12405SJérôme Glisse hmm->mm = mm; 93c0b12405SJérôme Glisse 94c0b12405SJérôme Glisse /* 95c0b12405SJérôme Glisse * We should only get here if hold the mmap_sem in write mode ie on 96c0b12405SJérôme Glisse * registration of first mirror through hmm_mirror_register() 97c0b12405SJérôme Glisse */ 98c0b12405SJérôme Glisse hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; 99c0b12405SJérôme Glisse if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) { 100c0b12405SJérôme Glisse kfree(hmm); 101c0b12405SJérôme Glisse return NULL; 102c0b12405SJérôme Glisse } 103c0b12405SJérôme Glisse 104c0b12405SJérôme Glisse spin_lock(&mm->page_table_lock); 105c0b12405SJérôme Glisse if (!mm->hmm) 106c0b12405SJérôme Glisse mm->hmm = hmm; 107c0b12405SJérôme Glisse else 108c0b12405SJérôme Glisse cleanup = true; 109c0b12405SJérôme Glisse spin_unlock(&mm->page_table_lock); 110c0b12405SJérôme Glisse 111c0b12405SJérôme Glisse if (cleanup) { 112c0b12405SJérôme Glisse mmu_notifier_unregister(&hmm->mmu_notifier, mm); 113c0b12405SJérôme Glisse kfree(hmm); 114c0b12405SJérôme Glisse } 115c0b12405SJérôme Glisse 116133ff0eaSJérôme Glisse return mm->hmm; 117133ff0eaSJérôme Glisse } 118133ff0eaSJérôme Glisse 119133ff0eaSJérôme Glisse void hmm_mm_destroy(struct mm_struct *mm) 120133ff0eaSJérôme Glisse { 121133ff0eaSJérôme Glisse kfree(mm->hmm); 122133ff0eaSJérôme Glisse } 123c0b12405SJérôme Glisse 124c0b12405SJérôme Glisse static void hmm_invalidate_range(struct hmm *hmm, 125c0b12405SJérôme Glisse enum hmm_update_type action, 126c0b12405SJérôme Glisse unsigned long start, 127c0b12405SJérôme Glisse unsigned long end) 128c0b12405SJérôme Glisse { 129c0b12405SJérôme Glisse struct hmm_mirror *mirror; 130da4c3c73SJérôme Glisse struct hmm_range *range; 131da4c3c73SJérôme Glisse 132da4c3c73SJérôme Glisse spin_lock(&hmm->lock); 133da4c3c73SJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 134da4c3c73SJérôme Glisse unsigned long addr, idx, npages; 135da4c3c73SJérôme Glisse 136da4c3c73SJérôme Glisse if (end < range->start || start >= range->end) 137da4c3c73SJérôme Glisse continue; 138da4c3c73SJérôme Glisse 139da4c3c73SJérôme Glisse range->valid = false; 140da4c3c73SJérôme Glisse addr = max(start, range->start); 141da4c3c73SJérôme Glisse idx = (addr - range->start) >> PAGE_SHIFT; 142da4c3c73SJérôme Glisse npages = (min(range->end, end) - addr) >> PAGE_SHIFT; 143da4c3c73SJérôme Glisse memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages); 144da4c3c73SJérôme Glisse } 145da4c3c73SJérôme Glisse spin_unlock(&hmm->lock); 146c0b12405SJérôme Glisse 147c0b12405SJérôme Glisse down_read(&hmm->mirrors_sem); 148c0b12405SJérôme Glisse list_for_each_entry(mirror, &hmm->mirrors, list) 149c0b12405SJérôme Glisse mirror->ops->sync_cpu_device_pagetables(mirror, action, 150c0b12405SJérôme Glisse start, end); 151c0b12405SJérôme Glisse up_read(&hmm->mirrors_sem); 152c0b12405SJérôme Glisse } 153c0b12405SJérôme Glisse 154e1401513SRalph Campbell static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm) 155e1401513SRalph Campbell { 156e1401513SRalph Campbell struct hmm_mirror *mirror; 157e1401513SRalph Campbell struct hmm *hmm = mm->hmm; 158e1401513SRalph Campbell 159e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 160e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror, 161e1401513SRalph Campbell list); 162e1401513SRalph Campbell while (mirror) { 163e1401513SRalph Campbell list_del_init(&mirror->list); 164e1401513SRalph Campbell if (mirror->ops->release) { 165e1401513SRalph Campbell /* 166e1401513SRalph Campbell * Drop mirrors_sem so callback can wait on any pending 167e1401513SRalph Campbell * work that might itself trigger mmu_notifier callback 168e1401513SRalph Campbell * and thus would deadlock with us. 169e1401513SRalph Campbell */ 170e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 171e1401513SRalph Campbell mirror->ops->release(mirror); 172e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 173e1401513SRalph Campbell } 174e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, 175e1401513SRalph Campbell struct hmm_mirror, list); 176e1401513SRalph Campbell } 177e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 178e1401513SRalph Campbell } 179e1401513SRalph Campbell 180c0b12405SJérôme Glisse static void hmm_invalidate_range_start(struct mmu_notifier *mn, 181c0b12405SJérôme Glisse struct mm_struct *mm, 182c0b12405SJérôme Glisse unsigned long start, 183c0b12405SJérôme Glisse unsigned long end) 184c0b12405SJérôme Glisse { 185c0b12405SJérôme Glisse struct hmm *hmm = mm->hmm; 186c0b12405SJérôme Glisse 187c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 188c0b12405SJérôme Glisse 189c0b12405SJérôme Glisse atomic_inc(&hmm->sequence); 190c0b12405SJérôme Glisse } 191c0b12405SJérôme Glisse 192c0b12405SJérôme Glisse static void hmm_invalidate_range_end(struct mmu_notifier *mn, 193c0b12405SJérôme Glisse struct mm_struct *mm, 194c0b12405SJérôme Glisse unsigned long start, 195c0b12405SJérôme Glisse unsigned long end) 196c0b12405SJérôme Glisse { 197c0b12405SJérôme Glisse struct hmm *hmm = mm->hmm; 198c0b12405SJérôme Glisse 199c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 200c0b12405SJérôme Glisse 201c0b12405SJérôme Glisse hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end); 202c0b12405SJérôme Glisse } 203c0b12405SJérôme Glisse 204c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { 205e1401513SRalph Campbell .release = hmm_release, 206c0b12405SJérôme Glisse .invalidate_range_start = hmm_invalidate_range_start, 207c0b12405SJérôme Glisse .invalidate_range_end = hmm_invalidate_range_end, 208c0b12405SJérôme Glisse }; 209c0b12405SJérôme Glisse 210c0b12405SJérôme Glisse /* 211c0b12405SJérôme Glisse * hmm_mirror_register() - register a mirror against an mm 212c0b12405SJérôme Glisse * 213c0b12405SJérôme Glisse * @mirror: new mirror struct to register 214c0b12405SJérôme Glisse * @mm: mm to register against 215c0b12405SJérôme Glisse * 216c0b12405SJérôme Glisse * To start mirroring a process address space, the device driver must register 217c0b12405SJérôme Glisse * an HMM mirror struct. 218c0b12405SJérôme Glisse * 219c0b12405SJérôme Glisse * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! 220c0b12405SJérôme Glisse */ 221c0b12405SJérôme Glisse int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) 222c0b12405SJérôme Glisse { 223c0b12405SJérôme Glisse /* Sanity check */ 224c0b12405SJérôme Glisse if (!mm || !mirror || !mirror->ops) 225c0b12405SJérôme Glisse return -EINVAL; 226c0b12405SJérôme Glisse 227c01cbba2SJérôme Glisse again: 228c0b12405SJérôme Glisse mirror->hmm = hmm_register(mm); 229c0b12405SJérôme Glisse if (!mirror->hmm) 230c0b12405SJérôme Glisse return -ENOMEM; 231c0b12405SJérôme Glisse 232c0b12405SJérôme Glisse down_write(&mirror->hmm->mirrors_sem); 233c01cbba2SJérôme Glisse if (mirror->hmm->mm == NULL) { 234c01cbba2SJérôme Glisse /* 235c01cbba2SJérôme Glisse * A racing hmm_mirror_unregister() is about to destroy the hmm 236c01cbba2SJérôme Glisse * struct. Try again to allocate a new one. 237c01cbba2SJérôme Glisse */ 238c01cbba2SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 239c01cbba2SJérôme Glisse mirror->hmm = NULL; 240c01cbba2SJérôme Glisse goto again; 241c01cbba2SJérôme Glisse } else { 242c0b12405SJérôme Glisse list_add(&mirror->list, &mirror->hmm->mirrors); 243c0b12405SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 244c01cbba2SJérôme Glisse } 245c0b12405SJérôme Glisse 246c0b12405SJérôme Glisse return 0; 247c0b12405SJérôme Glisse } 248c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_register); 249c0b12405SJérôme Glisse 250c0b12405SJérôme Glisse /* 251c0b12405SJérôme Glisse * hmm_mirror_unregister() - unregister a mirror 252c0b12405SJérôme Glisse * 253c0b12405SJérôme Glisse * @mirror: new mirror struct to register 254c0b12405SJérôme Glisse * 255c0b12405SJérôme Glisse * Stop mirroring a process address space, and cleanup. 256c0b12405SJérôme Glisse */ 257c0b12405SJérôme Glisse void hmm_mirror_unregister(struct hmm_mirror *mirror) 258c0b12405SJérôme Glisse { 259c01cbba2SJérôme Glisse bool should_unregister = false; 260c01cbba2SJérôme Glisse struct mm_struct *mm; 261c01cbba2SJérôme Glisse struct hmm *hmm; 262c0b12405SJérôme Glisse 263c01cbba2SJérôme Glisse if (mirror->hmm == NULL) 264c01cbba2SJérôme Glisse return; 265c01cbba2SJérôme Glisse 266c01cbba2SJérôme Glisse hmm = mirror->hmm; 267c0b12405SJérôme Glisse down_write(&hmm->mirrors_sem); 268e1401513SRalph Campbell list_del_init(&mirror->list); 269c01cbba2SJérôme Glisse should_unregister = list_empty(&hmm->mirrors); 270c01cbba2SJérôme Glisse mirror->hmm = NULL; 271c01cbba2SJérôme Glisse mm = hmm->mm; 272c01cbba2SJérôme Glisse hmm->mm = NULL; 273c0b12405SJérôme Glisse up_write(&hmm->mirrors_sem); 274c01cbba2SJérôme Glisse 275c01cbba2SJérôme Glisse if (!should_unregister || mm == NULL) 276c01cbba2SJérôme Glisse return; 277c01cbba2SJérôme Glisse 278c01cbba2SJérôme Glisse spin_lock(&mm->page_table_lock); 279c01cbba2SJérôme Glisse if (mm->hmm == hmm) 280c01cbba2SJérôme Glisse mm->hmm = NULL; 281c01cbba2SJérôme Glisse spin_unlock(&mm->page_table_lock); 282c01cbba2SJérôme Glisse 283c01cbba2SJérôme Glisse mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm); 284c01cbba2SJérôme Glisse kfree(hmm); 285c0b12405SJérôme Glisse } 286c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_unregister); 287da4c3c73SJérôme Glisse 28874eee180SJérôme Glisse struct hmm_vma_walk { 28974eee180SJérôme Glisse struct hmm_range *range; 29074eee180SJérôme Glisse unsigned long last; 29174eee180SJérôme Glisse bool fault; 29274eee180SJérôme Glisse bool block; 29374eee180SJérôme Glisse }; 29474eee180SJérôme Glisse 2952aee09d8SJérôme Glisse static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr, 2962aee09d8SJérôme Glisse bool write_fault, uint64_t *pfn) 29774eee180SJérôme Glisse { 29874eee180SJérôme Glisse unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE; 29974eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 300f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 30174eee180SJérôme Glisse struct vm_area_struct *vma = walk->vma; 30274eee180SJérôme Glisse int r; 30374eee180SJérôme Glisse 30474eee180SJérôme Glisse flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; 3052aee09d8SJérôme Glisse flags |= write_fault ? FAULT_FLAG_WRITE : 0; 30674eee180SJérôme Glisse r = handle_mm_fault(vma, addr, flags); 30774eee180SJérôme Glisse if (r & VM_FAULT_RETRY) 30874eee180SJérôme Glisse return -EBUSY; 30974eee180SJérôme Glisse if (r & VM_FAULT_ERROR) { 310f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 31174eee180SJérôme Glisse return -EFAULT; 31274eee180SJérôme Glisse } 31374eee180SJérôme Glisse 31474eee180SJérôme Glisse return -EAGAIN; 31574eee180SJérôme Glisse } 31674eee180SJérôme Glisse 317da4c3c73SJérôme Glisse static int hmm_pfns_bad(unsigned long addr, 318da4c3c73SJérôme Glisse unsigned long end, 319da4c3c73SJérôme Glisse struct mm_walk *walk) 320da4c3c73SJérôme Glisse { 321c719547fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 322c719547fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 323ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 324da4c3c73SJérôme Glisse unsigned long i; 325da4c3c73SJérôme Glisse 326da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 327da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) 328f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_ERROR]; 329da4c3c73SJérôme Glisse 330da4c3c73SJérôme Glisse return 0; 331da4c3c73SJérôme Glisse } 332da4c3c73SJérôme Glisse 3335504ed29SJérôme Glisse /* 3345504ed29SJérôme Glisse * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s) 3355504ed29SJérôme Glisse * @start: range virtual start address (inclusive) 3365504ed29SJérôme Glisse * @end: range virtual end address (exclusive) 3372aee09d8SJérôme Glisse * @fault: should we fault or not ? 3382aee09d8SJérôme Glisse * @write_fault: write fault ? 3395504ed29SJérôme Glisse * @walk: mm_walk structure 3405504ed29SJérôme Glisse * Returns: 0 on success, -EAGAIN after page fault, or page fault error 3415504ed29SJérôme Glisse * 3425504ed29SJérôme Glisse * This function will be called whenever pmd_none() or pte_none() returns true, 3435504ed29SJérôme Glisse * or whenever there is no page directory covering the virtual address range. 3445504ed29SJérôme Glisse */ 3452aee09d8SJérôme Glisse static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end, 3462aee09d8SJérôme Glisse bool fault, bool write_fault, 347da4c3c73SJérôme Glisse struct mm_walk *walk) 348da4c3c73SJérôme Glisse { 34974eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 35074eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 351ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 352da4c3c73SJérôme Glisse unsigned long i; 353da4c3c73SJérôme Glisse 35474eee180SJérôme Glisse hmm_vma_walk->last = addr; 355da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 35674eee180SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) { 357f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_NONE]; 3582aee09d8SJérôme Glisse if (fault || write_fault) { 35974eee180SJérôme Glisse int ret; 360da4c3c73SJérôme Glisse 3612aee09d8SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, write_fault, 3622aee09d8SJérôme Glisse &pfns[i]); 36374eee180SJérôme Glisse if (ret != -EAGAIN) 36474eee180SJérôme Glisse return ret; 36574eee180SJérôme Glisse } 36674eee180SJérôme Glisse } 36774eee180SJérôme Glisse 3682aee09d8SJérôme Glisse return (fault || write_fault) ? -EAGAIN : 0; 3692aee09d8SJérôme Glisse } 3702aee09d8SJérôme Glisse 3712aee09d8SJérôme Glisse static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 3722aee09d8SJérôme Glisse uint64_t pfns, uint64_t cpu_flags, 3732aee09d8SJérôme Glisse bool *fault, bool *write_fault) 3742aee09d8SJérôme Glisse { 375f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 376f88a1e90SJérôme Glisse 3772aee09d8SJérôme Glisse *fault = *write_fault = false; 3782aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) 3792aee09d8SJérôme Glisse return; 3802aee09d8SJérôme Glisse 3812aee09d8SJérôme Glisse /* We aren't ask to do anything ... */ 382f88a1e90SJérôme Glisse if (!(pfns & range->flags[HMM_PFN_VALID])) 3832aee09d8SJérôme Glisse return; 384f88a1e90SJérôme Glisse /* If this is device memory than only fault if explicitly requested */ 385f88a1e90SJérôme Glisse if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) { 386f88a1e90SJérôme Glisse /* Do we fault on device memory ? */ 387f88a1e90SJérôme Glisse if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) { 388f88a1e90SJérôme Glisse *write_fault = pfns & range->flags[HMM_PFN_WRITE]; 389f88a1e90SJérôme Glisse *fault = true; 390f88a1e90SJérôme Glisse } 3912aee09d8SJérôme Glisse return; 3922aee09d8SJérôme Glisse } 393f88a1e90SJérôme Glisse 394f88a1e90SJérôme Glisse /* If CPU page table is not valid then we need to fault */ 395f88a1e90SJérôme Glisse *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]); 396f88a1e90SJérôme Glisse /* Need to write fault ? */ 397f88a1e90SJérôme Glisse if ((pfns & range->flags[HMM_PFN_WRITE]) && 398f88a1e90SJérôme Glisse !(cpu_flags & range->flags[HMM_PFN_WRITE])) { 399f88a1e90SJérôme Glisse *write_fault = true; 4002aee09d8SJérôme Glisse *fault = true; 4012aee09d8SJérôme Glisse } 4022aee09d8SJérôme Glisse } 4032aee09d8SJérôme Glisse 4042aee09d8SJérôme Glisse static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4052aee09d8SJérôme Glisse const uint64_t *pfns, unsigned long npages, 4062aee09d8SJérôme Glisse uint64_t cpu_flags, bool *fault, 4072aee09d8SJérôme Glisse bool *write_fault) 4082aee09d8SJérôme Glisse { 4092aee09d8SJérôme Glisse unsigned long i; 4102aee09d8SJérôme Glisse 4112aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) { 4122aee09d8SJérôme Glisse *fault = *write_fault = false; 4132aee09d8SJérôme Glisse return; 4142aee09d8SJérôme Glisse } 4152aee09d8SJérôme Glisse 4162aee09d8SJérôme Glisse for (i = 0; i < npages; ++i) { 4172aee09d8SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags, 4182aee09d8SJérôme Glisse fault, write_fault); 4192aee09d8SJérôme Glisse if ((*fault) || (*write_fault)) 4202aee09d8SJérôme Glisse return; 4212aee09d8SJérôme Glisse } 4222aee09d8SJérôme Glisse } 4232aee09d8SJérôme Glisse 4242aee09d8SJérôme Glisse static int hmm_vma_walk_hole(unsigned long addr, unsigned long end, 4252aee09d8SJérôme Glisse struct mm_walk *walk) 4262aee09d8SJérôme Glisse { 4272aee09d8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 4282aee09d8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4292aee09d8SJérôme Glisse bool fault, write_fault; 4302aee09d8SJérôme Glisse unsigned long i, npages; 4312aee09d8SJérôme Glisse uint64_t *pfns; 4322aee09d8SJérôme Glisse 4332aee09d8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 4342aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 4352aee09d8SJérôme Glisse pfns = &range->pfns[i]; 4362aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 4372aee09d8SJérôme Glisse 0, &fault, &write_fault); 4382aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 4392aee09d8SJérôme Glisse } 4402aee09d8SJérôme Glisse 441f88a1e90SJérôme Glisse static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd) 4422aee09d8SJérôme Glisse { 4432aee09d8SJérôme Glisse if (pmd_protnone(pmd)) 4442aee09d8SJérôme Glisse return 0; 445f88a1e90SJérôme Glisse return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] | 446f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 447f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 448da4c3c73SJérôme Glisse } 449da4c3c73SJérôme Glisse 45053f5c3f4SJérôme Glisse static int hmm_vma_handle_pmd(struct mm_walk *walk, 45153f5c3f4SJérôme Glisse unsigned long addr, 45253f5c3f4SJérôme Glisse unsigned long end, 45353f5c3f4SJérôme Glisse uint64_t *pfns, 45453f5c3f4SJérôme Glisse pmd_t pmd) 45553f5c3f4SJérôme Glisse { 45653f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 457f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4582aee09d8SJérôme Glisse unsigned long pfn, npages, i; 4592aee09d8SJérôme Glisse bool fault, write_fault; 460f88a1e90SJérôme Glisse uint64_t cpu_flags; 46153f5c3f4SJérôme Glisse 4622aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 463f88a1e90SJérôme Glisse cpu_flags = pmd_to_hmm_pfn_flags(range, pmd); 4642aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags, 4652aee09d8SJérôme Glisse &fault, &write_fault); 46653f5c3f4SJérôme Glisse 4672aee09d8SJérôme Glisse if (pmd_protnone(pmd) || fault || write_fault) 4682aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 46953f5c3f4SJérôme Glisse 47053f5c3f4SJérôme Glisse pfn = pmd_pfn(pmd) + pte_index(addr); 47153f5c3f4SJérôme Glisse for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) 472f88a1e90SJérôme Glisse pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags; 47353f5c3f4SJérôme Glisse hmm_vma_walk->last = end; 47453f5c3f4SJérôme Glisse return 0; 47553f5c3f4SJérôme Glisse } 47653f5c3f4SJérôme Glisse 477f88a1e90SJérôme Glisse static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte) 4782aee09d8SJérôme Glisse { 4792aee09d8SJérôme Glisse if (pte_none(pte) || !pte_present(pte)) 4802aee09d8SJérôme Glisse return 0; 481f88a1e90SJérôme Glisse return pte_write(pte) ? range->flags[HMM_PFN_VALID] | 482f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 483f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 4842aee09d8SJérôme Glisse } 4852aee09d8SJérôme Glisse 48653f5c3f4SJérôme Glisse static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr, 48753f5c3f4SJérôme Glisse unsigned long end, pmd_t *pmdp, pte_t *ptep, 48853f5c3f4SJérôme Glisse uint64_t *pfn) 48953f5c3f4SJérôme Glisse { 49053f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 491f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 49253f5c3f4SJérôme Glisse struct vm_area_struct *vma = walk->vma; 4932aee09d8SJérôme Glisse bool fault, write_fault; 4942aee09d8SJérôme Glisse uint64_t cpu_flags; 49553f5c3f4SJérôme Glisse pte_t pte = *ptep; 496f88a1e90SJérôme Glisse uint64_t orig_pfn = *pfn; 49753f5c3f4SJérôme Glisse 498f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_NONE]; 499f88a1e90SJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, pte); 500f88a1e90SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 5012aee09d8SJérôme Glisse &fault, &write_fault); 50253f5c3f4SJérôme Glisse 50353f5c3f4SJérôme Glisse if (pte_none(pte)) { 5042aee09d8SJérôme Glisse if (fault || write_fault) 50553f5c3f4SJérôme Glisse goto fault; 50653f5c3f4SJérôme Glisse return 0; 50753f5c3f4SJérôme Glisse } 50853f5c3f4SJérôme Glisse 50953f5c3f4SJérôme Glisse if (!pte_present(pte)) { 51053f5c3f4SJérôme Glisse swp_entry_t entry = pte_to_swp_entry(pte); 51153f5c3f4SJérôme Glisse 51253f5c3f4SJérôme Glisse if (!non_swap_entry(entry)) { 5132aee09d8SJérôme Glisse if (fault || write_fault) 51453f5c3f4SJérôme Glisse goto fault; 51553f5c3f4SJérôme Glisse return 0; 51653f5c3f4SJérôme Glisse } 51753f5c3f4SJérôme Glisse 51853f5c3f4SJérôme Glisse /* 51953f5c3f4SJérôme Glisse * This is a special swap entry, ignore migration, use 52053f5c3f4SJérôme Glisse * device and report anything else as error. 52153f5c3f4SJérôme Glisse */ 52253f5c3f4SJérôme Glisse if (is_device_private_entry(entry)) { 523f88a1e90SJérôme Glisse cpu_flags = range->flags[HMM_PFN_VALID] | 524f88a1e90SJérôme Glisse range->flags[HMM_PFN_DEVICE_PRIVATE]; 5252aee09d8SJérôme Glisse cpu_flags |= is_write_device_private_entry(entry) ? 526f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 0; 527f88a1e90SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 528f88a1e90SJérôme Glisse &fault, &write_fault); 529f88a1e90SJérôme Glisse if (fault || write_fault) 530f88a1e90SJérôme Glisse goto fault; 531f88a1e90SJérôme Glisse *pfn = hmm_pfn_from_pfn(range, swp_offset(entry)); 532f88a1e90SJérôme Glisse *pfn |= cpu_flags; 53353f5c3f4SJérôme Glisse return 0; 53453f5c3f4SJérôme Glisse } 53553f5c3f4SJérôme Glisse 53653f5c3f4SJérôme Glisse if (is_migration_entry(entry)) { 5372aee09d8SJérôme Glisse if (fault || write_fault) { 53853f5c3f4SJérôme Glisse pte_unmap(ptep); 53953f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 54053f5c3f4SJérôme Glisse migration_entry_wait(vma->vm_mm, 54153f5c3f4SJérôme Glisse pmdp, addr); 54253f5c3f4SJérôme Glisse return -EAGAIN; 54353f5c3f4SJérôme Glisse } 54453f5c3f4SJérôme Glisse return 0; 54553f5c3f4SJérôme Glisse } 54653f5c3f4SJérôme Glisse 54753f5c3f4SJérôme Glisse /* Report error for everything else */ 548f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 54953f5c3f4SJérôme Glisse return -EFAULT; 55053f5c3f4SJérôme Glisse } 55153f5c3f4SJérôme Glisse 5522aee09d8SJérôme Glisse if (fault || write_fault) 55353f5c3f4SJérôme Glisse goto fault; 55453f5c3f4SJérôme Glisse 555f88a1e90SJérôme Glisse *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags; 55653f5c3f4SJérôme Glisse return 0; 55753f5c3f4SJérôme Glisse 55853f5c3f4SJérôme Glisse fault: 55953f5c3f4SJérôme Glisse pte_unmap(ptep); 56053f5c3f4SJérôme Glisse /* Fault any virtual address we were asked to fault */ 5612aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 56253f5c3f4SJérôme Glisse } 56353f5c3f4SJérôme Glisse 564da4c3c73SJérôme Glisse static int hmm_vma_walk_pmd(pmd_t *pmdp, 565da4c3c73SJérôme Glisse unsigned long start, 566da4c3c73SJérôme Glisse unsigned long end, 567da4c3c73SJérôme Glisse struct mm_walk *walk) 568da4c3c73SJérôme Glisse { 56974eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 57074eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 571ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 572da4c3c73SJérôme Glisse unsigned long addr = start, i; 573da4c3c73SJérôme Glisse pte_t *ptep; 574da4c3c73SJérôme Glisse 575da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 576da4c3c73SJérôme Glisse 577da4c3c73SJérôme Glisse again: 578da4c3c73SJérôme Glisse if (pmd_none(*pmdp)) 579da4c3c73SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 580da4c3c73SJérôme Glisse 58153f5c3f4SJérôme Glisse if (pmd_huge(*pmdp) && (range->vma->vm_flags & VM_HUGETLB)) 582da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 583da4c3c73SJérôme Glisse 584da4c3c73SJérôme Glisse if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) { 585da4c3c73SJérôme Glisse pmd_t pmd; 586da4c3c73SJérôme Glisse 587da4c3c73SJérôme Glisse /* 588da4c3c73SJérôme Glisse * No need to take pmd_lock here, even if some other threads 589da4c3c73SJérôme Glisse * is splitting the huge pmd we will get that event through 590da4c3c73SJérôme Glisse * mmu_notifier callback. 591da4c3c73SJérôme Glisse * 592da4c3c73SJérôme Glisse * So just read pmd value and check again its a transparent 593da4c3c73SJérôme Glisse * huge or device mapping one and compute corresponding pfn 594da4c3c73SJérôme Glisse * values. 595da4c3c73SJérôme Glisse */ 596da4c3c73SJérôme Glisse pmd = pmd_read_atomic(pmdp); 597da4c3c73SJérôme Glisse barrier(); 598da4c3c73SJérôme Glisse if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) 599da4c3c73SJérôme Glisse goto again; 600da4c3c73SJérôme Glisse 60153f5c3f4SJérôme Glisse return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd); 602da4c3c73SJérôme Glisse } 603da4c3c73SJérôme Glisse 604da4c3c73SJérôme Glisse if (pmd_bad(*pmdp)) 605da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 606da4c3c73SJérôme Glisse 607da4c3c73SJérôme Glisse ptep = pte_offset_map(pmdp, addr); 608da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { 60953f5c3f4SJérôme Glisse int r; 610da4c3c73SJérôme Glisse 61153f5c3f4SJérôme Glisse r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]); 61253f5c3f4SJérôme Glisse if (r) { 61353f5c3f4SJérôme Glisse /* hmm_vma_handle_pte() did unmap pte directory */ 61474eee180SJérôme Glisse hmm_vma_walk->last = addr; 61553f5c3f4SJérôme Glisse return r; 61674eee180SJérôme Glisse } 617da4c3c73SJérôme Glisse } 618da4c3c73SJérôme Glisse pte_unmap(ptep - 1); 619da4c3c73SJérôme Glisse 62053f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 621da4c3c73SJérôme Glisse return 0; 622da4c3c73SJérôme Glisse } 623da4c3c73SJérôme Glisse 624f88a1e90SJérôme Glisse static void hmm_pfns_clear(struct hmm_range *range, 625f88a1e90SJérôme Glisse uint64_t *pfns, 62633cd47dcSJérôme Glisse unsigned long addr, 62733cd47dcSJérôme Glisse unsigned long end) 62833cd47dcSJérôme Glisse { 62933cd47dcSJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 630f88a1e90SJérôme Glisse *pfns = range->values[HMM_PFN_NONE]; 63133cd47dcSJérôme Glisse } 63233cd47dcSJérôme Glisse 633855ce7d2SJérôme Glisse static void hmm_pfns_special(struct hmm_range *range) 634855ce7d2SJérôme Glisse { 635855ce7d2SJérôme Glisse unsigned long addr = range->start, i = 0; 636855ce7d2SJérôme Glisse 637855ce7d2SJérôme Glisse for (; addr < range->end; addr += PAGE_SIZE, i++) 638f88a1e90SJérôme Glisse range->pfns[i] = range->values[HMM_PFN_SPECIAL]; 639855ce7d2SJérôme Glisse } 640855ce7d2SJérôme Glisse 641da4c3c73SJérôme Glisse /* 642da4c3c73SJérôme Glisse * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses 64308232a45SJérôme Glisse * @range: range being snapshotted 64486586a41SJérôme Glisse * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid 64586586a41SJérôme Glisse * vma permission, 0 success 646da4c3c73SJérôme Glisse * 647da4c3c73SJérôme Glisse * This snapshots the CPU page table for a range of virtual addresses. Snapshot 648da4c3c73SJérôme Glisse * validity is tracked by range struct. See hmm_vma_range_done() for further 649da4c3c73SJérôme Glisse * information. 650da4c3c73SJérôme Glisse * 651da4c3c73SJérôme Glisse * The range struct is initialized here. It tracks the CPU page table, but only 652da4c3c73SJérôme Glisse * if the function returns success (0), in which case the caller must then call 653da4c3c73SJérôme Glisse * hmm_vma_range_done() to stop CPU page table update tracking on this range. 654da4c3c73SJérôme Glisse * 655da4c3c73SJérôme Glisse * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS 656da4c3c73SJérôme Glisse * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED ! 657da4c3c73SJérôme Glisse */ 65808232a45SJérôme Glisse int hmm_vma_get_pfns(struct hmm_range *range) 659da4c3c73SJérôme Glisse { 66008232a45SJérôme Glisse struct vm_area_struct *vma = range->vma; 66174eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 662da4c3c73SJérôme Glisse struct mm_walk mm_walk; 663da4c3c73SJérôme Glisse struct hmm *hmm; 664da4c3c73SJérôme Glisse 665da4c3c73SJérôme Glisse /* Sanity check, this really should not happen ! */ 66608232a45SJérôme Glisse if (range->start < vma->vm_start || range->start >= vma->vm_end) 667da4c3c73SJérôme Glisse return -EINVAL; 66808232a45SJérôme Glisse if (range->end < vma->vm_start || range->end > vma->vm_end) 669da4c3c73SJérôme Glisse return -EINVAL; 670da4c3c73SJérôme Glisse 671da4c3c73SJérôme Glisse hmm = hmm_register(vma->vm_mm); 672da4c3c73SJérôme Glisse if (!hmm) 673da4c3c73SJérôme Glisse return -ENOMEM; 674da4c3c73SJérôme Glisse /* Caller must have registered a mirror, via hmm_mirror_register() ! */ 675da4c3c73SJérôme Glisse if (!hmm->mmu_notifier.ops) 676da4c3c73SJérôme Glisse return -EINVAL; 677da4c3c73SJérôme Glisse 678855ce7d2SJérôme Glisse /* FIXME support hugetlb fs */ 679855ce7d2SJérôme Glisse if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) { 680855ce7d2SJérôme Glisse hmm_pfns_special(range); 681855ce7d2SJérôme Glisse return -EINVAL; 682855ce7d2SJérôme Glisse } 683855ce7d2SJérôme Glisse 68486586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 68586586a41SJérôme Glisse /* 68686586a41SJérôme Glisse * If vma do not allow read access, then assume that it does 68786586a41SJérôme Glisse * not allow write access, either. Architecture that allow 68886586a41SJérôme Glisse * write without read access are not supported by HMM, because 68986586a41SJérôme Glisse * operations such has atomic access would not work. 69086586a41SJérôme Glisse */ 691f88a1e90SJérôme Glisse hmm_pfns_clear(range, range->pfns, range->start, range->end); 69286586a41SJérôme Glisse return -EPERM; 69386586a41SJérôme Glisse } 69486586a41SJérôme Glisse 695da4c3c73SJérôme Glisse /* Initialize range to track CPU page table update */ 696da4c3c73SJérôme Glisse spin_lock(&hmm->lock); 697da4c3c73SJérôme Glisse range->valid = true; 698da4c3c73SJérôme Glisse list_add_rcu(&range->list, &hmm->ranges); 699da4c3c73SJérôme Glisse spin_unlock(&hmm->lock); 700da4c3c73SJérôme Glisse 70174eee180SJérôme Glisse hmm_vma_walk.fault = false; 70274eee180SJérôme Glisse hmm_vma_walk.range = range; 70374eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 70474eee180SJérôme Glisse 705da4c3c73SJérôme Glisse mm_walk.vma = vma; 706da4c3c73SJérôme Glisse mm_walk.mm = vma->vm_mm; 707da4c3c73SJérôme Glisse mm_walk.pte_entry = NULL; 708da4c3c73SJérôme Glisse mm_walk.test_walk = NULL; 709da4c3c73SJérôme Glisse mm_walk.hugetlb_entry = NULL; 710da4c3c73SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 711da4c3c73SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 712da4c3c73SJérôme Glisse 71308232a45SJérôme Glisse walk_page_range(range->start, range->end, &mm_walk); 714da4c3c73SJérôme Glisse return 0; 715da4c3c73SJérôme Glisse } 716da4c3c73SJérôme Glisse EXPORT_SYMBOL(hmm_vma_get_pfns); 717da4c3c73SJérôme Glisse 718da4c3c73SJérôme Glisse /* 719da4c3c73SJérôme Glisse * hmm_vma_range_done() - stop tracking change to CPU page table over a range 720da4c3c73SJérôme Glisse * @range: range being tracked 721da4c3c73SJérôme Glisse * Returns: false if range data has been invalidated, true otherwise 722da4c3c73SJérôme Glisse * 723da4c3c73SJérôme Glisse * Range struct is used to track updates to the CPU page table after a call to 724da4c3c73SJérôme Glisse * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done 725da4c3c73SJérôme Glisse * using the data, or wants to lock updates to the data it got from those 726da4c3c73SJérôme Glisse * functions, it must call the hmm_vma_range_done() function, which will then 727da4c3c73SJérôme Glisse * stop tracking CPU page table updates. 728da4c3c73SJérôme Glisse * 729da4c3c73SJérôme Glisse * Note that device driver must still implement general CPU page table update 730da4c3c73SJérôme Glisse * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using 731da4c3c73SJérôme Glisse * the mmu_notifier API directly. 732da4c3c73SJérôme Glisse * 733da4c3c73SJérôme Glisse * CPU page table update tracking done through hmm_range is only temporary and 734da4c3c73SJérôme Glisse * to be used while trying to duplicate CPU page table contents for a range of 735da4c3c73SJérôme Glisse * virtual addresses. 736da4c3c73SJérôme Glisse * 737da4c3c73SJérôme Glisse * There are two ways to use this : 738da4c3c73SJérôme Glisse * again: 73908232a45SJérôme Glisse * hmm_vma_get_pfns(range); or hmm_vma_fault(...); 740da4c3c73SJérôme Glisse * trans = device_build_page_table_update_transaction(pfns); 741da4c3c73SJérôme Glisse * device_page_table_lock(); 74208232a45SJérôme Glisse * if (!hmm_vma_range_done(range)) { 743da4c3c73SJérôme Glisse * device_page_table_unlock(); 744da4c3c73SJérôme Glisse * goto again; 745da4c3c73SJérôme Glisse * } 746da4c3c73SJérôme Glisse * device_commit_transaction(trans); 747da4c3c73SJérôme Glisse * device_page_table_unlock(); 748da4c3c73SJérôme Glisse * 749da4c3c73SJérôme Glisse * Or: 75008232a45SJérôme Glisse * hmm_vma_get_pfns(range); or hmm_vma_fault(...); 751da4c3c73SJérôme Glisse * device_page_table_lock(); 75208232a45SJérôme Glisse * hmm_vma_range_done(range); 75308232a45SJérôme Glisse * device_update_page_table(range->pfns); 754da4c3c73SJérôme Glisse * device_page_table_unlock(); 755da4c3c73SJérôme Glisse */ 75608232a45SJérôme Glisse bool hmm_vma_range_done(struct hmm_range *range) 757da4c3c73SJérôme Glisse { 758da4c3c73SJérôme Glisse unsigned long npages = (range->end - range->start) >> PAGE_SHIFT; 759da4c3c73SJérôme Glisse struct hmm *hmm; 760da4c3c73SJérôme Glisse 761da4c3c73SJérôme Glisse if (range->end <= range->start) { 762da4c3c73SJérôme Glisse BUG(); 763da4c3c73SJérôme Glisse return false; 764da4c3c73SJérôme Glisse } 765da4c3c73SJérôme Glisse 76608232a45SJérôme Glisse hmm = hmm_register(range->vma->vm_mm); 767da4c3c73SJérôme Glisse if (!hmm) { 768da4c3c73SJérôme Glisse memset(range->pfns, 0, sizeof(*range->pfns) * npages); 769da4c3c73SJérôme Glisse return false; 770da4c3c73SJérôme Glisse } 771da4c3c73SJérôme Glisse 772da4c3c73SJérôme Glisse spin_lock(&hmm->lock); 773da4c3c73SJérôme Glisse list_del_rcu(&range->list); 774da4c3c73SJérôme Glisse spin_unlock(&hmm->lock); 775da4c3c73SJérôme Glisse 776da4c3c73SJérôme Glisse return range->valid; 777da4c3c73SJérôme Glisse } 778da4c3c73SJérôme Glisse EXPORT_SYMBOL(hmm_vma_range_done); 77974eee180SJérôme Glisse 78074eee180SJérôme Glisse /* 78174eee180SJérôme Glisse * hmm_vma_fault() - try to fault some address in a virtual address range 78208232a45SJérôme Glisse * @range: range being faulted 78374eee180SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 78474eee180SJérôme Glisse * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop) 78574eee180SJérôme Glisse * 78674eee180SJérôme Glisse * This is similar to a regular CPU page fault except that it will not trigger 78774eee180SJérôme Glisse * any memory migration if the memory being faulted is not accessible by CPUs. 78874eee180SJérôme Glisse * 789ff05c0c6SJérôme Glisse * On error, for one virtual address in the range, the function will mark the 790ff05c0c6SJérôme Glisse * corresponding HMM pfn entry with an error flag. 79174eee180SJérôme Glisse * 79274eee180SJérôme Glisse * Expected use pattern: 79374eee180SJérôme Glisse * retry: 79474eee180SJérôme Glisse * down_read(&mm->mmap_sem); 79574eee180SJérôme Glisse * // Find vma and address device wants to fault, initialize hmm_pfn_t 79674eee180SJérôme Glisse * // array accordingly 79708232a45SJérôme Glisse * ret = hmm_vma_fault(range, write, block); 79874eee180SJérôme Glisse * switch (ret) { 79974eee180SJérôme Glisse * case -EAGAIN: 80008232a45SJérôme Glisse * hmm_vma_range_done(range); 80174eee180SJérôme Glisse * // You might want to rate limit or yield to play nicely, you may 80274eee180SJérôme Glisse * // also commit any valid pfn in the array assuming that you are 80374eee180SJérôme Glisse * // getting true from hmm_vma_range_monitor_end() 80474eee180SJérôme Glisse * goto retry; 80574eee180SJérôme Glisse * case 0: 80674eee180SJérôme Glisse * break; 80786586a41SJérôme Glisse * case -ENOMEM: 80886586a41SJérôme Glisse * case -EINVAL: 80986586a41SJérôme Glisse * case -EPERM: 81074eee180SJérôme Glisse * default: 81174eee180SJérôme Glisse * // Handle error ! 81274eee180SJérôme Glisse * up_read(&mm->mmap_sem) 81374eee180SJérôme Glisse * return; 81474eee180SJérôme Glisse * } 81574eee180SJérôme Glisse * // Take device driver lock that serialize device page table update 81674eee180SJérôme Glisse * driver_lock_device_page_table_update(); 81708232a45SJérôme Glisse * hmm_vma_range_done(range); 81874eee180SJérôme Glisse * // Commit pfns we got from hmm_vma_fault() 81974eee180SJérôme Glisse * driver_unlock_device_page_table_update(); 82074eee180SJérôme Glisse * up_read(&mm->mmap_sem) 82174eee180SJérôme Glisse * 82274eee180SJérôme Glisse * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0) 82374eee180SJérôme Glisse * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION ! 82474eee180SJérôme Glisse * 82574eee180SJérôme Glisse * YOU HAVE BEEN WARNED ! 82674eee180SJérôme Glisse */ 8272aee09d8SJérôme Glisse int hmm_vma_fault(struct hmm_range *range, bool block) 82874eee180SJérôme Glisse { 82908232a45SJérôme Glisse struct vm_area_struct *vma = range->vma; 83008232a45SJérôme Glisse unsigned long start = range->start; 83174eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 83274eee180SJérôme Glisse struct mm_walk mm_walk; 83374eee180SJérôme Glisse struct hmm *hmm; 83474eee180SJérôme Glisse int ret; 83574eee180SJérôme Glisse 83674eee180SJérôme Glisse /* Sanity check, this really should not happen ! */ 83708232a45SJérôme Glisse if (range->start < vma->vm_start || range->start >= vma->vm_end) 83874eee180SJérôme Glisse return -EINVAL; 83908232a45SJérôme Glisse if (range->end < vma->vm_start || range->end > vma->vm_end) 84074eee180SJérôme Glisse return -EINVAL; 84174eee180SJérôme Glisse 84274eee180SJérôme Glisse hmm = hmm_register(vma->vm_mm); 84374eee180SJérôme Glisse if (!hmm) { 844f88a1e90SJérôme Glisse hmm_pfns_clear(range, range->pfns, range->start, range->end); 84574eee180SJérôme Glisse return -ENOMEM; 84674eee180SJérôme Glisse } 84774eee180SJérôme Glisse /* Caller must have registered a mirror using hmm_mirror_register() */ 84874eee180SJérôme Glisse if (!hmm->mmu_notifier.ops) 84974eee180SJérôme Glisse return -EINVAL; 85074eee180SJérôme Glisse 851855ce7d2SJérôme Glisse /* FIXME support hugetlb fs */ 852855ce7d2SJérôme Glisse if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) { 853855ce7d2SJérôme Glisse hmm_pfns_special(range); 854855ce7d2SJérôme Glisse return -EINVAL; 855855ce7d2SJérôme Glisse } 856855ce7d2SJérôme Glisse 85786586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 85886586a41SJérôme Glisse /* 85986586a41SJérôme Glisse * If vma do not allow read access, then assume that it does 86086586a41SJérôme Glisse * not allow write access, either. Architecture that allow 86186586a41SJérôme Glisse * write without read access are not supported by HMM, because 86286586a41SJérôme Glisse * operations such has atomic access would not work. 86386586a41SJérôme Glisse */ 864f88a1e90SJérôme Glisse hmm_pfns_clear(range, range->pfns, range->start, range->end); 86586586a41SJérôme Glisse return -EPERM; 86686586a41SJérôme Glisse } 86774eee180SJérôme Glisse 86886586a41SJérôme Glisse /* Initialize range to track CPU page table update */ 86986586a41SJérôme Glisse spin_lock(&hmm->lock); 87086586a41SJérôme Glisse range->valid = true; 87186586a41SJérôme Glisse list_add_rcu(&range->list, &hmm->ranges); 87286586a41SJérôme Glisse spin_unlock(&hmm->lock); 87386586a41SJérôme Glisse 87474eee180SJérôme Glisse hmm_vma_walk.fault = true; 87574eee180SJérôme Glisse hmm_vma_walk.block = block; 87674eee180SJérôme Glisse hmm_vma_walk.range = range; 87774eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 87874eee180SJérôme Glisse hmm_vma_walk.last = range->start; 87974eee180SJérôme Glisse 88074eee180SJérôme Glisse mm_walk.vma = vma; 88174eee180SJérôme Glisse mm_walk.mm = vma->vm_mm; 88274eee180SJérôme Glisse mm_walk.pte_entry = NULL; 88374eee180SJérôme Glisse mm_walk.test_walk = NULL; 88474eee180SJérôme Glisse mm_walk.hugetlb_entry = NULL; 88574eee180SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 88674eee180SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 88774eee180SJérôme Glisse 88874eee180SJérôme Glisse do { 88908232a45SJérôme Glisse ret = walk_page_range(start, range->end, &mm_walk); 89074eee180SJérôme Glisse start = hmm_vma_walk.last; 89174eee180SJérôme Glisse } while (ret == -EAGAIN); 89274eee180SJérôme Glisse 89374eee180SJérôme Glisse if (ret) { 89474eee180SJérôme Glisse unsigned long i; 89574eee180SJérôme Glisse 89674eee180SJérôme Glisse i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 897f88a1e90SJérôme Glisse hmm_pfns_clear(range, &range->pfns[i], hmm_vma_walk.last, 898f88a1e90SJérôme Glisse range->end); 89908232a45SJérôme Glisse hmm_vma_range_done(range); 90074eee180SJérôme Glisse } 90174eee180SJérôme Glisse return ret; 90274eee180SJérôme Glisse } 90374eee180SJérôme Glisse EXPORT_SYMBOL(hmm_vma_fault); 904c0b12405SJérôme Glisse #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ 9054ef589dcSJérôme Glisse 9064ef589dcSJérôme Glisse 907df6ad698SJérôme Glisse #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC) 9084ef589dcSJérôme Glisse struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma, 9094ef589dcSJérôme Glisse unsigned long addr) 9104ef589dcSJérôme Glisse { 9114ef589dcSJérôme Glisse struct page *page; 9124ef589dcSJérôme Glisse 9134ef589dcSJérôme Glisse page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 9144ef589dcSJérôme Glisse if (!page) 9154ef589dcSJérôme Glisse return NULL; 9164ef589dcSJérôme Glisse lock_page(page); 9174ef589dcSJérôme Glisse return page; 9184ef589dcSJérôme Glisse } 9194ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_vma_alloc_locked_page); 9204ef589dcSJérôme Glisse 9214ef589dcSJérôme Glisse 9224ef589dcSJérôme Glisse static void hmm_devmem_ref_release(struct percpu_ref *ref) 9234ef589dcSJérôme Glisse { 9244ef589dcSJérôme Glisse struct hmm_devmem *devmem; 9254ef589dcSJérôme Glisse 9264ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 9274ef589dcSJérôme Glisse complete(&devmem->completion); 9284ef589dcSJérôme Glisse } 9294ef589dcSJérôme Glisse 9304ef589dcSJérôme Glisse static void hmm_devmem_ref_exit(void *data) 9314ef589dcSJérôme Glisse { 9324ef589dcSJérôme Glisse struct percpu_ref *ref = data; 9334ef589dcSJérôme Glisse struct hmm_devmem *devmem; 9344ef589dcSJérôme Glisse 9354ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 9364ef589dcSJérôme Glisse percpu_ref_exit(ref); 9374ef589dcSJérôme Glisse devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data); 9384ef589dcSJérôme Glisse } 9394ef589dcSJérôme Glisse 9404ef589dcSJérôme Glisse static void hmm_devmem_ref_kill(void *data) 9414ef589dcSJérôme Glisse { 9424ef589dcSJérôme Glisse struct percpu_ref *ref = data; 9434ef589dcSJérôme Glisse struct hmm_devmem *devmem; 9444ef589dcSJérôme Glisse 9454ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 9464ef589dcSJérôme Glisse percpu_ref_kill(ref); 9474ef589dcSJérôme Glisse wait_for_completion(&devmem->completion); 9484ef589dcSJérôme Glisse devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data); 9494ef589dcSJérôme Glisse } 9504ef589dcSJérôme Glisse 9514ef589dcSJérôme Glisse static int hmm_devmem_fault(struct vm_area_struct *vma, 9524ef589dcSJérôme Glisse unsigned long addr, 9534ef589dcSJérôme Glisse const struct page *page, 9544ef589dcSJérôme Glisse unsigned int flags, 9554ef589dcSJérôme Glisse pmd_t *pmdp) 9564ef589dcSJérôme Glisse { 9574ef589dcSJérôme Glisse struct hmm_devmem *devmem = page->pgmap->data; 9584ef589dcSJérôme Glisse 9594ef589dcSJérôme Glisse return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp); 9604ef589dcSJérôme Glisse } 9614ef589dcSJérôme Glisse 9624ef589dcSJérôme Glisse static void hmm_devmem_free(struct page *page, void *data) 9634ef589dcSJérôme Glisse { 9644ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 9654ef589dcSJérôme Glisse 966*2fa147bdSDan Williams page->mapping = NULL; 967*2fa147bdSDan Williams 9684ef589dcSJérôme Glisse devmem->ops->free(devmem, page); 9694ef589dcSJérôme Glisse } 9704ef589dcSJérôme Glisse 9714ef589dcSJérôme Glisse static DEFINE_MUTEX(hmm_devmem_lock); 9724ef589dcSJérôme Glisse static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL); 9734ef589dcSJérôme Glisse 9744ef589dcSJérôme Glisse static void hmm_devmem_radix_release(struct resource *resource) 9754ef589dcSJérôme Glisse { 976fec11bc0SColin Ian King resource_size_t key, align_start, align_size; 9774ef589dcSJérôme Glisse 9784ef589dcSJérôme Glisse align_start = resource->start & ~(PA_SECTION_SIZE - 1); 9794ef589dcSJérôme Glisse align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE); 9804ef589dcSJérôme Glisse 9814ef589dcSJérôme Glisse mutex_lock(&hmm_devmem_lock); 9824ef589dcSJérôme Glisse for (key = resource->start; 9834ef589dcSJérôme Glisse key <= resource->end; 9844ef589dcSJérôme Glisse key += PA_SECTION_SIZE) 9854ef589dcSJérôme Glisse radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT); 9864ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 9874ef589dcSJérôme Glisse } 9884ef589dcSJérôme Glisse 9894ef589dcSJérôme Glisse static void hmm_devmem_release(struct device *dev, void *data) 9904ef589dcSJérôme Glisse { 9914ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 9924ef589dcSJérôme Glisse struct resource *resource = devmem->resource; 9934ef589dcSJérôme Glisse unsigned long start_pfn, npages; 9944ef589dcSJérôme Glisse struct zone *zone; 9954ef589dcSJérôme Glisse struct page *page; 9964ef589dcSJérôme Glisse 9974ef589dcSJérôme Glisse if (percpu_ref_tryget_live(&devmem->ref)) { 9984ef589dcSJérôme Glisse dev_WARN(dev, "%s: page mapping is still live!\n", __func__); 9994ef589dcSJérôme Glisse percpu_ref_put(&devmem->ref); 10004ef589dcSJérôme Glisse } 10014ef589dcSJérôme Glisse 10024ef589dcSJérôme Glisse /* pages are dead and unused, undo the arch mapping */ 10034ef589dcSJérôme Glisse start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT; 10044ef589dcSJérôme Glisse npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT; 10054ef589dcSJérôme Glisse 10064ef589dcSJérôme Glisse page = pfn_to_page(start_pfn); 10074ef589dcSJérôme Glisse zone = page_zone(page); 10084ef589dcSJérôme Glisse 10094ef589dcSJérôme Glisse mem_hotplug_begin(); 1010d3df0a42SJérôme Glisse if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY) 1011da024512SChristoph Hellwig __remove_pages(zone, start_pfn, npages, NULL); 1012d3df0a42SJérôme Glisse else 1013d3df0a42SJérôme Glisse arch_remove_memory(start_pfn << PAGE_SHIFT, 1014da024512SChristoph Hellwig npages << PAGE_SHIFT, NULL); 10154ef589dcSJérôme Glisse mem_hotplug_done(); 10164ef589dcSJérôme Glisse 10174ef589dcSJérôme Glisse hmm_devmem_radix_release(resource); 10184ef589dcSJérôme Glisse } 10194ef589dcSJérôme Glisse 10204ef589dcSJérôme Glisse static int hmm_devmem_pages_create(struct hmm_devmem *devmem) 10214ef589dcSJérôme Glisse { 10224ef589dcSJérôme Glisse resource_size_t key, align_start, align_size, align_end; 10234ef589dcSJérôme Glisse struct device *device = devmem->device; 10244ef589dcSJérôme Glisse int ret, nid, is_ram; 10254ef589dcSJérôme Glisse unsigned long pfn; 10264ef589dcSJérôme Glisse 10274ef589dcSJérôme Glisse align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1); 10284ef589dcSJérôme Glisse align_size = ALIGN(devmem->resource->start + 10294ef589dcSJérôme Glisse resource_size(devmem->resource), 10304ef589dcSJérôme Glisse PA_SECTION_SIZE) - align_start; 10314ef589dcSJérôme Glisse 10324ef589dcSJérôme Glisse is_ram = region_intersects(align_start, align_size, 10334ef589dcSJérôme Glisse IORESOURCE_SYSTEM_RAM, 10344ef589dcSJérôme Glisse IORES_DESC_NONE); 10354ef589dcSJérôme Glisse if (is_ram == REGION_MIXED) { 10364ef589dcSJérôme Glisse WARN_ONCE(1, "%s attempted on mixed region %pr\n", 10374ef589dcSJérôme Glisse __func__, devmem->resource); 10384ef589dcSJérôme Glisse return -ENXIO; 10394ef589dcSJérôme Glisse } 10404ef589dcSJérôme Glisse if (is_ram == REGION_INTERSECTS) 10414ef589dcSJérôme Glisse return -ENXIO; 10424ef589dcSJérôme Glisse 1043d3df0a42SJérôme Glisse if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY) 1044d3df0a42SJérôme Glisse devmem->pagemap.type = MEMORY_DEVICE_PUBLIC; 1045d3df0a42SJérôme Glisse else 10464ef589dcSJérôme Glisse devmem->pagemap.type = MEMORY_DEVICE_PRIVATE; 1047d3df0a42SJérôme Glisse 1048e7744aa2SLogan Gunthorpe devmem->pagemap.res = *devmem->resource; 10494ef589dcSJérôme Glisse devmem->pagemap.page_fault = hmm_devmem_fault; 10504ef589dcSJérôme Glisse devmem->pagemap.page_free = hmm_devmem_free; 10514ef589dcSJérôme Glisse devmem->pagemap.dev = devmem->device; 10524ef589dcSJérôme Glisse devmem->pagemap.ref = &devmem->ref; 10534ef589dcSJérôme Glisse devmem->pagemap.data = devmem; 10544ef589dcSJérôme Glisse 10554ef589dcSJérôme Glisse mutex_lock(&hmm_devmem_lock); 10564ef589dcSJérôme Glisse align_end = align_start + align_size - 1; 10574ef589dcSJérôme Glisse for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) { 10584ef589dcSJérôme Glisse struct hmm_devmem *dup; 10594ef589dcSJérôme Glisse 106018be460eSTejun Heo dup = radix_tree_lookup(&hmm_devmem_radix, 106118be460eSTejun Heo key >> PA_SECTION_SHIFT); 10624ef589dcSJérôme Glisse if (dup) { 10634ef589dcSJérôme Glisse dev_err(device, "%s: collides with mapping for %s\n", 10644ef589dcSJérôme Glisse __func__, dev_name(dup->device)); 10654ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 10664ef589dcSJérôme Glisse ret = -EBUSY; 10674ef589dcSJérôme Glisse goto error; 10684ef589dcSJérôme Glisse } 10694ef589dcSJérôme Glisse ret = radix_tree_insert(&hmm_devmem_radix, 10704ef589dcSJérôme Glisse key >> PA_SECTION_SHIFT, 10714ef589dcSJérôme Glisse devmem); 10724ef589dcSJérôme Glisse if (ret) { 10734ef589dcSJérôme Glisse dev_err(device, "%s: failed: %d\n", __func__, ret); 10744ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 10754ef589dcSJérôme Glisse goto error_radix; 10764ef589dcSJérôme Glisse } 10774ef589dcSJérôme Glisse } 10784ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 10794ef589dcSJérôme Glisse 10804ef589dcSJérôme Glisse nid = dev_to_node(device); 10814ef589dcSJérôme Glisse if (nid < 0) 10824ef589dcSJérôme Glisse nid = numa_mem_id(); 10834ef589dcSJérôme Glisse 10844ef589dcSJérôme Glisse mem_hotplug_begin(); 10854ef589dcSJérôme Glisse /* 10864ef589dcSJérôme Glisse * For device private memory we call add_pages() as we only need to 10874ef589dcSJérôme Glisse * allocate and initialize struct page for the device memory. More- 10884ef589dcSJérôme Glisse * over the device memory is un-accessible thus we do not want to 10894ef589dcSJérôme Glisse * create a linear mapping for the memory like arch_add_memory() 10904ef589dcSJérôme Glisse * would do. 1091d3df0a42SJérôme Glisse * 1092d3df0a42SJérôme Glisse * For device public memory, which is accesible by the CPU, we do 1093d3df0a42SJérôme Glisse * want the linear mapping and thus use arch_add_memory(). 10944ef589dcSJérôme Glisse */ 1095d3df0a42SJérôme Glisse if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC) 109624e6d5a5SChristoph Hellwig ret = arch_add_memory(nid, align_start, align_size, NULL, 109724e6d5a5SChristoph Hellwig false); 1098d3df0a42SJérôme Glisse else 10994ef589dcSJérôme Glisse ret = add_pages(nid, align_start >> PAGE_SHIFT, 110024e6d5a5SChristoph Hellwig align_size >> PAGE_SHIFT, NULL, false); 11014ef589dcSJérôme Glisse if (ret) { 11024ef589dcSJérôme Glisse mem_hotplug_done(); 11034ef589dcSJérôme Glisse goto error_add_memory; 11044ef589dcSJérôme Glisse } 11054ef589dcSJérôme Glisse move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE], 11064ef589dcSJérôme Glisse align_start >> PAGE_SHIFT, 1107a99583e7SChristoph Hellwig align_size >> PAGE_SHIFT, NULL); 11084ef589dcSJérôme Glisse mem_hotplug_done(); 11094ef589dcSJérôme Glisse 11104ef589dcSJérôme Glisse for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) { 11114ef589dcSJérôme Glisse struct page *page = pfn_to_page(pfn); 11124ef589dcSJérôme Glisse 11134ef589dcSJérôme Glisse page->pgmap = &devmem->pagemap; 11144ef589dcSJérôme Glisse } 11154ef589dcSJérôme Glisse return 0; 11164ef589dcSJérôme Glisse 11174ef589dcSJérôme Glisse error_add_memory: 11184ef589dcSJérôme Glisse untrack_pfn(NULL, PHYS_PFN(align_start), align_size); 11194ef589dcSJérôme Glisse error_radix: 11204ef589dcSJérôme Glisse hmm_devmem_radix_release(devmem->resource); 11214ef589dcSJérôme Glisse error: 11224ef589dcSJérôme Glisse return ret; 11234ef589dcSJérôme Glisse } 11244ef589dcSJérôme Glisse 11254ef589dcSJérôme Glisse static int hmm_devmem_match(struct device *dev, void *data, void *match_data) 11264ef589dcSJérôme Glisse { 11274ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 11284ef589dcSJérôme Glisse 11294ef589dcSJérôme Glisse return devmem->resource == match_data; 11304ef589dcSJérôme Glisse } 11314ef589dcSJérôme Glisse 11324ef589dcSJérôme Glisse static void hmm_devmem_pages_remove(struct hmm_devmem *devmem) 11334ef589dcSJérôme Glisse { 11344ef589dcSJérôme Glisse devres_release(devmem->device, &hmm_devmem_release, 11354ef589dcSJérôme Glisse &hmm_devmem_match, devmem->resource); 11364ef589dcSJérôme Glisse } 11374ef589dcSJérôme Glisse 11384ef589dcSJérôme Glisse /* 11394ef589dcSJérôme Glisse * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory 11404ef589dcSJérôme Glisse * 11414ef589dcSJérôme Glisse * @ops: memory event device driver callback (see struct hmm_devmem_ops) 11424ef589dcSJérôme Glisse * @device: device struct to bind the resource too 11434ef589dcSJérôme Glisse * @size: size in bytes of the device memory to add 11444ef589dcSJérôme Glisse * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise 11454ef589dcSJérôme Glisse * 11464ef589dcSJérôme Glisse * This function first finds an empty range of physical address big enough to 11474ef589dcSJérôme Glisse * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which 11484ef589dcSJérôme Glisse * in turn allocates struct pages. It does not do anything beyond that; all 11494ef589dcSJérôme Glisse * events affecting the memory will go through the various callbacks provided 11504ef589dcSJérôme Glisse * by hmm_devmem_ops struct. 11514ef589dcSJérôme Glisse * 11524ef589dcSJérôme Glisse * Device driver should call this function during device initialization and 11534ef589dcSJérôme Glisse * is then responsible of memory management. HMM only provides helpers. 11544ef589dcSJérôme Glisse */ 11554ef589dcSJérôme Glisse struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, 11564ef589dcSJérôme Glisse struct device *device, 11574ef589dcSJérôme Glisse unsigned long size) 11584ef589dcSJérôme Glisse { 11594ef589dcSJérôme Glisse struct hmm_devmem *devmem; 11604ef589dcSJérôme Glisse resource_size_t addr; 11614ef589dcSJérôme Glisse int ret; 11624ef589dcSJérôme Glisse 1163e7638488SDan Williams dev_pagemap_get_ops(); 11644ef589dcSJérôme Glisse 11654ef589dcSJérôme Glisse devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem), 11664ef589dcSJérôme Glisse GFP_KERNEL, dev_to_node(device)); 11674ef589dcSJérôme Glisse if (!devmem) 11684ef589dcSJérôme Glisse return ERR_PTR(-ENOMEM); 11694ef589dcSJérôme Glisse 11704ef589dcSJérôme Glisse init_completion(&devmem->completion); 11714ef589dcSJérôme Glisse devmem->pfn_first = -1UL; 11724ef589dcSJérôme Glisse devmem->pfn_last = -1UL; 11734ef589dcSJérôme Glisse devmem->resource = NULL; 11744ef589dcSJérôme Glisse devmem->device = device; 11754ef589dcSJérôme Glisse devmem->ops = ops; 11764ef589dcSJérôme Glisse 11774ef589dcSJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 11784ef589dcSJérôme Glisse 0, GFP_KERNEL); 11794ef589dcSJérôme Glisse if (ret) 11804ef589dcSJérôme Glisse goto error_percpu_ref; 11814ef589dcSJérôme Glisse 11824ef589dcSJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref); 11834ef589dcSJérôme Glisse if (ret) 11844ef589dcSJérôme Glisse goto error_devm_add_action; 11854ef589dcSJérôme Glisse 11864ef589dcSJérôme Glisse size = ALIGN(size, PA_SECTION_SIZE); 11874ef589dcSJérôme Glisse addr = min((unsigned long)iomem_resource.end, 11884ef589dcSJérôme Glisse (1UL << MAX_PHYSMEM_BITS) - 1); 11894ef589dcSJérôme Glisse addr = addr - size + 1UL; 11904ef589dcSJérôme Glisse 11914ef589dcSJérôme Glisse /* 11924ef589dcSJérôme Glisse * FIXME add a new helper to quickly walk resource tree and find free 11934ef589dcSJérôme Glisse * range 11944ef589dcSJérôme Glisse * 11954ef589dcSJérôme Glisse * FIXME what about ioport_resource resource ? 11964ef589dcSJérôme Glisse */ 11974ef589dcSJérôme Glisse for (; addr > size && addr >= iomem_resource.start; addr -= size) { 11984ef589dcSJérôme Glisse ret = region_intersects(addr, size, 0, IORES_DESC_NONE); 11994ef589dcSJérôme Glisse if (ret != REGION_DISJOINT) 12004ef589dcSJérôme Glisse continue; 12014ef589dcSJérôme Glisse 12024ef589dcSJérôme Glisse devmem->resource = devm_request_mem_region(device, addr, size, 12034ef589dcSJérôme Glisse dev_name(device)); 12044ef589dcSJérôme Glisse if (!devmem->resource) { 12054ef589dcSJérôme Glisse ret = -ENOMEM; 12064ef589dcSJérôme Glisse goto error_no_resource; 12074ef589dcSJérôme Glisse } 12084ef589dcSJérôme Glisse break; 12094ef589dcSJérôme Glisse } 12104ef589dcSJérôme Glisse if (!devmem->resource) { 12114ef589dcSJérôme Glisse ret = -ERANGE; 12124ef589dcSJérôme Glisse goto error_no_resource; 12134ef589dcSJérôme Glisse } 12144ef589dcSJérôme Glisse 12154ef589dcSJérôme Glisse devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY; 12164ef589dcSJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 12174ef589dcSJérôme Glisse devmem->pfn_last = devmem->pfn_first + 12184ef589dcSJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 12194ef589dcSJérôme Glisse 12204ef589dcSJérôme Glisse ret = hmm_devmem_pages_create(devmem); 12214ef589dcSJérôme Glisse if (ret) 12224ef589dcSJérôme Glisse goto error_pages; 12234ef589dcSJérôme Glisse 12244ef589dcSJérôme Glisse devres_add(device, devmem); 12254ef589dcSJérôme Glisse 12264ef589dcSJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref); 12274ef589dcSJérôme Glisse if (ret) { 12284ef589dcSJérôme Glisse hmm_devmem_remove(devmem); 12294ef589dcSJérôme Glisse return ERR_PTR(ret); 12304ef589dcSJérôme Glisse } 12314ef589dcSJérôme Glisse 12324ef589dcSJérôme Glisse return devmem; 12334ef589dcSJérôme Glisse 12344ef589dcSJérôme Glisse error_pages: 12354ef589dcSJérôme Glisse devm_release_mem_region(device, devmem->resource->start, 12364ef589dcSJérôme Glisse resource_size(devmem->resource)); 12374ef589dcSJérôme Glisse error_no_resource: 12384ef589dcSJérôme Glisse error_devm_add_action: 12394ef589dcSJérôme Glisse hmm_devmem_ref_kill(&devmem->ref); 12404ef589dcSJérôme Glisse hmm_devmem_ref_exit(&devmem->ref); 12414ef589dcSJérôme Glisse error_percpu_ref: 12424ef589dcSJérôme Glisse devres_free(devmem); 12434ef589dcSJérôme Glisse return ERR_PTR(ret); 12444ef589dcSJérôme Glisse } 12454ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_devmem_add); 12464ef589dcSJérôme Glisse 1247d3df0a42SJérôme Glisse struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops, 1248d3df0a42SJérôme Glisse struct device *device, 1249d3df0a42SJérôme Glisse struct resource *res) 1250d3df0a42SJérôme Glisse { 1251d3df0a42SJérôme Glisse struct hmm_devmem *devmem; 1252d3df0a42SJérôme Glisse int ret; 1253d3df0a42SJérôme Glisse 1254d3df0a42SJérôme Glisse if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY) 1255d3df0a42SJérôme Glisse return ERR_PTR(-EINVAL); 1256d3df0a42SJérôme Glisse 1257e7638488SDan Williams dev_pagemap_get_ops(); 1258d3df0a42SJérôme Glisse 1259d3df0a42SJérôme Glisse devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem), 1260d3df0a42SJérôme Glisse GFP_KERNEL, dev_to_node(device)); 1261d3df0a42SJérôme Glisse if (!devmem) 1262d3df0a42SJérôme Glisse return ERR_PTR(-ENOMEM); 1263d3df0a42SJérôme Glisse 1264d3df0a42SJérôme Glisse init_completion(&devmem->completion); 1265d3df0a42SJérôme Glisse devmem->pfn_first = -1UL; 1266d3df0a42SJérôme Glisse devmem->pfn_last = -1UL; 1267d3df0a42SJérôme Glisse devmem->resource = res; 1268d3df0a42SJérôme Glisse devmem->device = device; 1269d3df0a42SJérôme Glisse devmem->ops = ops; 1270d3df0a42SJérôme Glisse 1271d3df0a42SJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 1272d3df0a42SJérôme Glisse 0, GFP_KERNEL); 1273d3df0a42SJérôme Glisse if (ret) 1274d3df0a42SJérôme Glisse goto error_percpu_ref; 1275d3df0a42SJérôme Glisse 1276d3df0a42SJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref); 1277d3df0a42SJérôme Glisse if (ret) 1278d3df0a42SJérôme Glisse goto error_devm_add_action; 1279d3df0a42SJérôme Glisse 1280d3df0a42SJérôme Glisse 1281d3df0a42SJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 1282d3df0a42SJérôme Glisse devmem->pfn_last = devmem->pfn_first + 1283d3df0a42SJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1284d3df0a42SJérôme Glisse 1285d3df0a42SJérôme Glisse ret = hmm_devmem_pages_create(devmem); 1286d3df0a42SJérôme Glisse if (ret) 1287d3df0a42SJérôme Glisse goto error_devm_add_action; 1288d3df0a42SJérôme Glisse 1289d3df0a42SJérôme Glisse devres_add(device, devmem); 1290d3df0a42SJérôme Glisse 1291d3df0a42SJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref); 1292d3df0a42SJérôme Glisse if (ret) { 1293d3df0a42SJérôme Glisse hmm_devmem_remove(devmem); 1294d3df0a42SJérôme Glisse return ERR_PTR(ret); 1295d3df0a42SJérôme Glisse } 1296d3df0a42SJérôme Glisse 1297d3df0a42SJérôme Glisse return devmem; 1298d3df0a42SJérôme Glisse 1299d3df0a42SJérôme Glisse error_devm_add_action: 1300d3df0a42SJérôme Glisse hmm_devmem_ref_kill(&devmem->ref); 1301d3df0a42SJérôme Glisse hmm_devmem_ref_exit(&devmem->ref); 1302d3df0a42SJérôme Glisse error_percpu_ref: 1303d3df0a42SJérôme Glisse devres_free(devmem); 1304d3df0a42SJérôme Glisse return ERR_PTR(ret); 1305d3df0a42SJérôme Glisse } 1306d3df0a42SJérôme Glisse EXPORT_SYMBOL(hmm_devmem_add_resource); 1307d3df0a42SJérôme Glisse 13084ef589dcSJérôme Glisse /* 13094ef589dcSJérôme Glisse * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE) 13104ef589dcSJérôme Glisse * 13114ef589dcSJérôme Glisse * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory 13124ef589dcSJérôme Glisse * 13134ef589dcSJérôme Glisse * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf 13144ef589dcSJérôme Glisse * of the device driver. It will free struct page and remove the resource that 13154ef589dcSJérôme Glisse * reserved the physical address range for this device memory. 13164ef589dcSJérôme Glisse */ 13174ef589dcSJérôme Glisse void hmm_devmem_remove(struct hmm_devmem *devmem) 13184ef589dcSJérôme Glisse { 13194ef589dcSJérôme Glisse resource_size_t start, size; 13204ef589dcSJérôme Glisse struct device *device; 1321d3df0a42SJérôme Glisse bool cdm = false; 13224ef589dcSJérôme Glisse 13234ef589dcSJérôme Glisse if (!devmem) 13244ef589dcSJérôme Glisse return; 13254ef589dcSJérôme Glisse 13264ef589dcSJérôme Glisse device = devmem->device; 13274ef589dcSJérôme Glisse start = devmem->resource->start; 13284ef589dcSJérôme Glisse size = resource_size(devmem->resource); 13294ef589dcSJérôme Glisse 1330d3df0a42SJérôme Glisse cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY; 13314ef589dcSJérôme Glisse hmm_devmem_ref_kill(&devmem->ref); 13324ef589dcSJérôme Glisse hmm_devmem_ref_exit(&devmem->ref); 13334ef589dcSJérôme Glisse hmm_devmem_pages_remove(devmem); 13344ef589dcSJérôme Glisse 1335d3df0a42SJérôme Glisse if (!cdm) 13364ef589dcSJérôme Glisse devm_release_mem_region(device, start, size); 13374ef589dcSJérôme Glisse } 13384ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_devmem_remove); 1339858b54daSJérôme Glisse 1340858b54daSJérôme Glisse /* 1341858b54daSJérôme Glisse * A device driver that wants to handle multiple devices memory through a 1342858b54daSJérôme Glisse * single fake device can use hmm_device to do so. This is purely a helper 1343858b54daSJérôme Glisse * and it is not needed to make use of any HMM functionality. 1344858b54daSJérôme Glisse */ 1345858b54daSJérôme Glisse #define HMM_DEVICE_MAX 256 1346858b54daSJérôme Glisse 1347858b54daSJérôme Glisse static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX); 1348858b54daSJérôme Glisse static DEFINE_SPINLOCK(hmm_device_lock); 1349858b54daSJérôme Glisse static struct class *hmm_device_class; 1350858b54daSJérôme Glisse static dev_t hmm_device_devt; 1351858b54daSJérôme Glisse 1352858b54daSJérôme Glisse static void hmm_device_release(struct device *device) 1353858b54daSJérôme Glisse { 1354858b54daSJérôme Glisse struct hmm_device *hmm_device; 1355858b54daSJérôme Glisse 1356858b54daSJérôme Glisse hmm_device = container_of(device, struct hmm_device, device); 1357858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1358858b54daSJérôme Glisse clear_bit(hmm_device->minor, hmm_device_mask); 1359858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1360858b54daSJérôme Glisse 1361858b54daSJérôme Glisse kfree(hmm_device); 1362858b54daSJérôme Glisse } 1363858b54daSJérôme Glisse 1364858b54daSJérôme Glisse struct hmm_device *hmm_device_new(void *drvdata) 1365858b54daSJérôme Glisse { 1366858b54daSJérôme Glisse struct hmm_device *hmm_device; 1367858b54daSJérôme Glisse 1368858b54daSJérôme Glisse hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL); 1369858b54daSJérôme Glisse if (!hmm_device) 1370858b54daSJérôme Glisse return ERR_PTR(-ENOMEM); 1371858b54daSJérôme Glisse 1372858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1373858b54daSJérôme Glisse hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX); 1374858b54daSJérôme Glisse if (hmm_device->minor >= HMM_DEVICE_MAX) { 1375858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1376858b54daSJérôme Glisse kfree(hmm_device); 1377858b54daSJérôme Glisse return ERR_PTR(-EBUSY); 1378858b54daSJérôme Glisse } 1379858b54daSJérôme Glisse set_bit(hmm_device->minor, hmm_device_mask); 1380858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1381858b54daSJérôme Glisse 1382858b54daSJérôme Glisse dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor); 1383858b54daSJérôme Glisse hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt), 1384858b54daSJérôme Glisse hmm_device->minor); 1385858b54daSJérôme Glisse hmm_device->device.release = hmm_device_release; 1386858b54daSJérôme Glisse dev_set_drvdata(&hmm_device->device, drvdata); 1387858b54daSJérôme Glisse hmm_device->device.class = hmm_device_class; 1388858b54daSJérôme Glisse device_initialize(&hmm_device->device); 1389858b54daSJérôme Glisse 1390858b54daSJérôme Glisse return hmm_device; 1391858b54daSJérôme Glisse } 1392858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_new); 1393858b54daSJérôme Glisse 1394858b54daSJérôme Glisse void hmm_device_put(struct hmm_device *hmm_device) 1395858b54daSJérôme Glisse { 1396858b54daSJérôme Glisse put_device(&hmm_device->device); 1397858b54daSJérôme Glisse } 1398858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_put); 1399858b54daSJérôme Glisse 1400858b54daSJérôme Glisse static int __init hmm_init(void) 1401858b54daSJérôme Glisse { 1402858b54daSJérôme Glisse int ret; 1403858b54daSJérôme Glisse 1404858b54daSJérôme Glisse ret = alloc_chrdev_region(&hmm_device_devt, 0, 1405858b54daSJérôme Glisse HMM_DEVICE_MAX, 1406858b54daSJérôme Glisse "hmm_device"); 1407858b54daSJérôme Glisse if (ret) 1408858b54daSJérôme Glisse return ret; 1409858b54daSJérôme Glisse 1410858b54daSJérôme Glisse hmm_device_class = class_create(THIS_MODULE, "hmm_device"); 1411858b54daSJérôme Glisse if (IS_ERR(hmm_device_class)) { 1412858b54daSJérôme Glisse unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX); 1413858b54daSJérôme Glisse return PTR_ERR(hmm_device_class); 1414858b54daSJérôme Glisse } 1415858b54daSJérôme Glisse return 0; 1416858b54daSJérôme Glisse } 1417858b54daSJérôme Glisse 1418858b54daSJérôme Glisse device_initcall(hmm_init); 1419df6ad698SJérôme Glisse #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ 1420