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 defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC) 397b2d55d2SJérôme Glisse /* 407b2d55d2SJérôme Glisse * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h 417b2d55d2SJérôme Glisse */ 427b2d55d2SJérôme Glisse DEFINE_STATIC_KEY_FALSE(device_private_key); 437b2d55d2SJérôme Glisse EXPORT_SYMBOL(device_private_key); 446b368cd4SJérôme Glisse #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ 457b2d55d2SJérôme Glisse 467b2d55d2SJérôme Glisse 476b368cd4SJérôme Glisse #if IS_ENABLED(CONFIG_HMM_MIRROR) 48c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops; 49c0b12405SJérôme Glisse 50133ff0eaSJérôme Glisse /* 51133ff0eaSJérôme Glisse * struct hmm - HMM per mm struct 52133ff0eaSJérôme Glisse * 53133ff0eaSJérôme Glisse * @mm: mm struct this HMM struct is bound to 54da4c3c73SJérôme Glisse * @lock: lock protecting ranges list 55c0b12405SJérôme Glisse * @sequence: we track updates to the CPU page table with a sequence number 56da4c3c73SJérôme Glisse * @ranges: list of range being snapshotted 57c0b12405SJérôme Glisse * @mirrors: list of mirrors for this mm 58c0b12405SJérôme Glisse * @mmu_notifier: mmu notifier to track updates to CPU page table 59c0b12405SJérôme Glisse * @mirrors_sem: read/write semaphore protecting the mirrors list 60133ff0eaSJérôme Glisse */ 61133ff0eaSJérôme Glisse struct hmm { 62133ff0eaSJérôme Glisse struct mm_struct *mm; 63da4c3c73SJérôme Glisse spinlock_t lock; 64c0b12405SJérôme Glisse atomic_t sequence; 65da4c3c73SJérôme Glisse struct list_head ranges; 66c0b12405SJérôme Glisse struct list_head mirrors; 67c0b12405SJérôme Glisse struct mmu_notifier mmu_notifier; 68c0b12405SJérôme Glisse struct rw_semaphore mirrors_sem; 69133ff0eaSJérôme Glisse }; 70133ff0eaSJérôme Glisse 71133ff0eaSJérôme Glisse /* 72133ff0eaSJérôme Glisse * hmm_register - register HMM against an mm (HMM internal) 73133ff0eaSJérôme Glisse * 74133ff0eaSJérôme Glisse * @mm: mm struct to attach to 75133ff0eaSJérôme Glisse * 76133ff0eaSJérôme Glisse * This is not intended to be used directly by device drivers. It allocates an 77133ff0eaSJérôme Glisse * HMM struct if mm does not have one, and initializes it. 78133ff0eaSJérôme Glisse */ 79133ff0eaSJérôme Glisse static struct hmm *hmm_register(struct mm_struct *mm) 80133ff0eaSJérôme Glisse { 81c0b12405SJérôme Glisse struct hmm *hmm = READ_ONCE(mm->hmm); 82c0b12405SJérôme Glisse bool cleanup = false; 83133ff0eaSJérôme Glisse 84133ff0eaSJérôme Glisse /* 85133ff0eaSJérôme Glisse * The hmm struct can only be freed once the mm_struct goes away, 86133ff0eaSJérôme Glisse * hence we should always have pre-allocated an new hmm struct 87133ff0eaSJérôme Glisse * above. 88133ff0eaSJérôme Glisse */ 89c0b12405SJérôme Glisse if (hmm) 90c0b12405SJérôme Glisse return hmm; 91c0b12405SJérôme Glisse 92c0b12405SJérôme Glisse hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); 93c0b12405SJérôme Glisse if (!hmm) 94c0b12405SJérôme Glisse return NULL; 95c0b12405SJérôme Glisse INIT_LIST_HEAD(&hmm->mirrors); 96c0b12405SJérôme Glisse init_rwsem(&hmm->mirrors_sem); 97c0b12405SJérôme Glisse atomic_set(&hmm->sequence, 0); 98c0b12405SJérôme Glisse hmm->mmu_notifier.ops = NULL; 99da4c3c73SJérôme Glisse INIT_LIST_HEAD(&hmm->ranges); 100da4c3c73SJérôme Glisse spin_lock_init(&hmm->lock); 101c0b12405SJérôme Glisse hmm->mm = mm; 102c0b12405SJérôme Glisse 103c0b12405SJérôme Glisse /* 104c0b12405SJérôme Glisse * We should only get here if hold the mmap_sem in write mode ie on 105c0b12405SJérôme Glisse * registration of first mirror through hmm_mirror_register() 106c0b12405SJérôme Glisse */ 107c0b12405SJérôme Glisse hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; 108c0b12405SJérôme Glisse if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) { 109c0b12405SJérôme Glisse kfree(hmm); 110c0b12405SJérôme Glisse return NULL; 111c0b12405SJérôme Glisse } 112c0b12405SJérôme Glisse 113c0b12405SJérôme Glisse spin_lock(&mm->page_table_lock); 114c0b12405SJérôme Glisse if (!mm->hmm) 115c0b12405SJérôme Glisse mm->hmm = hmm; 116c0b12405SJérôme Glisse else 117c0b12405SJérôme Glisse cleanup = true; 118c0b12405SJérôme Glisse spin_unlock(&mm->page_table_lock); 119c0b12405SJérôme Glisse 120c0b12405SJérôme Glisse if (cleanup) { 121c0b12405SJérôme Glisse mmu_notifier_unregister(&hmm->mmu_notifier, mm); 122c0b12405SJérôme Glisse kfree(hmm); 123c0b12405SJérôme Glisse } 124c0b12405SJérôme Glisse 125133ff0eaSJérôme Glisse return mm->hmm; 126133ff0eaSJérôme Glisse } 127133ff0eaSJérôme Glisse 128133ff0eaSJérôme Glisse void hmm_mm_destroy(struct mm_struct *mm) 129133ff0eaSJérôme Glisse { 130133ff0eaSJérôme Glisse kfree(mm->hmm); 131133ff0eaSJérôme Glisse } 132c0b12405SJérôme Glisse 133c0b12405SJérôme Glisse static void hmm_invalidate_range(struct hmm *hmm, 134c0b12405SJérôme Glisse enum hmm_update_type action, 135c0b12405SJérôme Glisse unsigned long start, 136c0b12405SJérôme Glisse unsigned long end) 137c0b12405SJérôme Glisse { 138c0b12405SJérôme Glisse struct hmm_mirror *mirror; 139da4c3c73SJérôme Glisse struct hmm_range *range; 140da4c3c73SJérôme Glisse 141da4c3c73SJérôme Glisse spin_lock(&hmm->lock); 142da4c3c73SJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 143da4c3c73SJérôme Glisse unsigned long addr, idx, npages; 144da4c3c73SJérôme Glisse 145da4c3c73SJérôme Glisse if (end < range->start || start >= range->end) 146da4c3c73SJérôme Glisse continue; 147da4c3c73SJérôme Glisse 148da4c3c73SJérôme Glisse range->valid = false; 149da4c3c73SJérôme Glisse addr = max(start, range->start); 150da4c3c73SJérôme Glisse idx = (addr - range->start) >> PAGE_SHIFT; 151da4c3c73SJérôme Glisse npages = (min(range->end, end) - addr) >> PAGE_SHIFT; 152da4c3c73SJérôme Glisse memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages); 153da4c3c73SJérôme Glisse } 154da4c3c73SJérôme Glisse spin_unlock(&hmm->lock); 155c0b12405SJérôme Glisse 156c0b12405SJérôme Glisse down_read(&hmm->mirrors_sem); 157c0b12405SJérôme Glisse list_for_each_entry(mirror, &hmm->mirrors, list) 158c0b12405SJérôme Glisse mirror->ops->sync_cpu_device_pagetables(mirror, action, 159c0b12405SJérôme Glisse start, end); 160c0b12405SJérôme Glisse up_read(&hmm->mirrors_sem); 161c0b12405SJérôme Glisse } 162c0b12405SJérôme Glisse 163c0b12405SJérôme Glisse static void hmm_invalidate_range_start(struct mmu_notifier *mn, 164c0b12405SJérôme Glisse struct mm_struct *mm, 165c0b12405SJérôme Glisse unsigned long start, 166c0b12405SJérôme Glisse unsigned long end) 167c0b12405SJérôme Glisse { 168c0b12405SJérôme Glisse struct hmm *hmm = mm->hmm; 169c0b12405SJérôme Glisse 170c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 171c0b12405SJérôme Glisse 172c0b12405SJérôme Glisse atomic_inc(&hmm->sequence); 173c0b12405SJérôme Glisse } 174c0b12405SJérôme Glisse 175c0b12405SJérôme Glisse static void hmm_invalidate_range_end(struct mmu_notifier *mn, 176c0b12405SJérôme Glisse struct mm_struct *mm, 177c0b12405SJérôme Glisse unsigned long start, 178c0b12405SJérôme Glisse unsigned long end) 179c0b12405SJérôme Glisse { 180c0b12405SJérôme Glisse struct hmm *hmm = mm->hmm; 181c0b12405SJérôme Glisse 182c0b12405SJérôme Glisse VM_BUG_ON(!hmm); 183c0b12405SJérôme Glisse 184c0b12405SJérôme Glisse hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end); 185c0b12405SJérôme Glisse } 186c0b12405SJérôme Glisse 187c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { 188c0b12405SJérôme Glisse .invalidate_range_start = hmm_invalidate_range_start, 189c0b12405SJérôme Glisse .invalidate_range_end = hmm_invalidate_range_end, 190c0b12405SJérôme Glisse }; 191c0b12405SJérôme Glisse 192c0b12405SJérôme Glisse /* 193c0b12405SJérôme Glisse * hmm_mirror_register() - register a mirror against an mm 194c0b12405SJérôme Glisse * 195c0b12405SJérôme Glisse * @mirror: new mirror struct to register 196c0b12405SJérôme Glisse * @mm: mm to register against 197c0b12405SJérôme Glisse * 198c0b12405SJérôme Glisse * To start mirroring a process address space, the device driver must register 199c0b12405SJérôme Glisse * an HMM mirror struct. 200c0b12405SJérôme Glisse * 201c0b12405SJérôme Glisse * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! 202c0b12405SJérôme Glisse */ 203c0b12405SJérôme Glisse int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) 204c0b12405SJérôme Glisse { 205c0b12405SJérôme Glisse /* Sanity check */ 206c0b12405SJérôme Glisse if (!mm || !mirror || !mirror->ops) 207c0b12405SJérôme Glisse return -EINVAL; 208c0b12405SJérôme Glisse 209c0b12405SJérôme Glisse mirror->hmm = hmm_register(mm); 210c0b12405SJérôme Glisse if (!mirror->hmm) 211c0b12405SJérôme Glisse return -ENOMEM; 212c0b12405SJérôme Glisse 213c0b12405SJérôme Glisse down_write(&mirror->hmm->mirrors_sem); 214c0b12405SJérôme Glisse list_add(&mirror->list, &mirror->hmm->mirrors); 215c0b12405SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 216c0b12405SJérôme Glisse 217c0b12405SJérôme Glisse return 0; 218c0b12405SJérôme Glisse } 219c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_register); 220c0b12405SJérôme Glisse 221c0b12405SJérôme Glisse /* 222c0b12405SJérôme Glisse * hmm_mirror_unregister() - unregister a mirror 223c0b12405SJérôme Glisse * 224c0b12405SJérôme Glisse * @mirror: new mirror struct to register 225c0b12405SJérôme Glisse * 226c0b12405SJérôme Glisse * Stop mirroring a process address space, and cleanup. 227c0b12405SJérôme Glisse */ 228c0b12405SJérôme Glisse void hmm_mirror_unregister(struct hmm_mirror *mirror) 229c0b12405SJérôme Glisse { 230c0b12405SJérôme Glisse struct hmm *hmm = mirror->hmm; 231c0b12405SJérôme Glisse 232c0b12405SJérôme Glisse down_write(&hmm->mirrors_sem); 233c0b12405SJérôme Glisse list_del(&mirror->list); 234c0b12405SJérôme Glisse up_write(&hmm->mirrors_sem); 235c0b12405SJérôme Glisse } 236c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_unregister); 237da4c3c73SJérôme Glisse 23874eee180SJérôme Glisse struct hmm_vma_walk { 23974eee180SJérôme Glisse struct hmm_range *range; 24074eee180SJérôme Glisse unsigned long last; 24174eee180SJérôme Glisse bool fault; 24274eee180SJérôme Glisse bool block; 24374eee180SJérôme Glisse bool write; 24474eee180SJérôme Glisse }; 24574eee180SJérôme Glisse 24674eee180SJérôme Glisse static int hmm_vma_do_fault(struct mm_walk *walk, 24774eee180SJérôme Glisse unsigned long addr, 24874eee180SJérôme Glisse hmm_pfn_t *pfn) 24974eee180SJérôme Glisse { 25074eee180SJérôme Glisse unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE; 25174eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 25274eee180SJérôme Glisse struct vm_area_struct *vma = walk->vma; 25374eee180SJérôme Glisse int r; 25474eee180SJérôme Glisse 25574eee180SJérôme Glisse flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; 25674eee180SJérôme Glisse flags |= hmm_vma_walk->write ? FAULT_FLAG_WRITE : 0; 25774eee180SJérôme Glisse r = handle_mm_fault(vma, addr, flags); 25874eee180SJérôme Glisse if (r & VM_FAULT_RETRY) 25974eee180SJérôme Glisse return -EBUSY; 26074eee180SJérôme Glisse if (r & VM_FAULT_ERROR) { 26174eee180SJérôme Glisse *pfn = HMM_PFN_ERROR; 26274eee180SJérôme Glisse return -EFAULT; 26374eee180SJérôme Glisse } 26474eee180SJérôme Glisse 26574eee180SJérôme Glisse return -EAGAIN; 26674eee180SJérôme Glisse } 26774eee180SJérôme Glisse 268da4c3c73SJérôme Glisse static void hmm_pfns_special(hmm_pfn_t *pfns, 269da4c3c73SJérôme Glisse unsigned long addr, 270da4c3c73SJérôme Glisse unsigned long end) 271da4c3c73SJérôme Glisse { 272da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 273da4c3c73SJérôme Glisse *pfns = HMM_PFN_SPECIAL; 274da4c3c73SJérôme Glisse } 275da4c3c73SJérôme Glisse 276da4c3c73SJérôme Glisse static int hmm_pfns_bad(unsigned long addr, 277da4c3c73SJérôme Glisse unsigned long end, 278da4c3c73SJérôme Glisse struct mm_walk *walk) 279da4c3c73SJérôme Glisse { 280da4c3c73SJérôme Glisse struct hmm_range *range = walk->private; 281da4c3c73SJérôme Glisse hmm_pfn_t *pfns = range->pfns; 282da4c3c73SJérôme Glisse unsigned long i; 283da4c3c73SJérôme Glisse 284da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 285da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) 286da4c3c73SJérôme Glisse pfns[i] = HMM_PFN_ERROR; 287da4c3c73SJérôme Glisse 288da4c3c73SJérôme Glisse return 0; 289da4c3c73SJérôme Glisse } 290da4c3c73SJérôme Glisse 29174eee180SJérôme Glisse static void hmm_pfns_clear(hmm_pfn_t *pfns, 29274eee180SJérôme Glisse unsigned long addr, 29374eee180SJérôme Glisse unsigned long end) 29474eee180SJérôme Glisse { 29574eee180SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 29674eee180SJérôme Glisse *pfns = 0; 29774eee180SJérôme Glisse } 29874eee180SJérôme Glisse 299da4c3c73SJérôme Glisse static int hmm_vma_walk_hole(unsigned long addr, 300da4c3c73SJérôme Glisse unsigned long end, 301da4c3c73SJérôme Glisse struct mm_walk *walk) 302da4c3c73SJérôme Glisse { 30374eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 30474eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 305da4c3c73SJérôme Glisse hmm_pfn_t *pfns = range->pfns; 306da4c3c73SJérôme Glisse unsigned long i; 307da4c3c73SJérôme Glisse 30874eee180SJérôme Glisse hmm_vma_walk->last = addr; 309da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 31074eee180SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) { 311da4c3c73SJérôme Glisse pfns[i] = HMM_PFN_EMPTY; 31274eee180SJérôme Glisse if (hmm_vma_walk->fault) { 31374eee180SJérôme Glisse int ret; 314da4c3c73SJérôme Glisse 31574eee180SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, &pfns[i]); 31674eee180SJérôme Glisse if (ret != -EAGAIN) 31774eee180SJérôme Glisse return ret; 31874eee180SJérôme Glisse } 31974eee180SJérôme Glisse } 32074eee180SJérôme Glisse 32174eee180SJérôme Glisse return hmm_vma_walk->fault ? -EAGAIN : 0; 322da4c3c73SJérôme Glisse } 323da4c3c73SJérôme Glisse 324da4c3c73SJérôme Glisse static int hmm_vma_walk_clear(unsigned long addr, 325da4c3c73SJérôme Glisse unsigned long end, 326da4c3c73SJérôme Glisse struct mm_walk *walk) 327da4c3c73SJérôme Glisse { 32874eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 32974eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 330da4c3c73SJérôme Glisse hmm_pfn_t *pfns = range->pfns; 331da4c3c73SJérôme Glisse unsigned long i; 332da4c3c73SJérôme Glisse 33374eee180SJérôme Glisse hmm_vma_walk->last = addr; 334da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 33574eee180SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) { 336da4c3c73SJérôme Glisse pfns[i] = 0; 33774eee180SJérôme Glisse if (hmm_vma_walk->fault) { 33874eee180SJérôme Glisse int ret; 339da4c3c73SJérôme Glisse 34074eee180SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, &pfns[i]); 34174eee180SJérôme Glisse if (ret != -EAGAIN) 34274eee180SJérôme Glisse return ret; 34374eee180SJérôme Glisse } 34474eee180SJérôme Glisse } 34574eee180SJérôme Glisse 34674eee180SJérôme Glisse return hmm_vma_walk->fault ? -EAGAIN : 0; 347da4c3c73SJérôme Glisse } 348da4c3c73SJérôme Glisse 349da4c3c73SJérôme Glisse static int hmm_vma_walk_pmd(pmd_t *pmdp, 350da4c3c73SJérôme Glisse unsigned long start, 351da4c3c73SJérôme Glisse unsigned long end, 352da4c3c73SJérôme Glisse struct mm_walk *walk) 353da4c3c73SJérôme Glisse { 35474eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 35574eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 356da4c3c73SJérôme Glisse struct vm_area_struct *vma = walk->vma; 357da4c3c73SJérôme Glisse hmm_pfn_t *pfns = range->pfns; 358da4c3c73SJérôme Glisse unsigned long addr = start, i; 35974eee180SJérôme Glisse bool write_fault; 360da4c3c73SJérôme Glisse hmm_pfn_t flag; 361da4c3c73SJérôme Glisse pte_t *ptep; 362da4c3c73SJérôme Glisse 363da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 364da4c3c73SJérôme Glisse flag = vma->vm_flags & VM_READ ? HMM_PFN_READ : 0; 36574eee180SJérôme Glisse write_fault = hmm_vma_walk->fault & hmm_vma_walk->write; 366da4c3c73SJérôme Glisse 367da4c3c73SJérôme Glisse again: 368da4c3c73SJérôme Glisse if (pmd_none(*pmdp)) 369da4c3c73SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 370da4c3c73SJérôme Glisse 371da4c3c73SJérôme Glisse if (pmd_huge(*pmdp) && vma->vm_flags & VM_HUGETLB) 372da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 373da4c3c73SJérôme Glisse 374da4c3c73SJérôme Glisse if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) { 375da4c3c73SJérôme Glisse unsigned long pfn; 376da4c3c73SJérôme Glisse pmd_t pmd; 377da4c3c73SJérôme Glisse 378da4c3c73SJérôme Glisse /* 379da4c3c73SJérôme Glisse * No need to take pmd_lock here, even if some other threads 380da4c3c73SJérôme Glisse * is splitting the huge pmd we will get that event through 381da4c3c73SJérôme Glisse * mmu_notifier callback. 382da4c3c73SJérôme Glisse * 383da4c3c73SJérôme Glisse * So just read pmd value and check again its a transparent 384da4c3c73SJérôme Glisse * huge or device mapping one and compute corresponding pfn 385da4c3c73SJérôme Glisse * values. 386da4c3c73SJérôme Glisse */ 387da4c3c73SJérôme Glisse pmd = pmd_read_atomic(pmdp); 388da4c3c73SJérôme Glisse barrier(); 389da4c3c73SJérôme Glisse if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) 390da4c3c73SJérôme Glisse goto again; 391da4c3c73SJérôme Glisse if (pmd_protnone(pmd)) 392da4c3c73SJérôme Glisse return hmm_vma_walk_clear(start, end, walk); 393da4c3c73SJérôme Glisse 394f6f37321SLinus Torvalds if (write_fault && !pmd_write(pmd)) 39574eee180SJérôme Glisse return hmm_vma_walk_clear(start, end, walk); 39674eee180SJérôme Glisse 397da4c3c73SJérôme Glisse pfn = pmd_pfn(pmd) + pte_index(addr); 398f6f37321SLinus Torvalds flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0; 399da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++, pfn++) 400da4c3c73SJérôme Glisse pfns[i] = hmm_pfn_t_from_pfn(pfn) | flag; 401da4c3c73SJérôme Glisse return 0; 402da4c3c73SJérôme Glisse } 403da4c3c73SJérôme Glisse 404da4c3c73SJérôme Glisse if (pmd_bad(*pmdp)) 405da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 406da4c3c73SJérôme Glisse 407da4c3c73SJérôme Glisse ptep = pte_offset_map(pmdp, addr); 408da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { 409da4c3c73SJérôme Glisse pte_t pte = *ptep; 410da4c3c73SJérôme Glisse 411da4c3c73SJérôme Glisse pfns[i] = 0; 412da4c3c73SJérôme Glisse 41374eee180SJérôme Glisse if (pte_none(pte)) { 414da4c3c73SJérôme Glisse pfns[i] = HMM_PFN_EMPTY; 41574eee180SJérôme Glisse if (hmm_vma_walk->fault) 41674eee180SJérôme Glisse goto fault; 417da4c3c73SJérôme Glisse continue; 418da4c3c73SJérôme Glisse } 419da4c3c73SJérôme Glisse 42074eee180SJérôme Glisse if (!pte_present(pte)) { 42174eee180SJérôme Glisse swp_entry_t entry; 42274eee180SJérôme Glisse 42374eee180SJérôme Glisse if (!non_swap_entry(entry)) { 42474eee180SJérôme Glisse if (hmm_vma_walk->fault) 42574eee180SJérôme Glisse goto fault; 42674eee180SJérôme Glisse continue; 42774eee180SJérôme Glisse } 42874eee180SJérôme Glisse 42974eee180SJérôme Glisse entry = pte_to_swp_entry(pte); 43074eee180SJérôme Glisse 43174eee180SJérôme Glisse /* 43274eee180SJérôme Glisse * This is a special swap entry, ignore migration, use 43374eee180SJérôme Glisse * device and report anything else as error. 43474eee180SJérôme Glisse */ 4354ef589dcSJérôme Glisse if (is_device_private_entry(entry)) { 4364ef589dcSJérôme Glisse pfns[i] = hmm_pfn_t_from_pfn(swp_offset(entry)); 4374ef589dcSJérôme Glisse if (is_write_device_private_entry(entry)) { 4384ef589dcSJérôme Glisse pfns[i] |= HMM_PFN_WRITE; 4394ef589dcSJérôme Glisse } else if (write_fault) 4404ef589dcSJérôme Glisse goto fault; 4414ef589dcSJérôme Glisse pfns[i] |= HMM_PFN_DEVICE_UNADDRESSABLE; 4424ef589dcSJérôme Glisse pfns[i] |= flag; 4434ef589dcSJérôme Glisse } else if (is_migration_entry(entry)) { 44474eee180SJérôme Glisse if (hmm_vma_walk->fault) { 44574eee180SJérôme Glisse pte_unmap(ptep); 44674eee180SJérôme Glisse hmm_vma_walk->last = addr; 44774eee180SJérôme Glisse migration_entry_wait(vma->vm_mm, 44874eee180SJérôme Glisse pmdp, addr); 44974eee180SJérôme Glisse return -EAGAIN; 45074eee180SJérôme Glisse } 45174eee180SJérôme Glisse continue; 45274eee180SJérôme Glisse } else { 45374eee180SJérôme Glisse /* Report error for everything else */ 45474eee180SJérôme Glisse pfns[i] = HMM_PFN_ERROR; 45574eee180SJérôme Glisse } 45674eee180SJérôme Glisse continue; 45774eee180SJérôme Glisse } 45874eee180SJérôme Glisse 459f6f37321SLinus Torvalds if (write_fault && !pte_write(pte)) 46074eee180SJérôme Glisse goto fault; 46174eee180SJérôme Glisse 462da4c3c73SJérôme Glisse pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag; 463f6f37321SLinus Torvalds pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0; 46474eee180SJérôme Glisse continue; 46574eee180SJérôme Glisse 46674eee180SJérôme Glisse fault: 46774eee180SJérôme Glisse pte_unmap(ptep); 46874eee180SJérôme Glisse /* Fault all pages in range */ 46974eee180SJérôme Glisse return hmm_vma_walk_clear(start, end, walk); 470da4c3c73SJérôme Glisse } 471da4c3c73SJérôme Glisse pte_unmap(ptep - 1); 472da4c3c73SJérôme Glisse 473da4c3c73SJérôme Glisse return 0; 474da4c3c73SJérôme Glisse } 475da4c3c73SJérôme Glisse 476da4c3c73SJérôme Glisse /* 477da4c3c73SJérôme Glisse * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses 478da4c3c73SJérôme Glisse * @vma: virtual memory area containing the virtual address range 479da4c3c73SJérôme Glisse * @range: used to track snapshot validity 480da4c3c73SJérôme Glisse * @start: range virtual start address (inclusive) 481da4c3c73SJérôme Glisse * @end: range virtual end address (exclusive) 482da4c3c73SJérôme Glisse * @entries: array of hmm_pfn_t: provided by the caller, filled in by function 483da4c3c73SJérôme Glisse * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success 484da4c3c73SJérôme Glisse * 485da4c3c73SJérôme Glisse * This snapshots the CPU page table for a range of virtual addresses. Snapshot 486da4c3c73SJérôme Glisse * validity is tracked by range struct. See hmm_vma_range_done() for further 487da4c3c73SJérôme Glisse * information. 488da4c3c73SJérôme Glisse * 489da4c3c73SJérôme Glisse * The range struct is initialized here. It tracks the CPU page table, but only 490da4c3c73SJérôme Glisse * if the function returns success (0), in which case the caller must then call 491da4c3c73SJérôme Glisse * hmm_vma_range_done() to stop CPU page table update tracking on this range. 492da4c3c73SJérôme Glisse * 493da4c3c73SJérôme Glisse * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS 494da4c3c73SJérôme Glisse * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED ! 495da4c3c73SJérôme Glisse */ 496da4c3c73SJérôme Glisse int hmm_vma_get_pfns(struct vm_area_struct *vma, 497da4c3c73SJérôme Glisse struct hmm_range *range, 498da4c3c73SJérôme Glisse unsigned long start, 499da4c3c73SJérôme Glisse unsigned long end, 500da4c3c73SJérôme Glisse hmm_pfn_t *pfns) 501da4c3c73SJérôme Glisse { 50274eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 503da4c3c73SJérôme Glisse struct mm_walk mm_walk; 504da4c3c73SJérôme Glisse struct hmm *hmm; 505da4c3c73SJérôme Glisse 506da4c3c73SJérôme Glisse /* FIXME support hugetlb fs */ 507da4c3c73SJérôme Glisse if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) { 508da4c3c73SJérôme Glisse hmm_pfns_special(pfns, start, end); 509da4c3c73SJérôme Glisse return -EINVAL; 510da4c3c73SJérôme Glisse } 511da4c3c73SJérôme Glisse 512da4c3c73SJérôme Glisse /* Sanity check, this really should not happen ! */ 513da4c3c73SJérôme Glisse if (start < vma->vm_start || start >= vma->vm_end) 514da4c3c73SJérôme Glisse return -EINVAL; 515da4c3c73SJérôme Glisse if (end < vma->vm_start || end > vma->vm_end) 516da4c3c73SJérôme Glisse return -EINVAL; 517da4c3c73SJérôme Glisse 518da4c3c73SJérôme Glisse hmm = hmm_register(vma->vm_mm); 519da4c3c73SJérôme Glisse if (!hmm) 520da4c3c73SJérôme Glisse return -ENOMEM; 521da4c3c73SJérôme Glisse /* Caller must have registered a mirror, via hmm_mirror_register() ! */ 522da4c3c73SJérôme Glisse if (!hmm->mmu_notifier.ops) 523da4c3c73SJérôme Glisse return -EINVAL; 524da4c3c73SJérôme Glisse 525da4c3c73SJérôme Glisse /* Initialize range to track CPU page table update */ 526da4c3c73SJérôme Glisse range->start = start; 527da4c3c73SJérôme Glisse range->pfns = pfns; 528da4c3c73SJérôme Glisse range->end = end; 529da4c3c73SJérôme Glisse spin_lock(&hmm->lock); 530da4c3c73SJérôme Glisse range->valid = true; 531da4c3c73SJérôme Glisse list_add_rcu(&range->list, &hmm->ranges); 532da4c3c73SJérôme Glisse spin_unlock(&hmm->lock); 533da4c3c73SJérôme Glisse 53474eee180SJérôme Glisse hmm_vma_walk.fault = false; 53574eee180SJérôme Glisse hmm_vma_walk.range = range; 53674eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 53774eee180SJérôme Glisse 538da4c3c73SJérôme Glisse mm_walk.vma = vma; 539da4c3c73SJérôme Glisse mm_walk.mm = vma->vm_mm; 540da4c3c73SJérôme Glisse mm_walk.pte_entry = NULL; 541da4c3c73SJérôme Glisse mm_walk.test_walk = NULL; 542da4c3c73SJérôme Glisse mm_walk.hugetlb_entry = NULL; 543da4c3c73SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 544da4c3c73SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 545da4c3c73SJérôme Glisse 546da4c3c73SJérôme Glisse walk_page_range(start, end, &mm_walk); 547da4c3c73SJérôme Glisse return 0; 548da4c3c73SJérôme Glisse } 549da4c3c73SJérôme Glisse EXPORT_SYMBOL(hmm_vma_get_pfns); 550da4c3c73SJérôme Glisse 551da4c3c73SJérôme Glisse /* 552da4c3c73SJérôme Glisse * hmm_vma_range_done() - stop tracking change to CPU page table over a range 553da4c3c73SJérôme Glisse * @vma: virtual memory area containing the virtual address range 554da4c3c73SJérôme Glisse * @range: range being tracked 555da4c3c73SJérôme Glisse * Returns: false if range data has been invalidated, true otherwise 556da4c3c73SJérôme Glisse * 557da4c3c73SJérôme Glisse * Range struct is used to track updates to the CPU page table after a call to 558da4c3c73SJérôme Glisse * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done 559da4c3c73SJérôme Glisse * using the data, or wants to lock updates to the data it got from those 560da4c3c73SJérôme Glisse * functions, it must call the hmm_vma_range_done() function, which will then 561da4c3c73SJérôme Glisse * stop tracking CPU page table updates. 562da4c3c73SJérôme Glisse * 563da4c3c73SJérôme Glisse * Note that device driver must still implement general CPU page table update 564da4c3c73SJérôme Glisse * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using 565da4c3c73SJérôme Glisse * the mmu_notifier API directly. 566da4c3c73SJérôme Glisse * 567da4c3c73SJérôme Glisse * CPU page table update tracking done through hmm_range is only temporary and 568da4c3c73SJérôme Glisse * to be used while trying to duplicate CPU page table contents for a range of 569da4c3c73SJérôme Glisse * virtual addresses. 570da4c3c73SJérôme Glisse * 571da4c3c73SJérôme Glisse * There are two ways to use this : 572da4c3c73SJérôme Glisse * again: 57374eee180SJérôme Glisse * hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...); 574da4c3c73SJérôme Glisse * trans = device_build_page_table_update_transaction(pfns); 575da4c3c73SJérôme Glisse * device_page_table_lock(); 576da4c3c73SJérôme Glisse * if (!hmm_vma_range_done(vma, range)) { 577da4c3c73SJérôme Glisse * device_page_table_unlock(); 578da4c3c73SJérôme Glisse * goto again; 579da4c3c73SJérôme Glisse * } 580da4c3c73SJérôme Glisse * device_commit_transaction(trans); 581da4c3c73SJérôme Glisse * device_page_table_unlock(); 582da4c3c73SJérôme Glisse * 583da4c3c73SJérôme Glisse * Or: 58474eee180SJérôme Glisse * hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...); 585da4c3c73SJérôme Glisse * device_page_table_lock(); 586da4c3c73SJérôme Glisse * hmm_vma_range_done(vma, range); 587da4c3c73SJérôme Glisse * device_update_page_table(pfns); 588da4c3c73SJérôme Glisse * device_page_table_unlock(); 589da4c3c73SJérôme Glisse */ 590da4c3c73SJérôme Glisse bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range) 591da4c3c73SJérôme Glisse { 592da4c3c73SJérôme Glisse unsigned long npages = (range->end - range->start) >> PAGE_SHIFT; 593da4c3c73SJérôme Glisse struct hmm *hmm; 594da4c3c73SJérôme Glisse 595da4c3c73SJérôme Glisse if (range->end <= range->start) { 596da4c3c73SJérôme Glisse BUG(); 597da4c3c73SJérôme Glisse return false; 598da4c3c73SJérôme Glisse } 599da4c3c73SJérôme Glisse 600da4c3c73SJérôme Glisse hmm = hmm_register(vma->vm_mm); 601da4c3c73SJérôme Glisse if (!hmm) { 602da4c3c73SJérôme Glisse memset(range->pfns, 0, sizeof(*range->pfns) * npages); 603da4c3c73SJérôme Glisse return false; 604da4c3c73SJérôme Glisse } 605da4c3c73SJérôme Glisse 606da4c3c73SJérôme Glisse spin_lock(&hmm->lock); 607da4c3c73SJérôme Glisse list_del_rcu(&range->list); 608da4c3c73SJérôme Glisse spin_unlock(&hmm->lock); 609da4c3c73SJérôme Glisse 610da4c3c73SJérôme Glisse return range->valid; 611da4c3c73SJérôme Glisse } 612da4c3c73SJérôme Glisse EXPORT_SYMBOL(hmm_vma_range_done); 61374eee180SJérôme Glisse 61474eee180SJérôme Glisse /* 61574eee180SJérôme Glisse * hmm_vma_fault() - try to fault some address in a virtual address range 61674eee180SJérôme Glisse * @vma: virtual memory area containing the virtual address range 61774eee180SJérôme Glisse * @range: use to track pfns array content validity 61874eee180SJérôme Glisse * @start: fault range virtual start address (inclusive) 61974eee180SJérôme Glisse * @end: fault range virtual end address (exclusive) 62074eee180SJérôme Glisse * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted 62174eee180SJérôme Glisse * @write: is it a write fault 62274eee180SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 62374eee180SJérôme Glisse * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop) 62474eee180SJérôme Glisse * 62574eee180SJérôme Glisse * This is similar to a regular CPU page fault except that it will not trigger 62674eee180SJérôme Glisse * any memory migration if the memory being faulted is not accessible by CPUs. 62774eee180SJérôme Glisse * 62874eee180SJérôme Glisse * On error, for one virtual address in the range, the function will set the 62974eee180SJérôme Glisse * hmm_pfn_t error flag for the corresponding pfn entry. 63074eee180SJérôme Glisse * 63174eee180SJérôme Glisse * Expected use pattern: 63274eee180SJérôme Glisse * retry: 63374eee180SJérôme Glisse * down_read(&mm->mmap_sem); 63474eee180SJérôme Glisse * // Find vma and address device wants to fault, initialize hmm_pfn_t 63574eee180SJérôme Glisse * // array accordingly 63674eee180SJérôme Glisse * ret = hmm_vma_fault(vma, start, end, pfns, allow_retry); 63774eee180SJérôme Glisse * switch (ret) { 63874eee180SJérôme Glisse * case -EAGAIN: 63974eee180SJérôme Glisse * hmm_vma_range_done(vma, range); 64074eee180SJérôme Glisse * // You might want to rate limit or yield to play nicely, you may 64174eee180SJérôme Glisse * // also commit any valid pfn in the array assuming that you are 64274eee180SJérôme Glisse * // getting true from hmm_vma_range_monitor_end() 64374eee180SJérôme Glisse * goto retry; 64474eee180SJérôme Glisse * case 0: 64574eee180SJérôme Glisse * break; 64674eee180SJérôme Glisse * default: 64774eee180SJérôme Glisse * // Handle error ! 64874eee180SJérôme Glisse * up_read(&mm->mmap_sem) 64974eee180SJérôme Glisse * return; 65074eee180SJérôme Glisse * } 65174eee180SJérôme Glisse * // Take device driver lock that serialize device page table update 65274eee180SJérôme Glisse * driver_lock_device_page_table_update(); 65374eee180SJérôme Glisse * hmm_vma_range_done(vma, range); 65474eee180SJérôme Glisse * // Commit pfns we got from hmm_vma_fault() 65574eee180SJérôme Glisse * driver_unlock_device_page_table_update(); 65674eee180SJérôme Glisse * up_read(&mm->mmap_sem) 65774eee180SJérôme Glisse * 65874eee180SJérôme Glisse * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0) 65974eee180SJérôme Glisse * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION ! 66074eee180SJérôme Glisse * 66174eee180SJérôme Glisse * YOU HAVE BEEN WARNED ! 66274eee180SJérôme Glisse */ 66374eee180SJérôme Glisse int hmm_vma_fault(struct vm_area_struct *vma, 66474eee180SJérôme Glisse struct hmm_range *range, 66574eee180SJérôme Glisse unsigned long start, 66674eee180SJérôme Glisse unsigned long end, 66774eee180SJérôme Glisse hmm_pfn_t *pfns, 66874eee180SJérôme Glisse bool write, 66974eee180SJérôme Glisse bool block) 67074eee180SJérôme Glisse { 67174eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 67274eee180SJérôme Glisse struct mm_walk mm_walk; 67374eee180SJérôme Glisse struct hmm *hmm; 67474eee180SJérôme Glisse int ret; 67574eee180SJérôme Glisse 67674eee180SJérôme Glisse /* Sanity check, this really should not happen ! */ 67774eee180SJérôme Glisse if (start < vma->vm_start || start >= vma->vm_end) 67874eee180SJérôme Glisse return -EINVAL; 67974eee180SJérôme Glisse if (end < vma->vm_start || end > vma->vm_end) 68074eee180SJérôme Glisse return -EINVAL; 68174eee180SJérôme Glisse 68274eee180SJérôme Glisse hmm = hmm_register(vma->vm_mm); 68374eee180SJérôme Glisse if (!hmm) { 68474eee180SJérôme Glisse hmm_pfns_clear(pfns, start, end); 68574eee180SJérôme Glisse return -ENOMEM; 68674eee180SJérôme Glisse } 68774eee180SJérôme Glisse /* Caller must have registered a mirror using hmm_mirror_register() */ 68874eee180SJérôme Glisse if (!hmm->mmu_notifier.ops) 68974eee180SJérôme Glisse return -EINVAL; 69074eee180SJérôme Glisse 69174eee180SJérôme Glisse /* Initialize range to track CPU page table update */ 69274eee180SJérôme Glisse range->start = start; 69374eee180SJérôme Glisse range->pfns = pfns; 69474eee180SJérôme Glisse range->end = end; 69574eee180SJérôme Glisse spin_lock(&hmm->lock); 69674eee180SJérôme Glisse range->valid = true; 69774eee180SJérôme Glisse list_add_rcu(&range->list, &hmm->ranges); 69874eee180SJérôme Glisse spin_unlock(&hmm->lock); 69974eee180SJérôme Glisse 70074eee180SJérôme Glisse /* FIXME support hugetlb fs */ 70174eee180SJérôme Glisse if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) { 70274eee180SJérôme Glisse hmm_pfns_special(pfns, start, end); 70374eee180SJérôme Glisse return 0; 70474eee180SJérôme Glisse } 70574eee180SJérôme Glisse 70674eee180SJérôme Glisse hmm_vma_walk.fault = true; 70774eee180SJérôme Glisse hmm_vma_walk.write = write; 70874eee180SJérôme Glisse hmm_vma_walk.block = block; 70974eee180SJérôme Glisse hmm_vma_walk.range = range; 71074eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 71174eee180SJérôme Glisse hmm_vma_walk.last = range->start; 71274eee180SJérôme Glisse 71374eee180SJérôme Glisse mm_walk.vma = vma; 71474eee180SJérôme Glisse mm_walk.mm = vma->vm_mm; 71574eee180SJérôme Glisse mm_walk.pte_entry = NULL; 71674eee180SJérôme Glisse mm_walk.test_walk = NULL; 71774eee180SJérôme Glisse mm_walk.hugetlb_entry = NULL; 71874eee180SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 71974eee180SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 72074eee180SJérôme Glisse 72174eee180SJérôme Glisse do { 72274eee180SJérôme Glisse ret = walk_page_range(start, end, &mm_walk); 72374eee180SJérôme Glisse start = hmm_vma_walk.last; 72474eee180SJérôme Glisse } while (ret == -EAGAIN); 72574eee180SJérôme Glisse 72674eee180SJérôme Glisse if (ret) { 72774eee180SJérôme Glisse unsigned long i; 72874eee180SJérôme Glisse 72974eee180SJérôme Glisse i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 73074eee180SJérôme Glisse hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end); 73174eee180SJérôme Glisse hmm_vma_range_done(vma, range); 73274eee180SJérôme Glisse } 73374eee180SJérôme Glisse return ret; 73474eee180SJérôme Glisse } 73574eee180SJérôme Glisse EXPORT_SYMBOL(hmm_vma_fault); 736c0b12405SJérôme Glisse #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ 7374ef589dcSJérôme Glisse 7384ef589dcSJérôme Glisse 739df6ad698SJérôme Glisse #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC) 7404ef589dcSJérôme Glisse struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma, 7414ef589dcSJérôme Glisse unsigned long addr) 7424ef589dcSJérôme Glisse { 7434ef589dcSJérôme Glisse struct page *page; 7444ef589dcSJérôme Glisse 7454ef589dcSJérôme Glisse page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 7464ef589dcSJérôme Glisse if (!page) 7474ef589dcSJérôme Glisse return NULL; 7484ef589dcSJérôme Glisse lock_page(page); 7494ef589dcSJérôme Glisse return page; 7504ef589dcSJérôme Glisse } 7514ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_vma_alloc_locked_page); 7524ef589dcSJérôme Glisse 7534ef589dcSJérôme Glisse 7544ef589dcSJérôme Glisse static void hmm_devmem_ref_release(struct percpu_ref *ref) 7554ef589dcSJérôme Glisse { 7564ef589dcSJérôme Glisse struct hmm_devmem *devmem; 7574ef589dcSJérôme Glisse 7584ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 7594ef589dcSJérôme Glisse complete(&devmem->completion); 7604ef589dcSJérôme Glisse } 7614ef589dcSJérôme Glisse 7624ef589dcSJérôme Glisse static void hmm_devmem_ref_exit(void *data) 7634ef589dcSJérôme Glisse { 7644ef589dcSJérôme Glisse struct percpu_ref *ref = data; 7654ef589dcSJérôme Glisse struct hmm_devmem *devmem; 7664ef589dcSJérôme Glisse 7674ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 7684ef589dcSJérôme Glisse percpu_ref_exit(ref); 7694ef589dcSJérôme Glisse devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data); 7704ef589dcSJérôme Glisse } 7714ef589dcSJérôme Glisse 7724ef589dcSJérôme Glisse static void hmm_devmem_ref_kill(void *data) 7734ef589dcSJérôme Glisse { 7744ef589dcSJérôme Glisse struct percpu_ref *ref = data; 7754ef589dcSJérôme Glisse struct hmm_devmem *devmem; 7764ef589dcSJérôme Glisse 7774ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 7784ef589dcSJérôme Glisse percpu_ref_kill(ref); 7794ef589dcSJérôme Glisse wait_for_completion(&devmem->completion); 7804ef589dcSJérôme Glisse devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data); 7814ef589dcSJérôme Glisse } 7824ef589dcSJérôme Glisse 7834ef589dcSJérôme Glisse static int hmm_devmem_fault(struct vm_area_struct *vma, 7844ef589dcSJérôme Glisse unsigned long addr, 7854ef589dcSJérôme Glisse const struct page *page, 7864ef589dcSJérôme Glisse unsigned int flags, 7874ef589dcSJérôme Glisse pmd_t *pmdp) 7884ef589dcSJérôme Glisse { 7894ef589dcSJérôme Glisse struct hmm_devmem *devmem = page->pgmap->data; 7904ef589dcSJérôme Glisse 7914ef589dcSJérôme Glisse return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp); 7924ef589dcSJérôme Glisse } 7934ef589dcSJérôme Glisse 7944ef589dcSJérôme Glisse static void hmm_devmem_free(struct page *page, void *data) 7954ef589dcSJérôme Glisse { 7964ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 7974ef589dcSJérôme Glisse 7984ef589dcSJérôme Glisse devmem->ops->free(devmem, page); 7994ef589dcSJérôme Glisse } 8004ef589dcSJérôme Glisse 8014ef589dcSJérôme Glisse static DEFINE_MUTEX(hmm_devmem_lock); 8024ef589dcSJérôme Glisse static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL); 8034ef589dcSJérôme Glisse 8044ef589dcSJérôme Glisse static void hmm_devmem_radix_release(struct resource *resource) 8054ef589dcSJérôme Glisse { 806fec11bc0SColin Ian King resource_size_t key, align_start, align_size; 8074ef589dcSJérôme Glisse 8084ef589dcSJérôme Glisse align_start = resource->start & ~(PA_SECTION_SIZE - 1); 8094ef589dcSJérôme Glisse align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE); 8104ef589dcSJérôme Glisse 8114ef589dcSJérôme Glisse mutex_lock(&hmm_devmem_lock); 8124ef589dcSJérôme Glisse for (key = resource->start; 8134ef589dcSJérôme Glisse key <= resource->end; 8144ef589dcSJérôme Glisse key += PA_SECTION_SIZE) 8154ef589dcSJérôme Glisse radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT); 8164ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 8174ef589dcSJérôme Glisse } 8184ef589dcSJérôme Glisse 8194ef589dcSJérôme Glisse static void hmm_devmem_release(struct device *dev, void *data) 8204ef589dcSJérôme Glisse { 8214ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 8224ef589dcSJérôme Glisse struct resource *resource = devmem->resource; 8234ef589dcSJérôme Glisse unsigned long start_pfn, npages; 8244ef589dcSJérôme Glisse struct zone *zone; 8254ef589dcSJérôme Glisse struct page *page; 8264ef589dcSJérôme Glisse 8274ef589dcSJérôme Glisse if (percpu_ref_tryget_live(&devmem->ref)) { 8284ef589dcSJérôme Glisse dev_WARN(dev, "%s: page mapping is still live!\n", __func__); 8294ef589dcSJérôme Glisse percpu_ref_put(&devmem->ref); 8304ef589dcSJérôme Glisse } 8314ef589dcSJérôme Glisse 8324ef589dcSJérôme Glisse /* pages are dead and unused, undo the arch mapping */ 8334ef589dcSJérôme Glisse start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT; 8344ef589dcSJérôme Glisse npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT; 8354ef589dcSJérôme Glisse 8364ef589dcSJérôme Glisse page = pfn_to_page(start_pfn); 8374ef589dcSJérôme Glisse zone = page_zone(page); 8384ef589dcSJérôme Glisse 8394ef589dcSJérôme Glisse mem_hotplug_begin(); 840d3df0a42SJérôme Glisse if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY) 841da024512SChristoph Hellwig __remove_pages(zone, start_pfn, npages, NULL); 842d3df0a42SJérôme Glisse else 843d3df0a42SJérôme Glisse arch_remove_memory(start_pfn << PAGE_SHIFT, 844da024512SChristoph Hellwig npages << PAGE_SHIFT, NULL); 8454ef589dcSJérôme Glisse mem_hotplug_done(); 8464ef589dcSJérôme Glisse 8474ef589dcSJérôme Glisse hmm_devmem_radix_release(resource); 8484ef589dcSJérôme Glisse } 8494ef589dcSJérôme Glisse 8504ef589dcSJérôme Glisse static struct hmm_devmem *hmm_devmem_find(resource_size_t phys) 8514ef589dcSJérôme Glisse { 8524ef589dcSJérôme Glisse WARN_ON_ONCE(!rcu_read_lock_held()); 8534ef589dcSJérôme Glisse 8544ef589dcSJérôme Glisse return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT); 8554ef589dcSJérôme Glisse } 8564ef589dcSJérôme Glisse 8574ef589dcSJérôme Glisse static int hmm_devmem_pages_create(struct hmm_devmem *devmem) 8584ef589dcSJérôme Glisse { 8594ef589dcSJérôme Glisse resource_size_t key, align_start, align_size, align_end; 8604ef589dcSJérôme Glisse struct device *device = devmem->device; 8614ef589dcSJérôme Glisse int ret, nid, is_ram; 8624ef589dcSJérôme Glisse unsigned long pfn; 8634ef589dcSJérôme Glisse 8644ef589dcSJérôme Glisse align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1); 8654ef589dcSJérôme Glisse align_size = ALIGN(devmem->resource->start + 8664ef589dcSJérôme Glisse resource_size(devmem->resource), 8674ef589dcSJérôme Glisse PA_SECTION_SIZE) - align_start; 8684ef589dcSJérôme Glisse 8694ef589dcSJérôme Glisse is_ram = region_intersects(align_start, align_size, 8704ef589dcSJérôme Glisse IORESOURCE_SYSTEM_RAM, 8714ef589dcSJérôme Glisse IORES_DESC_NONE); 8724ef589dcSJérôme Glisse if (is_ram == REGION_MIXED) { 8734ef589dcSJérôme Glisse WARN_ONCE(1, "%s attempted on mixed region %pr\n", 8744ef589dcSJérôme Glisse __func__, devmem->resource); 8754ef589dcSJérôme Glisse return -ENXIO; 8764ef589dcSJérôme Glisse } 8774ef589dcSJérôme Glisse if (is_ram == REGION_INTERSECTS) 8784ef589dcSJérôme Glisse return -ENXIO; 8794ef589dcSJérôme Glisse 880d3df0a42SJérôme Glisse if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY) 881d3df0a42SJérôme Glisse devmem->pagemap.type = MEMORY_DEVICE_PUBLIC; 882d3df0a42SJérôme Glisse else 8834ef589dcSJérôme Glisse devmem->pagemap.type = MEMORY_DEVICE_PRIVATE; 884d3df0a42SJérôme Glisse 885*e7744aa2SLogan Gunthorpe devmem->pagemap.res = *devmem->resource; 8864ef589dcSJérôme Glisse devmem->pagemap.page_fault = hmm_devmem_fault; 8874ef589dcSJérôme Glisse devmem->pagemap.page_free = hmm_devmem_free; 8884ef589dcSJérôme Glisse devmem->pagemap.dev = devmem->device; 8894ef589dcSJérôme Glisse devmem->pagemap.ref = &devmem->ref; 8904ef589dcSJérôme Glisse devmem->pagemap.data = devmem; 8914ef589dcSJérôme Glisse 8924ef589dcSJérôme Glisse mutex_lock(&hmm_devmem_lock); 8934ef589dcSJérôme Glisse align_end = align_start + align_size - 1; 8944ef589dcSJérôme Glisse for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) { 8954ef589dcSJérôme Glisse struct hmm_devmem *dup; 8964ef589dcSJérôme Glisse 8974ef589dcSJérôme Glisse rcu_read_lock(); 8984ef589dcSJérôme Glisse dup = hmm_devmem_find(key); 8994ef589dcSJérôme Glisse rcu_read_unlock(); 9004ef589dcSJérôme Glisse if (dup) { 9014ef589dcSJérôme Glisse dev_err(device, "%s: collides with mapping for %s\n", 9024ef589dcSJérôme Glisse __func__, dev_name(dup->device)); 9034ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 9044ef589dcSJérôme Glisse ret = -EBUSY; 9054ef589dcSJérôme Glisse goto error; 9064ef589dcSJérôme Glisse } 9074ef589dcSJérôme Glisse ret = radix_tree_insert(&hmm_devmem_radix, 9084ef589dcSJérôme Glisse key >> PA_SECTION_SHIFT, 9094ef589dcSJérôme Glisse devmem); 9104ef589dcSJérôme Glisse if (ret) { 9114ef589dcSJérôme Glisse dev_err(device, "%s: failed: %d\n", __func__, ret); 9124ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 9134ef589dcSJérôme Glisse goto error_radix; 9144ef589dcSJérôme Glisse } 9154ef589dcSJérôme Glisse } 9164ef589dcSJérôme Glisse mutex_unlock(&hmm_devmem_lock); 9174ef589dcSJérôme Glisse 9184ef589dcSJérôme Glisse nid = dev_to_node(device); 9194ef589dcSJérôme Glisse if (nid < 0) 9204ef589dcSJérôme Glisse nid = numa_mem_id(); 9214ef589dcSJérôme Glisse 9224ef589dcSJérôme Glisse mem_hotplug_begin(); 9234ef589dcSJérôme Glisse /* 9244ef589dcSJérôme Glisse * For device private memory we call add_pages() as we only need to 9254ef589dcSJérôme Glisse * allocate and initialize struct page for the device memory. More- 9264ef589dcSJérôme Glisse * over the device memory is un-accessible thus we do not want to 9274ef589dcSJérôme Glisse * create a linear mapping for the memory like arch_add_memory() 9284ef589dcSJérôme Glisse * would do. 929d3df0a42SJérôme Glisse * 930d3df0a42SJérôme Glisse * For device public memory, which is accesible by the CPU, we do 931d3df0a42SJérôme Glisse * want the linear mapping and thus use arch_add_memory(). 9324ef589dcSJérôme Glisse */ 933d3df0a42SJérôme Glisse if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC) 93424e6d5a5SChristoph Hellwig ret = arch_add_memory(nid, align_start, align_size, NULL, 93524e6d5a5SChristoph Hellwig false); 936d3df0a42SJérôme Glisse else 9374ef589dcSJérôme Glisse ret = add_pages(nid, align_start >> PAGE_SHIFT, 93824e6d5a5SChristoph Hellwig align_size >> PAGE_SHIFT, NULL, false); 9394ef589dcSJérôme Glisse if (ret) { 9404ef589dcSJérôme Glisse mem_hotplug_done(); 9414ef589dcSJérôme Glisse goto error_add_memory; 9424ef589dcSJérôme Glisse } 9434ef589dcSJérôme Glisse move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE], 9444ef589dcSJérôme Glisse align_start >> PAGE_SHIFT, 945a99583e7SChristoph Hellwig align_size >> PAGE_SHIFT, NULL); 9464ef589dcSJérôme Glisse mem_hotplug_done(); 9474ef589dcSJérôme Glisse 9484ef589dcSJérôme Glisse for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) { 9494ef589dcSJérôme Glisse struct page *page = pfn_to_page(pfn); 9504ef589dcSJérôme Glisse 9514ef589dcSJérôme Glisse page->pgmap = &devmem->pagemap; 9524ef589dcSJérôme Glisse } 9534ef589dcSJérôme Glisse return 0; 9544ef589dcSJérôme Glisse 9554ef589dcSJérôme Glisse error_add_memory: 9564ef589dcSJérôme Glisse untrack_pfn(NULL, PHYS_PFN(align_start), align_size); 9574ef589dcSJérôme Glisse error_radix: 9584ef589dcSJérôme Glisse hmm_devmem_radix_release(devmem->resource); 9594ef589dcSJérôme Glisse error: 9604ef589dcSJérôme Glisse return ret; 9614ef589dcSJérôme Glisse } 9624ef589dcSJérôme Glisse 9634ef589dcSJérôme Glisse static int hmm_devmem_match(struct device *dev, void *data, void *match_data) 9644ef589dcSJérôme Glisse { 9654ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 9664ef589dcSJérôme Glisse 9674ef589dcSJérôme Glisse return devmem->resource == match_data; 9684ef589dcSJérôme Glisse } 9694ef589dcSJérôme Glisse 9704ef589dcSJérôme Glisse static void hmm_devmem_pages_remove(struct hmm_devmem *devmem) 9714ef589dcSJérôme Glisse { 9724ef589dcSJérôme Glisse devres_release(devmem->device, &hmm_devmem_release, 9734ef589dcSJérôme Glisse &hmm_devmem_match, devmem->resource); 9744ef589dcSJérôme Glisse } 9754ef589dcSJérôme Glisse 9764ef589dcSJérôme Glisse /* 9774ef589dcSJérôme Glisse * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory 9784ef589dcSJérôme Glisse * 9794ef589dcSJérôme Glisse * @ops: memory event device driver callback (see struct hmm_devmem_ops) 9804ef589dcSJérôme Glisse * @device: device struct to bind the resource too 9814ef589dcSJérôme Glisse * @size: size in bytes of the device memory to add 9824ef589dcSJérôme Glisse * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise 9834ef589dcSJérôme Glisse * 9844ef589dcSJérôme Glisse * This function first finds an empty range of physical address big enough to 9854ef589dcSJérôme Glisse * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which 9864ef589dcSJérôme Glisse * in turn allocates struct pages. It does not do anything beyond that; all 9874ef589dcSJérôme Glisse * events affecting the memory will go through the various callbacks provided 9884ef589dcSJérôme Glisse * by hmm_devmem_ops struct. 9894ef589dcSJérôme Glisse * 9904ef589dcSJérôme Glisse * Device driver should call this function during device initialization and 9914ef589dcSJérôme Glisse * is then responsible of memory management. HMM only provides helpers. 9924ef589dcSJérôme Glisse */ 9934ef589dcSJérôme Glisse struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, 9944ef589dcSJérôme Glisse struct device *device, 9954ef589dcSJérôme Glisse unsigned long size) 9964ef589dcSJérôme Glisse { 9974ef589dcSJérôme Glisse struct hmm_devmem *devmem; 9984ef589dcSJérôme Glisse resource_size_t addr; 9994ef589dcSJérôme Glisse int ret; 10004ef589dcSJérôme Glisse 10014ef589dcSJérôme Glisse static_branch_enable(&device_private_key); 10024ef589dcSJérôme Glisse 10034ef589dcSJérôme Glisse devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem), 10044ef589dcSJérôme Glisse GFP_KERNEL, dev_to_node(device)); 10054ef589dcSJérôme Glisse if (!devmem) 10064ef589dcSJérôme Glisse return ERR_PTR(-ENOMEM); 10074ef589dcSJérôme Glisse 10084ef589dcSJérôme Glisse init_completion(&devmem->completion); 10094ef589dcSJérôme Glisse devmem->pfn_first = -1UL; 10104ef589dcSJérôme Glisse devmem->pfn_last = -1UL; 10114ef589dcSJérôme Glisse devmem->resource = NULL; 10124ef589dcSJérôme Glisse devmem->device = device; 10134ef589dcSJérôme Glisse devmem->ops = ops; 10144ef589dcSJérôme Glisse 10154ef589dcSJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 10164ef589dcSJérôme Glisse 0, GFP_KERNEL); 10174ef589dcSJérôme Glisse if (ret) 10184ef589dcSJérôme Glisse goto error_percpu_ref; 10194ef589dcSJérôme Glisse 10204ef589dcSJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref); 10214ef589dcSJérôme Glisse if (ret) 10224ef589dcSJérôme Glisse goto error_devm_add_action; 10234ef589dcSJérôme Glisse 10244ef589dcSJérôme Glisse size = ALIGN(size, PA_SECTION_SIZE); 10254ef589dcSJérôme Glisse addr = min((unsigned long)iomem_resource.end, 10264ef589dcSJérôme Glisse (1UL << MAX_PHYSMEM_BITS) - 1); 10274ef589dcSJérôme Glisse addr = addr - size + 1UL; 10284ef589dcSJérôme Glisse 10294ef589dcSJérôme Glisse /* 10304ef589dcSJérôme Glisse * FIXME add a new helper to quickly walk resource tree and find free 10314ef589dcSJérôme Glisse * range 10324ef589dcSJérôme Glisse * 10334ef589dcSJérôme Glisse * FIXME what about ioport_resource resource ? 10344ef589dcSJérôme Glisse */ 10354ef589dcSJérôme Glisse for (; addr > size && addr >= iomem_resource.start; addr -= size) { 10364ef589dcSJérôme Glisse ret = region_intersects(addr, size, 0, IORES_DESC_NONE); 10374ef589dcSJérôme Glisse if (ret != REGION_DISJOINT) 10384ef589dcSJérôme Glisse continue; 10394ef589dcSJérôme Glisse 10404ef589dcSJérôme Glisse devmem->resource = devm_request_mem_region(device, addr, size, 10414ef589dcSJérôme Glisse dev_name(device)); 10424ef589dcSJérôme Glisse if (!devmem->resource) { 10434ef589dcSJérôme Glisse ret = -ENOMEM; 10444ef589dcSJérôme Glisse goto error_no_resource; 10454ef589dcSJérôme Glisse } 10464ef589dcSJérôme Glisse break; 10474ef589dcSJérôme Glisse } 10484ef589dcSJérôme Glisse if (!devmem->resource) { 10494ef589dcSJérôme Glisse ret = -ERANGE; 10504ef589dcSJérôme Glisse goto error_no_resource; 10514ef589dcSJérôme Glisse } 10524ef589dcSJérôme Glisse 10534ef589dcSJérôme Glisse devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY; 10544ef589dcSJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 10554ef589dcSJérôme Glisse devmem->pfn_last = devmem->pfn_first + 10564ef589dcSJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 10574ef589dcSJérôme Glisse 10584ef589dcSJérôme Glisse ret = hmm_devmem_pages_create(devmem); 10594ef589dcSJérôme Glisse if (ret) 10604ef589dcSJérôme Glisse goto error_pages; 10614ef589dcSJérôme Glisse 10624ef589dcSJérôme Glisse devres_add(device, devmem); 10634ef589dcSJérôme Glisse 10644ef589dcSJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref); 10654ef589dcSJérôme Glisse if (ret) { 10664ef589dcSJérôme Glisse hmm_devmem_remove(devmem); 10674ef589dcSJérôme Glisse return ERR_PTR(ret); 10684ef589dcSJérôme Glisse } 10694ef589dcSJérôme Glisse 10704ef589dcSJérôme Glisse return devmem; 10714ef589dcSJérôme Glisse 10724ef589dcSJérôme Glisse error_pages: 10734ef589dcSJérôme Glisse devm_release_mem_region(device, devmem->resource->start, 10744ef589dcSJérôme Glisse resource_size(devmem->resource)); 10754ef589dcSJérôme Glisse error_no_resource: 10764ef589dcSJérôme Glisse error_devm_add_action: 10774ef589dcSJérôme Glisse hmm_devmem_ref_kill(&devmem->ref); 10784ef589dcSJérôme Glisse hmm_devmem_ref_exit(&devmem->ref); 10794ef589dcSJérôme Glisse error_percpu_ref: 10804ef589dcSJérôme Glisse devres_free(devmem); 10814ef589dcSJérôme Glisse return ERR_PTR(ret); 10824ef589dcSJérôme Glisse } 10834ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_devmem_add); 10844ef589dcSJérôme Glisse 1085d3df0a42SJérôme Glisse struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops, 1086d3df0a42SJérôme Glisse struct device *device, 1087d3df0a42SJérôme Glisse struct resource *res) 1088d3df0a42SJérôme Glisse { 1089d3df0a42SJérôme Glisse struct hmm_devmem *devmem; 1090d3df0a42SJérôme Glisse int ret; 1091d3df0a42SJérôme Glisse 1092d3df0a42SJérôme Glisse if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY) 1093d3df0a42SJérôme Glisse return ERR_PTR(-EINVAL); 1094d3df0a42SJérôme Glisse 1095d3df0a42SJérôme Glisse static_branch_enable(&device_private_key); 1096d3df0a42SJérôme Glisse 1097d3df0a42SJérôme Glisse devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem), 1098d3df0a42SJérôme Glisse GFP_KERNEL, dev_to_node(device)); 1099d3df0a42SJérôme Glisse if (!devmem) 1100d3df0a42SJérôme Glisse return ERR_PTR(-ENOMEM); 1101d3df0a42SJérôme Glisse 1102d3df0a42SJérôme Glisse init_completion(&devmem->completion); 1103d3df0a42SJérôme Glisse devmem->pfn_first = -1UL; 1104d3df0a42SJérôme Glisse devmem->pfn_last = -1UL; 1105d3df0a42SJérôme Glisse devmem->resource = res; 1106d3df0a42SJérôme Glisse devmem->device = device; 1107d3df0a42SJérôme Glisse devmem->ops = ops; 1108d3df0a42SJérôme Glisse 1109d3df0a42SJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 1110d3df0a42SJérôme Glisse 0, GFP_KERNEL); 1111d3df0a42SJérôme Glisse if (ret) 1112d3df0a42SJérôme Glisse goto error_percpu_ref; 1113d3df0a42SJérôme Glisse 1114d3df0a42SJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref); 1115d3df0a42SJérôme Glisse if (ret) 1116d3df0a42SJérôme Glisse goto error_devm_add_action; 1117d3df0a42SJérôme Glisse 1118d3df0a42SJérôme Glisse 1119d3df0a42SJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 1120d3df0a42SJérôme Glisse devmem->pfn_last = devmem->pfn_first + 1121d3df0a42SJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1122d3df0a42SJérôme Glisse 1123d3df0a42SJérôme Glisse ret = hmm_devmem_pages_create(devmem); 1124d3df0a42SJérôme Glisse if (ret) 1125d3df0a42SJérôme Glisse goto error_devm_add_action; 1126d3df0a42SJérôme Glisse 1127d3df0a42SJérôme Glisse devres_add(device, devmem); 1128d3df0a42SJérôme Glisse 1129d3df0a42SJérôme Glisse ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref); 1130d3df0a42SJérôme Glisse if (ret) { 1131d3df0a42SJérôme Glisse hmm_devmem_remove(devmem); 1132d3df0a42SJérôme Glisse return ERR_PTR(ret); 1133d3df0a42SJérôme Glisse } 1134d3df0a42SJérôme Glisse 1135d3df0a42SJérôme Glisse return devmem; 1136d3df0a42SJérôme Glisse 1137d3df0a42SJérôme Glisse error_devm_add_action: 1138d3df0a42SJérôme Glisse hmm_devmem_ref_kill(&devmem->ref); 1139d3df0a42SJérôme Glisse hmm_devmem_ref_exit(&devmem->ref); 1140d3df0a42SJérôme Glisse error_percpu_ref: 1141d3df0a42SJérôme Glisse devres_free(devmem); 1142d3df0a42SJérôme Glisse return ERR_PTR(ret); 1143d3df0a42SJérôme Glisse } 1144d3df0a42SJérôme Glisse EXPORT_SYMBOL(hmm_devmem_add_resource); 1145d3df0a42SJérôme Glisse 11464ef589dcSJérôme Glisse /* 11474ef589dcSJérôme Glisse * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE) 11484ef589dcSJérôme Glisse * 11494ef589dcSJérôme Glisse * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory 11504ef589dcSJérôme Glisse * 11514ef589dcSJérôme Glisse * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf 11524ef589dcSJérôme Glisse * of the device driver. It will free struct page and remove the resource that 11534ef589dcSJérôme Glisse * reserved the physical address range for this device memory. 11544ef589dcSJérôme Glisse */ 11554ef589dcSJérôme Glisse void hmm_devmem_remove(struct hmm_devmem *devmem) 11564ef589dcSJérôme Glisse { 11574ef589dcSJérôme Glisse resource_size_t start, size; 11584ef589dcSJérôme Glisse struct device *device; 1159d3df0a42SJérôme Glisse bool cdm = false; 11604ef589dcSJérôme Glisse 11614ef589dcSJérôme Glisse if (!devmem) 11624ef589dcSJérôme Glisse return; 11634ef589dcSJérôme Glisse 11644ef589dcSJérôme Glisse device = devmem->device; 11654ef589dcSJérôme Glisse start = devmem->resource->start; 11664ef589dcSJérôme Glisse size = resource_size(devmem->resource); 11674ef589dcSJérôme Glisse 1168d3df0a42SJérôme Glisse cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY; 11694ef589dcSJérôme Glisse hmm_devmem_ref_kill(&devmem->ref); 11704ef589dcSJérôme Glisse hmm_devmem_ref_exit(&devmem->ref); 11714ef589dcSJérôme Glisse hmm_devmem_pages_remove(devmem); 11724ef589dcSJérôme Glisse 1173d3df0a42SJérôme Glisse if (!cdm) 11744ef589dcSJérôme Glisse devm_release_mem_region(device, start, size); 11754ef589dcSJérôme Glisse } 11764ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_devmem_remove); 1177858b54daSJérôme Glisse 1178858b54daSJérôme Glisse /* 1179858b54daSJérôme Glisse * A device driver that wants to handle multiple devices memory through a 1180858b54daSJérôme Glisse * single fake device can use hmm_device to do so. This is purely a helper 1181858b54daSJérôme Glisse * and it is not needed to make use of any HMM functionality. 1182858b54daSJérôme Glisse */ 1183858b54daSJérôme Glisse #define HMM_DEVICE_MAX 256 1184858b54daSJérôme Glisse 1185858b54daSJérôme Glisse static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX); 1186858b54daSJérôme Glisse static DEFINE_SPINLOCK(hmm_device_lock); 1187858b54daSJérôme Glisse static struct class *hmm_device_class; 1188858b54daSJérôme Glisse static dev_t hmm_device_devt; 1189858b54daSJérôme Glisse 1190858b54daSJérôme Glisse static void hmm_device_release(struct device *device) 1191858b54daSJérôme Glisse { 1192858b54daSJérôme Glisse struct hmm_device *hmm_device; 1193858b54daSJérôme Glisse 1194858b54daSJérôme Glisse hmm_device = container_of(device, struct hmm_device, device); 1195858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1196858b54daSJérôme Glisse clear_bit(hmm_device->minor, hmm_device_mask); 1197858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1198858b54daSJérôme Glisse 1199858b54daSJérôme Glisse kfree(hmm_device); 1200858b54daSJérôme Glisse } 1201858b54daSJérôme Glisse 1202858b54daSJérôme Glisse struct hmm_device *hmm_device_new(void *drvdata) 1203858b54daSJérôme Glisse { 1204858b54daSJérôme Glisse struct hmm_device *hmm_device; 1205858b54daSJérôme Glisse 1206858b54daSJérôme Glisse hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL); 1207858b54daSJérôme Glisse if (!hmm_device) 1208858b54daSJérôme Glisse return ERR_PTR(-ENOMEM); 1209858b54daSJérôme Glisse 1210858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1211858b54daSJérôme Glisse hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX); 1212858b54daSJérôme Glisse if (hmm_device->minor >= HMM_DEVICE_MAX) { 1213858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1214858b54daSJérôme Glisse kfree(hmm_device); 1215858b54daSJérôme Glisse return ERR_PTR(-EBUSY); 1216858b54daSJérôme Glisse } 1217858b54daSJérôme Glisse set_bit(hmm_device->minor, hmm_device_mask); 1218858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1219858b54daSJérôme Glisse 1220858b54daSJérôme Glisse dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor); 1221858b54daSJérôme Glisse hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt), 1222858b54daSJérôme Glisse hmm_device->minor); 1223858b54daSJérôme Glisse hmm_device->device.release = hmm_device_release; 1224858b54daSJérôme Glisse dev_set_drvdata(&hmm_device->device, drvdata); 1225858b54daSJérôme Glisse hmm_device->device.class = hmm_device_class; 1226858b54daSJérôme Glisse device_initialize(&hmm_device->device); 1227858b54daSJérôme Glisse 1228858b54daSJérôme Glisse return hmm_device; 1229858b54daSJérôme Glisse } 1230858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_new); 1231858b54daSJérôme Glisse 1232858b54daSJérôme Glisse void hmm_device_put(struct hmm_device *hmm_device) 1233858b54daSJérôme Glisse { 1234858b54daSJérôme Glisse put_device(&hmm_device->device); 1235858b54daSJérôme Glisse } 1236858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_put); 1237858b54daSJérôme Glisse 1238858b54daSJérôme Glisse static int __init hmm_init(void) 1239858b54daSJérôme Glisse { 1240858b54daSJérôme Glisse int ret; 1241858b54daSJérôme Glisse 1242858b54daSJérôme Glisse ret = alloc_chrdev_region(&hmm_device_devt, 0, 1243858b54daSJérôme Glisse HMM_DEVICE_MAX, 1244858b54daSJérôme Glisse "hmm_device"); 1245858b54daSJérôme Glisse if (ret) 1246858b54daSJérôme Glisse return ret; 1247858b54daSJérôme Glisse 1248858b54daSJérôme Glisse hmm_device_class = class_create(THIS_MODULE, "hmm_device"); 1249858b54daSJérôme Glisse if (IS_ERR(hmm_device_class)) { 1250858b54daSJérôme Glisse unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX); 1251858b54daSJérôme Glisse return PTR_ERR(hmm_device_class); 1252858b54daSJérôme Glisse } 1253858b54daSJérôme Glisse return 0; 1254858b54daSJérôme Glisse } 1255858b54daSJérôme Glisse 1256858b54daSJérôme Glisse device_initcall(hmm_init); 1257df6ad698SJérôme Glisse #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ 1258