KVM: iommu: fix releasing unmapped page
[platform/adaptation/renesas_rcar/renesas_kernel.git] / virt / kvm / iommu.c
1 /*
2  * Copyright (c) 2006, Intel Corporation.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms and conditions of the GNU General Public License,
6  * version 2, as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
11  * more details.
12  *
13  * You should have received a copy of the GNU General Public License along with
14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
15  * Place - Suite 330, Boston, MA 02111-1307 USA.
16  *
17  * Copyright (C) 2006-2008 Intel Corporation
18  * Copyright IBM Corporation, 2008
19  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
20  *
21  * Author: Allen M. Kay <allen.m.kay@intel.com>
22  * Author: Weidong Han <weidong.han@intel.com>
23  * Author: Ben-Ami Yassour <benami@il.ibm.com>
24  */
25
26 #include <linux/list.h>
27 #include <linux/kvm_host.h>
28 #include <linux/module.h>
29 #include <linux/pci.h>
30 #include <linux/stat.h>
31 #include <linux/dmar.h>
32 #include <linux/iommu.h>
33 #include <linux/intel-iommu.h>
34
35 static bool allow_unsafe_assigned_interrupts;
36 module_param_named(allow_unsafe_assigned_interrupts,
37                    allow_unsafe_assigned_interrupts, bool, S_IRUGO | S_IWUSR);
38 MODULE_PARM_DESC(allow_unsafe_assigned_interrupts,
39  "Enable device assignment on platforms without interrupt remapping support.");
40
41 static int kvm_iommu_unmap_memslots(struct kvm *kvm);
42 static void kvm_iommu_put_pages(struct kvm *kvm,
43                                 gfn_t base_gfn, unsigned long npages);
44
45 static pfn_t kvm_pin_pages(struct kvm_memory_slot *slot, gfn_t gfn,
46                            unsigned long size)
47 {
48         gfn_t end_gfn;
49         pfn_t pfn;
50
51         pfn     = gfn_to_pfn_memslot(slot, gfn);
52         end_gfn = gfn + (size >> PAGE_SHIFT);
53         gfn    += 1;
54
55         if (is_error_pfn(pfn))
56                 return pfn;
57
58         while (gfn < end_gfn)
59                 gfn_to_pfn_memslot(slot, gfn++);
60
61         return pfn;
62 }
63
64 int kvm_iommu_map_pages(struct kvm *kvm, struct kvm_memory_slot *slot)
65 {
66         gfn_t gfn, end_gfn;
67         pfn_t pfn;
68         int r = 0;
69         struct iommu_domain *domain = kvm->arch.iommu_domain;
70         int flags;
71
72         /* check if iommu exists and in use */
73         if (!domain)
74                 return 0;
75
76         gfn     = slot->base_gfn;
77         end_gfn = gfn + slot->npages;
78
79         flags = IOMMU_READ | IOMMU_WRITE;
80         if (kvm->arch.iommu_flags & KVM_IOMMU_CACHE_COHERENCY)
81                 flags |= IOMMU_CACHE;
82
83
84         while (gfn < end_gfn) {
85                 unsigned long page_size;
86
87                 /* Check if already mapped */
88                 if (iommu_iova_to_phys(domain, gfn_to_gpa(gfn))) {
89                         gfn += 1;
90                         continue;
91                 }
92
93                 /* Get the page size we could use to map */
94                 page_size = kvm_host_page_size(kvm, gfn);
95
96                 /* Make sure the page_size does not exceed the memslot */
97                 while ((gfn + (page_size >> PAGE_SHIFT)) > end_gfn)
98                         page_size >>= 1;
99
100                 /* Make sure gfn is aligned to the page size we want to map */
101                 while ((gfn << PAGE_SHIFT) & (page_size - 1))
102                         page_size >>= 1;
103
104                 /*
105                  * Pin all pages we are about to map in memory. This is
106                  * important because we unmap and unpin in 4kb steps later.
107                  */
108                 pfn = kvm_pin_pages(slot, gfn, page_size);
109                 if (is_error_pfn(pfn)) {
110                         kvm_release_pfn_clean(pfn);
111                         gfn += 1;
112                         continue;
113                 }
114
115                 /* Map into IO address space */
116                 r = iommu_map(domain, gfn_to_gpa(gfn), pfn_to_hpa(pfn),
117                               page_size, flags);
118                 if (r) {
119                         printk(KERN_ERR "kvm_iommu_map_address:"
120                                "iommu failed to map pfn=%llx\n", pfn);
121                         goto unmap_pages;
122                 }
123
124                 gfn += page_size >> PAGE_SHIFT;
125
126
127         }
128
129         return 0;
130
131 unmap_pages:
132         kvm_iommu_put_pages(kvm, slot->base_gfn, gfn);
133         return r;
134 }
135
136 static int kvm_iommu_map_memslots(struct kvm *kvm)
137 {
138         int idx, r = 0;
139         struct kvm_memslots *slots;
140         struct kvm_memory_slot *memslot;
141
142         idx = srcu_read_lock(&kvm->srcu);
143         slots = kvm_memslots(kvm);
144
145         kvm_for_each_memslot(memslot, slots) {
146                 r = kvm_iommu_map_pages(kvm, memslot);
147                 if (r)
148                         break;
149         }
150         srcu_read_unlock(&kvm->srcu, idx);
151
152         return r;
153 }
154
155 int kvm_assign_device(struct kvm *kvm,
156                       struct kvm_assigned_dev_kernel *assigned_dev)
157 {
158         struct pci_dev *pdev = NULL;
159         struct iommu_domain *domain = kvm->arch.iommu_domain;
160         int r, last_flags;
161
162         /* check if iommu exists and in use */
163         if (!domain)
164                 return 0;
165
166         pdev = assigned_dev->dev;
167         if (pdev == NULL)
168                 return -ENODEV;
169
170         r = iommu_attach_device(domain, &pdev->dev);
171         if (r) {
172                 printk(KERN_ERR "assign device %x:%x:%x.%x failed",
173                         pci_domain_nr(pdev->bus),
174                         pdev->bus->number,
175                         PCI_SLOT(pdev->devfn),
176                         PCI_FUNC(pdev->devfn));
177                 return r;
178         }
179
180         last_flags = kvm->arch.iommu_flags;
181         if (iommu_domain_has_cap(kvm->arch.iommu_domain,
182                                  IOMMU_CAP_CACHE_COHERENCY))
183                 kvm->arch.iommu_flags |= KVM_IOMMU_CACHE_COHERENCY;
184
185         /* Check if need to update IOMMU page table for guest memory */
186         if ((last_flags ^ kvm->arch.iommu_flags) ==
187                         KVM_IOMMU_CACHE_COHERENCY) {
188                 kvm_iommu_unmap_memslots(kvm);
189                 r = kvm_iommu_map_memslots(kvm);
190                 if (r)
191                         goto out_unmap;
192         }
193
194         pdev->dev_flags |= PCI_DEV_FLAGS_ASSIGNED;
195
196         printk(KERN_DEBUG "assign device %x:%x:%x.%x\n",
197                 assigned_dev->host_segnr,
198                 assigned_dev->host_busnr,
199                 PCI_SLOT(assigned_dev->host_devfn),
200                 PCI_FUNC(assigned_dev->host_devfn));
201
202         return 0;
203 out_unmap:
204         kvm_iommu_unmap_memslots(kvm);
205         return r;
206 }
207
208 int kvm_deassign_device(struct kvm *kvm,
209                         struct kvm_assigned_dev_kernel *assigned_dev)
210 {
211         struct iommu_domain *domain = kvm->arch.iommu_domain;
212         struct pci_dev *pdev = NULL;
213
214         /* check if iommu exists and in use */
215         if (!domain)
216                 return 0;
217
218         pdev = assigned_dev->dev;
219         if (pdev == NULL)
220                 return -ENODEV;
221
222         iommu_detach_device(domain, &pdev->dev);
223
224         pdev->dev_flags &= ~PCI_DEV_FLAGS_ASSIGNED;
225
226         printk(KERN_DEBUG "deassign device %x:%x:%x.%x\n",
227                 assigned_dev->host_segnr,
228                 assigned_dev->host_busnr,
229                 PCI_SLOT(assigned_dev->host_devfn),
230                 PCI_FUNC(assigned_dev->host_devfn));
231
232         return 0;
233 }
234
235 int kvm_iommu_map_guest(struct kvm *kvm)
236 {
237         int r;
238
239         if (!iommu_present(&pci_bus_type)) {
240                 printk(KERN_ERR "%s: iommu not found\n", __func__);
241                 return -ENODEV;
242         }
243
244         mutex_lock(&kvm->slots_lock);
245
246         kvm->arch.iommu_domain = iommu_domain_alloc(&pci_bus_type);
247         if (!kvm->arch.iommu_domain) {
248                 r = -ENOMEM;
249                 goto out_unlock;
250         }
251
252         if (!allow_unsafe_assigned_interrupts &&
253             !iommu_domain_has_cap(kvm->arch.iommu_domain,
254                                   IOMMU_CAP_INTR_REMAP)) {
255                 printk(KERN_WARNING "%s: No interrupt remapping support,"
256                        " disallowing device assignment."
257                        " Re-enble with \"allow_unsafe_assigned_interrupts=1\""
258                        " module option.\n", __func__);
259                 iommu_domain_free(kvm->arch.iommu_domain);
260                 kvm->arch.iommu_domain = NULL;
261                 r = -EPERM;
262                 goto out_unlock;
263         }
264
265         r = kvm_iommu_map_memslots(kvm);
266         if (r)
267                 kvm_iommu_unmap_memslots(kvm);
268
269 out_unlock:
270         mutex_unlock(&kvm->slots_lock);
271         return r;
272 }
273
274 static void kvm_unpin_pages(struct kvm *kvm, pfn_t pfn, unsigned long npages)
275 {
276         unsigned long i;
277
278         for (i = 0; i < npages; ++i)
279                 kvm_release_pfn_clean(pfn + i);
280 }
281
282 static void kvm_iommu_put_pages(struct kvm *kvm,
283                                 gfn_t base_gfn, unsigned long npages)
284 {
285         struct iommu_domain *domain;
286         gfn_t end_gfn, gfn;
287         pfn_t pfn;
288         u64 phys;
289
290         domain  = kvm->arch.iommu_domain;
291         end_gfn = base_gfn + npages;
292         gfn     = base_gfn;
293
294         /* check if iommu exists and in use */
295         if (!domain)
296                 return;
297
298         while (gfn < end_gfn) {
299                 unsigned long unmap_pages;
300                 size_t size;
301
302                 /* Get physical address */
303                 phys = iommu_iova_to_phys(domain, gfn_to_gpa(gfn));
304
305                 if (!phys) {
306                         gfn++;
307                         continue;
308                 }
309
310                 pfn  = phys >> PAGE_SHIFT;
311
312                 /* Unmap address from IO address space */
313                 size       = iommu_unmap(domain, gfn_to_gpa(gfn), PAGE_SIZE);
314                 unmap_pages = 1ULL << get_order(size);
315
316                 /* Unpin all pages we just unmapped to not leak any memory */
317                 kvm_unpin_pages(kvm, pfn, unmap_pages);
318
319                 gfn += unmap_pages;
320         }
321 }
322
323 void kvm_iommu_unmap_pages(struct kvm *kvm, struct kvm_memory_slot *slot)
324 {
325         kvm_iommu_put_pages(kvm, slot->base_gfn, slot->npages);
326 }
327
328 static int kvm_iommu_unmap_memslots(struct kvm *kvm)
329 {
330         int idx;
331         struct kvm_memslots *slots;
332         struct kvm_memory_slot *memslot;
333
334         idx = srcu_read_lock(&kvm->srcu);
335         slots = kvm_memslots(kvm);
336
337         kvm_for_each_memslot(memslot, slots)
338                 kvm_iommu_unmap_pages(kvm, memslot);
339
340         srcu_read_unlock(&kvm->srcu, idx);
341
342         return 0;
343 }
344
345 int kvm_iommu_unmap_guest(struct kvm *kvm)
346 {
347         struct iommu_domain *domain = kvm->arch.iommu_domain;
348
349         /* check if iommu exists and in use */
350         if (!domain)
351                 return 0;
352
353         mutex_lock(&kvm->slots_lock);
354         kvm_iommu_unmap_memslots(kvm);
355         kvm->arch.iommu_domain = NULL;
356         mutex_unlock(&kvm->slots_lock);
357
358         iommu_domain_free(domain);
359         return 0;
360 }