]>
Commit | Line | Data |
---|---|---|
c942fddf | 1 | /* SPDX-License-Identifier: GPL-2.0-or-later */ |
133ff0ea JG |
2 | /* |
3 | * Copyright 2013 Red Hat Inc. | |
4 | * | |
f813f219 | 5 | * Authors: Jérôme Glisse <jglisse@redhat.com> |
133ff0ea JG |
6 | */ |
7 | /* | |
8 | * Heterogeneous Memory Management (HMM) | |
9 | * | |
ad56b738 | 10 | * See Documentation/vm/hmm.rst for reasons and overview of what HMM is and it |
133ff0ea JG |
11 | * is for. Here we focus on the HMM API description, with some explanation of |
12 | * the underlying implementation. | |
13 | * | |
14 | * Short description: HMM provides a set of helpers to share a virtual address | |
15 | * space between CPU and a device, so that the device can access any valid | |
16 | * address of the process (while still obeying memory protection). HMM also | |
17 | * provides helpers to migrate process memory to device memory, and back. Each | |
18 | * set of functionality (address space mirroring, and migration to and from | |
19 | * device memory) can be used independently of the other. | |
20 | * | |
21 | * | |
22 | * HMM address space mirroring API: | |
23 | * | |
085ea250 RC |
24 | * Use HMM address space mirroring if you want to mirror a range of the CPU |
25 | * page tables of a process into a device page table. Here, "mirror" means "keep | |
133ff0ea JG |
26 | * synchronized". Prerequisites: the device must provide the ability to write- |
27 | * protect its page tables (at PAGE_SIZE granularity), and must be able to | |
28 | * recover from the resulting potential page faults. | |
29 | * | |
30 | * HMM guarantees that at any point in time, a given virtual address points to | |
31 | * either the same memory in both CPU and device page tables (that is: CPU and | |
32 | * device page tables each point to the same pages), or that one page table (CPU | |
33 | * or device) points to no entry, while the other still points to the old page | |
34 | * for the address. The latter case happens when the CPU page table update | |
35 | * happens first, and then the update is mirrored over to the device page table. | |
36 | * This does not cause any issue, because the CPU page table cannot start | |
37 | * pointing to a new page until the device page table is invalidated. | |
38 | * | |
39 | * HMM uses mmu_notifiers to monitor the CPU page tables, and forwards any | |
40 | * updates to each device driver that has registered a mirror. It also provides | |
41 | * some API calls to help with taking a snapshot of the CPU page table, and to | |
42 | * synchronize with any updates that might happen concurrently. | |
43 | * | |
44 | * | |
45 | * HMM migration to and from device memory: | |
46 | * | |
47 | * HMM provides a set of helpers to hotplug device memory as ZONE_DEVICE, with | |
48 | * a new MEMORY_DEVICE_PRIVATE type. This provides a struct page for each page | |
49 | * of the device memory, and allows the device driver to manage its memory | |
50 | * using those struct pages. Having struct pages for device memory makes | |
51 | * migration easier. Because that memory is not addressable by the CPU it must | |
52 | * never be pinned to the device; in other words, any CPU page fault can always | |
53 | * cause the device memory to be migrated (copied/moved) back to regular memory. | |
54 | * | |
55 | * A new migrate helper (migrate_vma()) has been added (see mm/migrate.c) that | |
56 | * allows use of a device DMA engine to perform the copy operation between | |
57 | * regular system memory and device memory. | |
58 | */ | |
59 | #ifndef LINUX_HMM_H | |
60 | #define LINUX_HMM_H | |
61 | ||
62 | #include <linux/kconfig.h> | |
063a7d1d | 63 | #include <asm/pgtable.h> |
133ff0ea | 64 | |
858b54da | 65 | #include <linux/device.h> |
4ef589dc JG |
66 | #include <linux/migrate.h> |
67 | #include <linux/memremap.h> | |
68 | #include <linux/completion.h> | |
a3e0d41c | 69 | #include <linux/mmu_notifier.h> |
4ef589dc | 70 | |
133ff0ea | 71 | /* |
f88a1e90 JG |
72 | * hmm_pfn_flag_e - HMM flag enums |
73 | * | |
133ff0ea | 74 | * Flags: |
86586a41 | 75 | * HMM_PFN_VALID: pfn is valid. It has, at least, read permission. |
133ff0ea | 76 | * HMM_PFN_WRITE: CPU page table has write permission set |
f88a1e90 | 77 | * |
085ea250 RC |
78 | * The driver provides a flags array for mapping page protections to device |
79 | * PTE bits. If the driver valid bit for an entry is bit 3, | |
80 | * i.e., (entry & (1 << 3)), then the driver must provide | |
f88a1e90 | 81 | * an array in hmm_range.flags with hmm_range.flags[HMM_PFN_VALID] == 1 << 3. |
085ea250 | 82 | * Same logic apply to all flags. This is the same idea as vm_page_prot in vma |
f88a1e90 JG |
83 | * except that this is per device driver rather than per architecture. |
84 | */ | |
85 | enum hmm_pfn_flag_e { | |
86 | HMM_PFN_VALID = 0, | |
87 | HMM_PFN_WRITE, | |
f88a1e90 JG |
88 | HMM_PFN_FLAG_MAX |
89 | }; | |
90 | ||
91 | /* | |
92 | * hmm_pfn_value_e - HMM pfn special value | |
93 | * | |
94 | * Flags: | |
da4c3c73 | 95 | * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory |
f88a1e90 | 96 | * HMM_PFN_NONE: corresponding CPU page table entry is pte_none() |
da4c3c73 | 97 | * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the |
67fa1666 | 98 | * result of vmf_insert_pfn() or vm_insert_page(). Therefore, it should not |
da4c3c73 JG |
99 | * be mirrored by a device, because the entry will never have HMM_PFN_VALID |
100 | * set and the pfn value is undefined. | |
f88a1e90 | 101 | * |
085ea250 RC |
102 | * Driver provides values for none entry, error entry, and special entry. |
103 | * Driver can alias (i.e., use same value) error and special, but | |
104 | * it should not alias none with error or special. | |
f88a1e90 JG |
105 | * |
106 | * HMM pfn value returned by hmm_vma_get_pfns() or hmm_vma_fault() will be: | |
107 | * hmm_range.values[HMM_PFN_ERROR] if CPU page table entry is poisonous, | |
085ea250 | 108 | * hmm_range.values[HMM_PFN_NONE] if there is no CPU page table entry, |
f88a1e90 | 109 | * hmm_range.values[HMM_PFN_SPECIAL] if CPU page table entry is a special one |
133ff0ea | 110 | */ |
f88a1e90 JG |
111 | enum hmm_pfn_value_e { |
112 | HMM_PFN_ERROR, | |
113 | HMM_PFN_NONE, | |
114 | HMM_PFN_SPECIAL, | |
115 | HMM_PFN_VALUE_MAX | |
116 | }; | |
117 | ||
118 | /* | |
119 | * struct hmm_range - track invalidation lock on virtual address range | |
120 | * | |
a22dd506 JG |
121 | * @notifier: a mmu_interval_notifier that includes the start/end |
122 | * @notifier_seq: result of mmu_interval_read_begin() | |
704f3f2c | 123 | * @hmm: the core HMM structure this range is active against |
f88a1e90 JG |
124 | * @vma: the vm area struct for the range |
125 | * @list: all range lock are on a list | |
126 | * @start: range virtual start address (inclusive) | |
127 | * @end: range virtual end address (exclusive) | |
128 | * @pfns: array of pfns (big enough for the range) | |
129 | * @flags: pfn flags to match device driver page table | |
130 | * @values: pfn value for some special case (none, special, error, ...) | |
023a019a JG |
131 | * @default_flags: default flags for the range (write, read, ... see hmm doc) |
132 | * @pfn_flags_mask: allows to mask pfn flags so that only default_flags matter | |
f88a1e90 JG |
133 | * @pfn_shifts: pfn shift value (should be <= PAGE_SHIFT) |
134 | * @valid: pfns array did not change since it has been fill by an HMM function | |
08ddddda | 135 | * @dev_private_owner: owner of device private pages |
f88a1e90 JG |
136 | */ |
137 | struct hmm_range { | |
04ec32fb JG |
138 | struct mmu_interval_notifier *notifier; |
139 | unsigned long notifier_seq; | |
f88a1e90 JG |
140 | unsigned long start; |
141 | unsigned long end; | |
142 | uint64_t *pfns; | |
143 | const uint64_t *flags; | |
144 | const uint64_t *values; | |
023a019a JG |
145 | uint64_t default_flags; |
146 | uint64_t pfn_flags_mask; | |
f88a1e90 | 147 | uint8_t pfn_shift; |
08ddddda | 148 | void *dev_private_owner; |
f88a1e90 | 149 | }; |
133ff0ea JG |
150 | |
151 | /* | |
391aab11 JG |
152 | * hmm_device_entry_to_page() - return struct page pointed to by a device entry |
153 | * @range: range use to decode device entry value | |
154 | * @entry: device entry value to get corresponding struct page from | |
085ea250 | 155 | * Return: struct page pointer if entry is a valid, NULL otherwise |
133ff0ea | 156 | * |
391aab11 JG |
157 | * If the device entry is valid (ie valid flag set) then return the struct page |
158 | * matching the entry value. Otherwise return NULL. | |
133ff0ea | 159 | */ |
391aab11 JG |
160 | static inline struct page *hmm_device_entry_to_page(const struct hmm_range *range, |
161 | uint64_t entry) | |
133ff0ea | 162 | { |
391aab11 | 163 | if (entry == range->values[HMM_PFN_NONE]) |
f88a1e90 | 164 | return NULL; |
391aab11 | 165 | if (entry == range->values[HMM_PFN_ERROR]) |
f88a1e90 | 166 | return NULL; |
391aab11 | 167 | if (entry == range->values[HMM_PFN_SPECIAL]) |
133ff0ea | 168 | return NULL; |
391aab11 | 169 | if (!(entry & range->flags[HMM_PFN_VALID])) |
f88a1e90 | 170 | return NULL; |
391aab11 | 171 | return pfn_to_page(entry >> range->pfn_shift); |
133ff0ea JG |
172 | } |
173 | ||
174 | /* | |
391aab11 JG |
175 | * hmm_device_entry_to_pfn() - return pfn value store in a device entry |
176 | * @range: range use to decode device entry value | |
177 | * @entry: device entry to extract pfn from | |
085ea250 | 178 | * Return: pfn value if device entry is valid, -1UL otherwise |
133ff0ea | 179 | */ |
391aab11 JG |
180 | static inline unsigned long |
181 | hmm_device_entry_to_pfn(const struct hmm_range *range, uint64_t pfn) | |
133ff0ea | 182 | { |
f88a1e90 JG |
183 | if (pfn == range->values[HMM_PFN_NONE]) |
184 | return -1UL; | |
185 | if (pfn == range->values[HMM_PFN_ERROR]) | |
186 | return -1UL; | |
187 | if (pfn == range->values[HMM_PFN_SPECIAL]) | |
133ff0ea | 188 | return -1UL; |
f88a1e90 JG |
189 | if (!(pfn & range->flags[HMM_PFN_VALID])) |
190 | return -1UL; | |
191 | return (pfn >> range->pfn_shift); | |
133ff0ea JG |
192 | } |
193 | ||
194 | /* | |
391aab11 | 195 | * hmm_device_entry_from_page() - create a valid device entry for a page |
f88a1e90 | 196 | * @range: range use to encode HMM pfn value |
391aab11 | 197 | * @page: page for which to create the device entry |
085ea250 | 198 | * Return: valid device entry for the page |
133ff0ea | 199 | */ |
391aab11 JG |
200 | static inline uint64_t hmm_device_entry_from_page(const struct hmm_range *range, |
201 | struct page *page) | |
133ff0ea | 202 | { |
f88a1e90 JG |
203 | return (page_to_pfn(page) << range->pfn_shift) | |
204 | range->flags[HMM_PFN_VALID]; | |
133ff0ea JG |
205 | } |
206 | ||
207 | /* | |
391aab11 | 208 | * hmm_device_entry_from_pfn() - create a valid device entry value from pfn |
f88a1e90 | 209 | * @range: range use to encode HMM pfn value |
391aab11 | 210 | * @pfn: pfn value for which to create the device entry |
085ea250 | 211 | * Return: valid device entry for the pfn |
133ff0ea | 212 | */ |
391aab11 JG |
213 | static inline uint64_t hmm_device_entry_from_pfn(const struct hmm_range *range, |
214 | unsigned long pfn) | |
133ff0ea | 215 | { |
f88a1e90 JG |
216 | return (pfn << range->pfn_shift) | |
217 | range->flags[HMM_PFN_VALID]; | |
133ff0ea JG |
218 | } |
219 | ||
107e8998 JG |
220 | /* Don't fault in missing PTEs, just snapshot the current state. */ |
221 | #define HMM_FAULT_SNAPSHOT (1 << 1) | |
222 | ||
da4c3c73 | 223 | /* |
a3e0d41c | 224 | * Please see Documentation/vm/hmm.rst for how to use the range API. |
da4c3c73 | 225 | */ |
9a4903e4 | 226 | long hmm_range_fault(struct hmm_range *range, unsigned int flags); |
74eee180 JG |
227 | |
228 | /* | |
a3e0d41c | 229 | * HMM_RANGE_DEFAULT_TIMEOUT - default timeout (ms) when waiting for a range |
74eee180 | 230 | * |
a3e0d41c JG |
231 | * When waiting for mmu notifiers we need some kind of time out otherwise we |
232 | * could potentialy wait for ever, 1000ms ie 1s sounds like a long time to | |
233 | * wait already. | |
74eee180 | 234 | */ |
a3e0d41c JG |
235 | #define HMM_RANGE_DEFAULT_TIMEOUT 1000 |
236 | ||
133ff0ea | 237 | #endif /* LINUX_HMM_H */ |