xref: /openbmc/linux/include/linux/hmm.h (revision 6c33a6f4)
1 /* SPDX-License-Identifier: GPL-2.0-or-later */
2 /*
3  * Copyright 2013 Red Hat Inc.
4  *
5  * Authors: Jérôme Glisse <jglisse@redhat.com>
6  */
7 /*
8  * Heterogeneous Memory Management (HMM)
9  *
10  * See Documentation/vm/hmm.rst for reasons and overview of what HMM is and it
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  *
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
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>
63 #include <asm/pgtable.h>
64 
65 #include <linux/device.h>
66 #include <linux/migrate.h>
67 #include <linux/memremap.h>
68 #include <linux/completion.h>
69 #include <linux/mmu_notifier.h>
70 
71 /*
72  * hmm_pfn_flag_e - HMM flag enums
73  *
74  * Flags:
75  * HMM_PFN_VALID: pfn is valid. It has, at least, read permission.
76  * HMM_PFN_WRITE: CPU page table has write permission set
77  * HMM_PFN_DEVICE_PRIVATE: private device memory (ZONE_DEVICE)
78  *
79  * The driver provides a flags array for mapping page protections to device
80  * PTE bits. If the driver valid bit for an entry is bit 3,
81  * i.e., (entry & (1 << 3)), then the driver must provide
82  * an array in hmm_range.flags with hmm_range.flags[HMM_PFN_VALID] == 1 << 3.
83  * Same logic apply to all flags. This is the same idea as vm_page_prot in vma
84  * except that this is per device driver rather than per architecture.
85  */
86 enum hmm_pfn_flag_e {
87 	HMM_PFN_VALID = 0,
88 	HMM_PFN_WRITE,
89 	HMM_PFN_DEVICE_PRIVATE,
90 	HMM_PFN_FLAG_MAX
91 };
92 
93 /*
94  * hmm_pfn_value_e - HMM pfn special value
95  *
96  * Flags:
97  * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory
98  * HMM_PFN_NONE: corresponding CPU page table entry is pte_none()
99  * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the
100  *      result of vmf_insert_pfn() or vm_insert_page(). Therefore, it should not
101  *      be mirrored by a device, because the entry will never have HMM_PFN_VALID
102  *      set and the pfn value is undefined.
103  *
104  * Driver provides values for none entry, error entry, and special entry.
105  * Driver can alias (i.e., use same value) error and special, but
106  * it should not alias none with error or special.
107  *
108  * HMM pfn value returned by hmm_vma_get_pfns() or hmm_vma_fault() will be:
109  * hmm_range.values[HMM_PFN_ERROR] if CPU page table entry is poisonous,
110  * hmm_range.values[HMM_PFN_NONE] if there is no CPU page table entry,
111  * hmm_range.values[HMM_PFN_SPECIAL] if CPU page table entry is a special one
112  */
113 enum hmm_pfn_value_e {
114 	HMM_PFN_ERROR,
115 	HMM_PFN_NONE,
116 	HMM_PFN_SPECIAL,
117 	HMM_PFN_VALUE_MAX
118 };
119 
120 /*
121  * struct hmm_range - track invalidation lock on virtual address range
122  *
123  * @notifier: a mmu_interval_notifier that includes the start/end
124  * @notifier_seq: result of mmu_interval_read_begin()
125  * @hmm: the core HMM structure this range is active against
126  * @vma: the vm area struct for the range
127  * @list: all range lock are on a list
128  * @start: range virtual start address (inclusive)
129  * @end: range virtual end address (exclusive)
130  * @pfns: array of pfns (big enough for the range)
131  * @flags: pfn flags to match device driver page table
132  * @values: pfn value for some special case (none, special, error, ...)
133  * @default_flags: default flags for the range (write, read, ... see hmm doc)
134  * @pfn_flags_mask: allows to mask pfn flags so that only default_flags matter
135  * @pfn_shifts: pfn shift value (should be <= PAGE_SHIFT)
136  * @valid: pfns array did not change since it has been fill by an HMM function
137  */
138 struct hmm_range {
139 	struct mmu_interval_notifier *notifier;
140 	unsigned long		notifier_seq;
141 	unsigned long		start;
142 	unsigned long		end;
143 	uint64_t		*pfns;
144 	const uint64_t		*flags;
145 	const uint64_t		*values;
146 	uint64_t		default_flags;
147 	uint64_t		pfn_flags_mask;
148 	uint8_t			pfn_shift;
149 };
150 
151 /*
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
155  * Return: struct page pointer if entry is a valid, NULL otherwise
156  *
157  * If the device entry is valid (ie valid flag set) then return the struct page
158  * matching the entry value. Otherwise return NULL.
159  */
160 static inline struct page *hmm_device_entry_to_page(const struct hmm_range *range,
161 						    uint64_t entry)
162 {
163 	if (entry == range->values[HMM_PFN_NONE])
164 		return NULL;
165 	if (entry == range->values[HMM_PFN_ERROR])
166 		return NULL;
167 	if (entry == range->values[HMM_PFN_SPECIAL])
168 		return NULL;
169 	if (!(entry & range->flags[HMM_PFN_VALID]))
170 		return NULL;
171 	return pfn_to_page(entry >> range->pfn_shift);
172 }
173 
174 /*
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
178  * Return: pfn value if device entry is valid, -1UL otherwise
179  */
180 static inline unsigned long
181 hmm_device_entry_to_pfn(const struct hmm_range *range, uint64_t pfn)
182 {
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])
188 		return -1UL;
189 	if (!(pfn & range->flags[HMM_PFN_VALID]))
190 		return -1UL;
191 	return (pfn >> range->pfn_shift);
192 }
193 
194 /*
195  * hmm_device_entry_from_page() - create a valid device entry for a page
196  * @range: range use to encode HMM pfn value
197  * @page: page for which to create the device entry
198  * Return: valid device entry for the page
199  */
200 static inline uint64_t hmm_device_entry_from_page(const struct hmm_range *range,
201 						  struct page *page)
202 {
203 	return (page_to_pfn(page) << range->pfn_shift) |
204 		range->flags[HMM_PFN_VALID];
205 }
206 
207 /*
208  * hmm_device_entry_from_pfn() - create a valid device entry value from pfn
209  * @range: range use to encode HMM pfn value
210  * @pfn: pfn value for which to create the device entry
211  * Return: valid device entry for the pfn
212  */
213 static inline uint64_t hmm_device_entry_from_pfn(const struct hmm_range *range,
214 						 unsigned long pfn)
215 {
216 	return (pfn << range->pfn_shift) |
217 		range->flags[HMM_PFN_VALID];
218 }
219 
220 /*
221  * Retry fault if non-blocking, drop mmap_sem and return -EAGAIN in that case.
222  */
223 #define HMM_FAULT_ALLOW_RETRY		(1 << 0)
224 
225 /* Don't fault in missing PTEs, just snapshot the current state. */
226 #define HMM_FAULT_SNAPSHOT		(1 << 1)
227 
228 #ifdef CONFIG_HMM_MIRROR
229 /*
230  * Please see Documentation/vm/hmm.rst for how to use the range API.
231  */
232 long hmm_range_fault(struct hmm_range *range, unsigned int flags);
233 #else
234 static inline long hmm_range_fault(struct hmm_range *range, unsigned int flags)
235 {
236 	return -EOPNOTSUPP;
237 }
238 #endif
239 
240 /*
241  * HMM_RANGE_DEFAULT_TIMEOUT - default timeout (ms) when waiting for a range
242  *
243  * When waiting for mmu notifiers we need some kind of time out otherwise we
244  * could potentialy wait for ever, 1000ms ie 1s sounds like a long time to
245  * wait already.
246  */
247 #define HMM_RANGE_DEFAULT_TIMEOUT 1000
248 
249 #endif /* LINUX_HMM_H */
250