Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights
4 * reserved.
5 */
6
7/*
8 * SN Platform Special Memory (mspec) Support
9 *
10 * This driver exports the SN special memory (mspec) facility to user
11 * processes.
12 * There are two types of memory made available thru this driver:
13 * uncached and cached.
14 *
15 * Uncached are used for memory write combining feature of the ia64
16 * cpu.
17 *
18 * Cached are used for areas of memory that are used as cached addresses
19 * on our partition and used as uncached addresses from other partitions.
20 * Due to a design constraint of the SN2 Shub, you can not have processors
21 * on the same FSB perform both a cached and uncached reference to the
22 * same cache line. These special memory cached regions prevent the
23 * kernel from ever dropping in a TLB entry and therefore prevent the
24 * processor from ever speculating a cache line from this page.
25 */
26
27#include <linux/types.h>
28#include <linux/kernel.h>
29#include <linux/module.h>
30#include <linux/init.h>
31#include <linux/errno.h>
32#include <linux/miscdevice.h>
33#include <linux/spinlock.h>
34#include <linux/mm.h>
35#include <linux/fs.h>
36#include <linux/vmalloc.h>
37#include <linux/string.h>
38#include <linux/slab.h>
39#include <linux/numa.h>
40#include <linux/refcount.h>
41#include <asm/page.h>
42#include <linux/atomic.h>
43#include <asm/tlbflush.h>
44#include <asm/uncached.h>
45
46
47#define CACHED_ID "Cached,"
48#define UNCACHED_ID "Uncached"
49#define REVISION "4.0"
50#define MSPEC_BASENAME "mspec"
51
52/*
53 * Page types allocated by the device.
54 */
55enum mspec_page_type {
56 MSPEC_CACHED = 2,
57 MSPEC_UNCACHED
58};
59
60/*
61 * One of these structures is allocated when an mspec region is mmaped. The
62 * structure is pointed to by the vma->vm_private_data field in the vma struct.
63 * This structure is used to record the addresses of the mspec pages.
64 * This structure is shared by all vma's that are split off from the
65 * original vma when split_vma()'s are done.
66 *
67 * The refcnt is incremented atomically because mm->mmap_lock does not
68 * protect in fork case where multiple tasks share the vma_data.
69 */
70struct vma_data {
71 refcount_t refcnt; /* Number of vmas sharing the data. */
72 spinlock_t lock; /* Serialize access to this structure. */
73 int count; /* Number of pages allocated. */
74 enum mspec_page_type type; /* Type of pages allocated. */
75 unsigned long vm_start; /* Original (unsplit) base. */
76 unsigned long vm_end; /* Original (unsplit) end. */
77 unsigned long maddr[]; /* Array of MSPEC addresses. */
78};
79
80/*
81 * mspec_open
82 *
83 * Called when a device mapping is created by a means other than mmap
84 * (via fork, munmap, etc.). Increments the reference count on the
85 * underlying mspec data so it is not freed prematurely.
86 */
87static void
88mspec_open(struct vm_area_struct *vma)
89{
90 struct vma_data *vdata;
91
92 vdata = vma->vm_private_data;
93 refcount_inc(&vdata->refcnt);
94}
95
96/*
97 * mspec_close
98 *
99 * Called when unmapping a device mapping. Frees all mspec pages
100 * belonging to all the vma's sharing this vma_data structure.
101 */
102static void
103mspec_close(struct vm_area_struct *vma)
104{
105 struct vma_data *vdata;
106 int index, last_index;
107 unsigned long my_page;
108
109 vdata = vma->vm_private_data;
110
111 if (!refcount_dec_and_test(&vdata->refcnt))
112 return;
113
114 last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT;
115 for (index = 0; index < last_index; index++) {
116 if (vdata->maddr[index] == 0)
117 continue;
118 /*
119 * Clear the page before sticking it back
120 * into the pool.
121 */
122 my_page = vdata->maddr[index];
123 vdata->maddr[index] = 0;
124 memset((char *)my_page, 0, PAGE_SIZE);
125 uncached_free_page(my_page, 1);
126 }
127
128 kvfree(vdata);
129}
130
131/*
132 * mspec_fault
133 *
134 * Creates a mspec page and maps it to user space.
135 */
136static vm_fault_t
137mspec_fault(struct vm_fault *vmf)
138{
139 unsigned long paddr, maddr;
140 unsigned long pfn;
141 pgoff_t index = vmf->pgoff;
142 struct vma_data *vdata = vmf->vma->vm_private_data;
143
144 maddr = (volatile unsigned long) vdata->maddr[index];
145 if (maddr == 0) {
146 maddr = uncached_alloc_page(numa_node_id(), 1);
147 if (maddr == 0)
148 return VM_FAULT_OOM;
149
150 spin_lock(&vdata->lock);
151 if (vdata->maddr[index] == 0) {
152 vdata->count++;
153 vdata->maddr[index] = maddr;
154 } else {
155 uncached_free_page(maddr, 1);
156 maddr = vdata->maddr[index];
157 }
158 spin_unlock(&vdata->lock);
159 }
160
161 paddr = maddr & ~__IA64_UNCACHED_OFFSET;
162 pfn = paddr >> PAGE_SHIFT;
163
164 return vmf_insert_pfn(vmf->vma, vmf->address, pfn);
165}
166
167static const struct vm_operations_struct mspec_vm_ops = {
168 .open = mspec_open,
169 .close = mspec_close,
170 .fault = mspec_fault,
171};
172
173/*
174 * mspec_mmap
175 *
176 * Called when mmapping the device. Initializes the vma with a fault handler
177 * and private data structure necessary to allocate, track, and free the
178 * underlying pages.
179 */
180static int
181mspec_mmap(struct file *file, struct vm_area_struct *vma,
182 enum mspec_page_type type)
183{
184 struct vma_data *vdata;
185 int pages, vdata_size;
186
187 if (vma->vm_pgoff != 0)
188 return -EINVAL;
189
190 if ((vma->vm_flags & VM_SHARED) == 0)
191 return -EINVAL;
192
193 if ((vma->vm_flags & VM_WRITE) == 0)
194 return -EPERM;
195
196 pages = vma_pages(vma);
197 vdata_size = sizeof(struct vma_data) + pages * sizeof(long);
198 if (vdata_size <= PAGE_SIZE)
199 vdata = kzalloc(vdata_size, GFP_KERNEL);
200 else
201 vdata = vzalloc(vdata_size);
202 if (!vdata)
203 return -ENOMEM;
204
205 vdata->vm_start = vma->vm_start;
206 vdata->vm_end = vma->vm_end;
207 vdata->type = type;
208 spin_lock_init(&vdata->lock);
209 refcount_set(&vdata->refcnt, 1);
210 vma->vm_private_data = vdata;
211
212 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
213 if (vdata->type == MSPEC_UNCACHED)
214 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
215 vma->vm_ops = &mspec_vm_ops;
216
217 return 0;
218}
219
220static int
221cached_mmap(struct file *file, struct vm_area_struct *vma)
222{
223 return mspec_mmap(file, vma, MSPEC_CACHED);
224}
225
226static int
227uncached_mmap(struct file *file, struct vm_area_struct *vma)
228{
229 return mspec_mmap(file, vma, MSPEC_UNCACHED);
230}
231
232static const struct file_operations cached_fops = {
233 .owner = THIS_MODULE,
234 .mmap = cached_mmap,
235 .llseek = noop_llseek,
236};
237
238static struct miscdevice cached_miscdev = {
239 .minor = MISC_DYNAMIC_MINOR,
240 .name = "mspec_cached",
241 .fops = &cached_fops
242};
243
244static const struct file_operations uncached_fops = {
245 .owner = THIS_MODULE,
246 .mmap = uncached_mmap,
247 .llseek = noop_llseek,
248};
249
250static struct miscdevice uncached_miscdev = {
251 .minor = MISC_DYNAMIC_MINOR,
252 .name = "mspec_uncached",
253 .fops = &uncached_fops
254};
255
256/*
257 * mspec_init
258 *
259 * Called at boot time to initialize the mspec facility.
260 */
261static int __init
262mspec_init(void)
263{
264 int ret;
265
266 ret = misc_register(&cached_miscdev);
267 if (ret) {
268 printk(KERN_ERR "%s: failed to register device %i\n",
269 CACHED_ID, ret);
270 return ret;
271 }
272 ret = misc_register(&uncached_miscdev);
273 if (ret) {
274 printk(KERN_ERR "%s: failed to register device %i\n",
275 UNCACHED_ID, ret);
276 misc_deregister(&cached_miscdev);
277 return ret;
278 }
279
280 printk(KERN_INFO "%s %s initialized devices: %s %s\n",
281 MSPEC_BASENAME, REVISION, CACHED_ID, UNCACHED_ID);
282
283 return 0;
284}
285
286static void __exit
287mspec_exit(void)
288{
289 misc_deregister(&uncached_miscdev);
290 misc_deregister(&cached_miscdev);
291}
292
293module_init(mspec_init);
294module_exit(mspec_exit);
295
296MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>");
297MODULE_DESCRIPTION("Driver for SGI SN special memory operations");
298MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights
4 * reserved.
5 */
6
7/*
8 * SN Platform Special Memory (mspec) Support
9 *
10 * This driver exports the SN special memory (mspec) facility to user
11 * processes.
12 * There are two types of memory made available thru this driver:
13 * uncached and cached.
14 *
15 * Uncached are used for memory write combining feature of the ia64
16 * cpu.
17 *
18 * Cached are used for areas of memory that are used as cached addresses
19 * on our partition and used as uncached addresses from other partitions.
20 * Due to a design constraint of the SN2 Shub, you can not have processors
21 * on the same FSB perform both a cached and uncached reference to the
22 * same cache line. These special memory cached regions prevent the
23 * kernel from ever dropping in a TLB entry and therefore prevent the
24 * processor from ever speculating a cache line from this page.
25 */
26
27#include <linux/types.h>
28#include <linux/kernel.h>
29#include <linux/module.h>
30#include <linux/init.h>
31#include <linux/errno.h>
32#include <linux/miscdevice.h>
33#include <linux/spinlock.h>
34#include <linux/mm.h>
35#include <linux/fs.h>
36#include <linux/vmalloc.h>
37#include <linux/string.h>
38#include <linux/slab.h>
39#include <linux/numa.h>
40#include <linux/refcount.h>
41#include <asm/page.h>
42#include <asm/pgtable.h>
43#include <linux/atomic.h>
44#include <asm/tlbflush.h>
45#include <asm/uncached.h>
46
47
48#define CACHED_ID "Cached,"
49#define UNCACHED_ID "Uncached"
50#define REVISION "4.0"
51#define MSPEC_BASENAME "mspec"
52
53/*
54 * Page types allocated by the device.
55 */
56enum mspec_page_type {
57 MSPEC_CACHED = 2,
58 MSPEC_UNCACHED
59};
60
61/*
62 * One of these structures is allocated when an mspec region is mmaped. The
63 * structure is pointed to by the vma->vm_private_data field in the vma struct.
64 * This structure is used to record the addresses of the mspec pages.
65 * This structure is shared by all vma's that are split off from the
66 * original vma when split_vma()'s are done.
67 *
68 * The refcnt is incremented atomically because mm->mmap_sem does not
69 * protect in fork case where multiple tasks share the vma_data.
70 */
71struct vma_data {
72 refcount_t refcnt; /* Number of vmas sharing the data. */
73 spinlock_t lock; /* Serialize access to this structure. */
74 int count; /* Number of pages allocated. */
75 enum mspec_page_type type; /* Type of pages allocated. */
76 unsigned long vm_start; /* Original (unsplit) base. */
77 unsigned long vm_end; /* Original (unsplit) end. */
78 unsigned long maddr[0]; /* Array of MSPEC addresses. */
79};
80
81/*
82 * mspec_open
83 *
84 * Called when a device mapping is created by a means other than mmap
85 * (via fork, munmap, etc.). Increments the reference count on the
86 * underlying mspec data so it is not freed prematurely.
87 */
88static void
89mspec_open(struct vm_area_struct *vma)
90{
91 struct vma_data *vdata;
92
93 vdata = vma->vm_private_data;
94 refcount_inc(&vdata->refcnt);
95}
96
97/*
98 * mspec_close
99 *
100 * Called when unmapping a device mapping. Frees all mspec pages
101 * belonging to all the vma's sharing this vma_data structure.
102 */
103static void
104mspec_close(struct vm_area_struct *vma)
105{
106 struct vma_data *vdata;
107 int index, last_index;
108 unsigned long my_page;
109
110 vdata = vma->vm_private_data;
111
112 if (!refcount_dec_and_test(&vdata->refcnt))
113 return;
114
115 last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT;
116 for (index = 0; index < last_index; index++) {
117 if (vdata->maddr[index] == 0)
118 continue;
119 /*
120 * Clear the page before sticking it back
121 * into the pool.
122 */
123 my_page = vdata->maddr[index];
124 vdata->maddr[index] = 0;
125 memset((char *)my_page, 0, PAGE_SIZE);
126 uncached_free_page(my_page, 1);
127 }
128
129 kvfree(vdata);
130}
131
132/*
133 * mspec_fault
134 *
135 * Creates a mspec page and maps it to user space.
136 */
137static vm_fault_t
138mspec_fault(struct vm_fault *vmf)
139{
140 unsigned long paddr, maddr;
141 unsigned long pfn;
142 pgoff_t index = vmf->pgoff;
143 struct vma_data *vdata = vmf->vma->vm_private_data;
144
145 maddr = (volatile unsigned long) vdata->maddr[index];
146 if (maddr == 0) {
147 maddr = uncached_alloc_page(numa_node_id(), 1);
148 if (maddr == 0)
149 return VM_FAULT_OOM;
150
151 spin_lock(&vdata->lock);
152 if (vdata->maddr[index] == 0) {
153 vdata->count++;
154 vdata->maddr[index] = maddr;
155 } else {
156 uncached_free_page(maddr, 1);
157 maddr = vdata->maddr[index];
158 }
159 spin_unlock(&vdata->lock);
160 }
161
162 paddr = maddr & ~__IA64_UNCACHED_OFFSET;
163 pfn = paddr >> PAGE_SHIFT;
164
165 return vmf_insert_pfn(vmf->vma, vmf->address, pfn);
166}
167
168static const struct vm_operations_struct mspec_vm_ops = {
169 .open = mspec_open,
170 .close = mspec_close,
171 .fault = mspec_fault,
172};
173
174/*
175 * mspec_mmap
176 *
177 * Called when mmapping the device. Initializes the vma with a fault handler
178 * and private data structure necessary to allocate, track, and free the
179 * underlying pages.
180 */
181static int
182mspec_mmap(struct file *file, struct vm_area_struct *vma,
183 enum mspec_page_type type)
184{
185 struct vma_data *vdata;
186 int pages, vdata_size;
187
188 if (vma->vm_pgoff != 0)
189 return -EINVAL;
190
191 if ((vma->vm_flags & VM_SHARED) == 0)
192 return -EINVAL;
193
194 if ((vma->vm_flags & VM_WRITE) == 0)
195 return -EPERM;
196
197 pages = vma_pages(vma);
198 vdata_size = sizeof(struct vma_data) + pages * sizeof(long);
199 if (vdata_size <= PAGE_SIZE)
200 vdata = kzalloc(vdata_size, GFP_KERNEL);
201 else
202 vdata = vzalloc(vdata_size);
203 if (!vdata)
204 return -ENOMEM;
205
206 vdata->vm_start = vma->vm_start;
207 vdata->vm_end = vma->vm_end;
208 vdata->type = type;
209 spin_lock_init(&vdata->lock);
210 refcount_set(&vdata->refcnt, 1);
211 vma->vm_private_data = vdata;
212
213 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
214 if (vdata->type == MSPEC_UNCACHED)
215 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
216 vma->vm_ops = &mspec_vm_ops;
217
218 return 0;
219}
220
221static int
222cached_mmap(struct file *file, struct vm_area_struct *vma)
223{
224 return mspec_mmap(file, vma, MSPEC_CACHED);
225}
226
227static int
228uncached_mmap(struct file *file, struct vm_area_struct *vma)
229{
230 return mspec_mmap(file, vma, MSPEC_UNCACHED);
231}
232
233static const struct file_operations cached_fops = {
234 .owner = THIS_MODULE,
235 .mmap = cached_mmap,
236 .llseek = noop_llseek,
237};
238
239static struct miscdevice cached_miscdev = {
240 .minor = MISC_DYNAMIC_MINOR,
241 .name = "mspec_cached",
242 .fops = &cached_fops
243};
244
245static const struct file_operations uncached_fops = {
246 .owner = THIS_MODULE,
247 .mmap = uncached_mmap,
248 .llseek = noop_llseek,
249};
250
251static struct miscdevice uncached_miscdev = {
252 .minor = MISC_DYNAMIC_MINOR,
253 .name = "mspec_uncached",
254 .fops = &uncached_fops
255};
256
257/*
258 * mspec_init
259 *
260 * Called at boot time to initialize the mspec facility.
261 */
262static int __init
263mspec_init(void)
264{
265 int ret;
266
267 ret = misc_register(&cached_miscdev);
268 if (ret) {
269 printk(KERN_ERR "%s: failed to register device %i\n",
270 CACHED_ID, ret);
271 return ret;
272 }
273 ret = misc_register(&uncached_miscdev);
274 if (ret) {
275 printk(KERN_ERR "%s: failed to register device %i\n",
276 UNCACHED_ID, ret);
277 misc_deregister(&cached_miscdev);
278 return ret;
279 }
280
281 printk(KERN_INFO "%s %s initialized devices: %s %s\n",
282 MSPEC_BASENAME, REVISION, CACHED_ID, UNCACHED_ID);
283
284 return 0;
285}
286
287static void __exit
288mspec_exit(void)
289{
290 misc_deregister(&uncached_miscdev);
291 misc_deregister(&cached_miscdev);
292}
293
294module_init(mspec_init);
295module_exit(mspec_exit);
296
297MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>");
298MODULE_DESCRIPTION("Driver for SGI SN special memory operations");
299MODULE_LICENSE("GPL");