Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* IBM POWER Barrier Synchronization Register Driver
3 *
4 * Copyright IBM Corporation 2008
5 *
6 * Author: Sonny Rao <sonnyrao@us.ibm.com>
7 */
8
9#include <linux/device.h>
10#include <linux/kernel.h>
11#include <linux/of.h>
12#include <linux/of_address.h>
13#include <linux/fs.h>
14#include <linux/module.h>
15#include <linux/cdev.h>
16#include <linux/list.h>
17#include <linux/mm.h>
18#include <linux/slab.h>
19#include <asm/io.h>
20
21/*
22 This driver exposes a special register which can be used for fast
23 synchronization across a large SMP machine. The hardware is exposed
24 as an array of bytes where each process will write to one of the bytes to
25 indicate it has finished the current stage and this update is broadcast to
26 all processors without having to bounce a cacheline between them. In
27 POWER5 and POWER6 there is one of these registers per SMP, but it is
28 presented in two forms; first, it is given as a whole and then as a number
29 of smaller registers which alias to parts of the single whole register.
30 This can potentially allow multiple groups of processes to each have their
31 own private synchronization device.
32
33 Note that this hardware *must* be written to using *only* single byte writes.
34 It may be read using 1, 2, 4, or 8 byte loads which must be aligned since
35 this region is treated as cache-inhibited processes should also use a
36 full sync before and after writing to the BSR to ensure all stores and
37 the BSR update have made it to all chips in the system
38*/
39
40/* This is arbitrary number, up to Power6 it's been 17 or fewer */
41#define BSR_MAX_DEVS (32)
42
43struct bsr_dev {
44 u64 bsr_addr; /* Real address */
45 u64 bsr_len; /* length of mem region we can map */
46 unsigned bsr_bytes; /* size of the BSR reg itself */
47 unsigned bsr_stride; /* interval at which BSR repeats in the page */
48 unsigned bsr_type; /* maps to enum below */
49 unsigned bsr_num; /* bsr id number for its type */
50 int bsr_minor;
51
52 struct list_head bsr_list;
53
54 dev_t bsr_dev;
55 struct cdev bsr_cdev;
56 struct device *bsr_device;
57 char bsr_name[32];
58
59};
60
61static unsigned total_bsr_devs;
62static LIST_HEAD(bsr_devs);
63static int bsr_major;
64
65enum {
66 BSR_8 = 0,
67 BSR_16 = 1,
68 BSR_64 = 2,
69 BSR_128 = 3,
70 BSR_4096 = 4,
71 BSR_UNKNOWN = 5,
72 BSR_MAX = 6,
73};
74
75static unsigned bsr_types[BSR_MAX];
76
77static ssize_t
78bsr_size_show(struct device *dev, struct device_attribute *attr, char *buf)
79{
80 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
81 return sprintf(buf, "%u\n", bsr_dev->bsr_bytes);
82}
83static DEVICE_ATTR_RO(bsr_size);
84
85static ssize_t
86bsr_stride_show(struct device *dev, struct device_attribute *attr, char *buf)
87{
88 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
89 return sprintf(buf, "%u\n", bsr_dev->bsr_stride);
90}
91static DEVICE_ATTR_RO(bsr_stride);
92
93static ssize_t
94bsr_length_show(struct device *dev, struct device_attribute *attr, char *buf)
95{
96 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
97 return sprintf(buf, "%llu\n", bsr_dev->bsr_len);
98}
99static DEVICE_ATTR_RO(bsr_length);
100
101static struct attribute *bsr_dev_attrs[] = {
102 &dev_attr_bsr_size.attr,
103 &dev_attr_bsr_stride.attr,
104 &dev_attr_bsr_length.attr,
105 NULL,
106};
107ATTRIBUTE_GROUPS(bsr_dev);
108
109static const struct class bsr_class = {
110 .name = "bsr",
111 .dev_groups = bsr_dev_groups,
112};
113
114static int bsr_mmap(struct file *filp, struct vm_area_struct *vma)
115{
116 unsigned long size = vma->vm_end - vma->vm_start;
117 struct bsr_dev *dev = filp->private_data;
118 int ret;
119
120 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
121
122 /* check for the case of a small BSR device and map one 4k page for it*/
123 if (dev->bsr_len < PAGE_SIZE && size == PAGE_SIZE)
124 ret = remap_4k_pfn(vma, vma->vm_start, dev->bsr_addr >> 12,
125 vma->vm_page_prot);
126 else if (size <= dev->bsr_len)
127 ret = io_remap_pfn_range(vma, vma->vm_start,
128 dev->bsr_addr >> PAGE_SHIFT,
129 size, vma->vm_page_prot);
130 else
131 return -EINVAL;
132
133 if (ret)
134 return -EAGAIN;
135
136 return 0;
137}
138
139static int bsr_open(struct inode *inode, struct file *filp)
140{
141 struct cdev *cdev = inode->i_cdev;
142 struct bsr_dev *dev = container_of(cdev, struct bsr_dev, bsr_cdev);
143
144 filp->private_data = dev;
145 return 0;
146}
147
148static const struct file_operations bsr_fops = {
149 .owner = THIS_MODULE,
150 .mmap = bsr_mmap,
151 .open = bsr_open,
152 .llseek = noop_llseek,
153};
154
155static void bsr_cleanup_devs(void)
156{
157 struct bsr_dev *cur, *n;
158
159 list_for_each_entry_safe(cur, n, &bsr_devs, bsr_list) {
160 if (cur->bsr_device) {
161 cdev_del(&cur->bsr_cdev);
162 device_del(cur->bsr_device);
163 }
164 list_del(&cur->bsr_list);
165 kfree(cur);
166 }
167}
168
169static int bsr_add_node(struct device_node *bn)
170{
171 int bsr_stride_len, bsr_bytes_len, num_bsr_devs;
172 const u32 *bsr_stride;
173 const u32 *bsr_bytes;
174 unsigned i;
175 int ret = -ENODEV;
176
177 bsr_stride = of_get_property(bn, "ibm,lock-stride", &bsr_stride_len);
178 bsr_bytes = of_get_property(bn, "ibm,#lock-bytes", &bsr_bytes_len);
179
180 if (!bsr_stride || !bsr_bytes ||
181 (bsr_stride_len != bsr_bytes_len)) {
182 printk(KERN_ERR "bsr of-node has missing/incorrect property\n");
183 return ret;
184 }
185
186 num_bsr_devs = bsr_bytes_len / sizeof(u32);
187
188 for (i = 0 ; i < num_bsr_devs; i++) {
189 struct bsr_dev *cur = kzalloc(sizeof(struct bsr_dev),
190 GFP_KERNEL);
191 struct resource res;
192 int result;
193
194 if (!cur) {
195 printk(KERN_ERR "Unable to alloc bsr dev\n");
196 ret = -ENOMEM;
197 goto out_err;
198 }
199
200 result = of_address_to_resource(bn, i, &res);
201 if (result < 0) {
202 printk(KERN_ERR "bsr of-node has invalid reg property, skipping\n");
203 kfree(cur);
204 continue;
205 }
206
207 cur->bsr_minor = i + total_bsr_devs;
208 cur->bsr_addr = res.start;
209 cur->bsr_len = resource_size(&res);
210 cur->bsr_bytes = bsr_bytes[i];
211 cur->bsr_stride = bsr_stride[i];
212 cur->bsr_dev = MKDEV(bsr_major, i + total_bsr_devs);
213
214 /* if we have a bsr_len of > 4k and less then PAGE_SIZE (64k pages) */
215 /* we can only map 4k of it, so only advertise the 4k in sysfs */
216 if (cur->bsr_len > 4096 && cur->bsr_len < PAGE_SIZE)
217 cur->bsr_len = 4096;
218
219 switch(cur->bsr_bytes) {
220 case 8:
221 cur->bsr_type = BSR_8;
222 break;
223 case 16:
224 cur->bsr_type = BSR_16;
225 break;
226 case 64:
227 cur->bsr_type = BSR_64;
228 break;
229 case 128:
230 cur->bsr_type = BSR_128;
231 break;
232 case 4096:
233 cur->bsr_type = BSR_4096;
234 break;
235 default:
236 cur->bsr_type = BSR_UNKNOWN;
237 }
238
239 cur->bsr_num = bsr_types[cur->bsr_type];
240 snprintf(cur->bsr_name, 32, "bsr%d_%d",
241 cur->bsr_bytes, cur->bsr_num);
242
243 cdev_init(&cur->bsr_cdev, &bsr_fops);
244 result = cdev_add(&cur->bsr_cdev, cur->bsr_dev, 1);
245 if (result) {
246 kfree(cur);
247 goto out_err;
248 }
249
250 cur->bsr_device = device_create(&bsr_class, NULL, cur->bsr_dev,
251 cur, "%s", cur->bsr_name);
252 if (IS_ERR(cur->bsr_device)) {
253 printk(KERN_ERR "device_create failed for %s\n",
254 cur->bsr_name);
255 cdev_del(&cur->bsr_cdev);
256 kfree(cur);
257 goto out_err;
258 }
259
260 bsr_types[cur->bsr_type] = cur->bsr_num + 1;
261 list_add_tail(&cur->bsr_list, &bsr_devs);
262 }
263
264 total_bsr_devs += num_bsr_devs;
265
266 return 0;
267
268 out_err:
269
270 bsr_cleanup_devs();
271 return ret;
272}
273
274static int bsr_create_devs(struct device_node *bn)
275{
276 int ret;
277
278 while (bn) {
279 ret = bsr_add_node(bn);
280 if (ret) {
281 of_node_put(bn);
282 return ret;
283 }
284 bn = of_find_compatible_node(bn, NULL, "ibm,bsr");
285 }
286 return 0;
287}
288
289static int __init bsr_init(void)
290{
291 struct device_node *np;
292 dev_t bsr_dev;
293 int ret = -ENODEV;
294
295 np = of_find_compatible_node(NULL, NULL, "ibm,bsr");
296 if (!np)
297 goto out_err;
298
299 ret = class_register(&bsr_class);
300 if (ret)
301 goto out_err_1;
302
303 ret = alloc_chrdev_region(&bsr_dev, 0, BSR_MAX_DEVS, "bsr");
304 bsr_major = MAJOR(bsr_dev);
305 if (ret < 0) {
306 printk(KERN_ERR "alloc_chrdev_region() failed for bsr\n");
307 goto out_err_2;
308 }
309
310 ret = bsr_create_devs(np);
311 if (ret < 0) {
312 np = NULL;
313 goto out_err_3;
314 }
315
316 return 0;
317
318 out_err_3:
319 unregister_chrdev_region(bsr_dev, BSR_MAX_DEVS);
320
321 out_err_2:
322 class_unregister(&bsr_class);
323
324 out_err_1:
325 of_node_put(np);
326
327 out_err:
328
329 return ret;
330}
331
332static void __exit bsr_exit(void)
333{
334
335 bsr_cleanup_devs();
336
337 class_unregister(&bsr_class);
338
339 if (bsr_major)
340 unregister_chrdev_region(MKDEV(bsr_major, 0), BSR_MAX_DEVS);
341}
342
343module_init(bsr_init);
344module_exit(bsr_exit);
345MODULE_DESCRIPTION("IBM POWER Barrier Synchronization Register Driver");
346MODULE_LICENSE("GPL");
347MODULE_AUTHOR("Sonny Rao <sonnyrao@us.ibm.com>");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* IBM POWER Barrier Synchronization Register Driver
3 *
4 * Copyright IBM Corporation 2008
5 *
6 * Author: Sonny Rao <sonnyrao@us.ibm.com>
7 */
8
9#include <linux/device.h>
10#include <linux/kernel.h>
11#include <linux/of.h>
12#include <linux/of_address.h>
13#include <linux/fs.h>
14#include <linux/module.h>
15#include <linux/cdev.h>
16#include <linux/list.h>
17#include <linux/mm.h>
18#include <linux/slab.h>
19#include <asm/io.h>
20
21/*
22 This driver exposes a special register which can be used for fast
23 synchronization across a large SMP machine. The hardware is exposed
24 as an array of bytes where each process will write to one of the bytes to
25 indicate it has finished the current stage and this update is broadcast to
26 all processors without having to bounce a cacheline between them. In
27 POWER5 and POWER6 there is one of these registers per SMP, but it is
28 presented in two forms; first, it is given as a whole and then as a number
29 of smaller registers which alias to parts of the single whole register.
30 This can potentially allow multiple groups of processes to each have their
31 own private synchronization device.
32
33 Note that this hardware *must* be written to using *only* single byte writes.
34 It may be read using 1, 2, 4, or 8 byte loads which must be aligned since
35 this region is treated as cache-inhibited processes should also use a
36 full sync before and after writing to the BSR to ensure all stores and
37 the BSR update have made it to all chips in the system
38*/
39
40/* This is arbitrary number, up to Power6 it's been 17 or fewer */
41#define BSR_MAX_DEVS (32)
42
43struct bsr_dev {
44 u64 bsr_addr; /* Real address */
45 u64 bsr_len; /* length of mem region we can map */
46 unsigned bsr_bytes; /* size of the BSR reg itself */
47 unsigned bsr_stride; /* interval at which BSR repeats in the page */
48 unsigned bsr_type; /* maps to enum below */
49 unsigned bsr_num; /* bsr id number for its type */
50 int bsr_minor;
51
52 struct list_head bsr_list;
53
54 dev_t bsr_dev;
55 struct cdev bsr_cdev;
56 struct device *bsr_device;
57 char bsr_name[32];
58
59};
60
61static unsigned total_bsr_devs;
62static LIST_HEAD(bsr_devs);
63static int bsr_major;
64
65enum {
66 BSR_8 = 0,
67 BSR_16 = 1,
68 BSR_64 = 2,
69 BSR_128 = 3,
70 BSR_4096 = 4,
71 BSR_UNKNOWN = 5,
72 BSR_MAX = 6,
73};
74
75static unsigned bsr_types[BSR_MAX];
76
77static ssize_t
78bsr_size_show(struct device *dev, struct device_attribute *attr, char *buf)
79{
80 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
81 return sprintf(buf, "%u\n", bsr_dev->bsr_bytes);
82}
83static DEVICE_ATTR_RO(bsr_size);
84
85static ssize_t
86bsr_stride_show(struct device *dev, struct device_attribute *attr, char *buf)
87{
88 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
89 return sprintf(buf, "%u\n", bsr_dev->bsr_stride);
90}
91static DEVICE_ATTR_RO(bsr_stride);
92
93static ssize_t
94bsr_length_show(struct device *dev, struct device_attribute *attr, char *buf)
95{
96 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
97 return sprintf(buf, "%llu\n", bsr_dev->bsr_len);
98}
99static DEVICE_ATTR_RO(bsr_length);
100
101static struct attribute *bsr_dev_attrs[] = {
102 &dev_attr_bsr_size.attr,
103 &dev_attr_bsr_stride.attr,
104 &dev_attr_bsr_length.attr,
105 NULL,
106};
107ATTRIBUTE_GROUPS(bsr_dev);
108
109static const struct class bsr_class = {
110 .name = "bsr",
111 .dev_groups = bsr_dev_groups,
112};
113
114static int bsr_mmap(struct file *filp, struct vm_area_struct *vma)
115{
116 unsigned long size = vma->vm_end - vma->vm_start;
117 struct bsr_dev *dev = filp->private_data;
118 int ret;
119
120 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
121
122 /* check for the case of a small BSR device and map one 4k page for it*/
123 if (dev->bsr_len < PAGE_SIZE && size == PAGE_SIZE)
124 ret = remap_4k_pfn(vma, vma->vm_start, dev->bsr_addr >> 12,
125 vma->vm_page_prot);
126 else if (size <= dev->bsr_len)
127 ret = io_remap_pfn_range(vma, vma->vm_start,
128 dev->bsr_addr >> PAGE_SHIFT,
129 size, vma->vm_page_prot);
130 else
131 return -EINVAL;
132
133 if (ret)
134 return -EAGAIN;
135
136 return 0;
137}
138
139static int bsr_open(struct inode *inode, struct file *filp)
140{
141 struct cdev *cdev = inode->i_cdev;
142 struct bsr_dev *dev = container_of(cdev, struct bsr_dev, bsr_cdev);
143
144 filp->private_data = dev;
145 return 0;
146}
147
148static const struct file_operations bsr_fops = {
149 .owner = THIS_MODULE,
150 .mmap = bsr_mmap,
151 .open = bsr_open,
152 .llseek = noop_llseek,
153};
154
155static void bsr_cleanup_devs(void)
156{
157 struct bsr_dev *cur, *n;
158
159 list_for_each_entry_safe(cur, n, &bsr_devs, bsr_list) {
160 if (cur->bsr_device) {
161 cdev_del(&cur->bsr_cdev);
162 device_del(cur->bsr_device);
163 }
164 list_del(&cur->bsr_list);
165 kfree(cur);
166 }
167}
168
169static int bsr_add_node(struct device_node *bn)
170{
171 int bsr_stride_len, bsr_bytes_len, num_bsr_devs;
172 const u32 *bsr_stride;
173 const u32 *bsr_bytes;
174 unsigned i;
175 int ret = -ENODEV;
176
177 bsr_stride = of_get_property(bn, "ibm,lock-stride", &bsr_stride_len);
178 bsr_bytes = of_get_property(bn, "ibm,#lock-bytes", &bsr_bytes_len);
179
180 if (!bsr_stride || !bsr_bytes ||
181 (bsr_stride_len != bsr_bytes_len)) {
182 printk(KERN_ERR "bsr of-node has missing/incorrect property\n");
183 return ret;
184 }
185
186 num_bsr_devs = bsr_bytes_len / sizeof(u32);
187
188 for (i = 0 ; i < num_bsr_devs; i++) {
189 struct bsr_dev *cur = kzalloc(sizeof(struct bsr_dev),
190 GFP_KERNEL);
191 struct resource res;
192 int result;
193
194 if (!cur) {
195 printk(KERN_ERR "Unable to alloc bsr dev\n");
196 ret = -ENOMEM;
197 goto out_err;
198 }
199
200 result = of_address_to_resource(bn, i, &res);
201 if (result < 0) {
202 printk(KERN_ERR "bsr of-node has invalid reg property, skipping\n");
203 kfree(cur);
204 continue;
205 }
206
207 cur->bsr_minor = i + total_bsr_devs;
208 cur->bsr_addr = res.start;
209 cur->bsr_len = resource_size(&res);
210 cur->bsr_bytes = bsr_bytes[i];
211 cur->bsr_stride = bsr_stride[i];
212 cur->bsr_dev = MKDEV(bsr_major, i + total_bsr_devs);
213
214 /* if we have a bsr_len of > 4k and less then PAGE_SIZE (64k pages) */
215 /* we can only map 4k of it, so only advertise the 4k in sysfs */
216 if (cur->bsr_len > 4096 && cur->bsr_len < PAGE_SIZE)
217 cur->bsr_len = 4096;
218
219 switch(cur->bsr_bytes) {
220 case 8:
221 cur->bsr_type = BSR_8;
222 break;
223 case 16:
224 cur->bsr_type = BSR_16;
225 break;
226 case 64:
227 cur->bsr_type = BSR_64;
228 break;
229 case 128:
230 cur->bsr_type = BSR_128;
231 break;
232 case 4096:
233 cur->bsr_type = BSR_4096;
234 break;
235 default:
236 cur->bsr_type = BSR_UNKNOWN;
237 }
238
239 cur->bsr_num = bsr_types[cur->bsr_type];
240 snprintf(cur->bsr_name, 32, "bsr%d_%d",
241 cur->bsr_bytes, cur->bsr_num);
242
243 cdev_init(&cur->bsr_cdev, &bsr_fops);
244 result = cdev_add(&cur->bsr_cdev, cur->bsr_dev, 1);
245 if (result) {
246 kfree(cur);
247 goto out_err;
248 }
249
250 cur->bsr_device = device_create(&bsr_class, NULL, cur->bsr_dev,
251 cur, "%s", cur->bsr_name);
252 if (IS_ERR(cur->bsr_device)) {
253 printk(KERN_ERR "device_create failed for %s\n",
254 cur->bsr_name);
255 cdev_del(&cur->bsr_cdev);
256 kfree(cur);
257 goto out_err;
258 }
259
260 bsr_types[cur->bsr_type] = cur->bsr_num + 1;
261 list_add_tail(&cur->bsr_list, &bsr_devs);
262 }
263
264 total_bsr_devs += num_bsr_devs;
265
266 return 0;
267
268 out_err:
269
270 bsr_cleanup_devs();
271 return ret;
272}
273
274static int bsr_create_devs(struct device_node *bn)
275{
276 int ret;
277
278 while (bn) {
279 ret = bsr_add_node(bn);
280 if (ret) {
281 of_node_put(bn);
282 return ret;
283 }
284 bn = of_find_compatible_node(bn, NULL, "ibm,bsr");
285 }
286 return 0;
287}
288
289static int __init bsr_init(void)
290{
291 struct device_node *np;
292 dev_t bsr_dev;
293 int ret = -ENODEV;
294
295 np = of_find_compatible_node(NULL, NULL, "ibm,bsr");
296 if (!np)
297 goto out_err;
298
299 ret = class_register(&bsr_class);
300 if (ret)
301 goto out_err_1;
302
303 ret = alloc_chrdev_region(&bsr_dev, 0, BSR_MAX_DEVS, "bsr");
304 bsr_major = MAJOR(bsr_dev);
305 if (ret < 0) {
306 printk(KERN_ERR "alloc_chrdev_region() failed for bsr\n");
307 goto out_err_2;
308 }
309
310 ret = bsr_create_devs(np);
311 if (ret < 0) {
312 np = NULL;
313 goto out_err_3;
314 }
315
316 return 0;
317
318 out_err_3:
319 unregister_chrdev_region(bsr_dev, BSR_MAX_DEVS);
320
321 out_err_2:
322 class_unregister(&bsr_class);
323
324 out_err_1:
325 of_node_put(np);
326
327 out_err:
328
329 return ret;
330}
331
332static void __exit bsr_exit(void)
333{
334
335 bsr_cleanup_devs();
336
337 class_unregister(&bsr_class);
338
339 if (bsr_major)
340 unregister_chrdev_region(MKDEV(bsr_major, 0), BSR_MAX_DEVS);
341}
342
343module_init(bsr_init);
344module_exit(bsr_exit);
345MODULE_LICENSE("GPL");
346MODULE_AUTHOR("Sonny Rao <sonnyrao@us.ibm.com>");