1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* IBM POWER Barrier Synchronization Register Driver
4 * Copyright IBM Corporation 2008
6 * Author: Sonny Rao <sonnyrao@us.ibm.com>
9 #include <linux/kernel.h>
11 #include <linux/of_address.h>
12 #include <linux/of_device.h>
13 #include <linux/of_platform.h>
15 #include <linux/module.h>
16 #include <linux/cdev.h>
17 #include <linux/list.h>
19 #include <linux/slab.h>
23 This driver exposes a special register which can be used for fast
24 synchronization across a large SMP machine. The hardware is exposed
25 as an array of bytes where each process will write to one of the bytes to
26 indicate it has finished the current stage and this update is broadcast to
27 all processors without having to bounce a cacheline between them. In
28 POWER5 and POWER6 there is one of these registers per SMP, but it is
29 presented in two forms; first, it is given as a whole and then as a number
30 of smaller registers which alias to parts of the single whole register.
31 This can potentially allow multiple groups of processes to each have their
32 own private synchronization device.
34 Note that this hardware *must* be written to using *only* single byte writes.
35 It may be read using 1, 2, 4, or 8 byte loads which must be aligned since
36 this region is treated as cache-inhibited processes should also use a
37 full sync before and after writing to the BSR to ensure all stores and
38 the BSR update have made it to all chips in the system
41 /* This is arbitrary number, up to Power6 it's been 17 or fewer */
42 #define BSR_MAX_DEVS (32)
45 u64 bsr_addr; /* Real address */
46 u64 bsr_len; /* length of mem region we can map */
47 unsigned bsr_bytes; /* size of the BSR reg itself */
48 unsigned bsr_stride; /* interval at which BSR repeats in the page */
49 unsigned bsr_type; /* maps to enum below */
50 unsigned bsr_num; /* bsr id number for its type */
53 struct list_head bsr_list;
57 struct device *bsr_device;
62 static unsigned total_bsr_devs;
63 static LIST_HEAD(bsr_devs);
76 static unsigned bsr_types[BSR_MAX];
79 bsr_size_show(struct device *dev, struct device_attribute *attr, char *buf)
81 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
82 return sprintf(buf, "%u\n", bsr_dev->bsr_bytes);
84 static DEVICE_ATTR_RO(bsr_size);
87 bsr_stride_show(struct device *dev, struct device_attribute *attr, char *buf)
89 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
90 return sprintf(buf, "%u\n", bsr_dev->bsr_stride);
92 static DEVICE_ATTR_RO(bsr_stride);
95 bsr_length_show(struct device *dev, struct device_attribute *attr, char *buf)
97 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
98 return sprintf(buf, "%llu\n", bsr_dev->bsr_len);
100 static DEVICE_ATTR_RO(bsr_length);
102 static struct attribute *bsr_dev_attrs[] = {
103 &dev_attr_bsr_size.attr,
104 &dev_attr_bsr_stride.attr,
105 &dev_attr_bsr_length.attr,
108 ATTRIBUTE_GROUPS(bsr_dev);
110 static const struct class bsr_class = {
112 .dev_groups = bsr_dev_groups,
115 static int bsr_mmap(struct file *filp, struct vm_area_struct *vma)
117 unsigned long size = vma->vm_end - vma->vm_start;
118 struct bsr_dev *dev = filp->private_data;
121 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
123 /* check for the case of a small BSR device and map one 4k page for it*/
124 if (dev->bsr_len < PAGE_SIZE && size == PAGE_SIZE)
125 ret = remap_4k_pfn(vma, vma->vm_start, dev->bsr_addr >> 12,
127 else if (size <= dev->bsr_len)
128 ret = io_remap_pfn_range(vma, vma->vm_start,
129 dev->bsr_addr >> PAGE_SHIFT,
130 size, vma->vm_page_prot);
140 static int bsr_open(struct inode *inode, struct file *filp)
142 struct cdev *cdev = inode->i_cdev;
143 struct bsr_dev *dev = container_of(cdev, struct bsr_dev, bsr_cdev);
145 filp->private_data = dev;
149 static const struct file_operations bsr_fops = {
150 .owner = THIS_MODULE,
153 .llseek = noop_llseek,
156 static void bsr_cleanup_devs(void)
158 struct bsr_dev *cur, *n;
160 list_for_each_entry_safe(cur, n, &bsr_devs, bsr_list) {
161 if (cur->bsr_device) {
162 cdev_del(&cur->bsr_cdev);
163 device_del(cur->bsr_device);
165 list_del(&cur->bsr_list);
170 static int bsr_add_node(struct device_node *bn)
172 int bsr_stride_len, bsr_bytes_len, num_bsr_devs;
173 const u32 *bsr_stride;
174 const u32 *bsr_bytes;
178 bsr_stride = of_get_property(bn, "ibm,lock-stride", &bsr_stride_len);
179 bsr_bytes = of_get_property(bn, "ibm,#lock-bytes", &bsr_bytes_len);
181 if (!bsr_stride || !bsr_bytes ||
182 (bsr_stride_len != bsr_bytes_len)) {
183 printk(KERN_ERR "bsr of-node has missing/incorrect property\n");
187 num_bsr_devs = bsr_bytes_len / sizeof(u32);
189 for (i = 0 ; i < num_bsr_devs; i++) {
190 struct bsr_dev *cur = kzalloc(sizeof(struct bsr_dev),
196 printk(KERN_ERR "Unable to alloc bsr dev\n");
201 result = of_address_to_resource(bn, i, &res);
203 printk(KERN_ERR "bsr of-node has invalid reg property, skipping\n");
208 cur->bsr_minor = i + total_bsr_devs;
209 cur->bsr_addr = res.start;
210 cur->bsr_len = resource_size(&res);
211 cur->bsr_bytes = bsr_bytes[i];
212 cur->bsr_stride = bsr_stride[i];
213 cur->bsr_dev = MKDEV(bsr_major, i + total_bsr_devs);
215 /* if we have a bsr_len of > 4k and less then PAGE_SIZE (64k pages) */
216 /* we can only map 4k of it, so only advertise the 4k in sysfs */
217 if (cur->bsr_len > 4096 && cur->bsr_len < PAGE_SIZE)
220 switch(cur->bsr_bytes) {
222 cur->bsr_type = BSR_8;
225 cur->bsr_type = BSR_16;
228 cur->bsr_type = BSR_64;
231 cur->bsr_type = BSR_128;
234 cur->bsr_type = BSR_4096;
237 cur->bsr_type = BSR_UNKNOWN;
240 cur->bsr_num = bsr_types[cur->bsr_type];
241 snprintf(cur->bsr_name, 32, "bsr%d_%d",
242 cur->bsr_bytes, cur->bsr_num);
244 cdev_init(&cur->bsr_cdev, &bsr_fops);
245 result = cdev_add(&cur->bsr_cdev, cur->bsr_dev, 1);
251 cur->bsr_device = device_create(&bsr_class, NULL, cur->bsr_dev,
252 cur, "%s", cur->bsr_name);
253 if (IS_ERR(cur->bsr_device)) {
254 printk(KERN_ERR "device_create failed for %s\n",
256 cdev_del(&cur->bsr_cdev);
261 bsr_types[cur->bsr_type] = cur->bsr_num + 1;
262 list_add_tail(&cur->bsr_list, &bsr_devs);
265 total_bsr_devs += num_bsr_devs;
275 static int bsr_create_devs(struct device_node *bn)
280 ret = bsr_add_node(bn);
285 bn = of_find_compatible_node(bn, NULL, "ibm,bsr");
290 static int __init bsr_init(void)
292 struct device_node *np;
296 np = of_find_compatible_node(NULL, NULL, "ibm,bsr");
300 ret = class_register(&bsr_class);
304 ret = alloc_chrdev_region(&bsr_dev, 0, BSR_MAX_DEVS, "bsr");
305 bsr_major = MAJOR(bsr_dev);
307 printk(KERN_ERR "alloc_chrdev_region() failed for bsr\n");
311 ret = bsr_create_devs(np);
320 unregister_chrdev_region(bsr_dev, BSR_MAX_DEVS);
323 class_unregister(&bsr_class);
333 static void __exit bsr_exit(void)
338 class_unregister(&bsr_class);
341 unregister_chrdev_region(MKDEV(bsr_major, 0), BSR_MAX_DEVS);
344 module_init(bsr_init);
345 module_exit(bsr_exit);
346 MODULE_LICENSE("GPL");
347 MODULE_AUTHOR("Sonny Rao <sonnyrao@us.ibm.com>");