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/device.h>
10 #include <linux/kernel.h>
12 #include <linux/of_address.h>
14 #include <linux/module.h>
15 #include <linux/cdev.h>
16 #include <linux/list.h>
18 #include <linux/slab.h>
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.
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
40 /* This is arbitrary number, up to Power6 it's been 17 or fewer */
41 #define BSR_MAX_DEVS (32)
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 */
52 struct list_head bsr_list;
56 struct device *bsr_device;
61 static unsigned total_bsr_devs;
62 static LIST_HEAD(bsr_devs);
75 static unsigned bsr_types[BSR_MAX];
78 bsr_size_show(struct device *dev, struct device_attribute *attr, char *buf)
80 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
81 return sprintf(buf, "%u\n", bsr_dev->bsr_bytes);
83 static DEVICE_ATTR_RO(bsr_size);
86 bsr_stride_show(struct device *dev, struct device_attribute *attr, char *buf)
88 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
89 return sprintf(buf, "%u\n", bsr_dev->bsr_stride);
91 static DEVICE_ATTR_RO(bsr_stride);
94 bsr_length_show(struct device *dev, struct device_attribute *attr, char *buf)
96 struct bsr_dev *bsr_dev = dev_get_drvdata(dev);
97 return sprintf(buf, "%llu\n", bsr_dev->bsr_len);
99 static DEVICE_ATTR_RO(bsr_length);
101 static struct attribute *bsr_dev_attrs[] = {
102 &dev_attr_bsr_size.attr,
103 &dev_attr_bsr_stride.attr,
104 &dev_attr_bsr_length.attr,
107 ATTRIBUTE_GROUPS(bsr_dev);
109 static const struct class bsr_class = {
111 .dev_groups = bsr_dev_groups,
114 static int bsr_mmap(struct file *filp, struct vm_area_struct *vma)
116 unsigned long size = vma->vm_end - vma->vm_start;
117 struct bsr_dev *dev = filp->private_data;
120 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
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,
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);
139 static int bsr_open(struct inode *inode, struct file *filp)
141 struct cdev *cdev = inode->i_cdev;
142 struct bsr_dev *dev = container_of(cdev, struct bsr_dev, bsr_cdev);
144 filp->private_data = dev;
148 static const struct file_operations bsr_fops = {
149 .owner = THIS_MODULE,
152 .llseek = noop_llseek,
155 static void bsr_cleanup_devs(void)
157 struct bsr_dev *cur, *n;
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);
164 list_del(&cur->bsr_list);
169 static int bsr_add_node(struct device_node *bn)
171 int bsr_stride_len, bsr_bytes_len, num_bsr_devs;
172 const u32 *bsr_stride;
173 const u32 *bsr_bytes;
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);
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");
186 num_bsr_devs = bsr_bytes_len / sizeof(u32);
188 for (i = 0 ; i < num_bsr_devs; i++) {
189 struct bsr_dev *cur = kzalloc(sizeof(struct bsr_dev),
195 printk(KERN_ERR "Unable to alloc bsr dev\n");
200 result = of_address_to_resource(bn, i, &res);
202 printk(KERN_ERR "bsr of-node has invalid reg property, skipping\n");
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);
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)
219 switch(cur->bsr_bytes) {
221 cur->bsr_type = BSR_8;
224 cur->bsr_type = BSR_16;
227 cur->bsr_type = BSR_64;
230 cur->bsr_type = BSR_128;
233 cur->bsr_type = BSR_4096;
236 cur->bsr_type = BSR_UNKNOWN;
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);
243 cdev_init(&cur->bsr_cdev, &bsr_fops);
244 result = cdev_add(&cur->bsr_cdev, cur->bsr_dev, 1);
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",
255 cdev_del(&cur->bsr_cdev);
260 bsr_types[cur->bsr_type] = cur->bsr_num + 1;
261 list_add_tail(&cur->bsr_list, &bsr_devs);
264 total_bsr_devs += num_bsr_devs;
274 static int bsr_create_devs(struct device_node *bn)
279 ret = bsr_add_node(bn);
284 bn = of_find_compatible_node(bn, NULL, "ibm,bsr");
289 static int __init bsr_init(void)
291 struct device_node *np;
295 np = of_find_compatible_node(NULL, NULL, "ibm,bsr");
299 ret = class_register(&bsr_class);
303 ret = alloc_chrdev_region(&bsr_dev, 0, BSR_MAX_DEVS, "bsr");
304 bsr_major = MAJOR(bsr_dev);
306 printk(KERN_ERR "alloc_chrdev_region() failed for bsr\n");
310 ret = bsr_create_devs(np);
319 unregister_chrdev_region(bsr_dev, BSR_MAX_DEVS);
322 class_unregister(&bsr_class);
332 static void __exit bsr_exit(void)
337 class_unregister(&bsr_class);
340 unregister_chrdev_region(MKDEV(bsr_major, 0), BSR_MAX_DEVS);
343 module_init(bsr_init);
344 module_exit(bsr_exit);
345 MODULE_LICENSE("GPL");
346 MODULE_AUTHOR("Sonny Rao <sonnyrao@us.ibm.com>");