1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) IBM Corporation, 2014, 2017
4  * Anton Blanchard, Rashmica Gupta.
5  */
6 
7 #define pr_fmt(fmt) "memtrace: " fmt
8 
9 #include <linux/bitops.h>
10 #include <linux/string.h>
11 #include <linux/memblock.h>
12 #include <linux/init.h>
13 #include <linux/moduleparam.h>
14 #include <linux/fs.h>
15 #include <linux/debugfs.h>
16 #include <linux/slab.h>
17 #include <linux/memory.h>
18 #include <linux/memory_hotplug.h>
19 #include <linux/numa.h>
20 #include <asm/machdep.h>
21 #include <asm/debugfs.h>
22 
23 /* This enables us to keep track of the memory removed from each node. */
24 struct memtrace_entry {
25 	void *mem;
26 	u64 start;
27 	u64 size;
28 	u32 nid;
29 	struct dentry *dir;
30 	char name[16];
31 };
32 
33 static u64 memtrace_size;
34 
35 static struct memtrace_entry *memtrace_array;
36 static unsigned int memtrace_array_nr;
37 
38 
39 static ssize_t memtrace_read(struct file *filp, char __user *ubuf,
40 			     size_t count, loff_t *ppos)
41 {
42 	struct memtrace_entry *ent = filp->private_data;
43 
44 	return simple_read_from_buffer(ubuf, count, ppos, ent->mem, ent->size);
45 }
46 
47 static const struct file_operations memtrace_fops = {
48 	.llseek = default_llseek,
49 	.read	= memtrace_read,
50 	.open	= simple_open,
51 };
52 
53 static int check_memblock_online(struct memory_block *mem, void *arg)
54 {
55 	if (mem->state != MEM_ONLINE)
56 		return -1;
57 
58 	return 0;
59 }
60 
61 static int change_memblock_state(struct memory_block *mem, void *arg)
62 {
63 	unsigned long state = (unsigned long)arg;
64 
65 	mem->state = state;
66 
67 	return 0;
68 }
69 
70 /* called with device_hotplug_lock held */
71 static bool memtrace_offline_pages(u32 nid, u64 start_pfn, u64 nr_pages)
72 {
73 	const unsigned long start = PFN_PHYS(start_pfn);
74 	const unsigned long size = PFN_PHYS(nr_pages);
75 
76 	if (walk_memory_blocks(start, size, NULL, check_memblock_online))
77 		return false;
78 
79 	walk_memory_blocks(start, size, (void *)MEM_GOING_OFFLINE,
80 			   change_memblock_state);
81 
82 	if (offline_pages(start_pfn, nr_pages)) {
83 		walk_memory_blocks(start, size, (void *)MEM_ONLINE,
84 				   change_memblock_state);
85 		return false;
86 	}
87 
88 	walk_memory_blocks(start, size, (void *)MEM_OFFLINE,
89 			   change_memblock_state);
90 
91 
92 	return true;
93 }
94 
95 static u64 memtrace_alloc_node(u32 nid, u64 size)
96 {
97 	u64 start_pfn, end_pfn, nr_pages, pfn;
98 	u64 base_pfn;
99 	u64 bytes = memory_block_size_bytes();
100 
101 	if (!node_spanned_pages(nid))
102 		return 0;
103 
104 	start_pfn = node_start_pfn(nid);
105 	end_pfn = node_end_pfn(nid);
106 	nr_pages = size >> PAGE_SHIFT;
107 
108 	/* Trace memory needs to be aligned to the size */
109 	end_pfn = round_down(end_pfn - nr_pages, nr_pages);
110 
111 	lock_device_hotplug();
112 	for (base_pfn = end_pfn; base_pfn > start_pfn; base_pfn -= nr_pages) {
113 		if (memtrace_offline_pages(nid, base_pfn, nr_pages) == true) {
114 			/*
115 			 * Remove memory in memory block size chunks so that
116 			 * iomem resources are always split to the same size and
117 			 * we never try to remove memory that spans two iomem
118 			 * resources.
119 			 */
120 			end_pfn = base_pfn + nr_pages;
121 			for (pfn = base_pfn; pfn < end_pfn; pfn += bytes>> PAGE_SHIFT) {
122 				__remove_memory(nid, pfn << PAGE_SHIFT, bytes);
123 			}
124 			unlock_device_hotplug();
125 			return base_pfn << PAGE_SHIFT;
126 		}
127 	}
128 	unlock_device_hotplug();
129 
130 	return 0;
131 }
132 
133 static int memtrace_init_regions_runtime(u64 size)
134 {
135 	u32 nid;
136 	u64 m;
137 
138 	memtrace_array = kcalloc(num_online_nodes(),
139 				sizeof(struct memtrace_entry), GFP_KERNEL);
140 	if (!memtrace_array) {
141 		pr_err("Failed to allocate memtrace_array\n");
142 		return -EINVAL;
143 	}
144 
145 	for_each_online_node(nid) {
146 		m = memtrace_alloc_node(nid, size);
147 
148 		/*
149 		 * A node might not have any local memory, so warn but
150 		 * continue on.
151 		 */
152 		if (!m) {
153 			pr_err("Failed to allocate trace memory on node %d\n", nid);
154 			continue;
155 		}
156 
157 		pr_info("Allocated trace memory on node %d at 0x%016llx\n", nid, m);
158 
159 		memtrace_array[memtrace_array_nr].start = m;
160 		memtrace_array[memtrace_array_nr].size = size;
161 		memtrace_array[memtrace_array_nr].nid = nid;
162 		memtrace_array_nr++;
163 	}
164 
165 	return 0;
166 }
167 
168 static struct dentry *memtrace_debugfs_dir;
169 
170 static int memtrace_init_debugfs(void)
171 {
172 	int ret = 0;
173 	int i;
174 
175 	for (i = 0; i < memtrace_array_nr; i++) {
176 		struct dentry *dir;
177 		struct memtrace_entry *ent = &memtrace_array[i];
178 
179 		ent->mem = ioremap(ent->start, ent->size);
180 		/* Warn but continue on */
181 		if (!ent->mem) {
182 			pr_err("Failed to map trace memory at 0x%llx\n",
183 				 ent->start);
184 			ret = -1;
185 			continue;
186 		}
187 
188 		snprintf(ent->name, 16, "%08x", ent->nid);
189 		dir = debugfs_create_dir(ent->name, memtrace_debugfs_dir);
190 
191 		ent->dir = dir;
192 		debugfs_create_file("trace", 0400, dir, ent, &memtrace_fops);
193 		debugfs_create_x64("start", 0400, dir, &ent->start);
194 		debugfs_create_x64("size", 0400, dir, &ent->size);
195 	}
196 
197 	return ret;
198 }
199 
200 static int online_mem_block(struct memory_block *mem, void *arg)
201 {
202 	return device_online(&mem->dev);
203 }
204 
205 /*
206  * Iterate through the chunks of memory we have removed from the kernel
207  * and attempt to add them back to the kernel.
208  */
209 static int memtrace_online(void)
210 {
211 	int i, ret = 0;
212 	struct memtrace_entry *ent;
213 
214 	for (i = memtrace_array_nr - 1; i >= 0; i--) {
215 		ent = &memtrace_array[i];
216 
217 		/* We have onlined this chunk previously */
218 		if (ent->nid == NUMA_NO_NODE)
219 			continue;
220 
221 		/* Remove from io mappings */
222 		if (ent->mem) {
223 			iounmap(ent->mem);
224 			ent->mem = 0;
225 		}
226 
227 		if (add_memory(ent->nid, ent->start, ent->size)) {
228 			pr_err("Failed to add trace memory to node %d\n",
229 				ent->nid);
230 			ret += 1;
231 			continue;
232 		}
233 
234 		lock_device_hotplug();
235 		walk_memory_blocks(ent->start, ent->size, NULL,
236 				   online_mem_block);
237 		unlock_device_hotplug();
238 
239 		/*
240 		 * Memory was added successfully so clean up references to it
241 		 * so on reentry we can tell that this chunk was added.
242 		 */
243 		debugfs_remove_recursive(ent->dir);
244 		pr_info("Added trace memory back to node %d\n", ent->nid);
245 		ent->size = ent->start = ent->nid = NUMA_NO_NODE;
246 	}
247 	if (ret)
248 		return ret;
249 
250 	/* If all chunks of memory were added successfully, reset globals */
251 	kfree(memtrace_array);
252 	memtrace_array = NULL;
253 	memtrace_size = 0;
254 	memtrace_array_nr = 0;
255 	return 0;
256 }
257 
258 static int memtrace_enable_set(void *data, u64 val)
259 {
260 	u64 bytes;
261 
262 	/*
263 	 * Don't attempt to do anything if size isn't aligned to a memory
264 	 * block or equal to zero.
265 	 */
266 	bytes = memory_block_size_bytes();
267 	if (val & (bytes - 1)) {
268 		pr_err("Value must be aligned with 0x%llx\n", bytes);
269 		return -EINVAL;
270 	}
271 
272 	/* Re-add/online previously removed/offlined memory */
273 	if (memtrace_size) {
274 		if (memtrace_online())
275 			return -EAGAIN;
276 	}
277 
278 	if (!val)
279 		return 0;
280 
281 	/* Offline and remove memory */
282 	if (memtrace_init_regions_runtime(val))
283 		return -EINVAL;
284 
285 	if (memtrace_init_debugfs())
286 		return -EINVAL;
287 
288 	memtrace_size = val;
289 
290 	return 0;
291 }
292 
293 static int memtrace_enable_get(void *data, u64 *val)
294 {
295 	*val = memtrace_size;
296 	return 0;
297 }
298 
299 DEFINE_SIMPLE_ATTRIBUTE(memtrace_init_fops, memtrace_enable_get,
300 					memtrace_enable_set, "0x%016llx\n");
301 
302 static int memtrace_init(void)
303 {
304 	memtrace_debugfs_dir = debugfs_create_dir("memtrace",
305 						  powerpc_debugfs_root);
306 
307 	debugfs_create_file("enable", 0600, memtrace_debugfs_dir,
308 			    NULL, &memtrace_init_fops);
309 
310 	return 0;
311 }
312 machine_device_initcall(powernv, memtrace_init);
313