19d5171a8SRashmica Gupta /*
29d5171a8SRashmica Gupta  * Copyright (C) IBM Corporation, 2014, 2017
39d5171a8SRashmica Gupta  * Anton Blanchard, Rashmica Gupta.
49d5171a8SRashmica Gupta  *
59d5171a8SRashmica Gupta  * This program is free software; you can redistribute it and/or modify
69d5171a8SRashmica Gupta  * it under the terms of the GNU General Public License as published by
79d5171a8SRashmica Gupta  * the Free Software Foundation; either version 2 of the License, or
89d5171a8SRashmica Gupta  * (at your option) any later version.
99d5171a8SRashmica Gupta  */
109d5171a8SRashmica Gupta 
119d5171a8SRashmica Gupta #define pr_fmt(fmt) "memtrace: " fmt
129d5171a8SRashmica Gupta 
139d5171a8SRashmica Gupta #include <linux/bitops.h>
149d5171a8SRashmica Gupta #include <linux/string.h>
159d5171a8SRashmica Gupta #include <linux/memblock.h>
169d5171a8SRashmica Gupta #include <linux/init.h>
179d5171a8SRashmica Gupta #include <linux/moduleparam.h>
189d5171a8SRashmica Gupta #include <linux/fs.h>
199d5171a8SRashmica Gupta #include <linux/debugfs.h>
209d5171a8SRashmica Gupta #include <linux/slab.h>
219d5171a8SRashmica Gupta #include <linux/memory.h>
229d5171a8SRashmica Gupta #include <linux/memory_hotplug.h>
239d5171a8SRashmica Gupta #include <asm/machdep.h>
249d5171a8SRashmica Gupta #include <asm/debugfs.h>
259d5171a8SRashmica Gupta 
269d5171a8SRashmica Gupta /* This enables us to keep track of the memory removed from each node. */
279d5171a8SRashmica Gupta struct memtrace_entry {
289d5171a8SRashmica Gupta 	void *mem;
299d5171a8SRashmica Gupta 	u64 start;
309d5171a8SRashmica Gupta 	u64 size;
319d5171a8SRashmica Gupta 	u32 nid;
329d5171a8SRashmica Gupta 	struct dentry *dir;
339d5171a8SRashmica Gupta 	char name[16];
349d5171a8SRashmica Gupta };
359d5171a8SRashmica Gupta 
369d5171a8SRashmica Gupta static u64 memtrace_size;
379d5171a8SRashmica Gupta 
389d5171a8SRashmica Gupta static struct memtrace_entry *memtrace_array;
399d5171a8SRashmica Gupta static unsigned int memtrace_array_nr;
409d5171a8SRashmica Gupta 
419d5171a8SRashmica Gupta 
429d5171a8SRashmica Gupta static ssize_t memtrace_read(struct file *filp, char __user *ubuf,
439d5171a8SRashmica Gupta 			     size_t count, loff_t *ppos)
449d5171a8SRashmica Gupta {
459d5171a8SRashmica Gupta 	struct memtrace_entry *ent = filp->private_data;
469d5171a8SRashmica Gupta 
479d5171a8SRashmica Gupta 	return simple_read_from_buffer(ubuf, count, ppos, ent->mem, ent->size);
489d5171a8SRashmica Gupta }
499d5171a8SRashmica Gupta 
509d5171a8SRashmica Gupta static const struct file_operations memtrace_fops = {
519d5171a8SRashmica Gupta 	.llseek = default_llseek,
529d5171a8SRashmica Gupta 	.read	= memtrace_read,
539d5171a8SRashmica Gupta 	.open	= simple_open,
549d5171a8SRashmica Gupta };
559d5171a8SRashmica Gupta 
569d5171a8SRashmica Gupta static int check_memblock_online(struct memory_block *mem, void *arg)
579d5171a8SRashmica Gupta {
589d5171a8SRashmica Gupta 	if (mem->state != MEM_ONLINE)
599d5171a8SRashmica Gupta 		return -1;
609d5171a8SRashmica Gupta 
619d5171a8SRashmica Gupta 	return 0;
629d5171a8SRashmica Gupta }
639d5171a8SRashmica Gupta 
649d5171a8SRashmica Gupta static int change_memblock_state(struct memory_block *mem, void *arg)
659d5171a8SRashmica Gupta {
669d5171a8SRashmica Gupta 	unsigned long state = (unsigned long)arg;
679d5171a8SRashmica Gupta 
689d5171a8SRashmica Gupta 	mem->state = state;
699d5171a8SRashmica Gupta 
709d5171a8SRashmica Gupta 	return 0;
719d5171a8SRashmica Gupta }
729d5171a8SRashmica Gupta 
739d5171a8SRashmica Gupta static bool memtrace_offline_pages(u32 nid, u64 start_pfn, u64 nr_pages)
749d5171a8SRashmica Gupta {
759d5171a8SRashmica Gupta 	u64 end_pfn = start_pfn + nr_pages - 1;
769d5171a8SRashmica Gupta 
779d5171a8SRashmica Gupta 	if (walk_memory_range(start_pfn, end_pfn, NULL,
789d5171a8SRashmica Gupta 	    check_memblock_online))
799d5171a8SRashmica Gupta 		return false;
809d5171a8SRashmica Gupta 
819d5171a8SRashmica Gupta 	walk_memory_range(start_pfn, end_pfn, (void *)MEM_GOING_OFFLINE,
829d5171a8SRashmica Gupta 			  change_memblock_state);
839d5171a8SRashmica Gupta 
849d5171a8SRashmica Gupta 	if (offline_pages(start_pfn, nr_pages)) {
859d5171a8SRashmica Gupta 		walk_memory_range(start_pfn, end_pfn, (void *)MEM_ONLINE,
869d5171a8SRashmica Gupta 				  change_memblock_state);
879d5171a8SRashmica Gupta 		return false;
889d5171a8SRashmica Gupta 	}
899d5171a8SRashmica Gupta 
909d5171a8SRashmica Gupta 	walk_memory_range(start_pfn, end_pfn, (void *)MEM_OFFLINE,
919d5171a8SRashmica Gupta 			  change_memblock_state);
929d5171a8SRashmica Gupta 
939d5171a8SRashmica Gupta 	lock_device_hotplug();
949d5171a8SRashmica Gupta 	remove_memory(nid, start_pfn << PAGE_SHIFT, nr_pages << PAGE_SHIFT);
959d5171a8SRashmica Gupta 	unlock_device_hotplug();
969d5171a8SRashmica Gupta 
979d5171a8SRashmica Gupta 	return true;
989d5171a8SRashmica Gupta }
999d5171a8SRashmica Gupta 
1009d5171a8SRashmica Gupta static u64 memtrace_alloc_node(u32 nid, u64 size)
1019d5171a8SRashmica Gupta {
1029d5171a8SRashmica Gupta 	u64 start_pfn, end_pfn, nr_pages;
1039d5171a8SRashmica Gupta 	u64 base_pfn;
1049d5171a8SRashmica Gupta 
1058ccb442dSMichael Ellerman 	if (!node_spanned_pages(nid))
1069d5171a8SRashmica Gupta 		return 0;
1079d5171a8SRashmica Gupta 
1089d5171a8SRashmica Gupta 	start_pfn = node_start_pfn(nid);
1099d5171a8SRashmica Gupta 	end_pfn = node_end_pfn(nid);
1109d5171a8SRashmica Gupta 	nr_pages = size >> PAGE_SHIFT;
1119d5171a8SRashmica Gupta 
1129d5171a8SRashmica Gupta 	/* Trace memory needs to be aligned to the size */
1139d5171a8SRashmica Gupta 	end_pfn = round_down(end_pfn - nr_pages, nr_pages);
1149d5171a8SRashmica Gupta 
1159d5171a8SRashmica Gupta 	for (base_pfn = end_pfn; base_pfn > start_pfn; base_pfn -= nr_pages) {
1169d5171a8SRashmica Gupta 		if (memtrace_offline_pages(nid, base_pfn, nr_pages) == true)
1179d5171a8SRashmica Gupta 			return base_pfn << PAGE_SHIFT;
1189d5171a8SRashmica Gupta 	}
1199d5171a8SRashmica Gupta 
1209d5171a8SRashmica Gupta 	return 0;
1219d5171a8SRashmica Gupta }
1229d5171a8SRashmica Gupta 
1239d5171a8SRashmica Gupta static int memtrace_init_regions_runtime(u64 size)
1249d5171a8SRashmica Gupta {
1259d5171a8SRashmica Gupta 	u32 nid;
1269d5171a8SRashmica Gupta 	u64 m;
1279d5171a8SRashmica Gupta 
1289d5171a8SRashmica Gupta 	memtrace_array = kcalloc(num_online_nodes(),
1299d5171a8SRashmica Gupta 				sizeof(struct memtrace_entry), GFP_KERNEL);
1309d5171a8SRashmica Gupta 	if (!memtrace_array) {
1319d5171a8SRashmica Gupta 		pr_err("Failed to allocate memtrace_array\n");
1329d5171a8SRashmica Gupta 		return -EINVAL;
1339d5171a8SRashmica Gupta 	}
1349d5171a8SRashmica Gupta 
1359d5171a8SRashmica Gupta 	for_each_online_node(nid) {
1369d5171a8SRashmica Gupta 		m = memtrace_alloc_node(nid, size);
1379d5171a8SRashmica Gupta 
1389d5171a8SRashmica Gupta 		/*
1399d5171a8SRashmica Gupta 		 * A node might not have any local memory, so warn but
1409d5171a8SRashmica Gupta 		 * continue on.
1419d5171a8SRashmica Gupta 		 */
1429d5171a8SRashmica Gupta 		if (!m) {
1439d5171a8SRashmica Gupta 			pr_err("Failed to allocate trace memory on node %d\n", nid);
1449d5171a8SRashmica Gupta 			continue;
1459d5171a8SRashmica Gupta 		}
1469d5171a8SRashmica Gupta 
1479d5171a8SRashmica Gupta 		pr_info("Allocated trace memory on node %d at 0x%016llx\n", nid, m);
1489d5171a8SRashmica Gupta 
1499d5171a8SRashmica Gupta 		memtrace_array[memtrace_array_nr].start = m;
1509d5171a8SRashmica Gupta 		memtrace_array[memtrace_array_nr].size = size;
1519d5171a8SRashmica Gupta 		memtrace_array[memtrace_array_nr].nid = nid;
1529d5171a8SRashmica Gupta 		memtrace_array_nr++;
1539d5171a8SRashmica Gupta 	}
1549d5171a8SRashmica Gupta 
1559d5171a8SRashmica Gupta 	return 0;
1569d5171a8SRashmica Gupta }
1579d5171a8SRashmica Gupta 
1589d5171a8SRashmica Gupta static struct dentry *memtrace_debugfs_dir;
1599d5171a8SRashmica Gupta 
1609d5171a8SRashmica Gupta static int memtrace_init_debugfs(void)
1619d5171a8SRashmica Gupta {
1629d5171a8SRashmica Gupta 	int ret = 0;
1639d5171a8SRashmica Gupta 	int i;
1649d5171a8SRashmica Gupta 
1659d5171a8SRashmica Gupta 	for (i = 0; i < memtrace_array_nr; i++) {
1669d5171a8SRashmica Gupta 		struct dentry *dir;
1679d5171a8SRashmica Gupta 		struct memtrace_entry *ent = &memtrace_array[i];
1689d5171a8SRashmica Gupta 
1699d5171a8SRashmica Gupta 		ent->mem = ioremap(ent->start, ent->size);
1709d5171a8SRashmica Gupta 		/* Warn but continue on */
1719d5171a8SRashmica Gupta 		if (!ent->mem) {
1729d5171a8SRashmica Gupta 			pr_err("Failed to map trace memory at 0x%llx\n",
1739d5171a8SRashmica Gupta 				 ent->start);
1749d5171a8SRashmica Gupta 			ret = -1;
1759d5171a8SRashmica Gupta 			continue;
1769d5171a8SRashmica Gupta 		}
1779d5171a8SRashmica Gupta 
1789d5171a8SRashmica Gupta 		snprintf(ent->name, 16, "%08x", ent->nid);
1799d5171a8SRashmica Gupta 		dir = debugfs_create_dir(ent->name, memtrace_debugfs_dir);
180d3da701dSRashmica Gupta 		if (!dir) {
181d3da701dSRashmica Gupta 			pr_err("Failed to create debugfs directory for node %d\n",
182d3da701dSRashmica Gupta 				ent->nid);
1839d5171a8SRashmica Gupta 			return -1;
184d3da701dSRashmica Gupta 		}
1859d5171a8SRashmica Gupta 
1869d5171a8SRashmica Gupta 		ent->dir = dir;
1879d5171a8SRashmica Gupta 		debugfs_create_file("trace", 0400, dir, ent, &memtrace_fops);
1889d5171a8SRashmica Gupta 		debugfs_create_x64("start", 0400, dir, &ent->start);
1899d5171a8SRashmica Gupta 		debugfs_create_x64("size", 0400, dir, &ent->size);
1909d5171a8SRashmica Gupta 	}
1919d5171a8SRashmica Gupta 
1929d5171a8SRashmica Gupta 	return ret;
1939d5171a8SRashmica Gupta }
1949d5171a8SRashmica Gupta 
195d3da701dSRashmica Gupta static int online_mem_block(struct memory_block *mem, void *arg)
196d3da701dSRashmica Gupta {
197d3da701dSRashmica Gupta 	return device_online(&mem->dev);
198d3da701dSRashmica Gupta }
199d3da701dSRashmica Gupta 
200d3da701dSRashmica Gupta /*
201d3da701dSRashmica Gupta  * Iterate through the chunks of memory we have removed from the kernel
202d3da701dSRashmica Gupta  * and attempt to add them back to the kernel.
203d3da701dSRashmica Gupta  */
204d3da701dSRashmica Gupta static int memtrace_online(void)
205d3da701dSRashmica Gupta {
206d3da701dSRashmica Gupta 	int i, ret = 0;
207d3da701dSRashmica Gupta 	struct memtrace_entry *ent;
208d3da701dSRashmica Gupta 
209d3da701dSRashmica Gupta 	for (i = memtrace_array_nr - 1; i >= 0; i--) {
210d3da701dSRashmica Gupta 		ent = &memtrace_array[i];
211d3da701dSRashmica Gupta 
212d3da701dSRashmica Gupta 		/* We have onlined this chunk previously */
213d3da701dSRashmica Gupta 		if (ent->nid == -1)
214d3da701dSRashmica Gupta 			continue;
215d3da701dSRashmica Gupta 
216d3da701dSRashmica Gupta 		/* Remove from io mappings */
217d3da701dSRashmica Gupta 		if (ent->mem) {
218d3da701dSRashmica Gupta 			iounmap(ent->mem);
219d3da701dSRashmica Gupta 			ent->mem = 0;
220d3da701dSRashmica Gupta 		}
221d3da701dSRashmica Gupta 
222d3da701dSRashmica Gupta 		if (add_memory(ent->nid, ent->start, ent->size)) {
223d3da701dSRashmica Gupta 			pr_err("Failed to add trace memory to node %d\n",
224d3da701dSRashmica Gupta 				ent->nid);
225d3da701dSRashmica Gupta 			ret += 1;
226d3da701dSRashmica Gupta 			continue;
227d3da701dSRashmica Gupta 		}
228d3da701dSRashmica Gupta 
229d3da701dSRashmica Gupta 		/*
230d3da701dSRashmica Gupta 		 * If kernel isn't compiled with the auto online option
231d3da701dSRashmica Gupta 		 * we need to online the memory ourselves.
232d3da701dSRashmica Gupta 		 */
233d3da701dSRashmica Gupta 		if (!memhp_auto_online) {
234d3da701dSRashmica Gupta 			walk_memory_range(PFN_DOWN(ent->start),
235d3da701dSRashmica Gupta 					  PFN_UP(ent->start + ent->size - 1),
236d3da701dSRashmica Gupta 					  NULL, online_mem_block);
237d3da701dSRashmica Gupta 		}
238d3da701dSRashmica Gupta 
239d3da701dSRashmica Gupta 		/*
240d3da701dSRashmica Gupta 		 * Memory was added successfully so clean up references to it
241d3da701dSRashmica Gupta 		 * so on reentry we can tell that this chunk was added.
242d3da701dSRashmica Gupta 		 */
243d3da701dSRashmica Gupta 		debugfs_remove_recursive(ent->dir);
244d3da701dSRashmica Gupta 		pr_info("Added trace memory back to node %d\n", ent->nid);
245d3da701dSRashmica Gupta 		ent->size = ent->start = ent->nid = -1;
246d3da701dSRashmica Gupta 	}
247d3da701dSRashmica Gupta 	if (ret)
248d3da701dSRashmica Gupta 		return ret;
249d3da701dSRashmica Gupta 
250d3da701dSRashmica Gupta 	/* If all chunks of memory were added successfully, reset globals */
251d3da701dSRashmica Gupta 	kfree(memtrace_array);
252d3da701dSRashmica Gupta 	memtrace_array = NULL;
253d3da701dSRashmica Gupta 	memtrace_size = 0;
254d3da701dSRashmica Gupta 	memtrace_array_nr = 0;
255d3da701dSRashmica Gupta 	return 0;
256d3da701dSRashmica Gupta }
257d3da701dSRashmica Gupta 
2589d5171a8SRashmica Gupta static int memtrace_enable_set(void *data, u64 val)
2599d5171a8SRashmica Gupta {
260d3da701dSRashmica Gupta 	u64 bytes;
261d3da701dSRashmica Gupta 
262d3da701dSRashmica Gupta 	/*
263d3da701dSRashmica Gupta 	 * Don't attempt to do anything if size isn't aligned to a memory
264d3da701dSRashmica Gupta 	 * block or equal to zero.
265d3da701dSRashmica Gupta 	 */
266d3da701dSRashmica Gupta 	bytes = memory_block_size_bytes();
267d3da701dSRashmica Gupta 	if (val & (bytes - 1)) {
268d3da701dSRashmica Gupta 		pr_err("Value must be aligned with 0x%llx\n", bytes);
2699d5171a8SRashmica Gupta 		return -EINVAL;
270d3da701dSRashmica Gupta 	}
271d3da701dSRashmica Gupta 
272d3da701dSRashmica Gupta 	/* Re-add/online previously removed/offlined memory */
273d3da701dSRashmica Gupta 	if (memtrace_size) {
274d3da701dSRashmica Gupta 		if (memtrace_online())
275d3da701dSRashmica Gupta 			return -EAGAIN;
276d3da701dSRashmica Gupta 	}
2779d5171a8SRashmica Gupta 
2789d5171a8SRashmica Gupta 	if (!val)
279d3da701dSRashmica Gupta 		return 0;
2809d5171a8SRashmica Gupta 
281d3da701dSRashmica Gupta 	/* Offline and remove memory */
2829d5171a8SRashmica Gupta 	if (memtrace_init_regions_runtime(val))
2839d5171a8SRashmica Gupta 		return -EINVAL;
2849d5171a8SRashmica Gupta 
2859d5171a8SRashmica Gupta 	if (memtrace_init_debugfs())
2869d5171a8SRashmica Gupta 		return -EINVAL;
2879d5171a8SRashmica Gupta 
2889d5171a8SRashmica Gupta 	memtrace_size = val;
2899d5171a8SRashmica Gupta 
2909d5171a8SRashmica Gupta 	return 0;
2919d5171a8SRashmica Gupta }
2929d5171a8SRashmica Gupta 
2939d5171a8SRashmica Gupta static int memtrace_enable_get(void *data, u64 *val)
2949d5171a8SRashmica Gupta {
2959d5171a8SRashmica Gupta 	*val = memtrace_size;
2969d5171a8SRashmica Gupta 	return 0;
2979d5171a8SRashmica Gupta }
2989d5171a8SRashmica Gupta 
2999d5171a8SRashmica Gupta DEFINE_SIMPLE_ATTRIBUTE(memtrace_init_fops, memtrace_enable_get,
3009d5171a8SRashmica Gupta 					memtrace_enable_set, "0x%016llx\n");
3019d5171a8SRashmica Gupta 
3029d5171a8SRashmica Gupta static int memtrace_init(void)
3039d5171a8SRashmica Gupta {
3049d5171a8SRashmica Gupta 	memtrace_debugfs_dir = debugfs_create_dir("memtrace",
3059d5171a8SRashmica Gupta 						  powerpc_debugfs_root);
3069d5171a8SRashmica Gupta 	if (!memtrace_debugfs_dir)
3079d5171a8SRashmica Gupta 		return -1;
3089d5171a8SRashmica Gupta 
3099d5171a8SRashmica Gupta 	debugfs_create_file("enable", 0600, memtrace_debugfs_dir,
3109d5171a8SRashmica Gupta 			    NULL, &memtrace_init_fops);
3119d5171a8SRashmica Gupta 
3129d5171a8SRashmica Gupta 	return 0;
3139d5171a8SRashmica Gupta }
3149d5171a8SRashmica Gupta machine_device_initcall(powernv, memtrace_init);
315