xref: /illumos-gate/usr/src/uts/i86pc/os/memnode.c (revision a61ed2ce7a86a4d6428f2a83eb4739fae945447e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/systm.h>
27 #include <sys/sysmacros.h>
28 #include <sys/bootconf.h>
29 #include <sys/atomic.h>
30 #include <sys/lgrp.h>
31 #include <sys/memlist.h>
32 #include <sys/memnode.h>
33 #include <sys/platform_module.h>
34 #include <vm/vm_dep.h>
35 
36 int	max_mem_nodes = 1;
37 
38 struct mem_node_conf mem_node_config[MAX_MEM_NODES];
39 int mem_node_pfn_shift;
40 /*
41  * num_memnodes should be updated atomically and always >=
42  * the number of bits in memnodes_mask or the algorithm may fail.
43  */
44 uint16_t num_memnodes;
45 mnodeset_t memnodes_mask; /* assumes 8*(sizeof(mnodeset_t)) >= MAX_MEM_NODES */
46 
47 /*
48  * If set, mem_node_physalign should be a power of two, and
49  * should reflect the minimum address alignment of each node.
50  */
51 uint64_t mem_node_physalign;
52 
53 /*
54  * Platform hooks we will need.
55  */
56 
57 #pragma weak plat_build_mem_nodes
58 #pragma weak plat_slice_add
59 #pragma weak plat_slice_del
60 
61 /*
62  * Adjust the memnode config after a DR operation.
63  *
64  * It is rather tricky to do these updates since we can't
65  * protect the memnode structures with locks, so we must
66  * be mindful of the order in which updates and reads to
67  * these values can occur.
68  */
69 
70 void
71 mem_node_add_slice(pfn_t start, pfn_t end)
72 {
73 	int mnode;
74 	mnodeset_t newmask, oldmask;
75 
76 	/*
77 	 * DR will pass us the first pfn that is allocatable.
78 	 * We need to round down to get the real start of
79 	 * the slice.
80 	 */
81 	if (mem_node_physalign) {
82 		start &= ~(btop(mem_node_physalign) - 1);
83 		end = roundup(end, btop(mem_node_physalign)) - 1;
84 	}
85 
86 	mnode = PFN_2_MEM_NODE(start);
87 	ASSERT(mnode >= 0 && mnode < max_mem_nodes);
88 
89 	if (atomic_cas_32((uint32_t *)&mem_node_config[mnode].exists, 0, 1)) {
90 		/*
91 		 * Add slice to existing node.
92 		 */
93 		if (start < mem_node_config[mnode].physbase)
94 			mem_node_config[mnode].physbase = start;
95 		if (end > mem_node_config[mnode].physmax)
96 			mem_node_config[mnode].physmax = end;
97 	} else {
98 		mem_node_config[mnode].physbase = start;
99 		mem_node_config[mnode].physmax = end;
100 		atomic_inc_16(&num_memnodes);
101 		do {
102 			oldmask = memnodes_mask;
103 			newmask = memnodes_mask | (1ull << mnode);
104 		} while (atomic_cas_64(&memnodes_mask, oldmask, newmask) !=
105 		    oldmask);
106 	}
107 
108 	/*
109 	 * Inform the common lgrp framework about the new memory
110 	 */
111 	lgrp_config(LGRP_CONFIG_MEM_ADD, mnode, MEM_NODE_2_LGRPHAND(mnode));
112 }
113 
114 /*
115  * Remove a PFN range from a memnode.  On some platforms,
116  * the memnode will be created with physbase at the first
117  * allocatable PFN, but later deleted with the MC slice
118  * base address converted to a PFN, in which case we need
119  * to assume physbase and up.
120  */
121 void
122 mem_node_del_slice(pfn_t start, pfn_t end)
123 {
124 	int mnode;
125 	pgcnt_t delta_pgcnt, node_size;
126 	mnodeset_t omask, nmask;
127 
128 	if (mem_node_physalign) {
129 		start &= ~(btop(mem_node_physalign) - 1);
130 		end = roundup(end, btop(mem_node_physalign)) - 1;
131 	}
132 	mnode = PFN_2_MEM_NODE(start);
133 
134 	ASSERT(mnode >= 0 && mnode < max_mem_nodes);
135 	ASSERT(mem_node_config[mnode].exists == 1);
136 
137 	delta_pgcnt = end - start;
138 	node_size = mem_node_config[mnode].physmax -
139 	    mem_node_config[mnode].physbase;
140 
141 	if (node_size > delta_pgcnt) {
142 		/*
143 		 * Subtract the slice from the memnode.
144 		 */
145 		if (start <= mem_node_config[mnode].physbase)
146 			mem_node_config[mnode].physbase = end + 1;
147 		ASSERT(end <= mem_node_config[mnode].physmax);
148 		if (end == mem_node_config[mnode].physmax)
149 			mem_node_config[mnode].physmax = start - 1;
150 	} else {
151 		/*
152 		 * Let the common lgrp framework know this mnode is
153 		 * leaving
154 		 */
155 		lgrp_config(LGRP_CONFIG_MEM_DEL,
156 		    mnode, MEM_NODE_2_LGRPHAND(mnode));
157 
158 		/*
159 		 * Delete the whole node.
160 		 */
161 		ASSERT(MNODE_PGCNT(mnode) == 0);
162 		do {
163 			omask = memnodes_mask;
164 			nmask = omask & ~(1ull << mnode);
165 		} while (atomic_cas_64(&memnodes_mask, omask, nmask) != omask);
166 		atomic_dec_16(&num_memnodes);
167 		mem_node_config[mnode].exists = 0;
168 	}
169 }
170 
171 void
172 mem_node_add_range(pfn_t start, pfn_t end)
173 {
174 	if (&plat_slice_add)
175 		plat_slice_add(start, end);
176 	else
177 		mem_node_add_slice(start, end);
178 }
179 
180 void
181 mem_node_del_range(pfn_t start, pfn_t end)
182 {
183 	if (&plat_slice_del)
184 		plat_slice_del(start, end);
185 	else
186 		mem_node_del_slice(start, end);
187 }
188 
189 void
190 startup_build_mem_nodes(struct memlist *list)
191 {
192 	pfn_t	start, end;
193 
194 	/* LINTED: ASSERT will always true or false */
195 	ASSERT(NBBY * sizeof (mnodeset_t) >= max_mem_nodes);
196 
197 	if (&plat_build_mem_nodes) {
198 		plat_build_mem_nodes(list);
199 	} else {
200 		/*
201 		 * Boot install lists are arranged <addr, len>, ...
202 		 */
203 		while (list) {
204 			start = list->ml_address >> PAGESHIFT;
205 			if (start > physmax)
206 				continue;
207 			end =
208 			    (list->ml_address + list->ml_size - 1) >> PAGESHIFT;
209 			if (end > physmax)
210 				end = physmax;
211 			mem_node_add_range(start, end);
212 			list = list->ml_next;
213 		}
214 		mem_node_physalign = 0;
215 		mem_node_pfn_shift = 0;
216 	}
217 }
218 
219 /*
220  * Allocate an unassigned memnode.
221  */
222 int
223 mem_node_alloc()
224 {
225 	int mnode;
226 	mnodeset_t newmask, oldmask;
227 
228 	/*
229 	 * Find an unused memnode.  Update it atomically to prevent
230 	 * a first time memnode creation race.
231 	 */
232 	for (mnode = 0; mnode < max_mem_nodes; mnode++)
233 		if (atomic_cas_32((uint32_t *)&mem_node_config[mnode].exists,
234 		    0, 1) == 0)
235 			break;
236 
237 	if (mnode >= max_mem_nodes)
238 		panic("Out of free memnodes\n");
239 
240 	mem_node_config[mnode].physbase = (pfn_t)-1l;
241 	mem_node_config[mnode].physmax = 0;
242 	atomic_inc_16(&num_memnodes);
243 	do {
244 		oldmask = memnodes_mask;
245 		newmask = memnodes_mask | (1ull << mnode);
246 	} while (atomic_cas_64(&memnodes_mask, oldmask, newmask) != oldmask);
247 
248 	return (mnode);
249 }
250 
251 /*
252  * Find the intersection between a memnode and a memlist
253  * and returns the number of pages that overlap.
254  *
255  * Assumes the list is protected from DR operations by
256  * the memlist lock.
257  */
258 pgcnt_t
259 mem_node_memlist_pages(int mnode, struct memlist *mlist)
260 {
261 	pfn_t		base, end;
262 	pfn_t		cur_base, cur_end;
263 	pgcnt_t		npgs;
264 	struct memlist	*pmem;
265 
266 	base = mem_node_config[mnode].physbase;
267 	end = mem_node_config[mnode].physmax;
268 	npgs = 0;
269 
270 	memlist_read_lock();
271 
272 	for (pmem = mlist; pmem; pmem = pmem->ml_next) {
273 		cur_base = btop(pmem->ml_address);
274 		cur_end = cur_base + btop(pmem->ml_size) - 1;
275 		if (end < cur_base || base > cur_end)
276 			continue;
277 		npgs = npgs + (MIN(cur_end, end) -
278 		    MAX(cur_base, base)) + 1;
279 	}
280 
281 	memlist_read_unlock();
282 
283 	return (npgs);
284 }
285