xref: /titanic_41/usr/src/uts/i86pc/os/memnode.c (revision 16ba0fac26f672b18447f2e17a2f91f14ed3ce40)
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 < max_mem_nodes);
88 
89 	if (cas32((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_add_16(&num_memnodes, 1);
101 		do {
102 			oldmask = memnodes_mask;
103 			newmask = memnodes_mask | (1ull << mnode);
104 		} while (cas64(&memnodes_mask, oldmask, newmask) != oldmask);
105 	}
106 
107 	/*
108 	 * Inform the common lgrp framework about the new memory
109 	 */
110 	lgrp_config(LGRP_CONFIG_MEM_ADD, mnode, MEM_NODE_2_LGRPHAND(mnode));
111 }
112 
113 /*
114  * Remove a PFN range from a memnode.  On some platforms,
115  * the memnode will be created with physbase at the first
116  * allocatable PFN, but later deleted with the MC slice
117  * base address converted to a PFN, in which case we need
118  * to assume physbase and up.
119  */
120 void
121 mem_node_del_slice(pfn_t start, pfn_t end)
122 {
123 	int mnode;
124 	pgcnt_t delta_pgcnt, node_size;
125 	mnodeset_t omask, nmask;
126 
127 	if (mem_node_physalign) {
128 		start &= ~(btop(mem_node_physalign) - 1);
129 		end = roundup(end, btop(mem_node_physalign)) - 1;
130 	}
131 	mnode = PFN_2_MEM_NODE(start);
132 
133 	ASSERT(mnode < max_mem_nodes);
134 	ASSERT(mem_node_config[mnode].exists == 1);
135 
136 	delta_pgcnt = end - start;
137 	node_size = mem_node_config[mnode].physmax -
138 	    mem_node_config[mnode].physbase;
139 
140 	if (node_size > delta_pgcnt) {
141 		/*
142 		 * Subtract the slice from the memnode.
143 		 */
144 		if (start <= mem_node_config[mnode].physbase)
145 			mem_node_config[mnode].physbase = end + 1;
146 		ASSERT(end <= mem_node_config[mnode].physmax);
147 		if (end == mem_node_config[mnode].physmax)
148 			mem_node_config[mnode].physmax = start - 1;
149 	} else {
150 		/*
151 		 * Let the common lgrp framework know this mnode is
152 		 * leaving
153 		 */
154 		lgrp_config(LGRP_CONFIG_MEM_DEL,
155 		    mnode, MEM_NODE_2_LGRPHAND(mnode));
156 
157 		/*
158 		 * Delete the whole node.
159 		 */
160 		ASSERT(MNODE_PGCNT(mnode) == 0);
161 		do {
162 			omask = memnodes_mask;
163 			nmask = omask & ~(1ull << mnode);
164 		} while (cas64(&memnodes_mask, omask, nmask) != omask);
165 		atomic_add_16(&num_memnodes, -1);
166 		mem_node_config[mnode].exists = 0;
167 	}
168 }
169 
170 void
171 mem_node_add_range(pfn_t start, pfn_t end)
172 {
173 	if (&plat_slice_add)
174 		plat_slice_add(start, end);
175 	else
176 		mem_node_add_slice(start, end);
177 }
178 
179 void
180 mem_node_del_range(pfn_t start, pfn_t end)
181 {
182 	if (&plat_slice_del)
183 		plat_slice_del(start, end);
184 	else
185 		mem_node_del_slice(start, end);
186 }
187 
188 void
189 startup_build_mem_nodes(struct memlist *list)
190 {
191 	pfn_t	start, end;
192 
193 	/* LINTED: ASSERT will always true or false */
194 	ASSERT(NBBY * sizeof (mnodeset_t) >= max_mem_nodes);
195 
196 	if (&plat_build_mem_nodes) {
197 		plat_build_mem_nodes(list);
198 	} else {
199 		/*
200 		 * Boot install lists are arranged <addr, len>, ...
201 		 */
202 		while (list) {
203 			start = list->ml_address >> PAGESHIFT;
204 			if (start > physmax)
205 				continue;
206 			end =
207 			    (list->ml_address + list->ml_size - 1) >> PAGESHIFT;
208 			if (end > physmax)
209 				end = physmax;
210 			mem_node_add_range(start, end);
211 			list = list->ml_next;
212 		}
213 		mem_node_physalign = 0;
214 		mem_node_pfn_shift = 0;
215 	}
216 }
217 
218 /*
219  * Allocate an unassigned memnode.
220  */
221 int
222 mem_node_alloc()
223 {
224 	int mnode;
225 	mnodeset_t newmask, oldmask;
226 
227 	/*
228 	 * Find an unused memnode.  Update it atomically to prevent
229 	 * a first time memnode creation race.
230 	 */
231 	for (mnode = 0; mnode < max_mem_nodes; mnode++)
232 		if (cas32((uint32_t *)&mem_node_config[mnode].exists,
233 		    0, 1) == 0)
234 			break;
235 
236 	if (mnode >= max_mem_nodes)
237 		panic("Out of free memnodes\n");
238 
239 	mem_node_config[mnode].physbase = (pfn_t)-1l;
240 	mem_node_config[mnode].physmax = 0;
241 	atomic_add_16(&num_memnodes, 1);
242 	do {
243 		oldmask = memnodes_mask;
244 		newmask = memnodes_mask | (1ull << mnode);
245 	} while (cas64(&memnodes_mask, oldmask, newmask) != oldmask);
246 
247 	return (mnode);
248 }
249 
250 /*
251  * Find the intersection between a memnode and a memlist
252  * and returns the number of pages that overlap.
253  *
254  * Assumes the list is protected from DR operations by
255  * the memlist lock.
256  */
257 pgcnt_t
258 mem_node_memlist_pages(int mnode, struct memlist *mlist)
259 {
260 	pfn_t		base, end;
261 	pfn_t		cur_base, cur_end;
262 	pgcnt_t		npgs;
263 	struct memlist	*pmem;
264 
265 	base = mem_node_config[mnode].physbase;
266 	end = mem_node_config[mnode].physmax;
267 	npgs = 0;
268 
269 	memlist_read_lock();
270 
271 	for (pmem = mlist; pmem; pmem = pmem->ml_next) {
272 		cur_base = btop(pmem->ml_address);
273 		cur_end = cur_base + btop(pmem->ml_size) - 1;
274 		if (end < cur_base || base > cur_end)
275 			continue;
276 		npgs = npgs + (MIN(cur_end, end) -
277 		    MAX(cur_base, base)) + 1;
278 	}
279 
280 	memlist_read_unlock();
281 
282 	return (npgs);
283 }
284