1 // SPDX-License-Identifier: GPL-2.0
5 #include <linux/kernel.h>
6 #include <linux/errno.h>
7 #include <linux/topology.h>
8 #include <linux/memblock.h>
9 #include <linux/bootmem.h>
12 #include "numa_internal.h"
14 static int emu_nid_to_phys
[MAX_NUMNODES
];
15 static char *emu_cmdline __initdata
;
17 void __init
numa_emu_cmdline(char *str
)
22 static int __init
emu_find_memblk_by_nid(int nid
, const struct numa_meminfo
*mi
)
26 for (i
= 0; i
< mi
->nr_blks
; i
++)
27 if (mi
->blk
[i
].nid
== nid
)
32 static u64 __init
mem_hole_size(u64 start
, u64 end
)
34 unsigned long start_pfn
= PFN_UP(start
);
35 unsigned long end_pfn
= PFN_DOWN(end
);
37 if (start_pfn
< end_pfn
)
38 return PFN_PHYS(absent_pages_in_range(start_pfn
, end_pfn
));
43 * Sets up nid to range from @start to @end. The return value is -errno if
44 * something went wrong, 0 otherwise.
46 static int __init
emu_setup_memblk(struct numa_meminfo
*ei
,
47 struct numa_meminfo
*pi
,
48 int nid
, int phys_blk
, u64 size
)
50 struct numa_memblk
*eb
= &ei
->blk
[ei
->nr_blks
];
51 struct numa_memblk
*pb
= &pi
->blk
[phys_blk
];
53 if (ei
->nr_blks
>= NR_NODE_MEMBLKS
) {
54 pr_err("NUMA: Too many emulated memblks, failing emulation\n");
59 eb
->start
= pb
->start
;
60 eb
->end
= pb
->start
+ size
;
63 if (emu_nid_to_phys
[nid
] == NUMA_NO_NODE
)
64 emu_nid_to_phys
[nid
] = nid
;
67 if (pb
->start
>= pb
->end
) {
68 WARN_ON_ONCE(pb
->start
> pb
->end
);
69 numa_remove_memblk_from(phys_blk
, pi
);
72 printk(KERN_INFO
"Faking node %d at [mem %#018Lx-%#018Lx] (%LuMB)\n",
73 nid
, eb
->start
, eb
->end
- 1, (eb
->end
- eb
->start
) >> 20);
78 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
81 * Returns zero on success or negative on error.
83 static int __init
split_nodes_interleave(struct numa_meminfo
*ei
,
84 struct numa_meminfo
*pi
,
85 u64 addr
, u64 max_addr
, int nr_nodes
)
87 nodemask_t physnode_mask
= numa_nodes_parsed
;
95 if (nr_nodes
> MAX_NUMNODES
) {
96 pr_info("numa=fake=%d too large, reducing to %d\n",
97 nr_nodes
, MAX_NUMNODES
);
98 nr_nodes
= MAX_NUMNODES
;
102 * Calculate target node size. x86_32 freaks on __udivdi3() so do
103 * the division in ulong number of pages and convert back.
105 size
= max_addr
- addr
- mem_hole_size(addr
, max_addr
);
106 size
= PFN_PHYS((unsigned long)(size
>> PAGE_SHIFT
) / nr_nodes
);
109 * Calculate the number of big nodes that can be allocated as a result
110 * of consolidating the remainder.
112 big
= ((size
& ~FAKE_NODE_MIN_HASH_MASK
) * nr_nodes
) /
115 size
&= FAKE_NODE_MIN_HASH_MASK
;
117 pr_err("Not enough memory for each node. "
118 "NUMA emulation disabled.\n");
123 * Continue to fill physical nodes with fake nodes until there is no
124 * memory left on any of them.
126 while (nodes_weight(physnode_mask
)) {
127 for_each_node_mask(i
, physnode_mask
) {
128 u64 dma32_end
= PFN_PHYS(MAX_DMA32_PFN
);
129 u64 start
, limit
, end
;
132 phys_blk
= emu_find_memblk_by_nid(i
, pi
);
134 node_clear(i
, physnode_mask
);
137 start
= pi
->blk
[phys_blk
].start
;
138 limit
= pi
->blk
[phys_blk
].end
;
142 end
+= FAKE_NODE_MIN_SIZE
;
145 * Continue to add memory to this fake node if its
146 * non-reserved memory is less than the per-node size.
148 while (end
- start
- mem_hole_size(start
, end
) < size
) {
149 end
+= FAKE_NODE_MIN_SIZE
;
157 * If there won't be at least FAKE_NODE_MIN_SIZE of
158 * non-reserved memory in ZONE_DMA32 for the next node,
159 * this one must extend to the boundary.
161 if (end
< dma32_end
&& dma32_end
- end
-
162 mem_hole_size(end
, dma32_end
) < FAKE_NODE_MIN_SIZE
)
166 * If there won't be enough non-reserved memory for the
167 * next node, this one must extend to the end of the
170 if (limit
- end
- mem_hole_size(end
, limit
) < size
)
173 ret
= emu_setup_memblk(ei
, pi
, nid
++ % nr_nodes
,
175 min(end
, limit
) - start
);
184 * Returns the end address of a node so that there is at least `size' amount of
185 * non-reserved memory or `max_addr' is reached.
187 static u64 __init
find_end_of_node(u64 start
, u64 max_addr
, u64 size
)
189 u64 end
= start
+ size
;
191 while (end
- start
- mem_hole_size(start
, end
) < size
) {
192 end
+= FAKE_NODE_MIN_SIZE
;
193 if (end
> max_addr
) {
202 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
203 * `addr' to `max_addr'.
205 * Returns zero on success or negative on error.
207 static int __init
split_nodes_size_interleave(struct numa_meminfo
*ei
,
208 struct numa_meminfo
*pi
,
209 u64 addr
, u64 max_addr
, u64 size
)
211 nodemask_t physnode_mask
= numa_nodes_parsed
;
219 * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
220 * increased accordingly if the requested size is too small. This
221 * creates a uniform distribution of node sizes across the entire
222 * machine (but not necessarily over physical nodes).
224 min_size
= (max_addr
- addr
- mem_hole_size(addr
, max_addr
)) / MAX_NUMNODES
;
225 min_size
= max(min_size
, FAKE_NODE_MIN_SIZE
);
226 if ((min_size
& FAKE_NODE_MIN_HASH_MASK
) < min_size
)
227 min_size
= (min_size
+ FAKE_NODE_MIN_SIZE
) &
228 FAKE_NODE_MIN_HASH_MASK
;
229 if (size
< min_size
) {
230 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
231 size
>> 20, min_size
>> 20);
234 size
&= FAKE_NODE_MIN_HASH_MASK
;
237 * Fill physical nodes with fake nodes of size until there is no memory
238 * left on any of them.
240 while (nodes_weight(physnode_mask
)) {
241 for_each_node_mask(i
, physnode_mask
) {
242 u64 dma32_end
= PFN_PHYS(MAX_DMA32_PFN
);
243 u64 start
, limit
, end
;
246 phys_blk
= emu_find_memblk_by_nid(i
, pi
);
248 node_clear(i
, physnode_mask
);
251 start
= pi
->blk
[phys_blk
].start
;
252 limit
= pi
->blk
[phys_blk
].end
;
254 end
= find_end_of_node(start
, limit
, size
);
256 * If there won't be at least FAKE_NODE_MIN_SIZE of
257 * non-reserved memory in ZONE_DMA32 for the next node,
258 * this one must extend to the boundary.
260 if (end
< dma32_end
&& dma32_end
- end
-
261 mem_hole_size(end
, dma32_end
) < FAKE_NODE_MIN_SIZE
)
265 * If there won't be enough non-reserved memory for the
266 * next node, this one must extend to the end of the
269 if (limit
- end
- mem_hole_size(end
, limit
) < size
)
272 ret
= emu_setup_memblk(ei
, pi
, nid
++ % MAX_NUMNODES
,
274 min(end
, limit
) - start
);
282 int __init
setup_emu2phys_nid(int *dfl_phys_nid
)
284 int i
, max_emu_nid
= 0;
286 *dfl_phys_nid
= NUMA_NO_NODE
;
287 for (i
= 0; i
< ARRAY_SIZE(emu_nid_to_phys
); i
++) {
288 if (emu_nid_to_phys
[i
] != NUMA_NO_NODE
) {
290 if (*dfl_phys_nid
== NUMA_NO_NODE
)
291 *dfl_phys_nid
= emu_nid_to_phys
[i
];
299 * numa_emulation - Emulate NUMA nodes
300 * @numa_meminfo: NUMA configuration to massage
301 * @numa_dist_cnt: The size of the physical NUMA distance table
303 * Emulate NUMA nodes according to the numa=fake kernel parameter.
304 * @numa_meminfo contains the physical memory configuration and is modified
305 * to reflect the emulated configuration on success. @numa_dist_cnt is
306 * used to determine the size of the physical distance table.
308 * On success, the following modifications are made.
310 * - @numa_meminfo is updated to reflect the emulated nodes.
312 * - __apicid_to_node[] is updated such that APIC IDs are mapped to the
315 * - NUMA distance table is rebuilt to represent distances between emulated
316 * nodes. The distances are determined considering how emulated nodes
317 * are mapped to physical nodes and match the actual distances.
319 * - emu_nid_to_phys[] reflects how emulated nodes are mapped to physical
320 * nodes. This is used by numa_add_cpu() and numa_remove_cpu().
322 * If emulation is not enabled or fails, emu_nid_to_phys[] is filled with
323 * identity mapping and no other modification is made.
325 void __init
numa_emulation(struct numa_meminfo
*numa_meminfo
, int numa_dist_cnt
)
327 static struct numa_meminfo ei __initdata
;
328 static struct numa_meminfo pi __initdata
;
329 const u64 max_addr
= PFN_PHYS(max_pfn
);
330 u8
*phys_dist
= NULL
;
331 size_t phys_size
= numa_dist_cnt
* numa_dist_cnt
* sizeof(phys_dist
[0]);
332 int max_emu_nid
, dfl_phys_nid
;
338 memset(&ei
, 0, sizeof(ei
));
341 for (i
= 0; i
< MAX_NUMNODES
; i
++)
342 emu_nid_to_phys
[i
] = NUMA_NO_NODE
;
345 * If the numa=fake command-line contains a 'M' or 'G', it represents
346 * the fixed node size. Otherwise, if it is just a single number N,
347 * split the system RAM into N fake nodes.
349 if (strchr(emu_cmdline
, 'M') || strchr(emu_cmdline
, 'G')) {
352 size
= memparse(emu_cmdline
, &emu_cmdline
);
353 ret
= split_nodes_size_interleave(&ei
, &pi
, 0, max_addr
, size
);
357 n
= simple_strtoul(emu_cmdline
, &emu_cmdline
, 0);
358 ret
= split_nodes_interleave(&ei
, &pi
, 0, max_addr
, n
);
360 if (*emu_cmdline
== ':')
366 if (numa_cleanup_meminfo(&ei
) < 0) {
367 pr_warning("NUMA: Warning: constructed meminfo invalid, disabling emulation\n");
371 /* copy the physical distance table */
375 phys
= memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped
),
376 phys_size
, PAGE_SIZE
);
378 pr_warning("NUMA: Warning: can't allocate copy of distance table, disabling emulation\n");
381 memblock_reserve(phys
, phys_size
);
382 phys_dist
= __va(phys
);
384 for (i
= 0; i
< numa_dist_cnt
; i
++)
385 for (j
= 0; j
< numa_dist_cnt
; j
++)
386 phys_dist
[i
* numa_dist_cnt
+ j
] =
391 * Determine the max emulated nid and the default phys nid to use
392 * for unmapped nodes.
394 max_emu_nid
= setup_emu2phys_nid(&dfl_phys_nid
);
399 /* Make sure numa_nodes_parsed only contains emulated nodes */
400 nodes_clear(numa_nodes_parsed
);
401 for (i
= 0; i
< ARRAY_SIZE(ei
.blk
); i
++)
402 if (ei
.blk
[i
].start
!= ei
.blk
[i
].end
&&
403 ei
.blk
[i
].nid
!= NUMA_NO_NODE
)
404 node_set(ei
.blk
[i
].nid
, numa_nodes_parsed
);
407 * Transform __apicid_to_node table to use emulated nids by
408 * reverse-mapping phys_nid. The maps should always exist but fall
409 * back to zero just in case.
411 for (i
= 0; i
< ARRAY_SIZE(__apicid_to_node
); i
++) {
412 if (__apicid_to_node
[i
] == NUMA_NO_NODE
)
414 for (j
= 0; j
< ARRAY_SIZE(emu_nid_to_phys
); j
++)
415 if (__apicid_to_node
[i
] == emu_nid_to_phys
[j
])
417 __apicid_to_node
[i
] = j
< ARRAY_SIZE(emu_nid_to_phys
) ? j
: 0;
420 /* make sure all emulated nodes are mapped to a physical node */
421 for (i
= 0; i
< ARRAY_SIZE(emu_nid_to_phys
); i
++)
422 if (emu_nid_to_phys
[i
] == NUMA_NO_NODE
)
423 emu_nid_to_phys
[i
] = dfl_phys_nid
;
425 /* transform distance table */
426 numa_reset_distance();
427 for (i
= 0; i
< max_emu_nid
+ 1; i
++) {
428 for (j
= 0; j
< max_emu_nid
+ 1; j
++) {
429 int physi
= emu_nid_to_phys
[i
];
430 int physj
= emu_nid_to_phys
[j
];
433 if (get_option(&emu_cmdline
, &dist
) == 2)
435 else if (physi
>= numa_dist_cnt
|| physj
>= numa_dist_cnt
)
436 dist
= physi
== physj
?
437 LOCAL_DISTANCE
: REMOTE_DISTANCE
;
439 dist
= phys_dist
[physi
* numa_dist_cnt
+ physj
];
441 numa_set_distance(i
, j
, dist
);
445 /* free the copied physical distance table */
447 memblock_free(__pa(phys_dist
), phys_size
);
451 /* No emulation. Build identity emu_nid_to_phys[] for numa_add_cpu() */
452 for (i
= 0; i
< ARRAY_SIZE(emu_nid_to_phys
); i
++)
453 emu_nid_to_phys
[i
] = i
;
456 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
457 void numa_add_cpu(int cpu
)
461 nid
= early_cpu_to_node(cpu
);
462 BUG_ON(nid
== NUMA_NO_NODE
|| !node_online(nid
));
464 physnid
= emu_nid_to_phys
[nid
];
467 * Map the cpu to each emulated node that is allocated on the physical
468 * node of the cpu's apic id.
470 for_each_online_node(nid
)
471 if (emu_nid_to_phys
[nid
] == physnid
)
472 cpumask_set_cpu(cpu
, node_to_cpumask_map
[nid
]);
475 void numa_remove_cpu(int cpu
)
479 for_each_online_node(i
)
480 cpumask_clear_cpu(cpu
, node_to_cpumask_map
[i
]);
482 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
483 static void numa_set_cpumask(int cpu
, bool enable
)
487 nid
= early_cpu_to_node(cpu
);
488 if (nid
== NUMA_NO_NODE
) {
489 /* early_cpu_to_node() already emits a warning and trace */
493 physnid
= emu_nid_to_phys
[nid
];
495 for_each_online_node(nid
) {
496 if (emu_nid_to_phys
[nid
] != physnid
)
499 debug_cpumask_set_cpu(cpu
, nid
, enable
);
503 void numa_add_cpu(int cpu
)
505 numa_set_cpumask(cpu
, true);
508 void numa_remove_cpu(int cpu
)
510 numa_set_cpumask(cpu
, false);
512 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */