/root/src/xen/xen/common/page_alloc.c
Line | Count | Source (jump to first uncovered line) |
1 | | /****************************************************************************** |
2 | | * page_alloc.c |
3 | | * |
4 | | * Simple buddy heap allocator for Xen. |
5 | | * |
6 | | * Copyright (c) 2002-2004 K A Fraser |
7 | | * Copyright (c) 2006 IBM Ryan Harper <ryanh@us.ibm.com> |
8 | | * |
9 | | * This program is free software; you can redistribute it and/or modify |
10 | | * it under the terms of the GNU General Public License as published by |
11 | | * the Free Software Foundation; either version 2 of the License, or |
12 | | * (at your option) any later version. |
13 | | * |
14 | | * This program is distributed in the hope that it will be useful, |
15 | | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
16 | | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
17 | | * GNU General Public License for more details. |
18 | | * |
19 | | * You should have received a copy of the GNU General Public License |
20 | | * along with this program; If not, see <http://www.gnu.org/licenses/>. |
21 | | */ |
22 | | |
23 | | /* |
24 | | * In general Xen maintains two pools of memory: |
25 | | * |
26 | | * - Xen heap: Memory which is always mapped (i.e accessible by |
27 | | * virtual address), via a permanent and contiguous |
28 | | * "direct mapping". Macros like va() and pa() are valid |
29 | | * for such memory and it is always permissible to stash |
30 | | * pointers to Xen heap memory in data structures etc. |
31 | | * |
32 | | * Xen heap pages are always anonymous (that is, not tied |
33 | | * or accounted to any particular domain). |
34 | | * |
35 | | * - Dom heap: Memory which must be explicitly mapped, usually |
36 | | * transiently with map_domain_page(), in order to be |
37 | | * used. va() and pa() are not valid for such memory. Care |
38 | | * should be taken when stashing pointers to dom heap |
39 | | * pages that those mappings are permanent (e.g. vmap() or |
40 | | * map_domain_page_global()), it is not safe to stash |
41 | | * transient mappings such as those from map_domain_page() |
42 | | * |
43 | | * Dom heap pages are often tied to a particular domain, |
44 | | * but need not be (passing domain==NULL results in an |
45 | | * anonymous dom heap allocation). |
46 | | * |
47 | | * The exact nature of this split is a (sub)arch decision which can |
48 | | * select one of three main variants: |
49 | | * |
50 | | * CONFIG_SEPARATE_XENHEAP=y |
51 | | * |
52 | | * The xen heap is maintained as an entirely separate heap. |
53 | | * |
54 | | * Arch code arranges for some (perhaps small) amount of physical |
55 | | * memory to be covered by a direct mapping and registers that |
56 | | * memory as the Xen heap (via init_xenheap_pages()) and the |
57 | | * remainder as the dom heap. |
58 | | * |
59 | | * This mode of operation is most commonly used by 32-bit arches |
60 | | * where the virtual address space is insufficient to map all RAM. |
61 | | * |
62 | | * CONFIG_SEPARATE_XENHEAP=n W/ DIRECT MAP OF ALL RAM |
63 | | * |
64 | | * All of RAM is covered by a permanent contiguous mapping and there |
65 | | * is only a single heap. |
66 | | * |
67 | | * Memory allocated from the Xen heap is flagged (in |
68 | | * page_info.count_info) with PGC_xen_heap. Memory allocated from |
69 | | * the Dom heap must still be explicitly mapped before use |
70 | | * (e.g. with map_domain_page) in particular in common code. |
71 | | * |
72 | | * xenheap_max_mfn() should not be called by arch code. |
73 | | * |
74 | | * This mode of operation is most commonly used by 64-bit arches |
75 | | * which have sufficient free virtual address space to permanently |
76 | | * map the largest practical amount RAM currently expected on that |
77 | | * arch. |
78 | | * |
79 | | * CONFIG_SEPARATE_XENHEAP=n W/ DIRECT MAP OF ONLY PARTIAL RAM |
80 | | * |
81 | | * There is a single heap, but only the beginning (up to some |
82 | | * threshold) is covered by a permanent contiguous mapping. |
83 | | * |
84 | | * Memory allocated from the Xen heap is allocated from below the |
85 | | * threshold and flagged with PGC_xen_heap. Memory allocated from |
86 | | * the dom heap is allocated from anywhere in the heap (although it |
87 | | * will prefer to allocate from as high as possible to try and keep |
88 | | * Xen heap suitable memory available). |
89 | | * |
90 | | * Arch code must call xenheap_max_mfn() to signal the limit of the |
91 | | * direct mapping. |
92 | | * |
93 | | * This mode of operation is most commonly used by 64-bit arches |
94 | | * which have a restricted amount of virtual address space available |
95 | | * for a direct map (due to e.g. reservations for other purposes) |
96 | | * such that it is not possible to map all of RAM on systems with |
97 | | * the largest practical amount of RAM currently expected on that |
98 | | * arch. |
99 | | * |
100 | | * Boot Allocator |
101 | | * |
102 | | * In addition to the two primary pools (xen heap and dom heap) a |
103 | | * third "boot allocator" is used at start of day. This is a |
104 | | * simplified allocator which can be used. |
105 | | * |
106 | | * Typically all memory which is destined to be dom heap memory |
107 | | * (which is everything in the CONFIG_SEPARATE_XENHEAP=n |
108 | | * configurations) is first allocated to the boot allocator (with |
109 | | * init_boot_pages()) and is then handed over to the main dom heap in |
110 | | * end_boot_allocator(). |
111 | | * |
112 | | * "Contiguous" mappings |
113 | | * |
114 | | * Note that although the above talks about "contiguous" mappings |
115 | | * some architectures implement a scheme ("PDX compression") to |
116 | | * compress unused portions of the machine address space (i.e. large |
117 | | * gaps between distinct banks of memory) in order to avoid creating |
118 | | * enormous frame tables and direct maps which mostly map |
119 | | * nothing. Thus a contiguous mapping may still have distinct |
120 | | * regions within it. |
121 | | */ |
122 | | |
123 | | #include <xen/init.h> |
124 | | #include <xen/types.h> |
125 | | #include <xen/lib.h> |
126 | | #include <xen/sched.h> |
127 | | #include <xen/spinlock.h> |
128 | | #include <xen/mm.h> |
129 | | #include <xen/irq.h> |
130 | | #include <xen/softirq.h> |
131 | | #include <xen/domain_page.h> |
132 | | #include <xen/keyhandler.h> |
133 | | #include <xen/perfc.h> |
134 | | #include <xen/pfn.h> |
135 | | #include <xen/numa.h> |
136 | | #include <xen/nodemask.h> |
137 | | #include <xen/event.h> |
138 | | #include <xen/tmem.h> |
139 | | #include <xen/tmem_xen.h> |
140 | | #include <public/sysctl.h> |
141 | | #include <public/sched.h> |
142 | | #include <asm/page.h> |
143 | | #include <asm/numa.h> |
144 | | #include <asm/flushtlb.h> |
145 | | #ifdef CONFIG_X86 |
146 | | #include <asm/p2m.h> |
147 | | #include <asm/setup.h> /* for highmem_start only */ |
148 | | #else |
149 | | #define p2m_pod_offline_or_broken_hit(pg) 0 |
150 | | #define p2m_pod_offline_or_broken_replace(pg) BUG_ON(pg != NULL) |
151 | | #endif |
152 | | |
153 | | /* |
154 | | * Comma-separated list of hexadecimal page numbers containing bad bytes. |
155 | | * e.g. 'badpage=0x3f45,0x8a321'. |
156 | | */ |
157 | | static char __initdata opt_badpage[100] = ""; |
158 | | string_param("badpage", opt_badpage); |
159 | | |
160 | | /* |
161 | | * no-bootscrub -> Free pages are not zeroed during boot. |
162 | | */ |
163 | | static bool_t opt_bootscrub __initdata = 1; |
164 | | boolean_param("bootscrub", opt_bootscrub); |
165 | | |
166 | | /* |
167 | | * bootscrub_chunk -> Amount of bytes to scrub lockstep on non-SMT CPUs |
168 | | * on all NUMA nodes. |
169 | | */ |
170 | | static unsigned long __initdata opt_bootscrub_chunk = MB(128); |
171 | | size_param("bootscrub_chunk", opt_bootscrub_chunk); |
172 | | |
173 | | #ifdef CONFIG_SCRUB_DEBUG |
174 | | static bool __read_mostly scrub_debug; |
175 | | #else |
176 | | #define scrub_debug false |
177 | | #endif |
178 | | |
179 | | /* |
180 | | * Bit width of the DMA heap -- used to override NUMA-node-first. |
181 | | * allocation strategy, which can otherwise exhaust low memory. |
182 | | */ |
183 | | static unsigned int dma_bitsize; |
184 | | integer_param("dma_bits", dma_bitsize); |
185 | | |
186 | | /* Offlined page list, protected by heap_lock. */ |
187 | | PAGE_LIST_HEAD(page_offlined_list); |
188 | | /* Broken page list, protected by heap_lock. */ |
189 | | PAGE_LIST_HEAD(page_broken_list); |
190 | | |
191 | | /************************* |
192 | | * BOOT-TIME ALLOCATOR |
193 | | */ |
194 | | |
195 | | /* |
196 | | * first_valid_mfn is exported because it is use in ARM specific NUMA |
197 | | * helpers. See comment in asm-arm/numa.h. |
198 | | */ |
199 | | unsigned long first_valid_mfn = ~0UL; |
200 | | |
201 | | static struct bootmem_region { |
202 | | unsigned long s, e; /* MFNs @s through @e-1 inclusive are free */ |
203 | | } *__initdata bootmem_region_list; |
204 | | static unsigned int __initdata nr_bootmem_regions; |
205 | | |
206 | | struct scrub_region { |
207 | | unsigned long offset; |
208 | | unsigned long start; |
209 | | unsigned long per_cpu_sz; |
210 | | unsigned long rem; |
211 | | cpumask_t cpus; |
212 | | }; |
213 | | static struct scrub_region __initdata region[MAX_NUMNODES]; |
214 | | static unsigned long __initdata chunk_size; |
215 | | |
216 | | static void __init bootmem_region_add(unsigned long s, unsigned long e) |
217 | 93 | { |
218 | 93 | unsigned int i; |
219 | 93 | |
220 | 93 | if ( (bootmem_region_list == NULL) && (s < e) ) |
221 | 1 | bootmem_region_list = mfn_to_virt(s++); |
222 | 93 | |
223 | 93 | if ( s >= e ) |
224 | 77 | return; |
225 | 93 | |
226 | 134 | for ( i = 0; i < nr_bootmem_regions; i++ ) |
227 | 120 | if ( s < bootmem_region_list[i].e ) |
228 | 2 | break; |
229 | 16 | |
230 | 16 | BUG_ON((i < nr_bootmem_regions) && (e > bootmem_region_list[i].s)); |
231 | 16 | BUG_ON(nr_bootmem_regions == (PAGE_SIZE / sizeof(struct bootmem_region))); |
232 | 16 | |
233 | 16 | memmove(&bootmem_region_list[i+1], &bootmem_region_list[i], |
234 | 16 | (nr_bootmem_regions - i) * sizeof(*bootmem_region_list)); |
235 | 16 | bootmem_region_list[i] = (struct bootmem_region) { s, e }; |
236 | 16 | nr_bootmem_regions++; |
237 | 16 | } |
238 | | |
239 | | static void __init bootmem_region_zap(unsigned long s, unsigned long e) |
240 | 0 | { |
241 | 0 | unsigned int i; |
242 | 0 |
|
243 | 0 | for ( i = 0; i < nr_bootmem_regions; i++ ) |
244 | 0 | { |
245 | 0 | struct bootmem_region *r = &bootmem_region_list[i]; |
246 | 0 | if ( e <= r->s ) |
247 | 0 | break; |
248 | 0 | if ( s >= r->e ) |
249 | 0 | continue; |
250 | 0 | if ( s <= r->s ) |
251 | 0 | { |
252 | 0 | r->s = min(e, r->e); |
253 | 0 | } |
254 | 0 | else if ( e >= r->e ) |
255 | 0 | { |
256 | 0 | r->e = s; |
257 | 0 | } |
258 | 0 | else |
259 | 0 | { |
260 | 0 | unsigned long _e = r->e; |
261 | 0 | r->e = s; |
262 | 0 | bootmem_region_add(e, _e); |
263 | 0 | } |
264 | 0 | } |
265 | 0 | } |
266 | | |
267 | | void __init init_boot_pages(paddr_t ps, paddr_t pe) |
268 | 23 | { |
269 | 23 | unsigned long bad_spfn, bad_epfn; |
270 | 23 | const char *p; |
271 | 23 | #ifdef CONFIG_X86 |
272 | 23 | const unsigned long *badpage = NULL; |
273 | 23 | unsigned int i, array_size; |
274 | 23 | |
275 | 23 | BUILD_BUG_ON(8 * sizeof(frame_table->u.free.first_dirty) < |
276 | 23 | MAX_ORDER + 1); |
277 | 23 | #endif |
278 | 23 | BUILD_BUG_ON(sizeof(frame_table->u) != sizeof(unsigned long)); |
279 | 23 | |
280 | 23 | ps = round_pgup(ps); |
281 | 23 | pe = round_pgdown(pe); |
282 | 23 | if ( pe <= ps ) |
283 | 8 | return; |
284 | 23 | |
285 | 15 | first_valid_mfn = min_t(unsigned long, ps >> PAGE_SHIFT, first_valid_mfn); |
286 | 15 | |
287 | 15 | bootmem_region_add(ps >> PAGE_SHIFT, pe >> PAGE_SHIFT); |
288 | 15 | |
289 | 15 | #ifdef CONFIG_X86 |
290 | 15 | /* |
291 | 15 | * Here we put platform-specific memory range workarounds, i.e. |
292 | 15 | * memory known to be corrupt or otherwise in need to be reserved on |
293 | 15 | * specific platforms. |
294 | 15 | * We get these certain pages and remove them from memory region list. |
295 | 15 | */ |
296 | 15 | badpage = get_platform_badpages(&array_size); |
297 | 15 | if ( badpage ) |
298 | 0 | { |
299 | 0 | for ( i = 0; i < array_size; i++ ) |
300 | 0 | { |
301 | 0 | bootmem_region_zap(*badpage >> PAGE_SHIFT, |
302 | 0 | (*badpage >> PAGE_SHIFT) + 1); |
303 | 0 | badpage++; |
304 | 0 | } |
305 | 0 | } |
306 | 15 | #endif |
307 | 15 | |
308 | 15 | /* Check new pages against the bad-page list. */ |
309 | 15 | p = opt_badpage; |
310 | 15 | while ( *p != '\0' ) |
311 | 0 | { |
312 | 0 | bad_spfn = simple_strtoul(p, &p, 0); |
313 | 0 | bad_epfn = bad_spfn; |
314 | 0 |
|
315 | 0 | if ( *p == '-' ) |
316 | 0 | { |
317 | 0 | p++; |
318 | 0 | bad_epfn = simple_strtoul(p, &p, 0); |
319 | 0 | if ( bad_epfn < bad_spfn ) |
320 | 0 | bad_epfn = bad_spfn; |
321 | 0 | } |
322 | 0 |
|
323 | 0 | if ( *p == ',' ) |
324 | 0 | p++; |
325 | 0 | else if ( *p != '\0' ) |
326 | 0 | break; |
327 | 0 |
|
328 | 0 | bootmem_region_zap(bad_spfn, bad_epfn+1); |
329 | 0 | } |
330 | 15 | } |
331 | | |
332 | | mfn_t __init alloc_boot_pages(unsigned long nr_pfns, unsigned long pfn_align) |
333 | 78 | { |
334 | 78 | unsigned long pg, _e; |
335 | 78 | unsigned int i = nr_bootmem_regions; |
336 | 78 | |
337 | 78 | BUG_ON(!nr_bootmem_regions); |
338 | 78 | |
339 | 141 | while ( i-- ) |
340 | 141 | { |
341 | 141 | struct bootmem_region *r = &bootmem_region_list[i]; |
342 | 141 | |
343 | 141 | pg = (r->e - nr_pfns) & ~(pfn_align - 1); |
344 | 141 | if ( pg >= r->e || pg < r->s ) |
345 | 63 | continue; |
346 | 141 | |
347 | 141 | #if defined(CONFIG_X86) && !defined(NDEBUG) |
348 | 141 | /* |
349 | 141 | * Filtering pfn_align == 1 since the only allocations using a bigger |
350 | 141 | * alignment are the ones used for setting up the frame table chunks. |
351 | 141 | * Those allocations get remapped anyway, i.e. them not having 1:1 |
352 | 141 | * mappings always accessible is not a problem. |
353 | 141 | */ |
354 | 78 | if ( highmem_start && pfn_align == 1 && |
355 | 0 | r->e > PFN_DOWN(highmem_start) ) |
356 | 0 | { |
357 | 0 | pg = r->s; |
358 | 0 | if ( pg + nr_pfns > PFN_DOWN(highmem_start) ) |
359 | 0 | continue; |
360 | 0 | r->s = pg + nr_pfns; |
361 | 0 | return _mfn(pg); |
362 | 0 | } |
363 | 78 | #endif |
364 | 78 | |
365 | 78 | _e = r->e; |
366 | 78 | r->e = pg; |
367 | 78 | bootmem_region_add(pg + nr_pfns, _e); |
368 | 78 | return _mfn(pg); |
369 | 78 | } |
370 | 78 | |
371 | 0 | BUG(); |
372 | 0 | } |
373 | | |
374 | | |
375 | | |
376 | | /************************* |
377 | | * BINARY BUDDY ALLOCATOR |
378 | | */ |
379 | | |
380 | 43.2k | #define MEMZONE_XEN 0 |
381 | 18.4E | #define NR_ZONES (PADDR_BITS - PAGE_SHIFT + 1) |
382 | | |
383 | 0 | #define bits_to_zone(b) (((b) < (PAGE_SHIFT + 1)) ? 1 : ((b) - PAGE_SHIFT)) |
384 | 4.18M | #define page_to_zone(pg) (is_xen_heap_page(pg) ? MEMZONE_XEN : \ |
385 | 4.18M | (flsl(page_to_mfn(pg)) ? : 1)) |
386 | | |
387 | | typedef struct page_list_head heap_by_zone_and_order_t[NR_ZONES][MAX_ORDER+1]; |
388 | | static heap_by_zone_and_order_t *_heap[MAX_NUMNODES]; |
389 | 8.42M | #define heap(node, zone, order) ((*_heap[node])[zone][order]) |
390 | | |
391 | | static unsigned long node_need_scrub[MAX_NUMNODES]; |
392 | | |
393 | | static unsigned long *avail[MAX_NUMNODES]; |
394 | | static long total_avail_pages; |
395 | | |
396 | | /* TMEM: Reserve a fraction of memory for mid-size (0<order<9) allocations.*/ |
397 | | static long midsize_alloc_zone_pages; |
398 | | #define MIDSIZE_ALLOC_FRAC 128 |
399 | | |
400 | | static DEFINE_SPINLOCK(heap_lock); |
401 | | static long outstanding_claims; /* total outstanding claims by all domains */ |
402 | | |
403 | | unsigned long domain_adjust_tot_pages(struct domain *d, long pages) |
404 | 77 | { |
405 | 77 | long dom_before, dom_after, dom_claimed, sys_before, sys_after; |
406 | 77 | |
407 | 77 | ASSERT(spin_is_locked(&d->page_alloc_lock)); |
408 | 77 | d->tot_pages += pages; |
409 | 77 | |
410 | 77 | /* |
411 | 77 | * can test d->claimed_pages race-free because it can only change |
412 | 77 | * if d->page_alloc_lock and heap_lock are both held, see also |
413 | 77 | * domain_set_outstanding_pages below |
414 | 77 | */ |
415 | 77 | if ( !d->outstanding_pages ) |
416 | 77 | goto out; |
417 | 77 | |
418 | 0 | spin_lock(&heap_lock); |
419 | 0 | /* adjust domain outstanding pages; may not go negative */ |
420 | 0 | dom_before = d->outstanding_pages; |
421 | 0 | dom_after = dom_before - pages; |
422 | 0 | BUG_ON(dom_before < 0); |
423 | 0 | dom_claimed = dom_after < 0 ? 0 : dom_after; |
424 | 0 | d->outstanding_pages = dom_claimed; |
425 | 0 | /* flag accounting bug if system outstanding_claims would go negative */ |
426 | 0 | sys_before = outstanding_claims; |
427 | 0 | sys_after = sys_before - (dom_before - dom_claimed); |
428 | 0 | BUG_ON(sys_after < 0); |
429 | 0 | outstanding_claims = sys_after; |
430 | 0 | spin_unlock(&heap_lock); |
431 | 0 |
|
432 | 77 | out: |
433 | 77 | return d->tot_pages; |
434 | 0 | } |
435 | | |
436 | | int domain_set_outstanding_pages(struct domain *d, unsigned long pages) |
437 | 0 | { |
438 | 0 | int ret = -ENOMEM; |
439 | 0 | unsigned long claim, avail_pages; |
440 | 0 |
|
441 | 0 | /* |
442 | 0 | * take the domain's page_alloc_lock, else all d->tot_page adjustments |
443 | 0 | * must always take the global heap_lock rather than only in the much |
444 | 0 | * rarer case that d->outstanding_pages is non-zero |
445 | 0 | */ |
446 | 0 | spin_lock(&d->page_alloc_lock); |
447 | 0 | spin_lock(&heap_lock); |
448 | 0 |
|
449 | 0 | /* pages==0 means "unset" the claim. */ |
450 | 0 | if ( pages == 0 ) |
451 | 0 | { |
452 | 0 | outstanding_claims -= d->outstanding_pages; |
453 | 0 | d->outstanding_pages = 0; |
454 | 0 | ret = 0; |
455 | 0 | goto out; |
456 | 0 | } |
457 | 0 |
|
458 | 0 | /* only one active claim per domain please */ |
459 | 0 | if ( d->outstanding_pages ) |
460 | 0 | { |
461 | 0 | ret = -EINVAL; |
462 | 0 | goto out; |
463 | 0 | } |
464 | 0 |
|
465 | 0 | /* disallow a claim not exceeding current tot_pages or above max_pages */ |
466 | 0 | if ( (pages <= d->tot_pages) || (pages > d->max_pages) ) |
467 | 0 | { |
468 | 0 | ret = -EINVAL; |
469 | 0 | goto out; |
470 | 0 | } |
471 | 0 |
|
472 | 0 | /* how much memory is available? */ |
473 | 0 | avail_pages = total_avail_pages; |
474 | 0 |
|
475 | 0 | /* Note: The usage of claim means that allocation from a guest *might* |
476 | 0 | * have to come from freeable memory. Using free memory is always better, if |
477 | 0 | * it is available, than using freeable memory. |
478 | 0 | * |
479 | 0 | * But that is OK as once the claim has been made, it still can take minutes |
480 | 0 | * before the claim is fully satisfied. Tmem can make use of the unclaimed |
481 | 0 | * pages during this time (to store ephemeral/freeable pages only, |
482 | 0 | * not persistent pages). |
483 | 0 | */ |
484 | 0 | avail_pages += tmem_freeable_pages(); |
485 | 0 | avail_pages -= outstanding_claims; |
486 | 0 |
|
487 | 0 | /* |
488 | 0 | * Note, if domain has already allocated memory before making a claim |
489 | 0 | * then the claim must take tot_pages into account |
490 | 0 | */ |
491 | 0 | claim = pages - d->tot_pages; |
492 | 0 | if ( claim > avail_pages ) |
493 | 0 | goto out; |
494 | 0 |
|
495 | 0 | /* yay, claim fits in available memory, stake the claim, success! */ |
496 | 0 | d->outstanding_pages = claim; |
497 | 0 | outstanding_claims += d->outstanding_pages; |
498 | 0 | ret = 0; |
499 | 0 |
|
500 | 0 | out: |
501 | 0 | spin_unlock(&heap_lock); |
502 | 0 | spin_unlock(&d->page_alloc_lock); |
503 | 0 | return ret; |
504 | 0 | } |
505 | | |
506 | | void get_outstanding_claims(uint64_t *free_pages, uint64_t *outstanding_pages) |
507 | 0 | { |
508 | 0 | spin_lock(&heap_lock); |
509 | 0 | *outstanding_pages = outstanding_claims; |
510 | 0 | *free_pages = avail_domheap_pages(); |
511 | 0 | spin_unlock(&heap_lock); |
512 | 0 | } |
513 | | |
514 | | static bool_t __read_mostly first_node_initialised; |
515 | | #ifndef CONFIG_SEPARATE_XENHEAP |
516 | | static unsigned int __read_mostly xenheap_bits; |
517 | | #else |
518 | | #define xenheap_bits 0 |
519 | | #endif |
520 | | |
521 | | static unsigned long init_node_heap(int node, unsigned long mfn, |
522 | | unsigned long nr, bool_t *use_tail) |
523 | 1 | { |
524 | 1 | /* First node to be discovered has its heap metadata statically alloced. */ |
525 | 1 | static heap_by_zone_and_order_t _heap_static; |
526 | 1 | static unsigned long avail_static[NR_ZONES]; |
527 | 1 | unsigned long needed = (sizeof(**_heap) + |
528 | 1 | sizeof(**avail) * NR_ZONES + |
529 | 1 | PAGE_SIZE - 1) >> PAGE_SHIFT; |
530 | 1 | int i, j; |
531 | 1 | |
532 | 1 | if ( !first_node_initialised ) |
533 | 1 | { |
534 | 1 | _heap[node] = &_heap_static; |
535 | 1 | avail[node] = avail_static; |
536 | 1 | first_node_initialised = 1; |
537 | 1 | needed = 0; |
538 | 1 | } |
539 | 0 | else if ( *use_tail && nr >= needed && |
540 | 0 | arch_mfn_in_directmap(mfn + nr) && |
541 | 0 | (!xenheap_bits || |
542 | 0 | !((mfn + nr - 1) >> (xenheap_bits - PAGE_SHIFT))) ) |
543 | 0 | { |
544 | 0 | _heap[node] = mfn_to_virt(mfn + nr - needed); |
545 | 0 | avail[node] = mfn_to_virt(mfn + nr - 1) + |
546 | 0 | PAGE_SIZE - sizeof(**avail) * NR_ZONES; |
547 | 0 | } |
548 | 0 | else if ( nr >= needed && |
549 | 0 | arch_mfn_in_directmap(mfn + needed) && |
550 | 0 | (!xenheap_bits || |
551 | 0 | !((mfn + needed - 1) >> (xenheap_bits - PAGE_SHIFT))) ) |
552 | 0 | { |
553 | 0 | _heap[node] = mfn_to_virt(mfn); |
554 | 0 | avail[node] = mfn_to_virt(mfn + needed - 1) + |
555 | 0 | PAGE_SIZE - sizeof(**avail) * NR_ZONES; |
556 | 0 | *use_tail = 0; |
557 | 0 | } |
558 | 0 | else if ( get_order_from_bytes(sizeof(**_heap)) == |
559 | 0 | get_order_from_pages(needed) ) |
560 | 0 | { |
561 | 0 | _heap[node] = alloc_xenheap_pages(get_order_from_pages(needed), 0); |
562 | 0 | BUG_ON(!_heap[node]); |
563 | 0 | avail[node] = (void *)_heap[node] + (needed << PAGE_SHIFT) - |
564 | 0 | sizeof(**avail) * NR_ZONES; |
565 | 0 | needed = 0; |
566 | 0 | } |
567 | 0 | else |
568 | 0 | { |
569 | 0 | _heap[node] = xmalloc(heap_by_zone_and_order_t); |
570 | 0 | avail[node] = xmalloc_array(unsigned long, NR_ZONES); |
571 | 0 | BUG_ON(!_heap[node] || !avail[node]); |
572 | 0 | needed = 0; |
573 | 0 | } |
574 | 1 | |
575 | 1 | memset(avail[node], 0, NR_ZONES * sizeof(long)); |
576 | 1 | |
577 | 42 | for ( i = 0; i < NR_ZONES; i++ ) |
578 | 820 | for ( j = 0; j <= MAX_ORDER; j++ ) |
579 | 779 | INIT_PAGE_LIST_HEAD(&heap(node, i, j)); |
580 | 1 | |
581 | 1 | return needed; |
582 | 1 | } |
583 | | |
584 | | /* Default to 64 MiB */ |
585 | 1 | #define DEFAULT_LOW_MEM_VIRQ (((paddr_t) 64) << 20) |
586 | | #define MAX_LOW_MEM_VIRQ (((paddr_t) 1024) << 20) |
587 | | |
588 | | static paddr_t __read_mostly opt_low_mem_virq = ((paddr_t) -1); |
589 | | size_param("low_mem_virq_limit", opt_low_mem_virq); |
590 | | |
591 | | /* Thresholds to control hysteresis. In pages */ |
592 | | /* When memory grows above this threshold, reset hysteresis. |
593 | | * -1 initially to not reset until at least one virq issued. */ |
594 | | static unsigned long low_mem_virq_high = -1UL; |
595 | | /* Threshold at which we issue virq */ |
596 | | static unsigned long low_mem_virq_th = 0; |
597 | | /* Original threshold after all checks completed */ |
598 | | static unsigned long low_mem_virq_orig = 0; |
599 | | /* Order for current threshold */ |
600 | | static unsigned int low_mem_virq_th_order = 0; |
601 | | |
602 | | /* Perform bootstrapping checks and set bounds */ |
603 | | static void __init setup_low_mem_virq(void) |
604 | 1 | { |
605 | 1 | unsigned int order; |
606 | 1 | paddr_t threshold; |
607 | 1 | bool_t halve; |
608 | 1 | |
609 | 1 | /* If the user specifies zero, then he/she doesn't want this virq |
610 | 1 | * to ever trigger. */ |
611 | 1 | if ( opt_low_mem_virq == 0 ) |
612 | 0 | { |
613 | 0 | low_mem_virq_th = -1UL; |
614 | 0 | return; |
615 | 0 | } |
616 | 1 | |
617 | 1 | /* If the user did not specify a knob, remember that */ |
618 | 1 | halve = (opt_low_mem_virq == ((paddr_t) -1)); |
619 | 1 | threshold = halve ? DEFAULT_LOW_MEM_VIRQ : opt_low_mem_virq; |
620 | 1 | |
621 | 1 | /* Dom0 has already been allocated by now. So check we won't be |
622 | 1 | * complaining immediately with whatever's left of the heap. */ |
623 | 1 | threshold = min(threshold, |
624 | 1 | ((paddr_t) total_avail_pages) << PAGE_SHIFT); |
625 | 1 | |
626 | 1 | /* Then, cap to some predefined maximum */ |
627 | 1 | threshold = min(threshold, MAX_LOW_MEM_VIRQ); |
628 | 1 | |
629 | 1 | /* If the user specified no knob, and we are at the current available |
630 | 1 | * level, halve the threshold. */ |
631 | 1 | if ( halve && |
632 | 1 | (threshold == (((paddr_t) total_avail_pages) << PAGE_SHIFT)) ) |
633 | 0 | threshold >>= 1; |
634 | 1 | |
635 | 1 | /* Zero? Have to fire immediately */ |
636 | 1 | threshold = max(threshold, (paddr_t) PAGE_SIZE); |
637 | 1 | |
638 | 1 | /* Threshold bytes -> pages */ |
639 | 1 | low_mem_virq_th = threshold >> PAGE_SHIFT; |
640 | 1 | |
641 | 1 | /* Next, round the threshold down to the next order */ |
642 | 1 | order = get_order_from_pages(low_mem_virq_th); |
643 | 1 | if ( (1UL << order) > low_mem_virq_th ) |
644 | 0 | order--; |
645 | 1 | |
646 | 1 | /* Set bounds, ready to go */ |
647 | 1 | low_mem_virq_th = low_mem_virq_orig = 1UL << order; |
648 | 1 | low_mem_virq_th_order = order; |
649 | 1 | |
650 | 1 | printk("Initial low memory virq threshold set at %#lx pages.\n", |
651 | 1 | low_mem_virq_th); |
652 | 1 | } |
653 | | |
654 | | static void check_low_mem_virq(void) |
655 | 43.2k | { |
656 | 43.2k | unsigned long avail_pages = total_avail_pages + |
657 | 43.2k | tmem_freeable_pages() - outstanding_claims; |
658 | 43.2k | |
659 | 43.2k | if ( unlikely(avail_pages <= low_mem_virq_th) ) |
660 | 0 | { |
661 | 0 | send_global_virq(VIRQ_ENOMEM); |
662 | 0 |
|
663 | 0 | /* Update thresholds. Next warning will be when we drop below |
664 | 0 | * next order. However, we wait until we grow beyond one |
665 | 0 | * order above us to complain again at the current order */ |
666 | 0 | low_mem_virq_high = 1UL << (low_mem_virq_th_order + 1); |
667 | 0 | if ( low_mem_virq_th_order > 0 ) |
668 | 0 | low_mem_virq_th_order--; |
669 | 0 | low_mem_virq_th = 1UL << low_mem_virq_th_order; |
670 | 0 | return; |
671 | 0 | } |
672 | 43.2k | |
673 | 43.2k | if ( unlikely(avail_pages >= low_mem_virq_high) ) |
674 | 0 | { |
675 | 0 | /* Reset hysteresis. Bring threshold up one order. |
676 | 0 | * If we are back where originally set, set high |
677 | 0 | * threshold to -1 to avoid further growth of |
678 | 0 | * virq threshold. */ |
679 | 0 | low_mem_virq_th_order++; |
680 | 0 | low_mem_virq_th = 1UL << low_mem_virq_th_order; |
681 | 0 | if ( low_mem_virq_th == low_mem_virq_orig ) |
682 | 0 | low_mem_virq_high = -1UL; |
683 | 0 | else |
684 | 0 | low_mem_virq_high = 1UL << (low_mem_virq_th_order + 2); |
685 | 0 | } |
686 | 43.2k | } |
687 | | |
688 | | /* Pages that need a scrub are added to tail, otherwise to head. */ |
689 | | static void page_list_add_scrub(struct page_info *pg, unsigned int node, |
690 | | unsigned int zone, unsigned int order, |
691 | | unsigned int first_dirty) |
692 | 4.18M | { |
693 | 4.18M | PFN_ORDER(pg) = order; |
694 | 4.18M | pg->u.free.first_dirty = first_dirty; |
695 | 4.18M | pg->u.free.scrub_state = BUDDY_NOT_SCRUBBING; |
696 | 4.18M | |
697 | 4.18M | if ( first_dirty != INVALID_DIRTY_IDX ) |
698 | 127 | { |
699 | 127 | ASSERT(first_dirty < (1U << order)); |
700 | 127 | page_list_add_tail(pg, &heap(node, zone, order)); |
701 | 127 | } |
702 | 4.18M | else |
703 | 4.18M | page_list_add(pg, &heap(node, zone, order)); |
704 | 4.18M | } |
705 | | |
706 | | /* SCRUB_PATTERN needs to be a repeating series of bytes. */ |
707 | | #ifndef NDEBUG |
708 | 113 | #define SCRUB_PATTERN 0xc2c2c2c2c2c2c2c2ULL |
709 | | #else |
710 | | #define SCRUB_PATTERN 0ULL |
711 | | #endif |
712 | | #define SCRUB_BYTE_PATTERN (SCRUB_PATTERN & 0xff) |
713 | | |
714 | | static void poison_one_page(struct page_info *pg) |
715 | 228 | { |
716 | 228 | #ifdef CONFIG_SCRUB_DEBUG |
717 | 228 | mfn_t mfn = _mfn(page_to_mfn(pg)); |
718 | 228 | uint64_t *ptr; |
719 | 228 | |
720 | 228 | if ( !scrub_debug ) |
721 | 115 | return; |
722 | 228 | |
723 | 113 | ptr = map_domain_page(mfn); |
724 | 113 | *ptr = ~SCRUB_PATTERN; |
725 | 113 | unmap_domain_page(ptr); |
726 | 113 | #endif |
727 | 113 | } |
728 | | |
729 | | static void check_one_page(struct page_info *pg) |
730 | 0 | { |
731 | 0 | #ifdef CONFIG_SCRUB_DEBUG |
732 | 0 | mfn_t mfn = _mfn(page_to_mfn(pg)); |
733 | 0 | const uint64_t *ptr; |
734 | 0 | unsigned int i; |
735 | 0 |
|
736 | 0 | if ( !scrub_debug ) |
737 | 0 | return; |
738 | 0 |
|
739 | 0 | ptr = map_domain_page(mfn); |
740 | 0 | for ( i = 0; i < PAGE_SIZE / sizeof (*ptr); i++ ) |
741 | 0 | BUG_ON(ptr[i] != SCRUB_PATTERN); |
742 | 0 | unmap_domain_page(ptr); |
743 | 0 | #endif |
744 | 0 | } |
745 | | |
746 | | static void check_and_stop_scrub(struct page_info *head) |
747 | 4.14M | { |
748 | 4.14M | if ( head->u.free.scrub_state == BUDDY_SCRUBBING ) |
749 | 7 | { |
750 | 7 | typeof(head->u.free) pgfree; |
751 | 7 | |
752 | 7 | head->u.free.scrub_state = BUDDY_SCRUB_ABORT; |
753 | 7 | spin_lock_kick(); |
754 | 7 | for ( ; ; ) |
755 | 234 | { |
756 | 234 | /* Can't ACCESS_ONCE() a bitfield. */ |
757 | 234 | pgfree.val = ACCESS_ONCE(head->u.free.val); |
758 | 234 | if ( pgfree.scrub_state != BUDDY_SCRUB_ABORT ) |
759 | 7 | break; |
760 | 227 | cpu_relax(); |
761 | 227 | } |
762 | 7 | } |
763 | 4.14M | } |
764 | | |
765 | | static struct page_info *get_free_buddy(unsigned int zone_lo, |
766 | | unsigned int zone_hi, |
767 | | unsigned int order, unsigned int memflags, |
768 | | const struct domain *d) |
769 | 43.2k | { |
770 | 43.2k | nodeid_t first_node, node = MEMF_get_node(memflags), req_node = node; |
771 | 34.9k | nodemask_t nodemask = d ? d->node_affinity : node_online_map; |
772 | 43.2k | unsigned int j, zone, nodemask_retry = 0; |
773 | 43.2k | struct page_info *pg; |
774 | 43.2k | bool use_unscrubbed = (memflags & MEMF_no_scrub); |
775 | 43.2k | |
776 | 43.2k | if ( node == NUMA_NO_NODE ) |
777 | 43.2k | { |
778 | 43.2k | if ( d != NULL ) |
779 | 34.9k | { |
780 | 34.9k | node = next_node(d->last_alloc_node, nodemask); |
781 | 34.9k | if ( node >= MAX_NUMNODES ) |
782 | 34.6k | node = first_node(nodemask); |
783 | 34.9k | } |
784 | 43.2k | if ( node >= MAX_NUMNODES ) |
785 | 8.27k | node = cpu_to_node(smp_processor_id()); |
786 | 43.2k | } |
787 | 71 | else if ( unlikely(node >= MAX_NUMNODES) ) |
788 | 0 | { |
789 | 0 | ASSERT_UNREACHABLE(); |
790 | 0 | return NULL; |
791 | 0 | } |
792 | 43.2k | first_node = node; |
793 | 43.2k | |
794 | 43.2k | /* |
795 | 43.2k | * Start with requested node, but exhaust all node memory in requested |
796 | 43.2k | * zone before failing, only calc new node value if we fail to find memory |
797 | 43.2k | * in target node, this avoids needless computation on fast-path. |
798 | 43.2k | */ |
799 | 43.2k | for ( ; ; ) |
800 | 59.5k | { |
801 | 59.5k | zone = zone_hi; |
802 | 1.42M | do { |
803 | 1.42M | /* Check if target node can support the allocation. */ |
804 | 1.42M | if ( !avail[node] || (avail[node][zone] < (1UL << order)) ) |
805 | 1.38M | continue; |
806 | 1.42M | |
807 | 1.42M | /* Find smallest order which can satisfy the request. */ |
808 | 86.5k | for ( j = order; j <= MAX_ORDER; j++ ) |
809 | 86.5k | { |
810 | 86.5k | if ( (pg = page_list_remove_head(&heap(node, zone, j))) ) |
811 | 43.2k | { |
812 | 43.2k | if ( pg->u.free.first_dirty == INVALID_DIRTY_IDX ) |
813 | 43.2k | return pg; |
814 | 43.2k | /* |
815 | 43.2k | * We grab single pages (order=0) even if they are |
816 | 43.2k | * unscrubbed. Given that scrubbing one page is fairly quick |
817 | 43.2k | * it is not worth breaking higher orders. |
818 | 43.2k | */ |
819 | 13 | if ( (order == 0) || use_unscrubbed ) |
820 | 13 | { |
821 | 13 | check_and_stop_scrub(pg); |
822 | 13 | return pg; |
823 | 13 | } |
824 | 13 | |
825 | 0 | page_list_add_tail(pg, &heap(node, zone, j)); |
826 | 0 | } |
827 | 86.5k | } |
828 | 1.38M | } while ( zone-- > zone_lo ); /* careful: unsigned zone may wrap */ |
829 | 59.5k | |
830 | 16.2k | if ( (memflags & MEMF_exact_node) && req_node != NUMA_NO_NODE ) |
831 | 0 | return NULL; |
832 | 16.2k | |
833 | 16.2k | /* Pick next node. */ |
834 | 16.2k | if ( !node_isset(node, nodemask) ) |
835 | 0 | { |
836 | 0 | /* Very first node may be caller-specified and outside nodemask. */ |
837 | 0 | ASSERT(!nodemask_retry); |
838 | 0 | first_node = node = first_node(nodemask); |
839 | 0 | if ( node < MAX_NUMNODES ) |
840 | 0 | continue; |
841 | 0 | } |
842 | 16.2k | else if ( (node = next_node(node, nodemask)) >= MAX_NUMNODES ) |
843 | 270 | node = first_node(nodemask); |
844 | 16.2k | if ( node == first_node ) |
845 | 12 | { |
846 | 12 | /* When we have tried all in nodemask, we fall back to others. */ |
847 | 12 | if ( (memflags & MEMF_exact_node) || nodemask_retry++ ) |
848 | 0 | return NULL; |
849 | 12 | nodes_andnot(nodemask, node_online_map, nodemask); |
850 | 12 | first_node = node = first_node(nodemask); |
851 | 12 | if ( node >= MAX_NUMNODES ) |
852 | 12 | return NULL; |
853 | 12 | } |
854 | 16.2k | } |
855 | 43.2k | } |
856 | | |
857 | | /* Allocate 2^@order contiguous pages. */ |
858 | | static struct page_info *alloc_heap_pages( |
859 | | unsigned int zone_lo, unsigned int zone_hi, |
860 | | unsigned int order, unsigned int memflags, |
861 | | struct domain *d) |
862 | 43.2k | { |
863 | 43.2k | nodeid_t node; |
864 | 43.2k | unsigned int i, buddy_order, zone, first_dirty; |
865 | 43.2k | unsigned long request = 1UL << order; |
866 | 43.2k | struct page_info *pg; |
867 | 43.2k | bool need_tlbflush = false; |
868 | 43.2k | uint32_t tlbflush_timestamp = 0; |
869 | 43.2k | unsigned int dirty_cnt = 0; |
870 | 43.2k | |
871 | 43.2k | /* Make sure there are enough bits in memflags for nodeID. */ |
872 | 43.2k | BUILD_BUG_ON((_MEMF_bits - _MEMF_node) < (8 * sizeof(nodeid_t))); |
873 | 43.2k | |
874 | 43.2k | ASSERT(zone_lo <= zone_hi); |
875 | 43.2k | ASSERT(zone_hi < NR_ZONES); |
876 | 43.2k | |
877 | 43.2k | if ( unlikely(order > MAX_ORDER) ) |
878 | 0 | return NULL; |
879 | 43.2k | |
880 | 43.2k | spin_lock(&heap_lock); |
881 | 43.2k | |
882 | 43.2k | /* |
883 | 43.2k | * Claimed memory is considered unavailable unless the request |
884 | 43.2k | * is made by a domain with sufficient unclaimed pages. |
885 | 43.2k | */ |
886 | 43.2k | if ( (outstanding_claims + request > |
887 | 43.2k | total_avail_pages + tmem_freeable_pages()) && |
888 | 0 | ((memflags & MEMF_no_refcount) || |
889 | 0 | !d || d->outstanding_pages < request) ) |
890 | 0 | { |
891 | 0 | spin_unlock(&heap_lock); |
892 | 0 | return NULL; |
893 | 0 | } |
894 | 43.2k | |
895 | 43.2k | /* |
896 | 43.2k | * TMEM: When available memory is scarce due to tmem absorbing it, allow |
897 | 43.2k | * only mid-size allocations to avoid worst of fragmentation issues. |
898 | 43.2k | * Others try tmem pools then fail. This is a workaround until all |
899 | 43.2k | * post-dom0-creation-multi-page allocations can be eliminated. |
900 | 43.2k | */ |
901 | 43.2k | if ( ((order == 0) || (order >= 9)) && |
902 | 43.2k | (total_avail_pages <= midsize_alloc_zone_pages) && |
903 | 0 | tmem_freeable_pages() ) |
904 | 0 | { |
905 | 0 | /* Try to free memory from tmem. */ |
906 | 0 | pg = tmem_relinquish_pages(order, memflags); |
907 | 0 | spin_unlock(&heap_lock); |
908 | 0 | return pg; |
909 | 0 | } |
910 | 43.2k | |
911 | 43.2k | pg = get_free_buddy(zone_lo, zone_hi, order, memflags, d); |
912 | 43.2k | /* Try getting a dirty buddy if we couldn't get a clean one. */ |
913 | 43.2k | if ( !pg && !(memflags & MEMF_no_scrub) ) |
914 | 6 | pg = get_free_buddy(zone_lo, zone_hi, order, |
915 | 6 | memflags | MEMF_no_scrub, d); |
916 | 43.2k | if ( !pg ) |
917 | 6 | { |
918 | 6 | /* No suitable memory blocks. Fail the request. */ |
919 | 6 | spin_unlock(&heap_lock); |
920 | 6 | return NULL; |
921 | 6 | } |
922 | 43.2k | |
923 | 43.2k | node = phys_to_nid(page_to_maddr(pg)); |
924 | 43.2k | zone = page_to_zone(pg); |
925 | 43.2k | buddy_order = PFN_ORDER(pg); |
926 | 43.2k | |
927 | 43.2k | first_dirty = pg->u.free.first_dirty; |
928 | 43.2k | |
929 | 43.2k | /* We may have to halve the chunk a number of times. */ |
930 | 86.4k | while ( buddy_order != order ) |
931 | 43.2k | { |
932 | 43.2k | buddy_order--; |
933 | 43.2k | page_list_add_scrub(pg, node, zone, buddy_order, |
934 | 43.2k | (1U << buddy_order) > first_dirty ? |
935 | 43.2k | first_dirty : INVALID_DIRTY_IDX); |
936 | 43.2k | pg += 1U << buddy_order; |
937 | 43.2k | |
938 | 43.2k | if ( first_dirty != INVALID_DIRTY_IDX ) |
939 | 8 | { |
940 | 8 | /* Adjust first_dirty */ |
941 | 8 | if ( first_dirty >= 1U << buddy_order ) |
942 | 3 | first_dirty -= 1U << buddy_order; |
943 | 8 | else |
944 | 5 | first_dirty = 0; /* We've moved past original first_dirty */ |
945 | 8 | } |
946 | 43.2k | } |
947 | 43.2k | |
948 | 43.2k | ASSERT(avail[node][zone] >= request); |
949 | 43.2k | avail[node][zone] -= request; |
950 | 43.2k | total_avail_pages -= request; |
951 | 43.2k | ASSERT(total_avail_pages >= 0); |
952 | 43.2k | |
953 | 43.2k | check_low_mem_virq(); |
954 | 43.2k | |
955 | 43.2k | if ( d != NULL ) |
956 | 34.9k | d->last_alloc_node = node; |
957 | 43.2k | |
958 | 4.15M | for ( i = 0; i < (1 << order); i++ ) |
959 | 4.11M | { |
960 | 4.11M | /* Reference count must continuously be zero for free pages. */ |
961 | 4.11M | BUG_ON((pg[i].count_info & ~PGC_need_scrub) != PGC_state_free); |
962 | 4.11M | |
963 | 4.11M | /* PGC_need_scrub can only be set if first_dirty is valid */ |
964 | 4.11M | ASSERT(first_dirty != INVALID_DIRTY_IDX || !(pg[i].count_info & PGC_need_scrub)); |
965 | 4.11M | |
966 | 4.11M | /* Preserve PGC_need_scrub so we can check it after lock is dropped. */ |
967 | 4.11M | pg[i].count_info = PGC_state_inuse | (pg[i].count_info & PGC_need_scrub); |
968 | 4.11M | |
969 | 4.11M | if ( !(memflags & MEMF_no_tlbflush) ) |
970 | 4.11M | accumulate_tlbflush(&need_tlbflush, &pg[i], |
971 | 4.11M | &tlbflush_timestamp); |
972 | 4.11M | |
973 | 4.11M | /* Initialise fields which have other uses for free pages. */ |
974 | 4.11M | pg[i].u.inuse.type_info = 0; |
975 | 4.11M | page_set_owner(&pg[i], NULL); |
976 | 4.11M | |
977 | 4.11M | /* Ensure cache and RAM are consistent for platforms where the |
978 | 4.11M | * guest can control its own visibility of/through the cache. |
979 | 4.11M | */ |
980 | 4.11M | flush_page_to_ram(page_to_mfn(&pg[i]), !(memflags & MEMF_no_icache_flush)); |
981 | 4.11M | } |
982 | 43.2k | |
983 | 43.2k | spin_unlock(&heap_lock); |
984 | 43.2k | |
985 | 43.2k | if ( first_dirty != INVALID_DIRTY_IDX || |
986 | 43.2k | (scrub_debug && !(memflags & MEMF_no_scrub)) ) |
987 | 13 | { |
988 | 64 | for ( i = 0; i < (1U << order); i++ ) |
989 | 51 | { |
990 | 51 | if ( test_bit(_PGC_need_scrub, &pg[i].count_info) ) |
991 | 51 | { |
992 | 51 | if ( !(memflags & MEMF_no_scrub) ) |
993 | 2 | scrub_one_page(&pg[i]); |
994 | 51 | |
995 | 51 | dirty_cnt++; |
996 | 51 | |
997 | 51 | spin_lock(&heap_lock); |
998 | 51 | pg[i].count_info &= ~PGC_need_scrub; |
999 | 51 | spin_unlock(&heap_lock); |
1000 | 51 | } |
1001 | 0 | else if ( !(memflags & MEMF_no_scrub) ) |
1002 | 0 | check_one_page(&pg[i]); |
1003 | 51 | } |
1004 | 13 | |
1005 | 13 | if ( dirty_cnt ) |
1006 | 13 | { |
1007 | 13 | spin_lock(&heap_lock); |
1008 | 13 | node_need_scrub[node] -= dirty_cnt; |
1009 | 13 | spin_unlock(&heap_lock); |
1010 | 13 | } |
1011 | 13 | } |
1012 | 43.2k | |
1013 | 43.2k | if ( need_tlbflush ) |
1014 | 1 | filtered_flush_tlb_mask(tlbflush_timestamp); |
1015 | 43.2k | |
1016 | 43.2k | return pg; |
1017 | 43.2k | } |
1018 | | |
1019 | | /* Remove any offlined page in the buddy pointed to by head. */ |
1020 | | static int reserve_offlined_page(struct page_info *head) |
1021 | 0 | { |
1022 | 0 | unsigned int node = phys_to_nid(page_to_maddr(head)); |
1023 | 0 | int zone = page_to_zone(head), i, head_order = PFN_ORDER(head), count = 0; |
1024 | 0 | struct page_info *cur_head; |
1025 | 0 | unsigned int cur_order, first_dirty; |
1026 | 0 |
|
1027 | 0 | ASSERT(spin_is_locked(&heap_lock)); |
1028 | 0 |
|
1029 | 0 | cur_head = head; |
1030 | 0 |
|
1031 | 0 | check_and_stop_scrub(head); |
1032 | 0 | /* |
1033 | 0 | * We may break the buddy so let's mark the head as clean. Then, when |
1034 | 0 | * merging chunks back into the heap, we will see whether the chunk has |
1035 | 0 | * unscrubbed pages and set its first_dirty properly. |
1036 | 0 | */ |
1037 | 0 | first_dirty = head->u.free.first_dirty; |
1038 | 0 | head->u.free.first_dirty = INVALID_DIRTY_IDX; |
1039 | 0 |
|
1040 | 0 | page_list_del(head, &heap(node, zone, head_order)); |
1041 | 0 |
|
1042 | 0 | while ( cur_head < (head + (1 << head_order)) ) |
1043 | 0 | { |
1044 | 0 | struct page_info *pg; |
1045 | 0 | int next_order; |
1046 | 0 |
|
1047 | 0 | if ( page_state_is(cur_head, offlined) ) |
1048 | 0 | { |
1049 | 0 | cur_head++; |
1050 | 0 | if ( first_dirty != INVALID_DIRTY_IDX && first_dirty ) |
1051 | 0 | first_dirty--; |
1052 | 0 | continue; |
1053 | 0 | } |
1054 | 0 |
|
1055 | 0 | next_order = cur_order = 0; |
1056 | 0 |
|
1057 | 0 | while ( cur_order < head_order ) |
1058 | 0 | { |
1059 | 0 | next_order = cur_order + 1; |
1060 | 0 |
|
1061 | 0 | if ( (cur_head + (1 << next_order)) >= (head + ( 1 << head_order)) ) |
1062 | 0 | goto merge; |
1063 | 0 |
|
1064 | 0 | for ( i = (1 << cur_order), pg = cur_head + (1 << cur_order ); |
1065 | 0 | i < (1 << next_order); |
1066 | 0 | i++, pg++ ) |
1067 | 0 | if ( page_state_is(pg, offlined) ) |
1068 | 0 | break; |
1069 | 0 | if ( i == ( 1 << next_order) ) |
1070 | 0 | { |
1071 | 0 | cur_order = next_order; |
1072 | 0 | continue; |
1073 | 0 | } |
1074 | 0 | else |
1075 | 0 | { |
1076 | 0 | merge: |
1077 | 0 | /* We don't consider merging outside the head_order. */ |
1078 | 0 | page_list_add_scrub(cur_head, node, zone, cur_order, |
1079 | 0 | (1U << cur_order) > first_dirty ? |
1080 | 0 | first_dirty : INVALID_DIRTY_IDX); |
1081 | 0 | cur_head += (1 << cur_order); |
1082 | 0 |
|
1083 | 0 | /* Adjust first_dirty if needed. */ |
1084 | 0 | if ( first_dirty != INVALID_DIRTY_IDX ) |
1085 | 0 | { |
1086 | 0 | if ( first_dirty >= 1U << cur_order ) |
1087 | 0 | first_dirty -= 1U << cur_order; |
1088 | 0 | else |
1089 | 0 | first_dirty = 0; |
1090 | 0 | } |
1091 | 0 |
|
1092 | 0 | break; |
1093 | 0 | } |
1094 | 0 | } |
1095 | 0 | } |
1096 | 0 |
|
1097 | 0 | for ( cur_head = head; cur_head < head + ( 1UL << head_order); cur_head++ ) |
1098 | 0 | { |
1099 | 0 | if ( !page_state_is(cur_head, offlined) ) |
1100 | 0 | continue; |
1101 | 0 |
|
1102 | 0 | avail[node][zone]--; |
1103 | 0 | total_avail_pages--; |
1104 | 0 | ASSERT(total_avail_pages >= 0); |
1105 | 0 |
|
1106 | 0 | page_list_add_tail(cur_head, |
1107 | 0 | test_bit(_PGC_broken, &cur_head->count_info) ? |
1108 | 0 | &page_broken_list : &page_offlined_list); |
1109 | 0 |
|
1110 | 0 | count++; |
1111 | 0 | } |
1112 | 0 |
|
1113 | 0 | return count; |
1114 | 0 | } |
1115 | | |
1116 | | static nodemask_t node_scrubbing; |
1117 | | |
1118 | | /* |
1119 | | * If get_node is true this will return closest node that needs to be scrubbed, |
1120 | | * with appropriate bit in node_scrubbing set. |
1121 | | * If get_node is not set, this will return *a* node that needs to be scrubbed. |
1122 | | * node_scrubbing bitmask will no be updated. |
1123 | | * If no node needs scrubbing then NUMA_NO_NODE is returned. |
1124 | | */ |
1125 | | static unsigned int node_to_scrub(bool get_node) |
1126 | 1.97M | { |
1127 | 1.97M | nodeid_t node = cpu_to_node(smp_processor_id()), local_node; |
1128 | 1.97M | nodeid_t closest = NUMA_NO_NODE; |
1129 | 1.97M | u8 dist, shortest = 0xff; |
1130 | 1.97M | |
1131 | 1.97M | if ( node == NUMA_NO_NODE ) |
1132 | 0 | node = 0; |
1133 | 1.97M | |
1134 | 1.97M | if ( node_need_scrub[node] && |
1135 | 27 | (!get_node || !node_test_and_set(node, node_scrubbing)) ) |
1136 | 17 | return node; |
1137 | 1.97M | |
1138 | 1.97M | /* |
1139 | 1.97M | * See if there are memory-only nodes that need scrubbing and choose |
1140 | 1.97M | * the closest one. |
1141 | 1.97M | */ |
1142 | 1.97M | local_node = node; |
1143 | 1.97M | for ( ; ; ) |
1144 | 1.98M | { |
1145 | 1.98M | do { |
1146 | 1.98M | node = cycle_node(node, node_online_map); |
1147 | 1.98M | } while ( !cpumask_empty(&node_to_cpumask(node)) && |
1148 | 2.11M | (node != local_node) ); |
1149 | 1.98M | |
1150 | 1.98M | if ( node == local_node ) |
1151 | 2.12M | break; |
1152 | 1.98M | |
1153 | 18.4E | if ( node_need_scrub[node] ) |
1154 | 0 | { |
1155 | 0 | if ( !get_node ) |
1156 | 0 | return node; |
1157 | 0 |
|
1158 | 0 | dist = __node_distance(local_node, node); |
1159 | 0 |
|
1160 | 0 | /* |
1161 | 0 | * Grab the node right away. If we find a closer node later we will |
1162 | 0 | * release this one. While there is a chance that another CPU will |
1163 | 0 | * not be able to scrub that node when it is searching for scrub work |
1164 | 0 | * at the same time it will be able to do so next time it wakes up. |
1165 | 0 | * The alternative would be to perform this search under a lock but |
1166 | 0 | * then we'd need to take this lock every time we come in here. |
1167 | 0 | */ |
1168 | 0 | if ( (dist < shortest || closest == NUMA_NO_NODE) && |
1169 | 0 | !node_test_and_set(node, node_scrubbing) ) |
1170 | 0 | { |
1171 | 0 | if ( closest != NUMA_NO_NODE ) |
1172 | 0 | node_clear(closest, node_scrubbing); |
1173 | 0 | shortest = dist; |
1174 | 0 | closest = node; |
1175 | 0 | } |
1176 | 0 | } |
1177 | 18.4E | } |
1178 | 1.97M | |
1179 | 1.97M | return closest; |
1180 | 1.97M | } |
1181 | | |
1182 | | struct scrub_wait_state { |
1183 | | struct page_info *pg; |
1184 | | unsigned int first_dirty; |
1185 | | bool drop; |
1186 | | }; |
1187 | | |
1188 | | static void scrub_continue(void *data) |
1189 | 29 | { |
1190 | 29 | struct scrub_wait_state *st = data; |
1191 | 29 | |
1192 | 29 | if ( st->drop ) |
1193 | 18 | return; |
1194 | 29 | |
1195 | 11 | if ( st->pg->u.free.scrub_state == BUDDY_SCRUB_ABORT ) |
1196 | 1 | { |
1197 | 1 | /* There is a waiter for this buddy. Release it. */ |
1198 | 1 | st->drop = true; |
1199 | 1 | st->pg->u.free.first_dirty = st->first_dirty; |
1200 | 1 | smp_wmb(); |
1201 | 1 | st->pg->u.free.scrub_state = BUDDY_NOT_SCRUBBING; |
1202 | 1 | } |
1203 | 11 | } |
1204 | | |
1205 | | bool scrub_free_pages(void) |
1206 | 1.94M | { |
1207 | 1.94M | struct page_info *pg; |
1208 | 1.94M | unsigned int zone; |
1209 | 1.94M | unsigned int cpu = smp_processor_id(); |
1210 | 1.94M | bool preempt = false; |
1211 | 1.94M | nodeid_t node; |
1212 | 1.94M | unsigned int cnt = 0; |
1213 | 1.94M | |
1214 | 1.94M | node = node_to_scrub(true); |
1215 | 1.94M | if ( node == NUMA_NO_NODE ) |
1216 | 2.11M | return false; |
1217 | 1.94M | |
1218 | 18.4E | spin_lock(&heap_lock); |
1219 | 18.4E | |
1220 | 18.4E | for ( zone = 0; zone < NR_ZONES; zone++ ) |
1221 | 216 | { |
1222 | 216 | unsigned int order = MAX_ORDER; |
1223 | 216 | |
1224 | 4.07k | do { |
1225 | 4.08k | while ( !page_list_empty(&heap(node, zone, order)) ) |
1226 | 183 | { |
1227 | 183 | unsigned int i, dirty_cnt; |
1228 | 183 | struct scrub_wait_state st; |
1229 | 183 | |
1230 | 183 | /* Unscrubbed pages are always at the end of the list. */ |
1231 | 183 | pg = page_list_last(&heap(node, zone, order)); |
1232 | 183 | if ( pg->u.free.first_dirty == INVALID_DIRTY_IDX ) |
1233 | 167 | break; |
1234 | 183 | |
1235 | 16 | ASSERT(pg->u.free.scrub_state == BUDDY_NOT_SCRUBBING); |
1236 | 16 | pg->u.free.scrub_state = BUDDY_SCRUBBING; |
1237 | 16 | |
1238 | 16 | spin_unlock(&heap_lock); |
1239 | 16 | |
1240 | 16 | dirty_cnt = 0; |
1241 | 16 | |
1242 | 187 | for ( i = pg->u.free.first_dirty; i < (1U << order); i++) |
1243 | 177 | { |
1244 | 177 | if ( test_bit(_PGC_need_scrub, &pg[i].count_info) ) |
1245 | 177 | { |
1246 | 177 | scrub_one_page(&pg[i]); |
1247 | 177 | /* |
1248 | 177 | * We can modify count_info without holding heap |
1249 | 177 | * lock since we effectively locked this buddy by |
1250 | 177 | * setting its scrub_state. |
1251 | 177 | */ |
1252 | 177 | pg[i].count_info &= ~PGC_need_scrub; |
1253 | 177 | dirty_cnt++; |
1254 | 177 | cnt += 100; /* scrubbed pages add heavier weight. */ |
1255 | 177 | } |
1256 | 177 | else |
1257 | 0 | cnt++; |
1258 | 177 | |
1259 | 177 | if ( pg->u.free.scrub_state == BUDDY_SCRUB_ABORT ) |
1260 | 6 | { |
1261 | 6 | /* Someone wants this chunk. Drop everything. */ |
1262 | 6 | |
1263 | 6 | pg->u.free.first_dirty = (i == (1U << order) - 1) ? |
1264 | 5 | INVALID_DIRTY_IDX : i + 1; |
1265 | 6 | smp_wmb(); |
1266 | 6 | pg->u.free.scrub_state = BUDDY_NOT_SCRUBBING; |
1267 | 6 | |
1268 | 6 | spin_lock(&heap_lock); |
1269 | 6 | node_need_scrub[node] -= dirty_cnt; |
1270 | 6 | spin_unlock(&heap_lock); |
1271 | 6 | goto out_nolock; |
1272 | 6 | } |
1273 | 177 | |
1274 | 177 | /* |
1275 | 177 | * Scrub a few (8) pages before becoming eligible for |
1276 | 177 | * preemption. But also count non-scrubbing loop iterations |
1277 | 177 | * so that we don't get stuck here with an almost clean |
1278 | 177 | * heap. |
1279 | 177 | */ |
1280 | 171 | if ( cnt > 800 && softirq_pending(cpu) ) |
1281 | 0 | { |
1282 | 0 | preempt = true; |
1283 | 0 | break; |
1284 | 0 | } |
1285 | 171 | } |
1286 | 16 | |
1287 | 10 | st.pg = pg; |
1288 | 10 | /* |
1289 | 10 | * get_free_buddy() grabs a buddy with first_dirty set to |
1290 | 10 | * INVALID_DIRTY_IDX so we can't set pg's first_dirty here. |
1291 | 10 | * It will be set either below or in the lock callback (in |
1292 | 10 | * scrub_continue()). |
1293 | 10 | */ |
1294 | 10 | st.first_dirty = (i >= (1U << order) - 1) ? |
1295 | 10 | INVALID_DIRTY_IDX : i + 1; |
1296 | 10 | st.drop = false; |
1297 | 10 | spin_lock_cb(&heap_lock, scrub_continue, &st); |
1298 | 10 | |
1299 | 10 | node_need_scrub[node] -= dirty_cnt; |
1300 | 10 | |
1301 | 10 | if ( st.drop ) |
1302 | 1 | goto out; |
1303 | 10 | |
1304 | 9 | if ( i >= (1U << order) - 1 ) |
1305 | 9 | { |
1306 | 9 | page_list_del(pg, &heap(node, zone, order)); |
1307 | 9 | page_list_add_scrub(pg, node, zone, order, INVALID_DIRTY_IDX); |
1308 | 9 | } |
1309 | 9 | else |
1310 | 0 | pg->u.free.first_dirty = i + 1; |
1311 | 9 | |
1312 | 9 | pg->u.free.scrub_state = BUDDY_NOT_SCRUBBING; |
1313 | 9 | |
1314 | 9 | if ( preempt || (node_need_scrub[node] == 0) ) |
1315 | 3 | goto out; |
1316 | 9 | } |
1317 | 4.06k | } while ( order-- != 0 ); |
1318 | 216 | } |
1319 | 18.4E | |
1320 | 4 | out: |
1321 | 4 | spin_unlock(&heap_lock); |
1322 | 4 | |
1323 | 10 | out_nolock: |
1324 | 10 | node_clear(node, node_scrubbing); |
1325 | 10 | return node_to_scrub(false) != NUMA_NO_NODE; |
1326 | 4 | } |
1327 | | |
1328 | | /* Free 2^@order set of pages. */ |
1329 | | static void free_heap_pages( |
1330 | | struct page_info *pg, unsigned int order, bool need_scrub) |
1331 | 4.14M | { |
1332 | 4.14M | unsigned long mask, mfn = page_to_mfn(pg); |
1333 | 4.14M | unsigned int i, node = phys_to_nid(page_to_maddr(pg)), tainted = 0; |
1334 | 4.14M | unsigned int zone = page_to_zone(pg); |
1335 | 4.14M | |
1336 | 4.14M | ASSERT(order <= MAX_ORDER); |
1337 | 4.14M | ASSERT(node >= 0); |
1338 | 4.14M | |
1339 | 4.14M | spin_lock(&heap_lock); |
1340 | 4.14M | |
1341 | 8.29M | for ( i = 0; i < (1 << order); i++ ) |
1342 | 4.14M | { |
1343 | 4.14M | /* |
1344 | 4.14M | * Cannot assume that count_info == 0, as there are some corner cases |
1345 | 4.14M | * where it isn't the case and yet it isn't a bug: |
1346 | 4.14M | * 1. page_get_owner() is NULL |
1347 | 4.14M | * 2. page_get_owner() is a domain that was never accessible by |
1348 | 4.14M | * its domid (e.g., failed to fully construct the domain). |
1349 | 4.14M | * 3. page was never addressable by the guest (e.g., it's an |
1350 | 4.14M | * auto-translate-physmap guest and the page was never included |
1351 | 4.14M | * in its pseudophysical address space). |
1352 | 4.14M | * In all the above cases there can be no guest mappings of this page. |
1353 | 4.14M | */ |
1354 | 4.14M | ASSERT(!page_state_is(&pg[i], offlined)); |
1355 | 4.14M | pg[i].count_info = |
1356 | 4.14M | ((pg[i].count_info & PGC_broken) | |
1357 | 4.14M | (page_state_is(&pg[i], offlining) |
1358 | 4.14M | ? PGC_state_offlined : PGC_state_free)); |
1359 | 4.14M | if ( page_state_is(&pg[i], offlined) ) |
1360 | 0 | tainted = 1; |
1361 | 4.14M | |
1362 | 4.14M | /* If a page has no owner it will need no safety TLB flush. */ |
1363 | 4.14M | pg[i].u.free.need_tlbflush = (page_get_owner(&pg[i]) != NULL); |
1364 | 4.14M | if ( pg[i].u.free.need_tlbflush ) |
1365 | 1 | page_set_tlbflush_timestamp(&pg[i]); |
1366 | 4.14M | |
1367 | 4.14M | /* This page is not a guest frame any more. */ |
1368 | 4.14M | page_set_owner(&pg[i], NULL); /* set_gpfn_from_mfn snoops pg owner */ |
1369 | 4.14M | set_gpfn_from_mfn(mfn + i, INVALID_M2P_ENTRY); |
1370 | 4.14M | |
1371 | 4.14M | if ( need_scrub ) |
1372 | 228 | { |
1373 | 228 | pg[i].count_info |= PGC_need_scrub; |
1374 | 228 | poison_one_page(&pg[i]); |
1375 | 228 | } |
1376 | 4.14M | } |
1377 | 4.14M | |
1378 | 4.14M | avail[node][zone] += 1 << order; |
1379 | 4.14M | total_avail_pages += 1 << order; |
1380 | 4.14M | if ( need_scrub ) |
1381 | 122 | { |
1382 | 122 | node_need_scrub[node] += 1 << order; |
1383 | 122 | pg->u.free.first_dirty = 0; |
1384 | 122 | } |
1385 | 4.14M | else |
1386 | 4.14M | pg->u.free.first_dirty = INVALID_DIRTY_IDX; |
1387 | 4.14M | |
1388 | 4.14M | if ( tmem_enabled() ) |
1389 | 0 | midsize_alloc_zone_pages = max( |
1390 | 4.14M | midsize_alloc_zone_pages, total_avail_pages / MIDSIZE_ALLOC_FRAC); |
1391 | 4.14M | |
1392 | 4.14M | /* Merge chunks as far as possible. */ |
1393 | 8.29M | while ( order < MAX_ORDER ) |
1394 | 8.29M | { |
1395 | 8.29M | mask = 1UL << order; |
1396 | 8.29M | |
1397 | 8.29M | if ( (page_to_mfn(pg) & mask) ) |
1398 | 4.14M | { |
1399 | 4.14M | struct page_info *predecessor = pg - mask; |
1400 | 4.14M | |
1401 | 4.14M | /* Merge with predecessor block? */ |
1402 | 4.14M | if ( !mfn_valid(_mfn(page_to_mfn(predecessor))) || |
1403 | 4.14M | !page_state_is(predecessor, free) || |
1404 | 4.14M | (PFN_ORDER(predecessor) != order) || |
1405 | 4.14M | (phys_to_nid(page_to_maddr(predecessor)) != node) ) |
1406 | 46 | break; |
1407 | 4.14M | |
1408 | 4.14M | check_and_stop_scrub(predecessor); |
1409 | 4.14M | |
1410 | 4.14M | page_list_del(predecessor, &heap(node, zone, order)); |
1411 | 4.14M | |
1412 | 4.14M | /* Keep predecessor's first_dirty if it is already set. */ |
1413 | 4.14M | if ( predecessor->u.free.first_dirty == INVALID_DIRTY_IDX && |
1414 | 4.14M | pg->u.free.first_dirty != INVALID_DIRTY_IDX ) |
1415 | 9 | predecessor->u.free.first_dirty = (1U << order) + |
1416 | 9 | pg->u.free.first_dirty; |
1417 | 4.14M | |
1418 | 4.14M | pg = predecessor; |
1419 | 4.14M | } |
1420 | 8.29M | else |
1421 | 4.14M | { |
1422 | 4.14M | struct page_info *successor = pg + mask; |
1423 | 4.14M | |
1424 | 4.14M | /* Merge with successor block? */ |
1425 | 4.14M | if ( !mfn_valid(_mfn(page_to_mfn(successor))) || |
1426 | 4.14M | !page_state_is(successor, free) || |
1427 | 8 | (PFN_ORDER(successor) != order) || |
1428 | 8 | (phys_to_nid(page_to_maddr(successor)) != node) ) |
1429 | 4.14M | break; |
1430 | 4.14M | |
1431 | 8 | check_and_stop_scrub(successor); |
1432 | 8 | |
1433 | 8 | page_list_del(successor, &heap(node, zone, order)); |
1434 | 8 | } |
1435 | 8.29M | |
1436 | 4.14M | order++; |
1437 | 4.14M | } |
1438 | 4.14M | |
1439 | 4.14M | page_list_add_scrub(pg, node, zone, order, pg->u.free.first_dirty); |
1440 | 4.14M | |
1441 | 4.14M | if ( tainted ) |
1442 | 0 | reserve_offlined_page(pg); |
1443 | 4.14M | |
1444 | 4.14M | spin_unlock(&heap_lock); |
1445 | 4.14M | } |
1446 | | |
1447 | | |
1448 | | /* |
1449 | | * Following rules applied for page offline: |
1450 | | * Once a page is broken, it can't be assigned anymore |
1451 | | * A page will be offlined only if it is free |
1452 | | * return original count_info |
1453 | | */ |
1454 | | static unsigned long mark_page_offline(struct page_info *pg, int broken) |
1455 | 0 | { |
1456 | 0 | unsigned long nx, x, y = pg->count_info; |
1457 | 0 |
|
1458 | 0 | ASSERT(page_is_ram_type(page_to_mfn(pg), RAM_TYPE_CONVENTIONAL)); |
1459 | 0 | ASSERT(spin_is_locked(&heap_lock)); |
1460 | 0 |
|
1461 | 0 | do { |
1462 | 0 | nx = x = y; |
1463 | 0 |
|
1464 | 0 | if ( ((x & PGC_state) != PGC_state_offlined) && |
1465 | 0 | ((x & PGC_state) != PGC_state_offlining) ) |
1466 | 0 | { |
1467 | 0 | nx &= ~PGC_state; |
1468 | 0 | nx |= (((x & PGC_state) == PGC_state_free) |
1469 | 0 | ? PGC_state_offlined : PGC_state_offlining); |
1470 | 0 | } |
1471 | 0 |
|
1472 | 0 | if ( broken ) |
1473 | 0 | nx |= PGC_broken; |
1474 | 0 |
|
1475 | 0 | if ( x == nx ) |
1476 | 0 | break; |
1477 | 0 | } while ( (y = cmpxchg(&pg->count_info, x, nx)) != x ); |
1478 | 0 |
|
1479 | 0 | return y; |
1480 | 0 | } |
1481 | | |
1482 | | static int reserve_heap_page(struct page_info *pg) |
1483 | 0 | { |
1484 | 0 | struct page_info *head = NULL; |
1485 | 0 | unsigned int i, node = phys_to_nid(page_to_maddr(pg)); |
1486 | 0 | unsigned int zone = page_to_zone(pg); |
1487 | 0 |
|
1488 | 0 | for ( i = 0; i <= MAX_ORDER; i++ ) |
1489 | 0 | { |
1490 | 0 | struct page_info *tmp; |
1491 | 0 |
|
1492 | 0 | if ( page_list_empty(&heap(node, zone, i)) ) |
1493 | 0 | continue; |
1494 | 0 |
|
1495 | 0 | page_list_for_each_safe ( head, tmp, &heap(node, zone, i) ) |
1496 | 0 | { |
1497 | 0 | if ( (head <= pg) && |
1498 | 0 | (head + (1UL << i) > pg) ) |
1499 | 0 | return reserve_offlined_page(head); |
1500 | 0 | } |
1501 | 0 | } |
1502 | 0 |
|
1503 | 0 | return -EINVAL; |
1504 | 0 |
|
1505 | 0 | } |
1506 | | |
1507 | | int offline_page(unsigned long mfn, int broken, uint32_t *status) |
1508 | 0 | { |
1509 | 0 | unsigned long old_info = 0; |
1510 | 0 | struct domain *owner; |
1511 | 0 | struct page_info *pg; |
1512 | 0 |
|
1513 | 0 | if ( !mfn_valid(_mfn(mfn)) ) |
1514 | 0 | { |
1515 | 0 | dprintk(XENLOG_WARNING, |
1516 | 0 | "try to offline page out of range %lx\n", mfn); |
1517 | 0 | return -EINVAL; |
1518 | 0 | } |
1519 | 0 |
|
1520 | 0 | *status = 0; |
1521 | 0 | pg = mfn_to_page(mfn); |
1522 | 0 |
|
1523 | 0 | if ( is_xen_fixed_mfn(mfn) ) |
1524 | 0 | { |
1525 | 0 | *status = PG_OFFLINE_XENPAGE | PG_OFFLINE_FAILED | |
1526 | 0 | (DOMID_XEN << PG_OFFLINE_OWNER_SHIFT); |
1527 | 0 | return -EPERM; |
1528 | 0 | } |
1529 | 0 |
|
1530 | 0 | /* |
1531 | 0 | * N.B. xen's txt in x86_64 is marked reserved and handled already. |
1532 | 0 | * Also kexec range is reserved. |
1533 | 0 | */ |
1534 | 0 | if ( !page_is_ram_type(mfn, RAM_TYPE_CONVENTIONAL) ) |
1535 | 0 | { |
1536 | 0 | *status = PG_OFFLINE_FAILED | PG_OFFLINE_NOT_CONV_RAM; |
1537 | 0 | return -EINVAL; |
1538 | 0 | } |
1539 | 0 |
|
1540 | 0 | /* |
1541 | 0 | * NB. When broken page belong to guest, usually hypervisor will |
1542 | 0 | * notify the guest to handle the broken page. However, hypervisor |
1543 | 0 | * need to prevent malicious guest access the broken page again. |
1544 | 0 | * Under such case, hypervisor shutdown guest, preventing recursive mce. |
1545 | 0 | */ |
1546 | 0 | if ( (pg->count_info & PGC_broken) && (owner = page_get_owner(pg)) ) |
1547 | 0 | { |
1548 | 0 | *status = PG_OFFLINE_AGAIN; |
1549 | 0 | domain_shutdown(owner, SHUTDOWN_crash); |
1550 | 0 | return 0; |
1551 | 0 | } |
1552 | 0 |
|
1553 | 0 | spin_lock(&heap_lock); |
1554 | 0 |
|
1555 | 0 | old_info = mark_page_offline(pg, broken); |
1556 | 0 |
|
1557 | 0 | if ( page_state_is(pg, offlined) ) |
1558 | 0 | { |
1559 | 0 | reserve_heap_page(pg); |
1560 | 0 |
|
1561 | 0 | spin_unlock(&heap_lock); |
1562 | 0 |
|
1563 | 0 | *status = broken ? PG_OFFLINE_OFFLINED | PG_OFFLINE_BROKEN |
1564 | 0 | : PG_OFFLINE_OFFLINED; |
1565 | 0 | return 0; |
1566 | 0 | } |
1567 | 0 |
|
1568 | 0 | spin_unlock(&heap_lock); |
1569 | 0 |
|
1570 | 0 | if ( (owner = page_get_owner_and_reference(pg)) ) |
1571 | 0 | { |
1572 | 0 | if ( p2m_pod_offline_or_broken_hit(pg) ) |
1573 | 0 | { |
1574 | 0 | put_page(pg); |
1575 | 0 | p2m_pod_offline_or_broken_replace(pg); |
1576 | 0 | *status = PG_OFFLINE_OFFLINED; |
1577 | 0 | } |
1578 | 0 | else |
1579 | 0 | { |
1580 | 0 | *status = PG_OFFLINE_OWNED | PG_OFFLINE_PENDING | |
1581 | 0 | (owner->domain_id << PG_OFFLINE_OWNER_SHIFT); |
1582 | 0 | /* Release the reference since it will not be allocated anymore */ |
1583 | 0 | put_page(pg); |
1584 | 0 | } |
1585 | 0 | } |
1586 | 0 | else if ( old_info & PGC_xen_heap ) |
1587 | 0 | { |
1588 | 0 | *status = PG_OFFLINE_XENPAGE | PG_OFFLINE_PENDING | |
1589 | 0 | (DOMID_XEN << PG_OFFLINE_OWNER_SHIFT); |
1590 | 0 | } |
1591 | 0 | else |
1592 | 0 | { |
1593 | 0 | /* |
1594 | 0 | * assign_pages does not hold heap_lock, so small window that the owner |
1595 | 0 | * may be set later, but please notice owner will only change from |
1596 | 0 | * NULL to be set, not verse, since page is offlining now. |
1597 | 0 | * No windows If called from #MC handler, since all CPU are in softirq |
1598 | 0 | * If called from user space like CE handling, tools can wait some time |
1599 | 0 | * before call again. |
1600 | 0 | */ |
1601 | 0 | *status = PG_OFFLINE_ANONYMOUS | PG_OFFLINE_FAILED | |
1602 | 0 | (DOMID_INVALID << PG_OFFLINE_OWNER_SHIFT ); |
1603 | 0 | } |
1604 | 0 |
|
1605 | 0 | if ( broken ) |
1606 | 0 | *status |= PG_OFFLINE_BROKEN; |
1607 | 0 |
|
1608 | 0 | return 0; |
1609 | 0 | } |
1610 | | |
1611 | | /* |
1612 | | * Online the memory. |
1613 | | * The caller should make sure end_pfn <= max_page, |
1614 | | * if not, expand_pages() should be called prior to online_page(). |
1615 | | */ |
1616 | | unsigned int online_page(unsigned long mfn, uint32_t *status) |
1617 | 0 | { |
1618 | 0 | unsigned long x, nx, y; |
1619 | 0 | struct page_info *pg; |
1620 | 0 | int ret; |
1621 | 0 |
|
1622 | 0 | if ( !mfn_valid(_mfn(mfn)) ) |
1623 | 0 | { |
1624 | 0 | dprintk(XENLOG_WARNING, "call expand_pages() first\n"); |
1625 | 0 | return -EINVAL; |
1626 | 0 | } |
1627 | 0 |
|
1628 | 0 | pg = mfn_to_page(mfn); |
1629 | 0 |
|
1630 | 0 | spin_lock(&heap_lock); |
1631 | 0 |
|
1632 | 0 | y = pg->count_info; |
1633 | 0 | do { |
1634 | 0 | ret = *status = 0; |
1635 | 0 |
|
1636 | 0 | if ( y & PGC_broken ) |
1637 | 0 | { |
1638 | 0 | ret = -EINVAL; |
1639 | 0 | *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN; |
1640 | 0 | break; |
1641 | 0 | } |
1642 | 0 |
|
1643 | 0 | if ( (y & PGC_state) == PGC_state_offlined ) |
1644 | 0 | { |
1645 | 0 | page_list_del(pg, &page_offlined_list); |
1646 | 0 | *status = PG_ONLINE_ONLINED; |
1647 | 0 | } |
1648 | 0 | else if ( (y & PGC_state) == PGC_state_offlining ) |
1649 | 0 | { |
1650 | 0 | *status = PG_ONLINE_ONLINED; |
1651 | 0 | } |
1652 | 0 | else |
1653 | 0 | { |
1654 | 0 | break; |
1655 | 0 | } |
1656 | 0 |
|
1657 | 0 | x = y; |
1658 | 0 | nx = (x & ~PGC_state) | PGC_state_inuse; |
1659 | 0 | } while ( (y = cmpxchg(&pg->count_info, x, nx)) != x ); |
1660 | 0 |
|
1661 | 0 | spin_unlock(&heap_lock); |
1662 | 0 |
|
1663 | 0 | if ( (y & PGC_state) == PGC_state_offlined ) |
1664 | 0 | free_heap_pages(pg, 0, false); |
1665 | 0 |
|
1666 | 0 | return ret; |
1667 | 0 | } |
1668 | | |
1669 | | int query_page_offline(unsigned long mfn, uint32_t *status) |
1670 | 0 | { |
1671 | 0 | struct page_info *pg; |
1672 | 0 |
|
1673 | 0 | if ( !mfn_valid(_mfn(mfn)) || !page_is_ram_type(mfn, RAM_TYPE_CONVENTIONAL) ) |
1674 | 0 | { |
1675 | 0 | dprintk(XENLOG_WARNING, "call expand_pages() first\n"); |
1676 | 0 | return -EINVAL; |
1677 | 0 | } |
1678 | 0 |
|
1679 | 0 | *status = 0; |
1680 | 0 | spin_lock(&heap_lock); |
1681 | 0 |
|
1682 | 0 | pg = mfn_to_page(mfn); |
1683 | 0 |
|
1684 | 0 | if ( page_state_is(pg, offlining) ) |
1685 | 0 | *status |= PG_OFFLINE_STATUS_OFFLINE_PENDING; |
1686 | 0 | if ( pg->count_info & PGC_broken ) |
1687 | 0 | *status |= PG_OFFLINE_STATUS_BROKEN; |
1688 | 0 | if ( page_state_is(pg, offlined) ) |
1689 | 0 | *status |= PG_OFFLINE_STATUS_OFFLINED; |
1690 | 0 |
|
1691 | 0 | spin_unlock(&heap_lock); |
1692 | 0 |
|
1693 | 0 | return 0; |
1694 | 0 | } |
1695 | | |
1696 | | /* |
1697 | | * Hand the specified arbitrary page range to the specified heap zone |
1698 | | * checking the node_id of the previous page. If they differ and the |
1699 | | * latter is not on a MAX_ORDER boundary, then we reserve the page by |
1700 | | * not freeing it to the buddy allocator. |
1701 | | */ |
1702 | | static void init_heap_pages( |
1703 | | struct page_info *pg, unsigned long nr_pages) |
1704 | 20 | { |
1705 | 20 | unsigned long i; |
1706 | 20 | |
1707 | 20 | /* |
1708 | 20 | * Some pages may not go through the boot allocator (e.g reserved |
1709 | 20 | * memory at boot but released just after --- kernel, initramfs, |
1710 | 20 | * etc.). |
1711 | 20 | * Update first_valid_mfn to ensure those regions are covered. |
1712 | 20 | */ |
1713 | 20 | spin_lock(&heap_lock); |
1714 | 20 | first_valid_mfn = min_t(unsigned long, page_to_mfn(pg), first_valid_mfn); |
1715 | 20 | spin_unlock(&heap_lock); |
1716 | 20 | |
1717 | 4.14M | for ( i = 0; i < nr_pages; i++ ) |
1718 | 4.14M | { |
1719 | 4.14M | unsigned int nid = phys_to_nid(page_to_maddr(pg+i)); |
1720 | 4.14M | |
1721 | 4.14M | if ( unlikely(!avail[nid]) ) |
1722 | 1 | { |
1723 | 1 | unsigned long s = page_to_mfn(pg + i); |
1724 | 1 | unsigned long e = page_to_mfn(pg + nr_pages - 1) + 1; |
1725 | 1 | bool_t use_tail = (nid == phys_to_nid(pfn_to_paddr(e - 1))) && |
1726 | 1 | !(s & ((1UL << MAX_ORDER) - 1)) && |
1727 | 0 | (find_first_set_bit(e) <= find_first_set_bit(s)); |
1728 | 1 | unsigned long n; |
1729 | 1 | |
1730 | 1 | n = init_node_heap(nid, page_to_mfn(pg+i), nr_pages - i, |
1731 | 1 | &use_tail); |
1732 | 1 | BUG_ON(i + n > nr_pages); |
1733 | 1 | if ( n && !use_tail ) |
1734 | 0 | { |
1735 | 0 | i += n - 1; |
1736 | 0 | continue; |
1737 | 0 | } |
1738 | 1 | if ( i + n == nr_pages ) |
1739 | 0 | break; |
1740 | 1 | nr_pages -= n; |
1741 | 1 | } |
1742 | 4.14M | |
1743 | 4.14M | free_heap_pages(pg + i, 0, scrub_debug); |
1744 | 4.14M | } |
1745 | 20 | } |
1746 | | |
1747 | | static unsigned long avail_heap_pages( |
1748 | | unsigned int zone_lo, unsigned int zone_hi, unsigned int node) |
1749 | 1 | { |
1750 | 1 | unsigned int i, zone; |
1751 | 1 | unsigned long free_pages = 0; |
1752 | 1 | |
1753 | 1 | if ( zone_hi >= NR_ZONES ) |
1754 | 0 | zone_hi = NR_ZONES - 1; |
1755 | 1 | |
1756 | 1 | for_each_online_node(i) |
1757 | 1 | { |
1758 | 1 | if ( !avail[i] ) |
1759 | 0 | continue; |
1760 | 41 | for ( zone = zone_lo; zone <= zone_hi; zone++ ) |
1761 | 40 | if ( (node == -1) || (node == i) ) |
1762 | 40 | free_pages += avail[i][zone]; |
1763 | 1 | } |
1764 | 1 | |
1765 | 1 | return free_pages; |
1766 | 1 | } |
1767 | | |
1768 | | unsigned long total_free_pages(void) |
1769 | 0 | { |
1770 | 0 | return total_avail_pages - midsize_alloc_zone_pages; |
1771 | 0 | } |
1772 | | |
1773 | | void __init end_boot_allocator(void) |
1774 | 1 | { |
1775 | 1 | unsigned int i; |
1776 | 1 | |
1777 | 1 | /* Pages that are free now go to the domain sub-allocator. */ |
1778 | 1 | for ( i = 0; i < nr_bootmem_regions; i++ ) |
1779 | 1 | { |
1780 | 1 | struct bootmem_region *r = &bootmem_region_list[i]; |
1781 | 1 | if ( (r->s < r->e) && |
1782 | 1 | (phys_to_nid(pfn_to_paddr(r->s)) == cpu_to_node(0)) ) |
1783 | 1 | { |
1784 | 1 | init_heap_pages(mfn_to_page(r->s), r->e - r->s); |
1785 | 1 | r->e = r->s; |
1786 | 1 | break; |
1787 | 1 | } |
1788 | 1 | } |
1789 | 17 | for ( i = nr_bootmem_regions; i-- > 0; ) |
1790 | 16 | { |
1791 | 16 | struct bootmem_region *r = &bootmem_region_list[i]; |
1792 | 16 | if ( r->s < r->e ) |
1793 | 15 | init_heap_pages(mfn_to_page(r->s), r->e - r->s); |
1794 | 16 | } |
1795 | 1 | nr_bootmem_regions = 0; |
1796 | 1 | init_heap_pages(virt_to_page(bootmem_region_list), 1); |
1797 | 1 | |
1798 | 1 | if ( !dma_bitsize && (num_online_nodes() > 1) ) |
1799 | 0 | dma_bitsize = arch_get_dma_bitsize(); |
1800 | 1 | |
1801 | 1 | printk("Domain heap initialised"); |
1802 | 1 | if ( dma_bitsize ) |
1803 | 0 | printk(" DMA width %u bits", dma_bitsize); |
1804 | 1 | printk("\n"); |
1805 | 1 | } |
1806 | | |
1807 | | static void __init smp_scrub_heap_pages(void *data) |
1808 | 134 | { |
1809 | 134 | unsigned long mfn, start, end; |
1810 | 134 | struct page_info *pg; |
1811 | 134 | struct scrub_region *r; |
1812 | 134 | unsigned int temp_cpu, cpu_idx = 0; |
1813 | 134 | nodeid_t node; |
1814 | 134 | unsigned int cpu = smp_processor_id(); |
1815 | 134 | |
1816 | 134 | if ( data ) |
1817 | 0 | r = data; |
1818 | 134 | else |
1819 | 134 | { |
1820 | 134 | node = cpu_to_node(cpu); |
1821 | 134 | if ( node == NUMA_NO_NODE ) |
1822 | 0 | return; |
1823 | 134 | r = ®ion[node]; |
1824 | 134 | } |
1825 | 134 | |
1826 | 134 | /* Determine the current CPU's index into CPU's linked to this node. */ |
1827 | 134 | for_each_cpu ( temp_cpu, &r->cpus ) |
1828 | 338 | { |
1829 | 338 | if ( cpu == temp_cpu ) |
1830 | 115 | break; |
1831 | 223 | cpu_idx++; |
1832 | 223 | } |
1833 | 134 | |
1834 | 134 | /* Calculate the starting mfn for this CPU's memory block. */ |
1835 | 134 | start = r->start + (r->per_cpu_sz * cpu_idx) + r->offset; |
1836 | 134 | |
1837 | 134 | /* Calculate the end mfn into this CPU's memory block for this iteration. */ |
1838 | 134 | if ( r->offset + chunk_size >= r->per_cpu_sz ) |
1839 | 6 | { |
1840 | 6 | end = r->start + (r->per_cpu_sz * cpu_idx) + r->per_cpu_sz; |
1841 | 6 | |
1842 | 6 | if ( r->rem && (cpumask_weight(&r->cpus) - 1 == cpu_idx) ) |
1843 | 1 | end += r->rem; |
1844 | 6 | } |
1845 | 134 | else |
1846 | 128 | end = start + chunk_size; |
1847 | 134 | |
1848 | 830k | for ( mfn = start; mfn < end; mfn++ ) |
1849 | 830k | { |
1850 | 830k | pg = mfn_to_page(mfn); |
1851 | 830k | |
1852 | 830k | /* Check the mfn is valid and page is free. */ |
1853 | 830k | if ( !mfn_valid(_mfn(mfn)) || !page_state_is(pg, free) ) |
1854 | 820k | continue; |
1855 | 830k | |
1856 | 9.25k | scrub_one_page(pg); |
1857 | 9.25k | } |
1858 | 134 | } |
1859 | | |
1860 | | static int __init find_non_smt(unsigned int node, cpumask_t *dest) |
1861 | 1 | { |
1862 | 1 | cpumask_t node_cpus; |
1863 | 1 | unsigned int i, cpu; |
1864 | 1 | |
1865 | 1 | cpumask_and(&node_cpus, &node_to_cpumask(node), &cpu_online_map); |
1866 | 1 | cpumask_clear(dest); |
1867 | 1 | for_each_cpu ( i, &node_cpus ) |
1868 | 12 | { |
1869 | 12 | if ( cpumask_intersects(dest, per_cpu(cpu_sibling_mask, i)) ) |
1870 | 6 | continue; |
1871 | 6 | cpu = cpumask_first(per_cpu(cpu_sibling_mask, i)); |
1872 | 6 | __cpumask_set_cpu(cpu, dest); |
1873 | 6 | } |
1874 | 1 | return cpumask_weight(dest); |
1875 | 1 | } |
1876 | | |
1877 | | /* |
1878 | | * Scrub all unallocated pages in all heap zones. This function uses all |
1879 | | * online cpu's to scrub the memory in parallel. |
1880 | | */ |
1881 | | static void __init scrub_heap_pages(void) |
1882 | 1 | { |
1883 | 1 | cpumask_t node_cpus, all_worker_cpus; |
1884 | 1 | unsigned int i, j; |
1885 | 1 | unsigned long offset, max_per_cpu_sz = 0; |
1886 | 1 | unsigned long start, end; |
1887 | 1 | unsigned long rem = 0; |
1888 | 1 | int last_distance, best_node; |
1889 | 1 | int cpus; |
1890 | 1 | |
1891 | 1 | cpumask_clear(&all_worker_cpus); |
1892 | 1 | /* Scrub block size. */ |
1893 | 1 | chunk_size = opt_bootscrub_chunk >> PAGE_SHIFT; |
1894 | 1 | if ( chunk_size == 0 ) |
1895 | 0 | chunk_size = MB(128) >> PAGE_SHIFT; |
1896 | 1 | |
1897 | 1 | /* Round #0 - figure out amounts and which CPUs to use. */ |
1898 | 1 | for_each_online_node ( i ) |
1899 | 1 | { |
1900 | 1 | if ( !node_spanned_pages(i) ) |
1901 | 0 | continue; |
1902 | 1 | /* Calculate Node memory start and end address. */ |
1903 | 1 | start = max(node_start_pfn(i), first_valid_mfn); |
1904 | 1 | end = min(node_start_pfn(i) + node_spanned_pages(i), max_page); |
1905 | 1 | /* Just in case NODE has 1 page and starts below first_valid_mfn. */ |
1906 | 1 | end = max(end, start); |
1907 | 1 | /* CPUs that are online and on this node (if none, that it is OK). */ |
1908 | 1 | cpus = find_non_smt(i, &node_cpus); |
1909 | 1 | cpumask_or(&all_worker_cpus, &all_worker_cpus, &node_cpus); |
1910 | 1 | if ( cpus <= 0 ) |
1911 | 0 | { |
1912 | 0 | /* No CPUs on this node. Round #2 will take of it. */ |
1913 | 0 | rem = 0; |
1914 | 0 | region[i].per_cpu_sz = (end - start); |
1915 | 0 | } |
1916 | 1 | else |
1917 | 1 | { |
1918 | 1 | rem = (end - start) % cpus; |
1919 | 1 | region[i].per_cpu_sz = (end - start) / cpus; |
1920 | 1 | if ( region[i].per_cpu_sz > max_per_cpu_sz ) |
1921 | 1 | max_per_cpu_sz = region[i].per_cpu_sz; |
1922 | 1 | } |
1923 | 1 | region[i].start = start; |
1924 | 1 | region[i].rem = rem; |
1925 | 1 | cpumask_copy(®ion[i].cpus, &node_cpus); |
1926 | 1 | } |
1927 | 1 | |
1928 | 1 | printk("Scrubbing Free RAM on %d nodes using %d CPUs\n", num_online_nodes(), |
1929 | 1 | cpumask_weight(&all_worker_cpus)); |
1930 | 1 | |
1931 | 1 | /* Round: #1 - do NUMA nodes with CPUs. */ |
1932 | 25 | for ( offset = 0; offset < max_per_cpu_sz; offset += chunk_size ) |
1933 | 24 | { |
1934 | 24 | for_each_online_node ( i ) |
1935 | 24 | region[i].offset = offset; |
1936 | 24 | |
1937 | 24 | process_pending_softirqs(); |
1938 | 24 | |
1939 | 24 | spin_lock(&heap_lock); |
1940 | 24 | on_selected_cpus(&all_worker_cpus, smp_scrub_heap_pages, NULL, 1); |
1941 | 24 | spin_unlock(&heap_lock); |
1942 | 24 | |
1943 | 24 | printk("."); |
1944 | 24 | } |
1945 | 1 | |
1946 | 1 | /* |
1947 | 1 | * Round #2: NUMA nodes with no CPUs get scrubbed with CPUs on the node |
1948 | 1 | * closest to us and with CPUs. |
1949 | 1 | */ |
1950 | 1 | for_each_online_node ( i ) |
1951 | 1 | { |
1952 | 1 | node_cpus = node_to_cpumask(i); |
1953 | 1 | |
1954 | 1 | if ( !cpumask_empty(&node_cpus) ) |
1955 | 1 | continue; |
1956 | 1 | |
1957 | 0 | last_distance = INT_MAX; |
1958 | 0 | best_node = first_node(node_online_map); |
1959 | 0 | /* Figure out which NODE CPUs are close. */ |
1960 | 0 | for_each_online_node ( j ) |
1961 | 0 | { |
1962 | 0 | u8 distance; |
1963 | 0 |
|
1964 | 0 | if ( cpumask_empty(&node_to_cpumask(j)) ) |
1965 | 0 | continue; |
1966 | 0 |
|
1967 | 0 | distance = __node_distance(i, j); |
1968 | 0 | if ( (distance < last_distance) && (distance != NUMA_NO_DISTANCE) ) |
1969 | 0 | { |
1970 | 0 | last_distance = distance; |
1971 | 0 | best_node = j; |
1972 | 0 | } |
1973 | 0 | } |
1974 | 0 | /* |
1975 | 0 | * Use CPUs from best node, and if there are no CPUs on the |
1976 | 0 | * first node (the default) use the BSP. |
1977 | 0 | */ |
1978 | 0 | cpus = find_non_smt(best_node, &node_cpus); |
1979 | 0 | if ( cpus == 0 ) |
1980 | 0 | { |
1981 | 0 | __cpumask_set_cpu(smp_processor_id(), &node_cpus); |
1982 | 0 | cpus = 1; |
1983 | 0 | } |
1984 | 0 | /* We already have the node information from round #0. */ |
1985 | 0 | region[i].rem = region[i].per_cpu_sz % cpus; |
1986 | 0 | region[i].per_cpu_sz /= cpus; |
1987 | 0 | max_per_cpu_sz = region[i].per_cpu_sz; |
1988 | 0 | cpumask_copy(®ion[i].cpus, &node_cpus); |
1989 | 0 |
|
1990 | 0 | for ( offset = 0; offset < max_per_cpu_sz; offset += chunk_size ) |
1991 | 0 | { |
1992 | 0 | region[i].offset = offset; |
1993 | 0 |
|
1994 | 0 | process_pending_softirqs(); |
1995 | 0 |
|
1996 | 0 | spin_lock(&heap_lock); |
1997 | 0 | on_selected_cpus(&node_cpus, smp_scrub_heap_pages, ®ion[i], 1); |
1998 | 0 | spin_unlock(&heap_lock); |
1999 | 0 |
|
2000 | 0 | printk("."); |
2001 | 0 | } |
2002 | 0 | } |
2003 | 1 | |
2004 | 1 | printk("done.\n"); |
2005 | 1 | |
2006 | 1 | #ifdef CONFIG_SCRUB_DEBUG |
2007 | 1 | scrub_debug = true; |
2008 | 1 | #endif |
2009 | 1 | } |
2010 | | |
2011 | | void __init heap_init_late(void) |
2012 | 1 | { |
2013 | 1 | /* |
2014 | 1 | * Now that the heap is initialized set bounds |
2015 | 1 | * for the low mem virq algorithm. |
2016 | 1 | */ |
2017 | 1 | setup_low_mem_virq(); |
2018 | 1 | |
2019 | 1 | if ( opt_bootscrub ) |
2020 | 1 | scrub_heap_pages(); |
2021 | 1 | } |
2022 | | |
2023 | | |
2024 | | /************************* |
2025 | | * XEN-HEAP SUB-ALLOCATOR |
2026 | | */ |
2027 | | |
2028 | | #if defined(CONFIG_SEPARATE_XENHEAP) |
2029 | | |
2030 | | void init_xenheap_pages(paddr_t ps, paddr_t pe) |
2031 | | { |
2032 | | ps = round_pgup(ps); |
2033 | | pe = round_pgdown(pe); |
2034 | | if ( pe <= ps ) |
2035 | | return; |
2036 | | |
2037 | | /* |
2038 | | * Yuk! Ensure there is a one-page buffer between Xen and Dom zones, to |
2039 | | * prevent merging of power-of-two blocks across the zone boundary. |
2040 | | */ |
2041 | | if ( ps && !is_xen_heap_mfn(paddr_to_pfn(ps)-1) ) |
2042 | | ps += PAGE_SIZE; |
2043 | | if ( !is_xen_heap_mfn(paddr_to_pfn(pe)) ) |
2044 | | pe -= PAGE_SIZE; |
2045 | | |
2046 | | memguard_guard_range(maddr_to_virt(ps), pe - ps); |
2047 | | |
2048 | | init_heap_pages(maddr_to_page(ps), (pe - ps) >> PAGE_SHIFT); |
2049 | | } |
2050 | | |
2051 | | |
2052 | | void *alloc_xenheap_pages(unsigned int order, unsigned int memflags) |
2053 | | { |
2054 | | struct page_info *pg; |
2055 | | |
2056 | | ASSERT(!in_irq()); |
2057 | | |
2058 | | pg = alloc_heap_pages(MEMZONE_XEN, MEMZONE_XEN, |
2059 | | order, memflags | MEMF_no_scrub, NULL); |
2060 | | if ( unlikely(pg == NULL) ) |
2061 | | return NULL; |
2062 | | |
2063 | | memguard_unguard_range(page_to_virt(pg), 1 << (order + PAGE_SHIFT)); |
2064 | | |
2065 | | return page_to_virt(pg); |
2066 | | } |
2067 | | |
2068 | | |
2069 | | void free_xenheap_pages(void *v, unsigned int order) |
2070 | | { |
2071 | | ASSERT(!in_irq()); |
2072 | | |
2073 | | if ( v == NULL ) |
2074 | | return; |
2075 | | |
2076 | | memguard_guard_range(v, 1 << (order + PAGE_SHIFT)); |
2077 | | |
2078 | | free_heap_pages(virt_to_page(v), order, false); |
2079 | | } |
2080 | | |
2081 | | #else |
2082 | | |
2083 | | void __init xenheap_max_mfn(unsigned long mfn) |
2084 | 0 | { |
2085 | 0 | ASSERT(!first_node_initialised); |
2086 | 0 | ASSERT(!xenheap_bits); |
2087 | 0 | BUILD_BUG_ON(PADDR_BITS >= BITS_PER_LONG); |
2088 | 0 | xenheap_bits = min(flsl(mfn + 1) - 1 + PAGE_SHIFT, PADDR_BITS); |
2089 | 0 | printk(XENLOG_INFO "Xen heap: %u bits\n", xenheap_bits); |
2090 | 0 | } |
2091 | | |
2092 | | void init_xenheap_pages(paddr_t ps, paddr_t pe) |
2093 | 1 | { |
2094 | 1 | init_domheap_pages(ps, pe); |
2095 | 1 | } |
2096 | | |
2097 | | void *alloc_xenheap_pages(unsigned int order, unsigned int memflags) |
2098 | 197 | { |
2099 | 197 | struct page_info *pg; |
2100 | 197 | unsigned int i; |
2101 | 197 | |
2102 | 197 | ASSERT(!in_irq()); |
2103 | 197 | |
2104 | 197 | if ( xenheap_bits && (memflags >> _MEMF_bits) > xenheap_bits ) |
2105 | 0 | memflags &= ~MEMF_bits(~0U); |
2106 | 197 | if ( !(memflags >> _MEMF_bits) ) |
2107 | 166 | memflags |= MEMF_bits(xenheap_bits); |
2108 | 197 | |
2109 | 197 | pg = alloc_domheap_pages(NULL, order, memflags | MEMF_no_scrub); |
2110 | 197 | if ( unlikely(pg == NULL) ) |
2111 | 0 | return NULL; |
2112 | 197 | |
2113 | 817 | for ( i = 0; i < (1u << order); i++ ) |
2114 | 620 | pg[i].count_info |= PGC_xen_heap; |
2115 | 197 | |
2116 | 197 | return page_to_virt(pg); |
2117 | 197 | } |
2118 | | |
2119 | | void free_xenheap_pages(void *v, unsigned int order) |
2120 | 9 | { |
2121 | 9 | struct page_info *pg; |
2122 | 9 | unsigned int i; |
2123 | 9 | |
2124 | 9 | ASSERT(!in_irq()); |
2125 | 9 | |
2126 | 9 | if ( v == NULL ) |
2127 | 0 | return; |
2128 | 9 | |
2129 | 9 | pg = virt_to_page(v); |
2130 | 9 | |
2131 | 124 | for ( i = 0; i < (1u << order); i++ ) |
2132 | 115 | pg[i].count_info &= ~PGC_xen_heap; |
2133 | 9 | |
2134 | 9 | free_heap_pages(pg, order, true); |
2135 | 9 | } |
2136 | | |
2137 | | #endif |
2138 | | |
2139 | | |
2140 | | |
2141 | | /************************* |
2142 | | * DOMAIN-HEAP SUB-ALLOCATOR |
2143 | | */ |
2144 | | |
2145 | | void init_domheap_pages(paddr_t ps, paddr_t pe) |
2146 | 3 | { |
2147 | 3 | unsigned long smfn, emfn; |
2148 | 3 | |
2149 | 3 | ASSERT(!in_irq()); |
2150 | 3 | |
2151 | 3 | smfn = round_pgup(ps) >> PAGE_SHIFT; |
2152 | 3 | emfn = round_pgdown(pe) >> PAGE_SHIFT; |
2153 | 3 | |
2154 | 3 | if ( emfn <= smfn ) |
2155 | 0 | return; |
2156 | 3 | |
2157 | 3 | init_heap_pages(mfn_to_page(smfn), emfn - smfn); |
2158 | 3 | } |
2159 | | |
2160 | | |
2161 | | int assign_pages( |
2162 | | struct domain *d, |
2163 | | struct page_info *pg, |
2164 | | unsigned int order, |
2165 | | unsigned int memflags) |
2166 | 76 | { |
2167 | 76 | int rc = 0; |
2168 | 76 | unsigned long i; |
2169 | 76 | |
2170 | 76 | spin_lock(&d->page_alloc_lock); |
2171 | 76 | |
2172 | 76 | if ( unlikely(d->is_dying) ) |
2173 | 0 | { |
2174 | 0 | gdprintk(XENLOG_INFO, "Cannot assign page to domain%d -- dying.\n", |
2175 | 0 | d->domain_id); |
2176 | 0 | rc = -EINVAL; |
2177 | 0 | goto out; |
2178 | 0 | } |
2179 | 76 | |
2180 | 76 | if ( !(memflags & MEMF_no_refcount) ) |
2181 | 76 | { |
2182 | 76 | if ( unlikely((d->tot_pages + (1 << order)) > d->max_pages) ) |
2183 | 0 | { |
2184 | 0 | if ( !tmem_enabled() || order != 0 || d->tot_pages != d->max_pages ) |
2185 | 0 | gprintk(XENLOG_INFO, "Over-allocation for domain %u: " |
2186 | 0 | "%u > %u\n", d->domain_id, |
2187 | 0 | d->tot_pages + (1 << order), d->max_pages); |
2188 | 0 | rc = -E2BIG; |
2189 | 0 | goto out; |
2190 | 0 | } |
2191 | 76 | |
2192 | 76 | if ( unlikely(d->tot_pages == 0) ) |
2193 | 1 | get_knownalive_domain(d); |
2194 | 76 | |
2195 | 76 | domain_adjust_tot_pages(d, 1 << order); |
2196 | 76 | } |
2197 | 76 | |
2198 | 4.05M | for ( i = 0; i < (1 << order); i++ ) |
2199 | 4.05M | { |
2200 | 4.05M | ASSERT(page_get_owner(&pg[i]) == NULL); |
2201 | 4.05M | ASSERT((pg[i].count_info & ~(PGC_allocated | 1)) == 0); |
2202 | 4.05M | page_set_owner(&pg[i], d); |
2203 | 4.05M | smp_wmb(); /* Domain pointer must be visible before updating refcnt. */ |
2204 | 4.05M | pg[i].count_info = PGC_allocated | 1; |
2205 | 4.05M | page_list_add_tail(&pg[i], &d->page_list); |
2206 | 4.05M | } |
2207 | 76 | |
2208 | 76 | out: |
2209 | 76 | spin_unlock(&d->page_alloc_lock); |
2210 | 76 | return rc; |
2211 | 76 | } |
2212 | | |
2213 | | |
2214 | | struct page_info *alloc_domheap_pages( |
2215 | | struct domain *d, unsigned int order, unsigned int memflags) |
2216 | 43.2k | { |
2217 | 43.2k | struct page_info *pg = NULL; |
2218 | 43.2k | unsigned int bits = memflags >> _MEMF_bits, zone_hi = NR_ZONES - 1; |
2219 | 43.2k | unsigned int dma_zone; |
2220 | 43.2k | |
2221 | 43.2k | ASSERT(!in_irq()); |
2222 | 43.2k | |
2223 | 43.2k | bits = domain_clamp_alloc_bitsize(memflags & MEMF_no_owner ? NULL : d, |
2224 | 43.2k | bits ? : (BITS_PER_LONG+PAGE_SHIFT)); |
2225 | 43.2k | if ( (zone_hi = min_t(unsigned int, bits_to_zone(bits), zone_hi)) == 0 ) |
2226 | 0 | return NULL; |
2227 | 43.2k | |
2228 | 43.2k | if ( memflags & MEMF_no_owner ) |
2229 | 34.8k | memflags |= MEMF_no_refcount; |
2230 | 43.2k | |
2231 | 43.2k | if ( dma_bitsize && ((dma_zone = bits_to_zone(dma_bitsize)) < zone_hi) ) |
2232 | 0 | pg = alloc_heap_pages(dma_zone + 1, zone_hi, order, memflags, d); |
2233 | 43.2k | |
2234 | 43.2k | if ( (pg == NULL) && |
2235 | 43.2k | ((memflags & MEMF_no_dma) || |
2236 | 43.2k | ((pg = alloc_heap_pages(MEMZONE_XEN + 1, zone_hi, order, |
2237 | 43.2k | memflags, d)) == NULL)) ) |
2238 | 25 | return NULL; |
2239 | 43.2k | |
2240 | 43.2k | if ( d && !(memflags & MEMF_no_owner) && |
2241 | 76 | assign_pages(d, pg, order, memflags) ) |
2242 | 0 | { |
2243 | 0 | free_heap_pages(pg, order, memflags & MEMF_no_scrub); |
2244 | 0 | return NULL; |
2245 | 0 | } |
2246 | 43.2k | |
2247 | 43.2k | return pg; |
2248 | 43.2k | } |
2249 | | |
2250 | | void free_domheap_pages(struct page_info *pg, unsigned int order) |
2251 | 150 | { |
2252 | 150 | struct domain *d = page_get_owner(pg); |
2253 | 150 | unsigned int i; |
2254 | 150 | bool_t drop_dom_ref; |
2255 | 150 | |
2256 | 150 | ASSERT(!in_irq()); |
2257 | 150 | |
2258 | 150 | if ( unlikely(is_xen_heap_page(pg)) ) |
2259 | 149 | { |
2260 | 149 | /* NB. May recursively lock from relinquish_memory(). */ |
2261 | 149 | spin_lock_recursive(&d->page_alloc_lock); |
2262 | 149 | |
2263 | 298 | for ( i = 0; i < (1 << order); i++ ) |
2264 | 149 | arch_free_heap_page(d, &pg[i]); |
2265 | 149 | |
2266 | 149 | d->xenheap_pages -= 1 << order; |
2267 | 149 | drop_dom_ref = (d->xenheap_pages == 0); |
2268 | 149 | |
2269 | 149 | spin_unlock_recursive(&d->page_alloc_lock); |
2270 | 149 | } |
2271 | 150 | else |
2272 | 1 | { |
2273 | 1 | bool_t scrub; |
2274 | 1 | |
2275 | 1 | if ( likely(d) && likely(d != dom_cow) ) |
2276 | 1 | { |
2277 | 1 | /* NB. May recursively lock from relinquish_memory(). */ |
2278 | 1 | spin_lock_recursive(&d->page_alloc_lock); |
2279 | 1 | |
2280 | 2 | for ( i = 0; i < (1 << order); i++ ) |
2281 | 1 | { |
2282 | 1 | BUG_ON((pg[i].u.inuse.type_info & PGT_count_mask) != 0); |
2283 | 1 | arch_free_heap_page(d, &pg[i]); |
2284 | 1 | } |
2285 | 1 | |
2286 | 1 | drop_dom_ref = !domain_adjust_tot_pages(d, -(1 << order)); |
2287 | 1 | |
2288 | 1 | spin_unlock_recursive(&d->page_alloc_lock); |
2289 | 1 | |
2290 | 1 | /* |
2291 | 1 | * Normally we expect a domain to clear pages before freeing them, |
2292 | 1 | * if it cares about the secrecy of their contents. However, after |
2293 | 1 | * a domain has died we assume responsibility for erasure. |
2294 | 1 | */ |
2295 | 1 | scrub = d->is_dying || scrub_debug; |
2296 | 1 | } |
2297 | 1 | else |
2298 | 0 | { |
2299 | 0 | /* |
2300 | 0 | * All we need to check is that on dom_cow only order-0 chunks |
2301 | 0 | * make it here. Due to the if() above, the only two possible |
2302 | 0 | * cases right now are d == NULL and d == dom_cow. To protect |
2303 | 0 | * against relaxation of that if() condition without updating the |
2304 | 0 | * check here, don't check d != dom_cow for now. |
2305 | 0 | */ |
2306 | 0 | ASSERT(!d || !order); |
2307 | 0 | drop_dom_ref = 0; |
2308 | 0 | scrub = 1; |
2309 | 0 | } |
2310 | 1 | |
2311 | 1 | free_heap_pages(pg, order, scrub); |
2312 | 1 | } |
2313 | 150 | |
2314 | 150 | if ( drop_dom_ref ) |
2315 | 0 | put_domain(d); |
2316 | 150 | } |
2317 | | |
2318 | | unsigned long avail_domheap_pages_region( |
2319 | | unsigned int node, unsigned int min_width, unsigned int max_width) |
2320 | 1 | { |
2321 | 1 | int zone_lo, zone_hi; |
2322 | 1 | |
2323 | 1 | zone_lo = min_width ? bits_to_zone(min_width) : (MEMZONE_XEN + 1); |
2324 | 1 | zone_lo = max_t(int, MEMZONE_XEN + 1, min_t(int, NR_ZONES - 1, zone_lo)); |
2325 | 1 | |
2326 | 1 | zone_hi = max_width ? bits_to_zone(max_width) : (NR_ZONES - 1); |
2327 | 1 | zone_hi = max_t(int, MEMZONE_XEN + 1, min_t(int, NR_ZONES - 1, zone_hi)); |
2328 | 1 | |
2329 | 1 | return avail_heap_pages(zone_lo, zone_hi, node); |
2330 | 1 | } |
2331 | | |
2332 | | unsigned long avail_domheap_pages(void) |
2333 | 0 | { |
2334 | 0 | return avail_heap_pages(MEMZONE_XEN + 1, |
2335 | 0 | NR_ZONES - 1, |
2336 | 0 | -1); |
2337 | 0 | } |
2338 | | |
2339 | | unsigned long avail_node_heap_pages(unsigned int nodeid) |
2340 | 0 | { |
2341 | 0 | return avail_heap_pages(MEMZONE_XEN, NR_ZONES -1, nodeid); |
2342 | 0 | } |
2343 | | |
2344 | | |
2345 | | static void pagealloc_info(unsigned char key) |
2346 | 0 | { |
2347 | 0 | unsigned int zone = MEMZONE_XEN; |
2348 | 0 | unsigned long n, total = 0; |
2349 | 0 |
|
2350 | 0 | printk("Physical memory information:\n"); |
2351 | 0 | printk(" Xen heap: %lukB free\n", |
2352 | 0 | avail_heap_pages(zone, zone, -1) << (PAGE_SHIFT-10)); |
2353 | 0 |
|
2354 | 0 | while ( ++zone < NR_ZONES ) |
2355 | 0 | { |
2356 | 0 | if ( (zone + PAGE_SHIFT) == dma_bitsize ) |
2357 | 0 | { |
2358 | 0 | printk(" DMA heap: %lukB free\n", total << (PAGE_SHIFT-10)); |
2359 | 0 | total = 0; |
2360 | 0 | } |
2361 | 0 |
|
2362 | 0 | if ( (n = avail_heap_pages(zone, zone, -1)) != 0 ) |
2363 | 0 | { |
2364 | 0 | total += n; |
2365 | 0 | printk(" heap[%02u]: %lukB free\n", zone, n << (PAGE_SHIFT-10)); |
2366 | 0 | } |
2367 | 0 | } |
2368 | 0 |
|
2369 | 0 | printk(" Dom heap: %lukB free\n", total << (PAGE_SHIFT-10)); |
2370 | 0 | } |
2371 | | |
2372 | | static __init int pagealloc_keyhandler_init(void) |
2373 | 1 | { |
2374 | 1 | register_keyhandler('m', pagealloc_info, "memory info", 1); |
2375 | 1 | return 0; |
2376 | 1 | } |
2377 | | __initcall(pagealloc_keyhandler_init); |
2378 | | |
2379 | | |
2380 | | void scrub_one_page(struct page_info *pg) |
2381 | 34.7k | { |
2382 | 34.7k | if ( unlikely(pg->count_info & PGC_broken) ) |
2383 | 0 | return; |
2384 | 34.7k | |
2385 | 34.7k | #ifndef NDEBUG |
2386 | 34.7k | /* Avoid callers relying on allocations returning zeroed pages. */ |
2387 | 34.7k | unmap_domain_page(memset(__map_domain_page(pg), |
2388 | 34.7k | SCRUB_BYTE_PATTERN, PAGE_SIZE)); |
2389 | 34.7k | #else |
2390 | | /* For a production build, clear_page() is the fastest way to scrub. */ |
2391 | | clear_domain_page(_mfn(page_to_mfn(pg))); |
2392 | | #endif |
2393 | 34.7k | } |
2394 | | |
2395 | | static void dump_heap(unsigned char key) |
2396 | 0 | { |
2397 | 0 | s_time_t now = NOW(); |
2398 | 0 | int i, j; |
2399 | 0 |
|
2400 | 0 | printk("'%c' pressed -> dumping heap info (now-0x%X:%08X)\n", key, |
2401 | 0 | (u32)(now>>32), (u32)now); |
2402 | 0 |
|
2403 | 0 | for ( i = 0; i < MAX_NUMNODES; i++ ) |
2404 | 0 | { |
2405 | 0 | if ( !avail[i] ) |
2406 | 0 | continue; |
2407 | 0 | for ( j = 0; j < NR_ZONES; j++ ) |
2408 | 0 | printk("heap[node=%d][zone=%d] -> %lu pages\n", |
2409 | 0 | i, j, avail[i][j]); |
2410 | 0 | } |
2411 | 0 |
|
2412 | 0 | for ( i = 0; i < MAX_NUMNODES; i++ ) |
2413 | 0 | { |
2414 | 0 | if ( !node_need_scrub[i] ) |
2415 | 0 | continue; |
2416 | 0 | printk("Node %d has %lu unscrubbed pages\n", i, node_need_scrub[i]); |
2417 | 0 | } |
2418 | 0 | } |
2419 | | |
2420 | | static __init int register_heap_trigger(void) |
2421 | 1 | { |
2422 | 1 | register_keyhandler('H', dump_heap, "dump heap info", 1); |
2423 | 1 | return 0; |
2424 | 1 | } |
2425 | | __initcall(register_heap_trigger); |
2426 | | |
2427 | | /* |
2428 | | * Local variables: |
2429 | | * mode: C |
2430 | | * c-file-style: "BSD" |
2431 | | * c-basic-offset: 4 |
2432 | | * tab-width: 4 |
2433 | | * indent-tabs-mode: nil |
2434 | | * End: |
2435 | | */ |