Commit | Line | Data |
---|---|---|
453dc659 DT |
1 | /* |
2 | * VMware Balloon driver. | |
3 | * | |
4 | * Copyright (C) 2000-2010, VMware, Inc. All Rights Reserved. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify it | |
7 | * under the terms of the GNU General Public License as published by the | |
8 | * Free Software Foundation; version 2 of the License and no later version. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or | |
13 | * NON INFRINGEMENT. See the GNU General Public License for more | |
14 | * details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * along with this program; if not, write to the Free Software | |
18 | * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
19 | * | |
20 | * Maintained by: Dmitry Torokhov <dtor@vmware.com> | |
21 | */ | |
22 | ||
23 | /* | |
24 | * This is VMware physical memory management driver for Linux. The driver | |
25 | * acts like a "balloon" that can be inflated to reclaim physical pages by | |
26 | * reserving them in the guest and invalidating them in the monitor, | |
27 | * freeing up the underlying machine pages so they can be allocated to | |
28 | * other guests. The balloon can also be deflated to allow the guest to | |
29 | * use more physical memory. Higher level policies can control the sizes | |
30 | * of balloons in VMs in order to manage physical memory resources. | |
31 | */ | |
32 | ||
33 | //#define DEBUG | |
34 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
35 | ||
36 | #include <linux/types.h> | |
37 | #include <linux/kernel.h> | |
38 | #include <linux/mm.h> | |
39 | #include <linux/sched.h> | |
40 | #include <linux/module.h> | |
41 | #include <linux/workqueue.h> | |
42 | #include <linux/debugfs.h> | |
43 | #include <linux/seq_file.h> | |
a10a5698 | 44 | #include <asm/hypervisor.h> |
453dc659 DT |
45 | |
46 | MODULE_AUTHOR("VMware, Inc."); | |
47 | MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver"); | |
beda94da | 48 | MODULE_VERSION("1.2.1.3-k"); |
453dc659 DT |
49 | MODULE_ALIAS("dmi:*:svnVMware*:*"); |
50 | MODULE_ALIAS("vmware_vmmemctl"); | |
51 | MODULE_LICENSE("GPL"); | |
52 | ||
53 | /* | |
54 | * Various constants controlling rate of inflaint/deflating balloon, | |
55 | * measured in pages. | |
56 | */ | |
57 | ||
58 | /* | |
59 | * Rate of allocating memory when there is no memory pressure | |
60 | * (driver performs non-sleeping allocations). | |
61 | */ | |
62 | #define VMW_BALLOON_NOSLEEP_ALLOC_MAX 16384U | |
63 | ||
64 | /* | |
65 | * Rates of memory allocaton when guest experiences memory pressure | |
66 | * (driver performs sleeping allocations). | |
67 | */ | |
68 | #define VMW_BALLOON_RATE_ALLOC_MIN 512U | |
69 | #define VMW_BALLOON_RATE_ALLOC_MAX 2048U | |
70 | #define VMW_BALLOON_RATE_ALLOC_INC 16U | |
71 | ||
72 | /* | |
73 | * Rates for releasing pages while deflating balloon. | |
74 | */ | |
75 | #define VMW_BALLOON_RATE_FREE_MIN 512U | |
76 | #define VMW_BALLOON_RATE_FREE_MAX 16384U | |
77 | #define VMW_BALLOON_RATE_FREE_INC 16U | |
78 | ||
79 | /* | |
80 | * When guest is under memory pressure, use a reduced page allocation | |
81 | * rate for next several cycles. | |
82 | */ | |
83 | #define VMW_BALLOON_SLOW_CYCLES 4 | |
84 | ||
85 | /* | |
86 | * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't | |
87 | * allow wait (__GFP_WAIT) for NOSLEEP page allocations. Use | |
88 | * __GFP_NOWARN, to suppress page allocation failure warnings. | |
89 | */ | |
90 | #define VMW_PAGE_ALLOC_NOSLEEP (__GFP_HIGHMEM|__GFP_NOWARN) | |
91 | ||
92 | /* | |
93 | * Use GFP_HIGHUSER when executing in a separate kernel thread | |
94 | * context and allocation can sleep. This is less stressful to | |
95 | * the guest memory system, since it allows the thread to block | |
96 | * while memory is reclaimed, and won't take pages from emergency | |
97 | * low-memory pools. | |
98 | */ | |
99 | #define VMW_PAGE_ALLOC_CANSLEEP (GFP_HIGHUSER) | |
100 | ||
101 | /* Maximum number of page allocations without yielding processor */ | |
102 | #define VMW_BALLOON_YIELD_THRESHOLD 1024 | |
103 | ||
55adaa49 DT |
104 | /* Maximum number of refused pages we accumulate during inflation cycle */ |
105 | #define VMW_BALLOON_MAX_REFUSED 16 | |
453dc659 DT |
106 | |
107 | /* | |
108 | * Hypervisor communication port definitions. | |
109 | */ | |
110 | #define VMW_BALLOON_HV_PORT 0x5670 | |
111 | #define VMW_BALLOON_HV_MAGIC 0x456c6d6f | |
112 | #define VMW_BALLOON_PROTOCOL_VERSION 2 | |
113 | #define VMW_BALLOON_GUEST_ID 1 /* Linux */ | |
114 | ||
115 | #define VMW_BALLOON_CMD_START 0 | |
116 | #define VMW_BALLOON_CMD_GET_TARGET 1 | |
117 | #define VMW_BALLOON_CMD_LOCK 2 | |
118 | #define VMW_BALLOON_CMD_UNLOCK 3 | |
119 | #define VMW_BALLOON_CMD_GUEST_ID 4 | |
120 | ||
121 | /* error codes */ | |
122 | #define VMW_BALLOON_SUCCESS 0 | |
123 | #define VMW_BALLOON_FAILURE -1 | |
124 | #define VMW_BALLOON_ERROR_CMD_INVALID 1 | |
125 | #define VMW_BALLOON_ERROR_PPN_INVALID 2 | |
126 | #define VMW_BALLOON_ERROR_PPN_LOCKED 3 | |
127 | #define VMW_BALLOON_ERROR_PPN_UNLOCKED 4 | |
128 | #define VMW_BALLOON_ERROR_PPN_PINNED 5 | |
129 | #define VMW_BALLOON_ERROR_PPN_NOTNEEDED 6 | |
130 | #define VMW_BALLOON_ERROR_RESET 7 | |
131 | #define VMW_BALLOON_ERROR_BUSY 8 | |
132 | ||
133 | #define VMWARE_BALLOON_CMD(cmd, data, result) \ | |
134 | ({ \ | |
135 | unsigned long __stat, __dummy1, __dummy2; \ | |
e83736c8 | 136 | __asm__ __volatile__ ("inl %%dx" : \ |
453dc659 DT |
137 | "=a"(__stat), \ |
138 | "=c"(__dummy1), \ | |
139 | "=d"(__dummy2), \ | |
140 | "=b"(result) : \ | |
141 | "0"(VMW_BALLOON_HV_MAGIC), \ | |
142 | "1"(VMW_BALLOON_CMD_##cmd), \ | |
143 | "2"(VMW_BALLOON_HV_PORT), \ | |
144 | "3"(data) : \ | |
145 | "memory"); \ | |
146 | result &= -1UL; \ | |
147 | __stat & -1UL; \ | |
148 | }) | |
149 | ||
150 | #ifdef CONFIG_DEBUG_FS | |
151 | struct vmballoon_stats { | |
152 | unsigned int timer; | |
153 | ||
2ca02df6 | 154 | /* allocation statistics */ |
453dc659 DT |
155 | unsigned int alloc; |
156 | unsigned int alloc_fail; | |
157 | unsigned int sleep_alloc; | |
158 | unsigned int sleep_alloc_fail; | |
159 | unsigned int refused_alloc; | |
160 | unsigned int refused_free; | |
161 | unsigned int free; | |
162 | ||
163 | /* monitor operations */ | |
164 | unsigned int lock; | |
165 | unsigned int lock_fail; | |
166 | unsigned int unlock; | |
167 | unsigned int unlock_fail; | |
168 | unsigned int target; | |
169 | unsigned int target_fail; | |
170 | unsigned int start; | |
171 | unsigned int start_fail; | |
172 | unsigned int guest_type; | |
173 | unsigned int guest_type_fail; | |
174 | }; | |
175 | ||
176 | #define STATS_INC(stat) (stat)++ | |
177 | #else | |
178 | #define STATS_INC(stat) | |
179 | #endif | |
180 | ||
181 | struct vmballoon { | |
182 | ||
183 | /* list of reserved physical pages */ | |
184 | struct list_head pages; | |
185 | ||
186 | /* transient list of non-balloonable pages */ | |
187 | struct list_head refused_pages; | |
55adaa49 | 188 | unsigned int n_refused_pages; |
453dc659 DT |
189 | |
190 | /* balloon size in pages */ | |
191 | unsigned int size; | |
192 | unsigned int target; | |
193 | ||
194 | /* reset flag */ | |
195 | bool reset_required; | |
196 | ||
197 | /* adjustment rates (pages per second) */ | |
198 | unsigned int rate_alloc; | |
199 | unsigned int rate_free; | |
200 | ||
201 | /* slowdown page allocations for next few cycles */ | |
202 | unsigned int slow_allocation_cycles; | |
203 | ||
204 | #ifdef CONFIG_DEBUG_FS | |
205 | /* statistics */ | |
206 | struct vmballoon_stats stats; | |
207 | ||
208 | /* debugfs file exporting statistics */ | |
209 | struct dentry *dbg_entry; | |
210 | #endif | |
211 | ||
212 | struct sysinfo sysinfo; | |
213 | ||
214 | struct delayed_work dwork; | |
215 | }; | |
216 | ||
217 | static struct vmballoon balloon; | |
453dc659 DT |
218 | |
219 | /* | |
220 | * Send "start" command to the host, communicating supported version | |
221 | * of the protocol. | |
222 | */ | |
223 | static bool vmballoon_send_start(struct vmballoon *b) | |
224 | { | |
225 | unsigned long status, dummy; | |
226 | ||
227 | STATS_INC(b->stats.start); | |
228 | ||
229 | status = VMWARE_BALLOON_CMD(START, VMW_BALLOON_PROTOCOL_VERSION, dummy); | |
230 | if (status == VMW_BALLOON_SUCCESS) | |
231 | return true; | |
232 | ||
233 | pr_debug("%s - failed, hv returns %ld\n", __func__, status); | |
234 | STATS_INC(b->stats.start_fail); | |
235 | return false; | |
236 | } | |
237 | ||
238 | static bool vmballoon_check_status(struct vmballoon *b, unsigned long status) | |
239 | { | |
240 | switch (status) { | |
241 | case VMW_BALLOON_SUCCESS: | |
242 | return true; | |
243 | ||
244 | case VMW_BALLOON_ERROR_RESET: | |
245 | b->reset_required = true; | |
246 | /* fall through */ | |
247 | ||
248 | default: | |
249 | return false; | |
250 | } | |
251 | } | |
252 | ||
253 | /* | |
254 | * Communicate guest type to the host so that it can adjust ballooning | |
255 | * algorithm to the one most appropriate for the guest. This command | |
256 | * is normally issued after sending "start" command and is part of | |
257 | * standard reset sequence. | |
258 | */ | |
259 | static bool vmballoon_send_guest_id(struct vmballoon *b) | |
260 | { | |
261 | unsigned long status, dummy; | |
262 | ||
263 | status = VMWARE_BALLOON_CMD(GUEST_ID, VMW_BALLOON_GUEST_ID, dummy); | |
264 | ||
265 | STATS_INC(b->stats.guest_type); | |
266 | ||
267 | if (vmballoon_check_status(b, status)) | |
268 | return true; | |
269 | ||
270 | pr_debug("%s - failed, hv returns %ld\n", __func__, status); | |
271 | STATS_INC(b->stats.guest_type_fail); | |
272 | return false; | |
273 | } | |
274 | ||
275 | /* | |
276 | * Retrieve desired balloon size from the host. | |
277 | */ | |
278 | static bool vmballoon_send_get_target(struct vmballoon *b, u32 *new_target) | |
279 | { | |
280 | unsigned long status; | |
281 | unsigned long target; | |
282 | unsigned long limit; | |
283 | u32 limit32; | |
284 | ||
285 | /* | |
286 | * si_meminfo() is cheap. Moreover, we want to provide dynamic | |
287 | * max balloon size later. So let us call si_meminfo() every | |
288 | * iteration. | |
289 | */ | |
290 | si_meminfo(&b->sysinfo); | |
291 | limit = b->sysinfo.totalram; | |
292 | ||
293 | /* Ensure limit fits in 32-bits */ | |
294 | limit32 = (u32)limit; | |
295 | if (limit != limit32) | |
296 | return false; | |
297 | ||
298 | /* update stats */ | |
299 | STATS_INC(b->stats.target); | |
300 | ||
301 | status = VMWARE_BALLOON_CMD(GET_TARGET, limit, target); | |
302 | if (vmballoon_check_status(b, status)) { | |
303 | *new_target = target; | |
304 | return true; | |
305 | } | |
306 | ||
307 | pr_debug("%s - failed, hv returns %ld\n", __func__, status); | |
308 | STATS_INC(b->stats.target_fail); | |
309 | return false; | |
310 | } | |
311 | ||
312 | /* | |
313 | * Notify the host about allocated page so that host can use it without | |
314 | * fear that guest will need it. Host may reject some pages, we need to | |
315 | * check the return value and maybe submit a different page. | |
316 | */ | |
3e5ba466 | 317 | static int vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn, |
d27a0c06 | 318 | unsigned int *hv_status) |
453dc659 DT |
319 | { |
320 | unsigned long status, dummy; | |
321 | u32 pfn32; | |
322 | ||
323 | pfn32 = (u32)pfn; | |
324 | if (pfn32 != pfn) | |
3e5ba466 | 325 | return -1; |
453dc659 DT |
326 | |
327 | STATS_INC(b->stats.lock); | |
328 | ||
d27a0c06 | 329 | *hv_status = status = VMWARE_BALLOON_CMD(LOCK, pfn, dummy); |
453dc659 | 330 | if (vmballoon_check_status(b, status)) |
3e5ba466 | 331 | return 0; |
453dc659 DT |
332 | |
333 | pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status); | |
334 | STATS_INC(b->stats.lock_fail); | |
3e5ba466 | 335 | return 1; |
453dc659 DT |
336 | } |
337 | ||
338 | /* | |
339 | * Notify the host that guest intends to release given page back into | |
340 | * the pool of available (to the guest) pages. | |
341 | */ | |
342 | static bool vmballoon_send_unlock_page(struct vmballoon *b, unsigned long pfn) | |
343 | { | |
344 | unsigned long status, dummy; | |
345 | u32 pfn32; | |
346 | ||
347 | pfn32 = (u32)pfn; | |
348 | if (pfn32 != pfn) | |
349 | return false; | |
350 | ||
351 | STATS_INC(b->stats.unlock); | |
352 | ||
353 | status = VMWARE_BALLOON_CMD(UNLOCK, pfn, dummy); | |
354 | if (vmballoon_check_status(b, status)) | |
355 | return true; | |
356 | ||
357 | pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status); | |
358 | STATS_INC(b->stats.unlock_fail); | |
359 | return false; | |
360 | } | |
361 | ||
362 | /* | |
363 | * Quickly release all pages allocated for the balloon. This function is | |
364 | * called when host decides to "reset" balloon for one reason or another. | |
365 | * Unlike normal "deflate" we do not (shall not) notify host of the pages | |
366 | * being released. | |
367 | */ | |
368 | static void vmballoon_pop(struct vmballoon *b) | |
369 | { | |
370 | struct page *page, *next; | |
371 | unsigned int count = 0; | |
372 | ||
373 | list_for_each_entry_safe(page, next, &b->pages, lru) { | |
374 | list_del(&page->lru); | |
375 | __free_page(page); | |
376 | STATS_INC(b->stats.free); | |
377 | b->size--; | |
378 | ||
379 | if (++count >= b->rate_free) { | |
380 | count = 0; | |
381 | cond_resched(); | |
382 | } | |
383 | } | |
384 | } | |
385 | ||
386 | /* | |
387 | * Perform standard reset sequence by popping the balloon (in case it | |
388 | * is not empty) and then restarting protocol. This operation normally | |
389 | * happens when host responds with VMW_BALLOON_ERROR_RESET to a command. | |
390 | */ | |
391 | static void vmballoon_reset(struct vmballoon *b) | |
392 | { | |
393 | /* free all pages, skipping monitor unlock */ | |
394 | vmballoon_pop(b); | |
395 | ||
396 | if (vmballoon_send_start(b)) { | |
397 | b->reset_required = false; | |
398 | if (!vmballoon_send_guest_id(b)) | |
399 | pr_err("failed to send guest ID to the host\n"); | |
400 | } | |
401 | } | |
402 | ||
403 | /* | |
404 | * Allocate (or reserve) a page for the balloon and notify the host. If host | |
405 | * refuses the page put it on "refuse" list and allocate another one until host | |
406 | * is satisfied. "Refused" pages are released at the end of inflation cycle | |
407 | * (when we allocate b->rate_alloc pages). | |
408 | */ | |
409 | static int vmballoon_reserve_page(struct vmballoon *b, bool can_sleep) | |
410 | { | |
411 | struct page *page; | |
412 | gfp_t flags; | |
d27a0c06 | 413 | unsigned int hv_status; |
3e5ba466 | 414 | int locked; |
6d994a7e | 415 | flags = can_sleep ? VMW_PAGE_ALLOC_CANSLEEP : VMW_PAGE_ALLOC_NOSLEEP; |
453dc659 DT |
416 | |
417 | do { | |
418 | if (!can_sleep) | |
419 | STATS_INC(b->stats.alloc); | |
420 | else | |
421 | STATS_INC(b->stats.sleep_alloc); | |
422 | ||
453dc659 DT |
423 | page = alloc_page(flags); |
424 | if (!page) { | |
425 | if (!can_sleep) | |
426 | STATS_INC(b->stats.alloc_fail); | |
427 | else | |
428 | STATS_INC(b->stats.sleep_alloc_fail); | |
429 | return -ENOMEM; | |
430 | } | |
431 | ||
432 | /* inform monitor */ | |
d27a0c06 | 433 | locked = vmballoon_send_lock_page(b, page_to_pfn(page), &hv_status); |
3e5ba466 | 434 | if (locked > 0) { |
55adaa49 DT |
435 | STATS_INC(b->stats.refused_alloc); |
436 | ||
d27a0c06 DT |
437 | if (hv_status == VMW_BALLOON_ERROR_RESET || |
438 | hv_status == VMW_BALLOON_ERROR_PPN_NOTNEEDED) { | |
453dc659 DT |
439 | __free_page(page); |
440 | return -EIO; | |
441 | } | |
442 | ||
55adaa49 DT |
443 | /* |
444 | * Place page on the list of non-balloonable pages | |
445 | * and retry allocation, unless we already accumulated | |
446 | * too many of them, in which case take a breather. | |
447 | */ | |
453dc659 | 448 | list_add(&page->lru, &b->refused_pages); |
55adaa49 DT |
449 | if (++b->n_refused_pages >= VMW_BALLOON_MAX_REFUSED) |
450 | return -EIO; | |
453dc659 | 451 | } |
3e5ba466 | 452 | } while (locked != 0); |
453dc659 DT |
453 | |
454 | /* track allocated page */ | |
455 | list_add(&page->lru, &b->pages); | |
456 | ||
457 | /* update balloon size */ | |
458 | b->size++; | |
459 | ||
460 | return 0; | |
461 | } | |
462 | ||
463 | /* | |
464 | * Release the page allocated for the balloon. Note that we first notify | |
465 | * the host so it can make sure the page will be available for the guest | |
466 | * to use, if needed. | |
467 | */ | |
468 | static int vmballoon_release_page(struct vmballoon *b, struct page *page) | |
469 | { | |
470 | if (!vmballoon_send_unlock_page(b, page_to_pfn(page))) | |
471 | return -EIO; | |
472 | ||
473 | list_del(&page->lru); | |
474 | ||
475 | /* deallocate page */ | |
476 | __free_page(page); | |
477 | STATS_INC(b->stats.free); | |
478 | ||
479 | /* update balloon size */ | |
480 | b->size--; | |
481 | ||
482 | return 0; | |
483 | } | |
484 | ||
485 | /* | |
486 | * Release pages that were allocated while attempting to inflate the | |
487 | * balloon but were refused by the host for one reason or another. | |
488 | */ | |
489 | static void vmballoon_release_refused_pages(struct vmballoon *b) | |
490 | { | |
491 | struct page *page, *next; | |
492 | ||
493 | list_for_each_entry_safe(page, next, &b->refused_pages, lru) { | |
494 | list_del(&page->lru); | |
495 | __free_page(page); | |
496 | STATS_INC(b->stats.refused_free); | |
497 | } | |
55adaa49 DT |
498 | |
499 | b->n_refused_pages = 0; | |
453dc659 DT |
500 | } |
501 | ||
502 | /* | |
503 | * Inflate the balloon towards its target size. Note that we try to limit | |
504 | * the rate of allocation to make sure we are not choking the rest of the | |
505 | * system. | |
506 | */ | |
507 | static void vmballoon_inflate(struct vmballoon *b) | |
508 | { | |
509 | unsigned int goal; | |
510 | unsigned int rate; | |
511 | unsigned int i; | |
512 | unsigned int allocations = 0; | |
513 | int error = 0; | |
514 | bool alloc_can_sleep = false; | |
515 | ||
516 | pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target); | |
517 | ||
518 | /* | |
519 | * First try NOSLEEP page allocations to inflate balloon. | |
520 | * | |
521 | * If we do not throttle nosleep allocations, we can drain all | |
522 | * free pages in the guest quickly (if the balloon target is high). | |
523 | * As a side-effect, draining free pages helps to inform (force) | |
524 | * the guest to start swapping if balloon target is not met yet, | |
525 | * which is a desired behavior. However, balloon driver can consume | |
526 | * all available CPU cycles if too many pages are allocated in a | |
527 | * second. Therefore, we throttle nosleep allocations even when | |
528 | * the guest is not under memory pressure. OTOH, if we have already | |
529 | * predicted that the guest is under memory pressure, then we | |
530 | * slowdown page allocations considerably. | |
531 | */ | |
532 | ||
533 | goal = b->target - b->size; | |
534 | /* | |
535 | * Start with no sleep allocation rate which may be higher | |
536 | * than sleeping allocation rate. | |
537 | */ | |
538 | rate = b->slow_allocation_cycles ? | |
539 | b->rate_alloc : VMW_BALLOON_NOSLEEP_ALLOC_MAX; | |
540 | ||
541 | pr_debug("%s - goal: %d, no-sleep rate: %d, sleep rate: %d\n", | |
542 | __func__, goal, rate, b->rate_alloc); | |
543 | ||
544 | for (i = 0; i < goal; i++) { | |
545 | ||
546 | error = vmballoon_reserve_page(b, alloc_can_sleep); | |
547 | if (error) { | |
548 | if (error != -ENOMEM) { | |
549 | /* | |
550 | * Not a page allocation failure, stop this | |
551 | * cycle. Maybe we'll get new target from | |
552 | * the host soon. | |
553 | */ | |
554 | break; | |
555 | } | |
556 | ||
557 | if (alloc_can_sleep) { | |
558 | /* | |
559 | * CANSLEEP page allocation failed, so guest | |
560 | * is under severe memory pressure. Quickly | |
561 | * decrease allocation rate. | |
562 | */ | |
563 | b->rate_alloc = max(b->rate_alloc / 2, | |
564 | VMW_BALLOON_RATE_ALLOC_MIN); | |
565 | break; | |
566 | } | |
567 | ||
568 | /* | |
569 | * NOSLEEP page allocation failed, so the guest is | |
570 | * under memory pressure. Let us slow down page | |
571 | * allocations for next few cycles so that the guest | |
572 | * gets out of memory pressure. Also, if we already | |
573 | * allocated b->rate_alloc pages, let's pause, | |
574 | * otherwise switch to sleeping allocations. | |
575 | */ | |
576 | b->slow_allocation_cycles = VMW_BALLOON_SLOW_CYCLES; | |
577 | ||
578 | if (i >= b->rate_alloc) | |
579 | break; | |
580 | ||
581 | alloc_can_sleep = true; | |
582 | /* Lower rate for sleeping allocations. */ | |
583 | rate = b->rate_alloc; | |
584 | } | |
585 | ||
586 | if (++allocations > VMW_BALLOON_YIELD_THRESHOLD) { | |
587 | cond_resched(); | |
588 | allocations = 0; | |
589 | } | |
590 | ||
591 | if (i >= rate) { | |
592 | /* We allocated enough pages, let's take a break. */ | |
593 | break; | |
594 | } | |
595 | } | |
596 | ||
597 | /* | |
598 | * We reached our goal without failures so try increasing | |
599 | * allocation rate. | |
600 | */ | |
601 | if (error == 0 && i >= b->rate_alloc) { | |
602 | unsigned int mult = i / b->rate_alloc; | |
603 | ||
604 | b->rate_alloc = | |
605 | min(b->rate_alloc + mult * VMW_BALLOON_RATE_ALLOC_INC, | |
606 | VMW_BALLOON_RATE_ALLOC_MAX); | |
607 | } | |
608 | ||
609 | vmballoon_release_refused_pages(b); | |
610 | } | |
611 | ||
612 | /* | |
613 | * Decrease the size of the balloon allowing guest to use more memory. | |
614 | */ | |
615 | static void vmballoon_deflate(struct vmballoon *b) | |
616 | { | |
617 | struct page *page, *next; | |
618 | unsigned int i = 0; | |
619 | unsigned int goal; | |
620 | int error; | |
621 | ||
622 | pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target); | |
623 | ||
624 | /* limit deallocation rate */ | |
625 | goal = min(b->size - b->target, b->rate_free); | |
626 | ||
627 | pr_debug("%s - goal: %d, rate: %d\n", __func__, goal, b->rate_free); | |
628 | ||
629 | /* free pages to reach target */ | |
630 | list_for_each_entry_safe(page, next, &b->pages, lru) { | |
631 | error = vmballoon_release_page(b, page); | |
632 | if (error) { | |
633 | /* quickly decrease rate in case of error */ | |
634 | b->rate_free = max(b->rate_free / 2, | |
635 | VMW_BALLOON_RATE_FREE_MIN); | |
636 | return; | |
637 | } | |
638 | ||
639 | if (++i >= goal) | |
640 | break; | |
641 | } | |
642 | ||
643 | /* slowly increase rate if there were no errors */ | |
644 | b->rate_free = min(b->rate_free + VMW_BALLOON_RATE_FREE_INC, | |
645 | VMW_BALLOON_RATE_FREE_MAX); | |
646 | } | |
647 | ||
648 | /* | |
649 | * Balloon work function: reset protocol, if needed, get the new size and | |
650 | * adjust balloon as needed. Repeat in 1 sec. | |
651 | */ | |
652 | static void vmballoon_work(struct work_struct *work) | |
653 | { | |
654 | struct delayed_work *dwork = to_delayed_work(work); | |
655 | struct vmballoon *b = container_of(dwork, struct vmballoon, dwork); | |
656 | unsigned int target; | |
657 | ||
658 | STATS_INC(b->stats.timer); | |
659 | ||
660 | if (b->reset_required) | |
661 | vmballoon_reset(b); | |
662 | ||
663 | if (b->slow_allocation_cycles > 0) | |
664 | b->slow_allocation_cycles--; | |
665 | ||
666 | if (vmballoon_send_get_target(b, &target)) { | |
667 | /* update target, adjust size */ | |
668 | b->target = target; | |
669 | ||
670 | if (b->size < target) | |
671 | vmballoon_inflate(b); | |
672 | else if (b->size > target) | |
673 | vmballoon_deflate(b); | |
674 | } | |
675 | ||
beda94da DT |
676 | /* |
677 | * We are using a freezable workqueue so that balloon operations are | |
678 | * stopped while the system transitions to/from sleep/hibernation. | |
679 | */ | |
680 | queue_delayed_work(system_freezable_wq, | |
681 | dwork, round_jiffies_relative(HZ)); | |
453dc659 DT |
682 | } |
683 | ||
684 | /* | |
685 | * DEBUGFS Interface | |
686 | */ | |
687 | #ifdef CONFIG_DEBUG_FS | |
688 | ||
689 | static int vmballoon_debug_show(struct seq_file *f, void *offset) | |
690 | { | |
691 | struct vmballoon *b = f->private; | |
692 | struct vmballoon_stats *stats = &b->stats; | |
693 | ||
694 | /* format size info */ | |
695 | seq_printf(f, | |
696 | "target: %8d pages\n" | |
697 | "current: %8d pages\n", | |
698 | b->target, b->size); | |
699 | ||
700 | /* format rate info */ | |
701 | seq_printf(f, | |
702 | "rateNoSleepAlloc: %8d pages/sec\n" | |
703 | "rateSleepAlloc: %8d pages/sec\n" | |
704 | "rateFree: %8d pages/sec\n", | |
705 | VMW_BALLOON_NOSLEEP_ALLOC_MAX, | |
706 | b->rate_alloc, b->rate_free); | |
707 | ||
708 | seq_printf(f, | |
709 | "\n" | |
710 | "timer: %8u\n" | |
711 | "start: %8u (%4u failed)\n" | |
712 | "guestType: %8u (%4u failed)\n" | |
713 | "lock: %8u (%4u failed)\n" | |
714 | "unlock: %8u (%4u failed)\n" | |
715 | "target: %8u (%4u failed)\n" | |
716 | "primNoSleepAlloc: %8u (%4u failed)\n" | |
717 | "primCanSleepAlloc: %8u (%4u failed)\n" | |
718 | "primFree: %8u\n" | |
719 | "errAlloc: %8u\n" | |
720 | "errFree: %8u\n", | |
721 | stats->timer, | |
722 | stats->start, stats->start_fail, | |
723 | stats->guest_type, stats->guest_type_fail, | |
724 | stats->lock, stats->lock_fail, | |
725 | stats->unlock, stats->unlock_fail, | |
726 | stats->target, stats->target_fail, | |
727 | stats->alloc, stats->alloc_fail, | |
728 | stats->sleep_alloc, stats->sleep_alloc_fail, | |
729 | stats->free, | |
730 | stats->refused_alloc, stats->refused_free); | |
731 | ||
732 | return 0; | |
733 | } | |
734 | ||
735 | static int vmballoon_debug_open(struct inode *inode, struct file *file) | |
736 | { | |
737 | return single_open(file, vmballoon_debug_show, inode->i_private); | |
738 | } | |
739 | ||
740 | static const struct file_operations vmballoon_debug_fops = { | |
741 | .owner = THIS_MODULE, | |
742 | .open = vmballoon_debug_open, | |
743 | .read = seq_read, | |
744 | .llseek = seq_lseek, | |
745 | .release = single_release, | |
746 | }; | |
747 | ||
748 | static int __init vmballoon_debugfs_init(struct vmballoon *b) | |
749 | { | |
750 | int error; | |
751 | ||
752 | b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b, | |
753 | &vmballoon_debug_fops); | |
754 | if (IS_ERR(b->dbg_entry)) { | |
755 | error = PTR_ERR(b->dbg_entry); | |
756 | pr_err("failed to create debugfs entry, error: %d\n", error); | |
757 | return error; | |
758 | } | |
759 | ||
760 | return 0; | |
761 | } | |
762 | ||
763 | static void __exit vmballoon_debugfs_exit(struct vmballoon *b) | |
764 | { | |
765 | debugfs_remove(b->dbg_entry); | |
766 | } | |
767 | ||
768 | #else | |
769 | ||
770 | static inline int vmballoon_debugfs_init(struct vmballoon *b) | |
771 | { | |
772 | return 0; | |
773 | } | |
774 | ||
775 | static inline void vmballoon_debugfs_exit(struct vmballoon *b) | |
776 | { | |
777 | } | |
778 | ||
779 | #endif /* CONFIG_DEBUG_FS */ | |
780 | ||
781 | static int __init vmballoon_init(void) | |
782 | { | |
783 | int error; | |
784 | ||
785 | /* | |
786 | * Check if we are running on VMware's hypervisor and bail out | |
787 | * if we are not. | |
788 | */ | |
a10a5698 | 789 | if (x86_hyper != &x86_hyper_vmware) |
453dc659 DT |
790 | return -ENODEV; |
791 | ||
453dc659 DT |
792 | INIT_LIST_HEAD(&balloon.pages); |
793 | INIT_LIST_HEAD(&balloon.refused_pages); | |
794 | ||
795 | /* initialize rates */ | |
796 | balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX; | |
797 | balloon.rate_free = VMW_BALLOON_RATE_FREE_MAX; | |
798 | ||
799 | INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work); | |
800 | ||
801 | /* | |
802 | * Start balloon. | |
803 | */ | |
804 | if (!vmballoon_send_start(&balloon)) { | |
805 | pr_err("failed to send start command to the host\n"); | |
beda94da | 806 | return -EIO; |
453dc659 DT |
807 | } |
808 | ||
809 | if (!vmballoon_send_guest_id(&balloon)) { | |
810 | pr_err("failed to send guest ID to the host\n"); | |
beda94da | 811 | return -EIO; |
453dc659 DT |
812 | } |
813 | ||
814 | error = vmballoon_debugfs_init(&balloon); | |
815 | if (error) | |
beda94da | 816 | return error; |
453dc659 | 817 | |
beda94da | 818 | queue_delayed_work(system_freezable_wq, &balloon.dwork, 0); |
453dc659 DT |
819 | |
820 | return 0; | |
453dc659 DT |
821 | } |
822 | module_init(vmballoon_init); | |
823 | ||
824 | static void __exit vmballoon_exit(void) | |
825 | { | |
826 | cancel_delayed_work_sync(&balloon.dwork); | |
453dc659 DT |
827 | |
828 | vmballoon_debugfs_exit(&balloon); | |
829 | ||
830 | /* | |
831 | * Deallocate all reserved memory, and reset connection with monitor. | |
832 | * Reset connection before deallocating memory to avoid potential for | |
833 | * additional spurious resets from guest touching deallocated pages. | |
834 | */ | |
835 | vmballoon_send_start(&balloon); | |
836 | vmballoon_pop(&balloon); | |
837 | } | |
838 | module_exit(vmballoon_exit); |