Linux Audio

Check our new training course

Loading...
v3.1
 
  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>
 
 
 
 
 
 
 
 
 44#include <asm/hypervisor.h>
 45
 46MODULE_AUTHOR("VMware, Inc.");
 47MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
 48MODULE_VERSION("1.2.1.3-k");
 49MODULE_ALIAS("dmi:*:svnVMware*:*");
 50MODULE_ALIAS("vmware_vmmemctl");
 51MODULE_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
104/* Maximum number of refused pages we accumulate during inflation cycle */
105#define VMW_BALLOON_MAX_REFUSED		16
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;	\
136	__asm__ __volatile__ ("inl (%%dx)" :		\
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
151struct vmballoon_stats {
152	unsigned int timer;
 
 
 
153
154	/* allocation statustics */
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
181struct 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;
 
188	unsigned int n_refused_pages;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
217static struct vmballoon balloon;
218
219/*
220 * Send "start" command to the host, communicating supported version
221 * of the protocol.
222 */
223static 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
238static 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 */
259static 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 */
278static 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 */
317static bool vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn,
318				     unsigned int *hv_status)
 
319{
320	unsigned long status, dummy;
321	u32 pfn32;
 
322
323	pfn32 = (u32)pfn;
324	if (pfn32 != pfn)
325		return false;
326
327	STATS_INC(b->stats.lock);
 
 
328
329	*hv_status = status = VMWARE_BALLOON_CMD(LOCK, pfn, dummy);
330	if (vmballoon_check_status(b, status))
331		return true;
332
333	pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
334	STATS_INC(b->stats.lock_fail);
335	return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 */
342static 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 */
368static 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 */
391static 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 */
409static int vmballoon_reserve_page(struct vmballoon *b, bool can_sleep)
410{
411	struct page *page;
412	gfp_t flags;
413	unsigned int hv_status;
414	bool locked = false;
415
416	do {
417		if (!can_sleep)
418			STATS_INC(b->stats.alloc);
419		else
420			STATS_INC(b->stats.sleep_alloc);
421
422		flags = can_sleep ? VMW_PAGE_ALLOC_CANSLEEP : VMW_PAGE_ALLOC_NOSLEEP;
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 */
433		locked = vmballoon_send_lock_page(b, page_to_pfn(page), &hv_status);
434		if (!locked) {
435			STATS_INC(b->stats.refused_alloc);
436
437			if (hv_status == VMW_BALLOON_ERROR_RESET ||
438			    hv_status == VMW_BALLOON_ERROR_PPN_NOTNEEDED) {
439				__free_page(page);
440				return -EIO;
441			}
442
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			 */
448			list_add(&page->lru, &b->refused_pages);
449			if (++b->n_refused_pages >= VMW_BALLOON_MAX_REFUSED)
450				return -EIO;
451		}
452	} while (!locked);
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 */
468static 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 */
489static 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	}
498
499	b->n_refused_pages = 0;
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 */
507static 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 */
615static 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 */
652static 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
 
 
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));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
682}
683
684/*
685 * DEBUGFS Interface
686 */
687#ifdef CONFIG_DEBUG_FS
688
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
689static 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
735static int vmballoon_debug_open(struct inode *inode, struct file *file)
736{
737	return single_open(file, vmballoon_debug_show, inode->i_private);
 
 
 
 
 
 
 
 
 
 
 
738}
739
740static 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
748static 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
763static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
764{
 
765	debugfs_remove(b->dbg_entry);
 
 
766}
767
768#else
769
770static inline int vmballoon_debugfs_init(struct vmballoon *b)
771{
772	return 0;
773}
774
775static inline void vmballoon_debugfs_exit(struct vmballoon *b)
776{
777}
778
779#endif	/* CONFIG_DEBUG_FS */
780
781static 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	 */
789	if (x86_hyper != &x86_hyper_vmware)
790		return -ENODEV;
791
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");
806		return -EIO;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
807	}
808
809	if (!vmballoon_send_guest_id(&balloon)) {
810		pr_err("failed to send guest ID to the host\n");
811		return -EIO;
 
 
 
 
 
 
 
 
812	}
813
814	error = vmballoon_debugfs_init(&balloon);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
815	if (error)
816		return error;
 
 
 
 
 
 
 
 
817
818	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
819
 
 
820	return 0;
 
 
 
 
821}
822module_init(vmballoon_init);
 
 
 
 
 
 
 
823
824static void __exit vmballoon_exit(void)
825{
 
 
826	cancel_delayed_work_sync(&balloon.dwork);
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}
838module_exit(vmballoon_exit);
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * VMware Balloon driver.
   4 *
   5 * Copyright (C) 2000-2018, VMware, Inc. All Rights Reserved.
   6 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   7 * This is VMware physical memory management driver for Linux. The driver
   8 * acts like a "balloon" that can be inflated to reclaim physical pages by
   9 * reserving them in the guest and invalidating them in the monitor,
  10 * freeing up the underlying machine pages so they can be allocated to
  11 * other guests.  The balloon can also be deflated to allow the guest to
  12 * use more physical memory. Higher level policies can control the sizes
  13 * of balloons in VMs in order to manage physical memory resources.
  14 */
  15
  16//#define DEBUG
  17#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  18
  19#include <linux/types.h>
  20#include <linux/io.h>
  21#include <linux/kernel.h>
  22#include <linux/mm.h>
  23#include <linux/vmalloc.h>
  24#include <linux/sched.h>
  25#include <linux/module.h>
  26#include <linux/workqueue.h>
  27#include <linux/debugfs.h>
  28#include <linux/seq_file.h>
  29#include <linux/rwsem.h>
  30#include <linux/slab.h>
  31#include <linux/spinlock.h>
  32#include <linux/mount.h>
  33#include <linux/pseudo_fs.h>
  34#include <linux/balloon_compaction.h>
  35#include <linux/vmw_vmci_defs.h>
  36#include <linux/vmw_vmci_api.h>
  37#include <asm/hypervisor.h>
  38
  39MODULE_AUTHOR("VMware, Inc.");
  40MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
 
  41MODULE_ALIAS("dmi:*:svnVMware*:*");
  42MODULE_ALIAS("vmware_vmmemctl");
  43MODULE_LICENSE("GPL");
  44
  45static bool __read_mostly vmwballoon_shrinker_enable;
  46module_param(vmwballoon_shrinker_enable, bool, 0444);
  47MODULE_PARM_DESC(vmwballoon_shrinker_enable,
  48	"Enable non-cooperative out-of-memory protection. Disabled by default as it may degrade performance.");
  49
  50/* Delay in seconds after shrink before inflation. */
  51#define VMBALLOON_SHRINK_DELAY		(5)
 
 
 
  52
  53/* Maximum number of refused pages we accumulate during inflation cycle */
  54#define VMW_BALLOON_MAX_REFUSED		16
 
 
 
 
 
  55
  56/* Magic number for the balloon mount-point */
  57#define BALLOON_VMW_MAGIC		0x0ba11007
 
 
 
 
  58
  59/*
  60 * Hypervisor communication port definitions.
 
  61 */
  62#define VMW_BALLOON_HV_PORT		0x5670
  63#define VMW_BALLOON_HV_MAGIC		0x456c6d6f
  64#define VMW_BALLOON_GUEST_ID		1	/* Linux */
  65
  66enum vmwballoon_capabilities {
  67	/*
  68	 * Bit 0 is reserved and not associated to any capability.
  69	 */
  70	VMW_BALLOON_BASIC_CMDS			= (1 << 1),
  71	VMW_BALLOON_BATCHED_CMDS		= (1 << 2),
  72	VMW_BALLOON_BATCHED_2M_CMDS		= (1 << 3),
  73	VMW_BALLOON_SIGNALLED_WAKEUP_CMD	= (1 << 4),
  74	VMW_BALLOON_64_BIT_TARGET		= (1 << 5)
  75};
  76
  77#define VMW_BALLOON_CAPABILITIES_COMMON	(VMW_BALLOON_BASIC_CMDS \
  78					| VMW_BALLOON_BATCHED_CMDS \
  79					| VMW_BALLOON_BATCHED_2M_CMDS \
  80					| VMW_BALLOON_SIGNALLED_WAKEUP_CMD)
  81
  82#define VMW_BALLOON_2M_ORDER		(PMD_SHIFT - PAGE_SHIFT)
  83
  84/*
  85 * 64-bit targets are only supported in 64-bit
 
 
 
 
  86 */
  87#ifdef CONFIG_64BIT
  88#define VMW_BALLOON_CAPABILITIES	(VMW_BALLOON_CAPABILITIES_COMMON \
  89					| VMW_BALLOON_64_BIT_TARGET)
  90#else
  91#define VMW_BALLOON_CAPABILITIES	VMW_BALLOON_CAPABILITIES_COMMON
  92#endif
  93
  94enum vmballoon_page_size_type {
  95	VMW_BALLOON_4K_PAGE,
  96	VMW_BALLOON_2M_PAGE,
  97	VMW_BALLOON_LAST_SIZE = VMW_BALLOON_2M_PAGE
  98};
  99
 100#define VMW_BALLOON_NUM_PAGE_SIZES	(VMW_BALLOON_LAST_SIZE + 1)
 
 101
 102static const char * const vmballoon_page_size_names[] = {
 103	[VMW_BALLOON_4K_PAGE]			= "4k",
 104	[VMW_BALLOON_2M_PAGE]			= "2M"
 105};
 
 
 
 106
 107enum vmballoon_op {
 108	VMW_BALLOON_INFLATE,
 109	VMW_BALLOON_DEFLATE
 110};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 111
 112enum vmballoon_op_stat_type {
 113	VMW_BALLOON_OP_STAT,
 114	VMW_BALLOON_OP_FAIL_STAT
 115};
 116
 117#define VMW_BALLOON_OP_STAT_TYPES	(VMW_BALLOON_OP_FAIL_STAT + 1)
 118
 119/**
 120 * enum vmballoon_cmd_type - backdoor commands.
 121 *
 122 * Availability of the commands is as followed:
 123 *
 124 * %VMW_BALLOON_CMD_START, %VMW_BALLOON_CMD_GET_TARGET and
 125 * %VMW_BALLOON_CMD_GUEST_ID are always available.
 126 *
 127 * If the host reports %VMW_BALLOON_BASIC_CMDS are supported then
 128 * %VMW_BALLOON_CMD_LOCK and %VMW_BALLOON_CMD_UNLOCK commands are available.
 129 *
 130 * If the host reports %VMW_BALLOON_BATCHED_CMDS are supported then
 131 * %VMW_BALLOON_CMD_BATCHED_LOCK and VMW_BALLOON_CMD_BATCHED_UNLOCK commands
 132 * are available.
 133 *
 134 * If the host reports %VMW_BALLOON_BATCHED_2M_CMDS are supported then
 135 * %VMW_BALLOON_CMD_BATCHED_2M_LOCK and %VMW_BALLOON_CMD_BATCHED_2M_UNLOCK
 136 * are supported.
 137 *
 138 * If the host reports  VMW_BALLOON_SIGNALLED_WAKEUP_CMD is supported then
 139 * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command is supported.
 140 *
 141 * @VMW_BALLOON_CMD_START: Communicating supported version with the hypervisor.
 142 * @VMW_BALLOON_CMD_GET_TARGET: Gets the balloon target size.
 143 * @VMW_BALLOON_CMD_LOCK: Informs the hypervisor about a ballooned page.
 144 * @VMW_BALLOON_CMD_UNLOCK: Informs the hypervisor about a page that is about
 145 *			    to be deflated from the balloon.
 146 * @VMW_BALLOON_CMD_GUEST_ID: Informs the hypervisor about the type of OS that
 147 *			      runs in the VM.
 148 * @VMW_BALLOON_CMD_BATCHED_LOCK: Inform the hypervisor about a batch of
 149 *				  ballooned pages (up to 512).
 150 * @VMW_BALLOON_CMD_BATCHED_UNLOCK: Inform the hypervisor about a batch of
 151 *				  pages that are about to be deflated from the
 152 *				  balloon (up to 512).
 153 * @VMW_BALLOON_CMD_BATCHED_2M_LOCK: Similar to @VMW_BALLOON_CMD_BATCHED_LOCK
 154 *				     for 2MB pages.
 155 * @VMW_BALLOON_CMD_BATCHED_2M_UNLOCK: Similar to
 156 *				       @VMW_BALLOON_CMD_BATCHED_UNLOCK for 2MB
 157 *				       pages.
 158 * @VMW_BALLOON_CMD_VMCI_DOORBELL_SET: A command to set doorbell notification
 159 *				       that would be invoked when the balloon
 160 *				       size changes.
 161 * @VMW_BALLOON_CMD_LAST: Value of the last command.
 162 */
 163enum vmballoon_cmd_type {
 164	VMW_BALLOON_CMD_START,
 165	VMW_BALLOON_CMD_GET_TARGET,
 166	VMW_BALLOON_CMD_LOCK,
 167	VMW_BALLOON_CMD_UNLOCK,
 168	VMW_BALLOON_CMD_GUEST_ID,
 169	/* No command 5 */
 170	VMW_BALLOON_CMD_BATCHED_LOCK = 6,
 171	VMW_BALLOON_CMD_BATCHED_UNLOCK,
 172	VMW_BALLOON_CMD_BATCHED_2M_LOCK,
 173	VMW_BALLOON_CMD_BATCHED_2M_UNLOCK,
 174	VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
 175	VMW_BALLOON_CMD_LAST = VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
 176};
 177
 178#define VMW_BALLOON_CMD_NUM	(VMW_BALLOON_CMD_LAST + 1)
 
 
 
 179
 180enum vmballoon_error_codes {
 181	VMW_BALLOON_SUCCESS,
 182	VMW_BALLOON_ERROR_CMD_INVALID,
 183	VMW_BALLOON_ERROR_PPN_INVALID,
 184	VMW_BALLOON_ERROR_PPN_LOCKED,
 185	VMW_BALLOON_ERROR_PPN_UNLOCKED,
 186	VMW_BALLOON_ERROR_PPN_PINNED,
 187	VMW_BALLOON_ERROR_PPN_NOTNEEDED,
 188	VMW_BALLOON_ERROR_RESET,
 189	VMW_BALLOON_ERROR_BUSY
 190};
 191
 192#define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES	(0x03000000)
 193
 194#define VMW_BALLOON_CMD_WITH_TARGET_MASK			\
 195	((1UL << VMW_BALLOON_CMD_GET_TARGET)		|	\
 196	 (1UL << VMW_BALLOON_CMD_LOCK)			|	\
 197	 (1UL << VMW_BALLOON_CMD_UNLOCK)		|	\
 198	 (1UL << VMW_BALLOON_CMD_BATCHED_LOCK)		|	\
 199	 (1UL << VMW_BALLOON_CMD_BATCHED_UNLOCK)	|	\
 200	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_LOCK)	|	\
 201	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_UNLOCK))
 202
 203static const char * const vmballoon_cmd_names[] = {
 204	[VMW_BALLOON_CMD_START]			= "start",
 205	[VMW_BALLOON_CMD_GET_TARGET]		= "target",
 206	[VMW_BALLOON_CMD_LOCK]			= "lock",
 207	[VMW_BALLOON_CMD_UNLOCK]		= "unlock",
 208	[VMW_BALLOON_CMD_GUEST_ID]		= "guestType",
 209	[VMW_BALLOON_CMD_BATCHED_LOCK]		= "batchLock",
 210	[VMW_BALLOON_CMD_BATCHED_UNLOCK]	= "batchUnlock",
 211	[VMW_BALLOON_CMD_BATCHED_2M_LOCK]	= "2m-lock",
 212	[VMW_BALLOON_CMD_BATCHED_2M_UNLOCK]	= "2m-unlock",
 213	[VMW_BALLOON_CMD_VMCI_DOORBELL_SET]	= "doorbellSet"
 214};
 215
 216enum vmballoon_stat_page {
 217	VMW_BALLOON_PAGE_STAT_ALLOC,
 218	VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
 219	VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
 220	VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
 221	VMW_BALLOON_PAGE_STAT_FREE,
 222	VMW_BALLOON_PAGE_STAT_LAST = VMW_BALLOON_PAGE_STAT_FREE
 223};
 224
 225#define VMW_BALLOON_PAGE_STAT_NUM	(VMW_BALLOON_PAGE_STAT_LAST + 1)
 226
 227enum vmballoon_stat_general {
 228	VMW_BALLOON_STAT_TIMER,
 229	VMW_BALLOON_STAT_DOORBELL,
 230	VMW_BALLOON_STAT_RESET,
 231	VMW_BALLOON_STAT_SHRINK,
 232	VMW_BALLOON_STAT_SHRINK_FREE,
 233	VMW_BALLOON_STAT_LAST = VMW_BALLOON_STAT_SHRINK_FREE
 234};
 235
 236#define VMW_BALLOON_STAT_NUM		(VMW_BALLOON_STAT_LAST + 1)
 237
 238static DEFINE_STATIC_KEY_TRUE(vmw_balloon_batching);
 239static DEFINE_STATIC_KEY_FALSE(balloon_stat_enabled);
 240
 241struct vmballoon_ctl {
 242	struct list_head pages;
 243	struct list_head refused_pages;
 244	struct list_head prealloc_pages;
 245	unsigned int n_refused_pages;
 246	unsigned int n_pages;
 247	enum vmballoon_page_size_type page_size;
 248	enum vmballoon_op op;
 249};
 250
 251/**
 252 * struct vmballoon_batch_entry - a batch entry for lock or unlock.
 253 *
 254 * @status: the status of the operation, which is written by the hypervisor.
 255 * @reserved: reserved for future use. Must be set to zero.
 256 * @pfn: the physical frame number of the page to be locked or unlocked.
 257 */
 258struct vmballoon_batch_entry {
 259	u64 status : 5;
 260	u64 reserved : PAGE_SHIFT - 5;
 261	u64 pfn : 52;
 262} __packed;
 263
 264struct vmballoon {
 265	/**
 266	 * @max_page_size: maximum supported page size for ballooning.
 267	 *
 268	 * Protected by @conf_sem
 269	 */
 270	enum vmballoon_page_size_type max_page_size;
 271
 272	/**
 273	 * @size: balloon actual size in basic page size (frames).
 274	 *
 275	 * While we currently do not support size which is bigger than 32-bit,
 276	 * in preparation for future support, use 64-bits.
 277	 */
 278	atomic64_t size;
 279
 280	/**
 281	 * @target: balloon target size in basic page size (frames).
 282	 *
 283	 * We do not protect the target under the assumption that setting the
 284	 * value is always done through a single write. If this assumption ever
 285	 * breaks, we would have to use X_ONCE for accesses, and suffer the less
 286	 * optimized code. Although we may read stale target value if multiple
 287	 * accesses happen at once, the performance impact should be minor.
 288	 */
 289	unsigned long target;
 290
 291	/**
 292	 * @reset_required: reset flag
 293	 *
 294	 * Setting this flag may introduce races, but the code is expected to
 295	 * handle them gracefully. In the worst case, another operation will
 296	 * fail as reset did not take place. Clearing the flag is done while
 297	 * holding @conf_sem for write.
 298	 */
 299	bool reset_required;
 300
 301	/**
 302	 * @capabilities: hypervisor balloon capabilities.
 303	 *
 304	 * Protected by @conf_sem.
 305	 */
 306	unsigned long capabilities;
 307
 308	/**
 309	 * @batch_page: pointer to communication batch page.
 310	 *
 311	 * When batching is used, batch_page points to a page, which holds up to
 312	 * %VMW_BALLOON_BATCH_MAX_PAGES entries for locking or unlocking.
 313	 */
 314	struct vmballoon_batch_entry *batch_page;
 315
 316	/**
 317	 * @batch_max_pages: maximum pages that can be locked/unlocked.
 318	 *
 319	 * Indicates the number of pages that the hypervisor can lock or unlock
 320	 * at once, according to whether batching is enabled. If batching is
 321	 * disabled, only a single page can be locked/unlock on each operation.
 322	 *
 323	 * Protected by @conf_sem.
 324	 */
 325	unsigned int batch_max_pages;
 326
 327	/**
 328	 * @page: page to be locked/unlocked by the hypervisor
 329	 *
 330	 * @page is only used when batching is disabled and a single page is
 331	 * reclaimed on each iteration.
 332	 *
 333	 * Protected by @comm_lock.
 334	 */
 335	struct page *page;
 336
 337	/**
 338	 * @shrink_timeout: timeout until the next inflation.
 339	 *
 340	 * After an shrink event, indicates the time in jiffies after which
 341	 * inflation is allowed again. Can be written concurrently with reads,
 342	 * so must use READ_ONCE/WRITE_ONCE when accessing.
 343	 */
 344	unsigned long shrink_timeout;
 345
 
 346	/* statistics */
 347	struct vmballoon_stats *stats;
 348
 349#ifdef CONFIG_DEBUG_FS
 350	/* debugfs file exporting statistics */
 351	struct dentry *dbg_entry;
 352#endif
 353
 354	/**
 355	 * @b_dev_info: balloon device information descriptor.
 356	 */
 357	struct balloon_dev_info b_dev_info;
 358
 359	struct delayed_work dwork;
 360
 361	/**
 362	 * @huge_pages - list of the inflated 2MB pages.
 363	 *
 364	 * Protected by @b_dev_info.pages_lock .
 365	 */
 366	struct list_head huge_pages;
 367
 368	/**
 369	 * @vmci_doorbell.
 370	 *
 371	 * Protected by @conf_sem.
 372	 */
 373	struct vmci_handle vmci_doorbell;
 374
 375	/**
 376	 * @conf_sem: semaphore to protect the configuration and the statistics.
 377	 */
 378	struct rw_semaphore conf_sem;
 379
 380	/**
 381	 * @comm_lock: lock to protect the communication with the host.
 382	 *
 383	 * Lock ordering: @conf_sem -> @comm_lock .
 384	 */
 385	spinlock_t comm_lock;
 386
 387	/**
 388	 * @shrinker: shrinker interface that is used to avoid over-inflation.
 389	 */
 390	struct shrinker shrinker;
 391
 392	/**
 393	 * @shrinker_registered: whether the shrinker was registered.
 394	 *
 395	 * The shrinker interface does not handle gracefully the removal of
 396	 * shrinker that was not registered before. This indication allows to
 397	 * simplify the unregistration process.
 398	 */
 399	bool shrinker_registered;
 400};
 401
 402static struct vmballoon balloon;
 403
 404struct vmballoon_stats {
 405	/* timer / doorbell operations */
 406	atomic64_t general_stat[VMW_BALLOON_STAT_NUM];
 407
 408	/* allocation statistics for huge and small pages */
 409	atomic64_t
 410	       page_stat[VMW_BALLOON_PAGE_STAT_NUM][VMW_BALLOON_NUM_PAGE_SIZES];
 411
 412	/* Monitor operations: total operations, and failures */
 413	atomic64_t ops[VMW_BALLOON_CMD_NUM][VMW_BALLOON_OP_STAT_TYPES];
 414};
 415
 416static inline bool is_vmballoon_stats_on(void)
 417{
 418	return IS_ENABLED(CONFIG_DEBUG_FS) &&
 419		static_branch_unlikely(&balloon_stat_enabled);
 420}
 421
 422static inline void vmballoon_stats_op_inc(struct vmballoon *b, unsigned int op,
 423					  enum vmballoon_op_stat_type type)
 424{
 425	if (is_vmballoon_stats_on())
 426		atomic64_inc(&b->stats->ops[op][type]);
 427}
 428
 429static inline void vmballoon_stats_gen_inc(struct vmballoon *b,
 430					   enum vmballoon_stat_general stat)
 431{
 432	if (is_vmballoon_stats_on())
 433		atomic64_inc(&b->stats->general_stat[stat]);
 434}
 435
 436static inline void vmballoon_stats_gen_add(struct vmballoon *b,
 437					   enum vmballoon_stat_general stat,
 438					   unsigned int val)
 439{
 440	if (is_vmballoon_stats_on())
 441		atomic64_add(val, &b->stats->general_stat[stat]);
 442}
 443
 444static inline void vmballoon_stats_page_inc(struct vmballoon *b,
 445					    enum vmballoon_stat_page stat,
 446					    enum vmballoon_page_size_type size)
 447{
 448	if (is_vmballoon_stats_on())
 449		atomic64_inc(&b->stats->page_stat[stat][size]);
 450}
 451
 452static inline void vmballoon_stats_page_add(struct vmballoon *b,
 453					    enum vmballoon_stat_page stat,
 454					    enum vmballoon_page_size_type size,
 455					    unsigned int val)
 456{
 457	if (is_vmballoon_stats_on())
 458		atomic64_add(val, &b->stats->page_stat[stat][size]);
 459}
 460
 461static inline unsigned long
 462__vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
 463		unsigned long arg2, unsigned long *result)
 464{
 465	unsigned long status, dummy1, dummy2, dummy3, local_result;
 466
 467	vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_STAT);
 468
 469	asm volatile ("inl %%dx" :
 470		"=a"(status),
 471		"=c"(dummy1),
 472		"=d"(dummy2),
 473		"=b"(local_result),
 474		"=S"(dummy3) :
 475		"0"(VMW_BALLOON_HV_MAGIC),
 476		"1"(cmd),
 477		"2"(VMW_BALLOON_HV_PORT),
 478		"3"(arg1),
 479		"4"(arg2) :
 480		"memory");
 481
 482	/* update the result if needed */
 483	if (result)
 484		*result = (cmd == VMW_BALLOON_CMD_START) ? dummy1 :
 485							   local_result;
 486
 487	/* update target when applicable */
 488	if (status == VMW_BALLOON_SUCCESS &&
 489	    ((1ul << cmd) & VMW_BALLOON_CMD_WITH_TARGET_MASK))
 490		WRITE_ONCE(b->target, local_result);
 491
 492	if (status != VMW_BALLOON_SUCCESS &&
 493	    status != VMW_BALLOON_SUCCESS_WITH_CAPABILITIES) {
 494		vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_FAIL_STAT);
 495		pr_debug("%s: %s [0x%lx,0x%lx) failed, returned %ld\n",
 496			 __func__, vmballoon_cmd_names[cmd], arg1, arg2,
 497			 status);
 498	}
 499
 500	/* mark reset required accordingly */
 501	if (status == VMW_BALLOON_ERROR_RESET)
 502		b->reset_required = true;
 
 503
 504	return status;
 505}
 506
 507static __always_inline unsigned long
 508vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
 509	      unsigned long arg2)
 510{
 511	unsigned long dummy;
 512
 513	return __vmballoon_cmd(b, cmd, arg1, arg2, &dummy);
 514}
 515
 516/*
 517 * Send "start" command to the host, communicating supported version
 518 * of the protocol.
 519 */
 520static int vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
 521{
 522	unsigned long status, capabilities;
 523
 524	status = __vmballoon_cmd(b, VMW_BALLOON_CMD_START, req_caps, 0,
 525				 &capabilities);
 526
 527	switch (status) {
 528	case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
 529		b->capabilities = capabilities;
 530		break;
 531	case VMW_BALLOON_SUCCESS:
 532		b->capabilities = VMW_BALLOON_BASIC_CMDS;
 533		break;
 534	default:
 535		return -EIO;
 536	}
 537
 538	/*
 539	 * 2MB pages are only supported with batching. If batching is for some
 540	 * reason disabled, do not use 2MB pages, since otherwise the legacy
 541	 * mechanism is used with 2MB pages, causing a failure.
 542	 */
 543	b->max_page_size = VMW_BALLOON_4K_PAGE;
 544	if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
 545	    (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
 546		b->max_page_size = VMW_BALLOON_2M_PAGE;
 547
 548
 549	return 0;
 550}
 551
 552/**
 553 * vmballoon_send_guest_id - communicate guest type to the host.
 554 *
 555 * @b: pointer to the balloon.
 556 *
 557 * Communicate guest type to the host so that it can adjust ballooning
 558 * algorithm to the one most appropriate for the guest. This command
 559 * is normally issued after sending "start" command and is part of
 560 * standard reset sequence.
 561 *
 562 * Return: zero on success or appropriate error code.
 563 */
 564static int vmballoon_send_guest_id(struct vmballoon *b)
 565{
 566	unsigned long status;
 567
 568	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GUEST_ID,
 569			       VMW_BALLOON_GUEST_ID, 0);
 570
 571	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
 572}
 573
 574/**
 575 * vmballoon_page_order() - return the order of the page
 576 * @page_size: the size of the page.
 577 *
 578 * Return: the allocation order.
 579 */
 580static inline
 581unsigned int vmballoon_page_order(enum vmballoon_page_size_type page_size)
 582{
 583	return page_size == VMW_BALLOON_2M_PAGE ? VMW_BALLOON_2M_ORDER : 0;
 584}
 585
 586/**
 587 * vmballoon_page_in_frames() - returns the number of frames in a page.
 588 * @page_size: the size of the page.
 589 *
 590 * Return: the number of 4k frames.
 591 */
 592static inline unsigned int
 593vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)
 594{
 595	return 1 << vmballoon_page_order(page_size);
 596}
 597
 598/**
 599 * vmballoon_mark_page_offline() - mark a page as offline
 600 * @page: pointer for the page.
 601 * @page_size: the size of the page.
 602 */
 603static void
 604vmballoon_mark_page_offline(struct page *page,
 605			    enum vmballoon_page_size_type page_size)
 606{
 607	int i;
 608
 609	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
 610		__SetPageOffline(page + i);
 611}
 612
 613/**
 614 * vmballoon_mark_page_online() - mark a page as online
 615 * @page: pointer for the page.
 616 * @page_size: the size of the page.
 617 */
 618static void
 619vmballoon_mark_page_online(struct page *page,
 620			   enum vmballoon_page_size_type page_size)
 621{
 622	int i;
 623
 624	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
 625		__ClearPageOffline(page + i);
 626}
 627
 628/**
 629 * vmballoon_send_get_target() - Retrieve desired balloon size from the host.
 630 *
 631 * @b: pointer to the balloon.
 632 *
 633 * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
 634 * by the host-guest protocol and EIO if an error occurred in communicating with
 635 * the host.
 636 */
 637static int vmballoon_send_get_target(struct vmballoon *b)
 638{
 639	unsigned long status;
 
 640	unsigned long limit;
 
 641
 642	limit = totalram_pages();
 643
 644	/* Ensure limit fits in 32-bits if 64-bit targets are not supported */
 645	if (!(b->capabilities & VMW_BALLOON_64_BIT_TARGET) &&
 646	    limit != (u32)limit)
 647		return -EINVAL;
 648
 649	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
 650
 651	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
 652}
 653
 654/**
 655 * vmballoon_alloc_page_list - allocates a list of pages.
 656 *
 657 * @b: pointer to the balloon.
 658 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
 659 * @req_n_pages: the number of requested pages.
 660 *
 661 * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
 662 * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
 663 *
 664 * Return: zero on success or error code otherwise.
 665 */
 666static int vmballoon_alloc_page_list(struct vmballoon *b,
 667				     struct vmballoon_ctl *ctl,
 668				     unsigned int req_n_pages)
 669{
 670	struct page *page;
 671	unsigned int i;
 672
 673	for (i = 0; i < req_n_pages; i++) {
 674		/*
 675		 * First check if we happen to have pages that were allocated
 676		 * before. This happens when 2MB page rejected during inflation
 677		 * by the hypervisor, and then split into 4KB pages.
 678		 */
 679		if (!list_empty(&ctl->prealloc_pages)) {
 680			page = list_first_entry(&ctl->prealloc_pages,
 681						struct page, lru);
 682			list_del(&page->lru);
 683		} else {
 684			if (ctl->page_size == VMW_BALLOON_2M_PAGE)
 685				page = alloc_pages(__GFP_HIGHMEM|__GFP_NOWARN|
 686					__GFP_NOMEMALLOC, VMW_BALLOON_2M_ORDER);
 687			else
 688				page = balloon_page_alloc();
 689
 690			vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
 691						 ctl->page_size);
 692		}
 
 693
 694		if (page) {
 695			/* Success. Add the page to the list and continue. */
 696			list_add(&page->lru, &ctl->pages);
 697			continue;
 698		}
 699
 700		/* Allocation failed. Update statistics and stop. */
 701		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
 702					 ctl->page_size);
 703		break;
 704	}
 705
 706	ctl->n_pages = i;
 707
 708	return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
 709}
 710
 711/**
 712 * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
 713 *
 714 * @b: pointer for %struct vmballoon.
 715 * @page: pointer for the page whose result should be handled.
 716 * @page_size: size of the page.
 717 * @status: status of the operation as provided by the hypervisor.
 718 */
 719static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
 720				       enum vmballoon_page_size_type page_size,
 721				       unsigned long status)
 722{
 723	/* On success do nothing. The page is already on the balloon list. */
 724	if (likely(status == VMW_BALLOON_SUCCESS))
 725		return 0;
 726
 727	pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
 728		 page_to_pfn(page), status,
 729		 vmballoon_page_size_names[page_size]);
 730
 731	/* Error occurred */
 732	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
 733				 page_size);
 734
 735	return -EIO;
 736}
 
 737
 738/**
 739 * vmballoon_status_page - returns the status of (un)lock operation
 740 *
 741 * @b: pointer to the balloon.
 742 * @idx: index for the page for which the operation is performed.
 743 * @p: pointer to where the page struct is returned.
 744 *
 745 * Following a lock or unlock operation, returns the status of the operation for
 746 * an individual page. Provides the page that the operation was performed on on
 747 * the @page argument.
 748 *
 749 * Returns: The status of a lock or unlock operation for an individual page.
 750 */
 751static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
 752					   struct page **p)
 753{
 754	if (static_branch_likely(&vmw_balloon_batching)) {
 755		/* batching mode */
 756		*p = pfn_to_page(b->batch_page[idx].pfn);
 757		return b->batch_page[idx].status;
 758	}
 759
 760	/* non-batching mode */
 761	*p = b->page;
 762
 763	/*
 764	 * If a failure occurs, the indication will be provided in the status
 765	 * of the entire operation, which is considered before the individual
 766	 * page status. So for non-batching mode, the indication is always of
 767	 * success.
 768	 */
 769	return VMW_BALLOON_SUCCESS;
 770}
 771
 772/**
 773 * vmballoon_lock_op - notifies the host about inflated/deflated pages.
 774 * @b: pointer to the balloon.
 775 * @num_pages: number of inflated/deflated pages.
 776 * @page_size: size of the page.
 777 * @op: the type of operation (lock or unlock).
 778 *
 779 * Notify the host about page(s) that were ballooned (or removed from the
 780 * balloon) so that host can use it without fear that guest will need it (or
 781 * stop using them since the VM does). Host may reject some pages, we need to
 782 * check the return value and maybe submit a different page. The pages that are
 783 * inflated/deflated are pointed by @b->page.
 784 *
 785 * Return: result as provided by the hypervisor.
 786 */
 787static unsigned long vmballoon_lock_op(struct vmballoon *b,
 788				       unsigned int num_pages,
 789				       enum vmballoon_page_size_type page_size,
 790				       enum vmballoon_op op)
 791{
 792	unsigned long cmd, pfn;
 793
 794	lockdep_assert_held(&b->comm_lock);
 795
 796	if (static_branch_likely(&vmw_balloon_batching)) {
 797		if (op == VMW_BALLOON_INFLATE)
 798			cmd = page_size == VMW_BALLOON_2M_PAGE ?
 799				VMW_BALLOON_CMD_BATCHED_2M_LOCK :
 800				VMW_BALLOON_CMD_BATCHED_LOCK;
 801		else
 802			cmd = page_size == VMW_BALLOON_2M_PAGE ?
 803				VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
 804				VMW_BALLOON_CMD_BATCHED_UNLOCK;
 805
 806		pfn = PHYS_PFN(virt_to_phys(b->batch_page));
 807	} else {
 808		cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
 809						  VMW_BALLOON_CMD_UNLOCK;
 810		pfn = page_to_pfn(b->page);
 811
 812		/* In non-batching mode, PFNs must fit in 32-bit */
 813		if (unlikely(pfn != (u32)pfn))
 814			return VMW_BALLOON_ERROR_PPN_INVALID;
 815	}
 816
 817	return vmballoon_cmd(b, cmd, pfn, num_pages);
 818}
 819
 820/**
 821 * vmballoon_add_page - adds a page towards lock/unlock operation.
 822 *
 823 * @b: pointer to the balloon.
 824 * @idx: index of the page to be ballooned in this batch.
 825 * @p: pointer to the page that is about to be ballooned.
 826 *
 827 * Adds the page to be ballooned. Must be called while holding @comm_lock.
 828 */
 829static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
 830			       struct page *p)
 831{
 832	lockdep_assert_held(&b->comm_lock);
 
 833
 834	if (static_branch_likely(&vmw_balloon_batching))
 835		b->batch_page[idx] = (struct vmballoon_batch_entry)
 836					{ .pfn = page_to_pfn(p) };
 837	else
 838		b->page = p;
 839}
 840
 841/**
 842 * vmballoon_lock - lock or unlock a batch of pages.
 843 *
 844 * @b: pointer to the balloon.
 845 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
 846 *
 847 * Notifies the host of about ballooned pages (after inflation or deflation,
 848 * according to @ctl). If the host rejects the page put it on the
 849 * @ctl refuse list. These refused page are then released when moving to the
 850 * next size of pages.
 851 *
 852 * Note that we neither free any @page here nor put them back on the ballooned
 853 * pages list. Instead we queue it for later processing. We do that for several
 854 * reasons. First, we do not want to free the page under the lock. Second, it
 855 * allows us to unify the handling of lock and unlock. In the inflate case, the
 856 * caller will check if there are too many refused pages and release them.
 857 * Although it is not identical to the past behavior, it should not affect
 858 * performance.
 859 */
 860static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
 861{
 862	unsigned long batch_status;
 863	struct page *page;
 864	unsigned int i, num_pages;
 865
 866	num_pages = ctl->n_pages;
 867	if (num_pages == 0)
 868		return 0;
 869
 870	/* communication with the host is done under the communication lock */
 871	spin_lock(&b->comm_lock);
 872
 873	i = 0;
 874	list_for_each_entry(page, &ctl->pages, lru)
 875		vmballoon_add_page(b, i++, page);
 876
 877	batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
 878					 ctl->op);
 879
 880	/*
 881	 * Iterate over the pages in the provided list. Since we are changing
 882	 * @ctl->n_pages we are saving the original value in @num_pages and
 883	 * use this value to bound the loop.
 884	 */
 885	for (i = 0; i < num_pages; i++) {
 886		unsigned long status;
 887
 888		status = vmballoon_status_page(b, i, &page);
 889
 890		/*
 891		 * Failure of the whole batch overrides a single operation
 892		 * results.
 893		 */
 894		if (batch_status != VMW_BALLOON_SUCCESS)
 895			status = batch_status;
 896
 897		/* Continue if no error happened */
 898		if (!vmballoon_handle_one_result(b, page, ctl->page_size,
 899						 status))
 900			continue;
 901
 902		/*
 903		 * Error happened. Move the pages to the refused list and update
 904		 * the pages number.
 905		 */
 906		list_move(&page->lru, &ctl->refused_pages);
 907		ctl->n_pages--;
 908		ctl->n_refused_pages++;
 909	}
 910
 911	spin_unlock(&b->comm_lock);
 912
 913	return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
 914}
 915
 916/**
 917 * vmballoon_release_page_list() - Releases a page list
 918 *
 919 * @page_list: list of pages to release.
 920 * @n_pages: pointer to the number of pages.
 921 * @page_size: whether the pages in the list are 2MB (or else 4KB).
 922 *
 923 * Releases the list of pages and zeros the number of pages.
 924 */
 925static void vmballoon_release_page_list(struct list_head *page_list,
 926				       int *n_pages,
 927				       enum vmballoon_page_size_type page_size)
 928{
 929	struct page *page, *tmp;
 
 930
 931	list_for_each_entry_safe(page, tmp, page_list, lru) {
 932		list_del(&page->lru);
 933		__free_pages(page, vmballoon_page_order(page_size));
 
 
 
 
 
 
 
 934	}
 935
 936	if (n_pages)
 937		*n_pages = 0;
 938}
 939
 940
 941/*
 942 * Release pages that were allocated while attempting to inflate the
 943 * balloon but were refused by the host for one reason or another.
 
 944 */
 945static void vmballoon_release_refused_pages(struct vmballoon *b,
 946					    struct vmballoon_ctl *ctl)
 947{
 948	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
 949				 ctl->page_size);
 950
 951	vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
 952				    ctl->page_size);
 
 
 
 953}
 954
 955/**
 956 * vmballoon_change - retrieve the required balloon change
 957 *
 958 * @b: pointer for the balloon.
 959 *
 960 * Return: the required change for the balloon size. A positive number
 961 * indicates inflation, a negative number indicates a deflation.
 962 */
 963static int64_t vmballoon_change(struct vmballoon *b)
 964{
 965	int64_t size, target;
 
 
 
 
 
 
 
 
 
 966
 967	size = atomic64_read(&b->size);
 968	target = READ_ONCE(b->target);
 
 
 
 
 
 
 
 969
 970	/*
 971	 * We must cast first because of int sizes
 972	 * Otherwise we might get huge positives instead of negatives
 973	 */
 
 
 
 
 
 
 974
 975	if (b->reset_required)
 976		return 0;
 
 
 
 
 
 
 
 
 977
 978	/* consider a 2MB slack on deflate, unless the balloon is emptied */
 979	if (target < size && target != 0 &&
 980	    size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
 981		return 0;
 982
 983	/* If an out-of-memory recently occurred, inflation is disallowed. */
 984	if (target > size && time_before(jiffies, READ_ONCE(b->shrink_timeout)))
 985		return 0;
 986
 987	return target - size;
 988}
 989
 990/**
 991 * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
 992 *
 993 * @b: pointer to balloon.
 994 * @pages: list of pages to enqueue.
 995 * @n_pages: pointer to number of pages in list. The value is zeroed.
 996 * @page_size: whether the pages are 2MB or 4KB pages.
 997 *
 998 * Enqueues the provides list of pages in the ballooned page list, clears the
 999 * list and zeroes the number of pages that was provided.
1000 */
1001static void vmballoon_enqueue_page_list(struct vmballoon *b,
1002					struct list_head *pages,
1003					unsigned int *n_pages,
1004					enum vmballoon_page_size_type page_size)
1005{
1006	unsigned long flags;
1007	struct page *page;
1008
1009	if (page_size == VMW_BALLOON_4K_PAGE) {
1010		balloon_page_list_enqueue(&b->b_dev_info, pages);
1011	} else {
1012		/*
1013		 * Keep the huge pages in a local list which is not available
1014		 * for the balloon compaction mechanism.
1015		 */
1016		spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1017
1018		list_for_each_entry(page, pages, lru) {
1019			vmballoon_mark_page_offline(page, VMW_BALLOON_2M_PAGE);
1020		}
1021
1022		list_splice_init(pages, &b->huge_pages);
1023		__count_vm_events(BALLOON_INFLATE, *n_pages *
1024				  vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1025		spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1026	}
1027
1028	*n_pages = 0;
1029}
1030
1031/**
1032 * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
1033 *
1034 * @b: pointer to balloon.
1035 * @pages: list of pages to enqueue.
1036 * @n_pages: pointer to number of pages in list. The value is zeroed.
1037 * @page_size: whether the pages are 2MB or 4KB pages.
1038 * @n_req_pages: the number of requested pages.
1039 *
1040 * Dequeues the number of requested pages from the balloon for deflation. The
1041 * number of dequeued pages may be lower, if not enough pages in the requested
1042 * size are available.
1043 */
1044static void vmballoon_dequeue_page_list(struct vmballoon *b,
1045					struct list_head *pages,
1046					unsigned int *n_pages,
1047					enum vmballoon_page_size_type page_size,
1048					unsigned int n_req_pages)
1049{
1050	struct page *page, *tmp;
1051	unsigned int i = 0;
1052	unsigned long flags;
1053
1054	/* In the case of 4k pages, use the compaction infrastructure */
1055	if (page_size == VMW_BALLOON_4K_PAGE) {
1056		*n_pages = balloon_page_list_dequeue(&b->b_dev_info, pages,
1057						     n_req_pages);
1058		return;
1059	}
1060
1061	/* 2MB pages */
1062	spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1063	list_for_each_entry_safe(page, tmp, &b->huge_pages, lru) {
1064		vmballoon_mark_page_online(page, VMW_BALLOON_2M_PAGE);
1065
1066		list_move(&page->lru, pages);
1067		if (++i == n_req_pages)
1068			break;
1069	}
1070
1071	__count_vm_events(BALLOON_DEFLATE,
1072			  i * vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1073	spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1074	*n_pages = i;
1075}
1076
1077/**
1078 * vmballoon_split_refused_pages() - Split the 2MB refused pages to 4k.
1079 *
1080 * If inflation of 2MB pages was denied by the hypervisor, it is likely to be
1081 * due to one or few 4KB pages. These 2MB pages may keep being allocated and
1082 * then being refused. To prevent this case, this function splits the refused
1083 * pages into 4KB pages and adds them into @prealloc_pages list.
1084 *
1085 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
1086 */
1087static void vmballoon_split_refused_pages(struct vmballoon_ctl *ctl)
1088{
1089	struct page *page, *tmp;
1090	unsigned int i, order;
1091
1092	order = vmballoon_page_order(ctl->page_size);
1093
1094	list_for_each_entry_safe(page, tmp, &ctl->refused_pages, lru) {
1095		list_del(&page->lru);
1096		split_page(page, order);
1097		for (i = 0; i < (1 << order); i++)
1098			list_add(&page[i].lru, &ctl->prealloc_pages);
1099	}
1100	ctl->n_refused_pages = 0;
 
1101}
1102
1103/**
1104 * vmballoon_inflate() - Inflate the balloon towards its target size.
1105 *
1106 * @b: pointer to the balloon.
1107 */
1108static void vmballoon_inflate(struct vmballoon *b)
1109{
1110	int64_t to_inflate_frames;
1111	struct vmballoon_ctl ctl = {
1112		.pages = LIST_HEAD_INIT(ctl.pages),
1113		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1114		.prealloc_pages = LIST_HEAD_INIT(ctl.prealloc_pages),
1115		.page_size = b->max_page_size,
1116		.op = VMW_BALLOON_INFLATE
1117	};
1118
1119	while ((to_inflate_frames = vmballoon_change(b)) > 0) {
1120		unsigned int to_inflate_pages, page_in_frames;
1121		int alloc_error, lock_error = 0;
1122
1123		VM_BUG_ON(!list_empty(&ctl.pages));
1124		VM_BUG_ON(ctl.n_pages != 0);
1125
1126		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1127
1128		to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
1129					 DIV_ROUND_UP_ULL(to_inflate_frames,
1130							  page_in_frames));
1131
1132		/* Start by allocating */
1133		alloc_error = vmballoon_alloc_page_list(b, &ctl,
1134							to_inflate_pages);
1135
1136		/* Actually lock the pages by telling the hypervisor */
1137		lock_error = vmballoon_lock(b, &ctl);
1138
1139		/*
1140		 * If an error indicates that something serious went wrong,
1141		 * stop the inflation.
1142		 */
1143		if (lock_error)
1144			break;
1145
1146		/* Update the balloon size */
1147		atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1148
1149		vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1150					    ctl.page_size);
1151
1152		/*
1153		 * If allocation failed or the number of refused pages exceeds
1154		 * the maximum allowed, move to the next page size.
1155		 */
1156		if (alloc_error ||
1157		    ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1158			if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1159				break;
1160
1161			/*
1162			 * Split the refused pages to 4k. This will also empty
1163			 * the refused pages list.
1164			 */
1165			vmballoon_split_refused_pages(&ctl);
1166			ctl.page_size--;
1167		}
1168
1169		cond_resched();
1170	}
1171
 
1172	/*
1173	 * Release pages that were allocated while attempting to inflate the
1174	 * balloon but were refused by the host for one reason or another,
1175	 * and update the statistics.
1176	 */
1177	if (ctl.n_refused_pages != 0)
1178		vmballoon_release_refused_pages(b, &ctl);
1179
1180	vmballoon_release_page_list(&ctl.prealloc_pages, NULL, ctl.page_size);
1181}
1182
1183/**
1184 * vmballoon_deflate() - Decrease the size of the balloon.
1185 *
1186 * @b: pointer to the balloon
1187 * @n_frames: the number of frames to deflate. If zero, automatically
1188 * calculated according to the target size.
1189 * @coordinated: whether to coordinate with the host
1190 *
1191 * Decrease the size of the balloon allowing guest to use more memory.
1192 *
1193 * Return: The number of deflated frames (i.e., basic page size units)
1194 */
1195static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1196				       bool coordinated)
1197{
1198	unsigned long deflated_frames = 0;
1199	unsigned long tried_frames = 0;
1200	struct vmballoon_ctl ctl = {
1201		.pages = LIST_HEAD_INIT(ctl.pages),
1202		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1203		.page_size = VMW_BALLOON_4K_PAGE,
1204		.op = VMW_BALLOON_DEFLATE
1205	};
1206
1207	/* free pages to reach target */
1208	while (true) {
1209		unsigned int to_deflate_pages, n_unlocked_frames;
1210		unsigned int page_in_frames;
1211		int64_t to_deflate_frames;
1212		bool deflated_all;
1213
1214		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1215
1216		VM_BUG_ON(!list_empty(&ctl.pages));
1217		VM_BUG_ON(ctl.n_pages);
1218		VM_BUG_ON(!list_empty(&ctl.refused_pages));
1219		VM_BUG_ON(ctl.n_refused_pages);
1220
1221		/*
1222		 * If we were requested a specific number of frames, we try to
1223		 * deflate this number of frames. Otherwise, deflation is
1224		 * performed according to the target and balloon size.
1225		 */
1226		to_deflate_frames = n_frames ? n_frames - tried_frames :
1227					       -vmballoon_change(b);
1228
1229		/* break if no work to do */
1230		if (to_deflate_frames <= 0)
1231			break;
1232
1233		/*
1234		 * Calculate the number of frames based on current page size,
1235		 * but limit the deflated frames to a single chunk
1236		 */
1237		to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1238					 DIV_ROUND_UP_ULL(to_deflate_frames,
1239							  page_in_frames));
1240
1241		/* First take the pages from the balloon pages. */
1242		vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1243					    ctl.page_size, to_deflate_pages);
1244
1245		/*
1246		 * Before pages are moving to the refused list, count their
1247		 * frames as frames that we tried to deflate.
1248		 */
1249		tried_frames += ctl.n_pages * page_in_frames;
1250
1251		/*
1252		 * Unlock the pages by communicating with the hypervisor if the
1253		 * communication is coordinated (i.e., not pop). We ignore the
1254		 * return code. Instead we check if all the pages we manage to
1255		 * unlock all the pages. If we failed, we will move to the next
1256		 * page size, and would eventually try again later.
1257		 */
1258		if (coordinated)
1259			vmballoon_lock(b, &ctl);
1260
1261		/*
1262		 * Check if we deflated enough. We will move to the next page
1263		 * size if we did not manage to do so. This calculation takes
1264		 * place now, as once the pages are released, the number of
1265		 * pages is zeroed.
1266		 */
1267		deflated_all = (ctl.n_pages == to_deflate_pages);
1268
1269		/* Update local and global counters */
1270		n_unlocked_frames = ctl.n_pages * page_in_frames;
1271		atomic64_sub(n_unlocked_frames, &b->size);
1272		deflated_frames += n_unlocked_frames;
1273
1274		vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1275					 ctl.page_size, ctl.n_pages);
1276
1277		/* free the ballooned pages */
1278		vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1279					    ctl.page_size);
1280
1281		/* Return the refused pages to the ballooned list. */
1282		vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1283					    &ctl.n_refused_pages,
1284					    ctl.page_size);
1285
1286		/* If we failed to unlock all the pages, move to next size. */
1287		if (!deflated_all) {
1288			if (ctl.page_size == b->max_page_size)
1289				break;
1290			ctl.page_size++;
1291		}
1292
1293		cond_resched();
1294	}
 
 
 
 
 
 
 
1295
1296	return deflated_frames;
1297}
1298
1299/**
1300 * vmballoon_deinit_batching - disables batching mode.
1301 *
1302 * @b: pointer to &struct vmballoon.
1303 *
1304 * Disables batching, by deallocating the page for communication with the
1305 * hypervisor and disabling the static key to indicate that batching is off.
1306 */
1307static void vmballoon_deinit_batching(struct vmballoon *b)
1308{
1309	free_page((unsigned long)b->batch_page);
1310	b->batch_page = NULL;
1311	static_branch_disable(&vmw_balloon_batching);
1312	b->batch_max_pages = 1;
1313}
1314
1315/**
1316 * vmballoon_init_batching - enable batching mode.
1317 *
1318 * @b: pointer to &struct vmballoon.
1319 *
1320 * Enables batching, by allocating a page for communication with the hypervisor
1321 * and enabling the static_key to use batching.
1322 *
1323 * Return: zero on success or an appropriate error-code.
1324 */
1325static int vmballoon_init_batching(struct vmballoon *b)
1326{
1327	struct page *page;
1328
1329	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1330	if (!page)
1331		return -ENOMEM;
 
 
1332
1333	b->batch_page = page_address(page);
1334	b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
 
 
 
 
1335
1336	static_branch_enable(&vmw_balloon_batching);
1337
1338	return 0;
1339}
1340
1341/*
1342 * Receive notification and resize balloon
1343 */
1344static void vmballoon_doorbell(void *client_data)
1345{
1346	struct vmballoon *b = client_data;
1347
1348	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1349
1350	mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1351}
1352
1353/*
1354 * Clean up vmci doorbell
1355 */
1356static void vmballoon_vmci_cleanup(struct vmballoon *b)
1357{
1358	vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1359		      VMCI_INVALID_ID, VMCI_INVALID_ID);
1360
1361	if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1362		vmci_doorbell_destroy(b->vmci_doorbell);
1363		b->vmci_doorbell = VMCI_INVALID_HANDLE;
1364	}
1365}
1366
1367/**
1368 * vmballoon_vmci_init - Initialize vmci doorbell.
1369 *
1370 * @b: pointer to the balloon.
1371 *
1372 * Return: zero on success or when wakeup command not supported. Error-code
1373 * otherwise.
1374 *
1375 * Initialize vmci doorbell, to get notified as soon as balloon changes.
1376 */
1377static int vmballoon_vmci_init(struct vmballoon *b)
1378{
1379	unsigned long error;
1380
1381	if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1382		return 0;
1383
1384	error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1385				     VMCI_PRIVILEGE_FLAG_RESTRICTED,
1386				     vmballoon_doorbell, b);
1387
1388	if (error != VMCI_SUCCESS)
1389		goto fail;
1390
1391	error =	__vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1392				b->vmci_doorbell.context,
1393				b->vmci_doorbell.resource, NULL);
1394
1395	if (error != VMW_BALLOON_SUCCESS)
1396		goto fail;
1397
1398	return 0;
1399fail:
1400	vmballoon_vmci_cleanup(b);
1401	return -EIO;
1402}
1403
1404/**
1405 * vmballoon_pop - Quickly release all pages allocate for the balloon.
1406 *
1407 * @b: pointer to the balloon.
1408 *
1409 * This function is called when host decides to "reset" balloon for one reason
1410 * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1411 * pages being released.
1412 */
1413static void vmballoon_pop(struct vmballoon *b)
1414{
1415	unsigned long size;
1416
1417	while ((size = atomic64_read(&b->size)))
1418		vmballoon_deflate(b, size, false);
1419}
1420
1421/*
1422 * Perform standard reset sequence by popping the balloon (in case it
1423 * is not  empty) and then restarting protocol. This operation normally
1424 * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1425 */
1426static void vmballoon_reset(struct vmballoon *b)
1427{
 
 
 
1428	int error;
1429
1430	down_write(&b->conf_sem);
1431
1432	vmballoon_vmci_cleanup(b);
 
1433
1434	/* free all pages, skipping monitor unlock */
1435	vmballoon_pop(b);
1436
1437	if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1438		goto unlock;
 
 
 
 
 
 
 
1439
1440	if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1441		if (vmballoon_init_batching(b)) {
1442			/*
1443			 * We failed to initialize batching, inform the monitor
1444			 * about it by sending a null capability.
1445			 *
1446			 * The guest will retry in one second.
1447			 */
1448			vmballoon_send_start(b, 0);
1449			goto unlock;
1450		}
1451	} else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1452		vmballoon_deinit_batching(b);
1453	}
1454
1455	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1456	b->reset_required = false;
1457
1458	error = vmballoon_vmci_init(b);
1459	if (error)
1460		pr_err("failed to initialize vmci doorbell\n");
1461
1462	if (vmballoon_send_guest_id(b))
1463		pr_err("failed to send guest ID to the host\n");
1464
1465unlock:
1466	up_write(&b->conf_sem);
1467}
1468
1469/**
1470 * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1471 *
1472 * @work: pointer to the &work_struct which is provided by the workqueue.
1473 *
1474 * Resets the protocol if needed, gets the new size and adjusts balloon as
1475 * needed. Repeat in 1 sec.
1476 */
1477static void vmballoon_work(struct work_struct *work)
1478{
1479	struct delayed_work *dwork = to_delayed_work(work);
1480	struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1481	int64_t change = 0;
 
 
1482
1483	if (b->reset_required)
1484		vmballoon_reset(b);
1485
1486	down_read(&b->conf_sem);
 
1487
1488	/*
1489	 * Update the stats while holding the semaphore to ensure that
1490	 * @stats_enabled is consistent with whether the stats are actually
1491	 * enabled
1492	 */
1493	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1494
1495	if (!vmballoon_send_get_target(b))
1496		change = vmballoon_change(b);
1497
1498	if (change != 0) {
1499		pr_debug("%s - size: %llu, target %lu\n", __func__,
1500			 atomic64_read(&b->size), READ_ONCE(b->target));
1501
1502		if (change > 0)
1503			vmballoon_inflate(b);
1504		else  /* (change < 0) */
1505			vmballoon_deflate(b, 0, true);
1506	}
1507
1508	up_read(&b->conf_sem);
1509
1510	/*
1511	 * We are using a freezable workqueue so that balloon operations are
1512	 * stopped while the system transitions to/from sleep/hibernation.
1513	 */
1514	queue_delayed_work(system_freezable_wq,
1515			   dwork, round_jiffies_relative(HZ));
1516
1517}
1518
1519/**
1520 * vmballoon_shrinker_scan() - deflate the balloon due to memory pressure.
1521 * @shrinker: pointer to the balloon shrinker.
1522 * @sc: page reclaim information.
1523 *
1524 * Returns: number of pages that were freed during deflation.
1525 */
1526static unsigned long vmballoon_shrinker_scan(struct shrinker *shrinker,
1527					     struct shrink_control *sc)
1528{
1529	struct vmballoon *b = &balloon;
1530	unsigned long deflated_frames;
1531
1532	pr_debug("%s - size: %llu", __func__, atomic64_read(&b->size));
1533
1534	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_SHRINK);
1535
1536	/*
1537	 * If the lock is also contended for read, we cannot easily reclaim and
1538	 * we bail out.
1539	 */
1540	if (!down_read_trylock(&b->conf_sem))
1541		return 0;
1542
1543	deflated_frames = vmballoon_deflate(b, sc->nr_to_scan, true);
1544
1545	vmballoon_stats_gen_add(b, VMW_BALLOON_STAT_SHRINK_FREE,
1546				deflated_frames);
1547
1548	/*
1549	 * Delay future inflation for some time to mitigate the situations in
1550	 * which balloon continuously grows and shrinks. Use WRITE_ONCE() since
1551	 * the access is asynchronous.
1552	 */
1553	WRITE_ONCE(b->shrink_timeout, jiffies + HZ * VMBALLOON_SHRINK_DELAY);
1554
1555	up_read(&b->conf_sem);
1556
1557	return deflated_frames;
1558}
1559
1560/**
1561 * vmballoon_shrinker_count() - return the number of ballooned pages.
1562 * @shrinker: pointer to the balloon shrinker.
1563 * @sc: page reclaim information.
1564 *
1565 * Returns: number of 4k pages that are allocated for the balloon and can
1566 *	    therefore be reclaimed under pressure.
1567 */
1568static unsigned long vmballoon_shrinker_count(struct shrinker *shrinker,
1569					      struct shrink_control *sc)
1570{
1571	struct vmballoon *b = &balloon;
1572
1573	return atomic64_read(&b->size);
1574}
1575
1576static void vmballoon_unregister_shrinker(struct vmballoon *b)
1577{
1578	if (b->shrinker_registered)
1579		unregister_shrinker(&b->shrinker);
1580	b->shrinker_registered = false;
1581}
1582
1583static int vmballoon_register_shrinker(struct vmballoon *b)
1584{
1585	int r;
1586
1587	/* Do nothing if the shrinker is not enabled */
1588	if (!vmwballoon_shrinker_enable)
1589		return 0;
1590
1591	b->shrinker.scan_objects = vmballoon_shrinker_scan;
1592	b->shrinker.count_objects = vmballoon_shrinker_count;
1593	b->shrinker.seeks = DEFAULT_SEEKS;
1594
1595	r = register_shrinker(&b->shrinker);
1596
1597	if (r == 0)
1598		b->shrinker_registered = true;
1599
1600	return r;
1601}
1602
1603/*
1604 * DEBUGFS Interface
1605 */
1606#ifdef CONFIG_DEBUG_FS
1607
1608static const char * const vmballoon_stat_page_names[] = {
1609	[VMW_BALLOON_PAGE_STAT_ALLOC]		= "alloc",
1610	[VMW_BALLOON_PAGE_STAT_ALLOC_FAIL]	= "allocFail",
1611	[VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC]	= "errAlloc",
1612	[VMW_BALLOON_PAGE_STAT_REFUSED_FREE]	= "errFree",
1613	[VMW_BALLOON_PAGE_STAT_FREE]		= "free"
1614};
1615
1616static const char * const vmballoon_stat_names[] = {
1617	[VMW_BALLOON_STAT_TIMER]		= "timer",
1618	[VMW_BALLOON_STAT_DOORBELL]		= "doorbell",
1619	[VMW_BALLOON_STAT_RESET]		= "reset",
1620	[VMW_BALLOON_STAT_SHRINK]		= "shrink",
1621	[VMW_BALLOON_STAT_SHRINK_FREE]		= "shrinkFree"
1622};
1623
1624static int vmballoon_enable_stats(struct vmballoon *b)
1625{
1626	int r = 0;
1627
1628	down_write(&b->conf_sem);
1629
1630	/* did we somehow race with another reader which enabled stats? */
1631	if (b->stats)
1632		goto out;
1633
1634	b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1635
1636	if (!b->stats) {
1637		/* allocation failed */
1638		r = -ENOMEM;
1639		goto out;
1640	}
1641	static_key_enable(&balloon_stat_enabled.key);
1642out:
1643	up_write(&b->conf_sem);
1644	return r;
1645}
1646
1647/**
1648 * vmballoon_debug_show - shows statistics of balloon operations.
1649 * @f: pointer to the &struct seq_file.
1650 * @offset: ignored.
1651 *
1652 * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1653 * To avoid the overhead - mainly that of memory - of collecting the statistics,
1654 * we only collect statistics after the first time the counters are read.
1655 *
1656 * Return: zero on success or an error code.
1657 */
1658static int vmballoon_debug_show(struct seq_file *f, void *offset)
1659{
1660	struct vmballoon *b = f->private;
1661	int i, j;
1662
1663	/* enables stats if they are disabled */
1664	if (!b->stats) {
1665		int r = vmballoon_enable_stats(b);
1666
1667		if (r)
1668			return r;
1669	}
1670
1671	/* format capabilities info */
1672	seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1673		   VMW_BALLOON_CAPABILITIES);
1674	seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1675	seq_printf(f, "%-22s: %16s\n", "is resetting",
1676		   b->reset_required ? "y" : "n");
1677
1678	/* format size info */
1679	seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1680	seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1681
1682	for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1683		if (vmballoon_cmd_names[i] == NULL)
1684			continue;
1685
1686		seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1687			   vmballoon_cmd_names[i],
1688			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1689			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1690	}
1691
1692	for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1693		seq_printf(f, "%-22s: %16llu\n",
1694			   vmballoon_stat_names[i],
1695			   atomic64_read(&b->stats->general_stat[i]));
1696
1697	for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1698		for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1699			seq_printf(f, "%-18s(%s): %16llu\n",
1700				   vmballoon_stat_page_names[i],
1701				   vmballoon_page_size_names[j],
1702				   atomic64_read(&b->stats->page_stat[i][j]));
1703	}
1704
1705	return 0;
1706}
1707
1708DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
 
 
 
 
 
 
1709
1710static void __init vmballoon_debugfs_init(struct vmballoon *b)
1711{
 
 
1712	b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1713					   &vmballoon_debug_fops);
 
 
 
 
 
 
 
1714}
1715
1716static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1717{
1718	static_key_disable(&balloon_stat_enabled.key);
1719	debugfs_remove(b->dbg_entry);
1720	kfree(b->stats);
1721	b->stats = NULL;
1722}
1723
1724#else
1725
1726static inline void vmballoon_debugfs_init(struct vmballoon *b)
1727{
 
1728}
1729
1730static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1731{
1732}
1733
1734#endif	/* CONFIG_DEBUG_FS */
1735
1736
1737#ifdef CONFIG_BALLOON_COMPACTION
1738
1739static int vmballoon_init_fs_context(struct fs_context *fc)
1740{
1741	return init_pseudo(fc, BALLOON_VMW_MAGIC) ? 0 : -ENOMEM;
1742}
1743
1744static struct file_system_type vmballoon_fs = {
1745	.name           	= "balloon-vmware",
1746	.init_fs_context	= vmballoon_init_fs_context,
1747	.kill_sb        	= kill_anon_super,
1748};
1749
1750static struct vfsmount *vmballoon_mnt;
1751
1752/**
1753 * vmballoon_migratepage() - migrates a balloon page.
1754 * @b_dev_info: balloon device information descriptor.
1755 * @newpage: the page to which @page should be migrated.
1756 * @page: a ballooned page that should be migrated.
1757 * @mode: migration mode, ignored.
1758 *
1759 * This function is really open-coded, but that is according to the interface
1760 * that balloon_compaction provides.
1761 *
1762 * Return: zero on success, -EAGAIN when migration cannot be performed
1763 *	   momentarily, and -EBUSY if migration failed and should be retried
1764 *	   with that specific page.
1765 */
1766static int vmballoon_migratepage(struct balloon_dev_info *b_dev_info,
1767				 struct page *newpage, struct page *page,
1768				 enum migrate_mode mode)
1769{
1770	unsigned long status, flags;
1771	struct vmballoon *b;
1772	int ret;
1773
1774	b = container_of(b_dev_info, struct vmballoon, b_dev_info);
1775
1776	/*
1777	 * If the semaphore is taken, there is ongoing configuration change
1778	 * (i.e., balloon reset), so try again.
1779	 */
1780	if (!down_read_trylock(&b->conf_sem))
1781		return -EAGAIN;
1782
1783	spin_lock(&b->comm_lock);
1784	/*
1785	 * We must start by deflating and not inflating, as otherwise the
1786	 * hypervisor may tell us that it has enough memory and the new page is
1787	 * not needed. Since the old page is isolated, we cannot use the list
1788	 * interface to unlock it, as the LRU field is used for isolation.
1789	 * Instead, we use the native interface directly.
1790	 */
1791	vmballoon_add_page(b, 0, page);
1792	status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1793				   VMW_BALLOON_DEFLATE);
1794
1795	if (status == VMW_BALLOON_SUCCESS)
1796		status = vmballoon_status_page(b, 0, &page);
 
1797
1798	/*
1799	 * If a failure happened, let the migration mechanism know that it
1800	 * should not retry.
1801	 */
1802	if (status != VMW_BALLOON_SUCCESS) {
1803		spin_unlock(&b->comm_lock);
1804		ret = -EBUSY;
1805		goto out_unlock;
1806	}
1807
1808	/*
1809	 * The page is isolated, so it is safe to delete it without holding
1810	 * @pages_lock . We keep holding @comm_lock since we will need it in a
1811	 * second.
1812	 */
1813	balloon_page_delete(page);
1814
1815	put_page(page);
1816
1817	/* Inflate */
1818	vmballoon_add_page(b, 0, newpage);
1819	status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1820				   VMW_BALLOON_INFLATE);
1821
1822	if (status == VMW_BALLOON_SUCCESS)
1823		status = vmballoon_status_page(b, 0, &newpage);
1824
1825	spin_unlock(&b->comm_lock);
1826
1827	if (status != VMW_BALLOON_SUCCESS) {
1828		/*
1829		 * A failure happened. While we can deflate the page we just
1830		 * inflated, this deflation can also encounter an error. Instead
1831		 * we will decrease the size of the balloon to reflect the
1832		 * change and report failure.
1833		 */
1834		atomic64_dec(&b->size);
1835		ret = -EBUSY;
1836	} else {
1837		/*
1838		 * Success. Take a reference for the page, and we will add it to
1839		 * the list after acquiring the lock.
1840		 */
1841		get_page(newpage);
1842		ret = MIGRATEPAGE_SUCCESS;
1843	}
1844
1845	/* Update the balloon list under the @pages_lock */
1846	spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1847
1848	/*
1849	 * On inflation success, we already took a reference for the @newpage.
1850	 * If we succeed just insert it to the list and update the statistics
1851	 * under the lock.
1852	 */
1853	if (ret == MIGRATEPAGE_SUCCESS) {
1854		balloon_page_insert(&b->b_dev_info, newpage);
1855		__count_vm_event(BALLOON_MIGRATE);
1856	}
1857
1858	/*
1859	 * We deflated successfully, so regardless to the inflation success, we
1860	 * need to reduce the number of isolated_pages.
1861	 */
1862	b->b_dev_info.isolated_pages--;
1863	spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1864
1865out_unlock:
1866	up_read(&b->conf_sem);
1867	return ret;
1868}
1869
1870/**
1871 * vmballoon_compaction_deinit() - removes compaction related data.
1872 *
1873 * @b: pointer to the balloon.
1874 */
1875static void vmballoon_compaction_deinit(struct vmballoon *b)
1876{
1877	if (!IS_ERR(b->b_dev_info.inode))
1878		iput(b->b_dev_info.inode);
1879
1880	b->b_dev_info.inode = NULL;
1881	kern_unmount(vmballoon_mnt);
1882	vmballoon_mnt = NULL;
1883}
1884
1885/**
1886 * vmballoon_compaction_init() - initialized compaction for the balloon.
1887 *
1888 * @b: pointer to the balloon.
1889 *
1890 * If during the initialization a failure occurred, this function does not
1891 * perform cleanup. The caller must call vmballoon_compaction_deinit() in this
1892 * case.
1893 *
1894 * Return: zero on success or error code on failure.
1895 */
1896static __init int vmballoon_compaction_init(struct vmballoon *b)
1897{
1898	vmballoon_mnt = kern_mount(&vmballoon_fs);
1899	if (IS_ERR(vmballoon_mnt))
1900		return PTR_ERR(vmballoon_mnt);
1901
1902	b->b_dev_info.migratepage = vmballoon_migratepage;
1903	b->b_dev_info.inode = alloc_anon_inode(vmballoon_mnt->mnt_sb);
1904
1905	if (IS_ERR(b->b_dev_info.inode))
1906		return PTR_ERR(b->b_dev_info.inode);
1907
1908	b->b_dev_info.inode->i_mapping->a_ops = &balloon_aops;
1909	return 0;
1910}
1911
1912#else /* CONFIG_BALLOON_COMPACTION */
1913
1914static void vmballoon_compaction_deinit(struct vmballoon *b)
1915{
1916}
1917
1918static int vmballoon_compaction_init(struct vmballoon *b)
1919{
1920	return 0;
1921}
1922
1923#endif /* CONFIG_BALLOON_COMPACTION */
1924
1925static int __init vmballoon_init(void)
1926{
1927	int error;
1928
1929	/*
1930	 * Check if we are running on VMware's hypervisor and bail out
1931	 * if we are not.
1932	 */
1933	if (x86_hyper_type != X86_HYPER_VMWARE)
1934		return -ENODEV;
1935
1936	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1937
1938	error = vmballoon_register_shrinker(&balloon);
1939	if (error)
1940		goto fail;
1941
1942	/*
1943	 * Initialization of compaction must be done after the call to
1944	 * balloon_devinfo_init() .
1945	 */
1946	balloon_devinfo_init(&balloon.b_dev_info);
1947	error = vmballoon_compaction_init(&balloon);
1948	if (error)
1949		goto fail;
1950
1951	INIT_LIST_HEAD(&balloon.huge_pages);
1952	spin_lock_init(&balloon.comm_lock);
1953	init_rwsem(&balloon.conf_sem);
1954	balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1955	balloon.batch_page = NULL;
1956	balloon.page = NULL;
1957	balloon.reset_required = true;
1958
1959	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1960
1961	vmballoon_debugfs_init(&balloon);
1962
1963	return 0;
1964fail:
1965	vmballoon_unregister_shrinker(&balloon);
1966	vmballoon_compaction_deinit(&balloon);
1967	return error;
1968}
1969
1970/*
1971 * Using late_initcall() instead of module_init() allows the balloon to use the
1972 * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1973 * VMCI is probed only after the balloon is initialized. If the balloon is used
1974 * as a module, late_initcall() is equivalent to module_init().
1975 */
1976late_initcall(vmballoon_init);
1977
1978static void __exit vmballoon_exit(void)
1979{
1980	vmballoon_unregister_shrinker(&balloon);
1981	vmballoon_vmci_cleanup(&balloon);
1982	cancel_delayed_work_sync(&balloon.dwork);
1983
1984	vmballoon_debugfs_exit(&balloon);
1985
1986	/*
1987	 * Deallocate all reserved memory, and reset connection with monitor.
1988	 * Reset connection before deallocating memory to avoid potential for
1989	 * additional spurious resets from guest touching deallocated pages.
1990	 */
1991	vmballoon_send_start(&balloon, 0);
1992	vmballoon_pop(&balloon);
1993
1994	/* Only once we popped the balloon, compaction can be deinit */
1995	vmballoon_compaction_deinit(&balloon);
1996}
1997module_exit(vmballoon_exit);