Linux Audio

Check our new training course

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