2 * VMware Balloon driver.
4 * Copyright (C) 2000-2014, VMware, Inc. All Rights Reserved.
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.
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
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.
20 * Maintained by: Xavier Deguillard <xdeguillard@vmware.com>
21 * Philip Moltmann <moltmann@vmware.com>
25 * This is VMware physical memory management driver for Linux. The driver
26 * acts like a "balloon" that can be inflated to reclaim physical pages by
27 * reserving them in the guest and invalidating them in the monitor,
28 * freeing up the underlying machine pages so they can be allocated to
29 * other guests. The balloon can also be deflated to allow the guest to
30 * use more physical memory. Higher level policies can control the sizes
31 * of balloons in VMs in order to manage physical memory resources.
35 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
37 #include <linux/types.h>
38 #include <linux/kernel.h>
40 #include <linux/vmalloc.h>
41 #include <linux/sched.h>
42 #include <linux/module.h>
43 #include <linux/workqueue.h>
44 #include <linux/debugfs.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmw_vmci_defs.h>
47 #include <linux/vmw_vmci_api.h>
49 #include <asm/hypervisor.h>
51 MODULE_AUTHOR("VMware, Inc.");
52 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
53 MODULE_VERSION("1.5.0.0-k");
54 MODULE_ALIAS("dmi:*:svnVMware*:*");
55 MODULE_ALIAS("vmware_vmmemctl");
56 MODULE_LICENSE("GPL");
59 * Various constants controlling rate of inflaint/deflating balloon,
64 * Rates of memory allocaton when guest experiences memory pressure
65 * (driver performs sleeping allocations).
67 #define VMW_BALLOON_RATE_ALLOC_MIN 512U
68 #define VMW_BALLOON_RATE_ALLOC_MAX 2048U
69 #define VMW_BALLOON_RATE_ALLOC_INC 16U
72 * When guest is under memory pressure, use a reduced page allocation
73 * rate for next several cycles.
75 #define VMW_BALLOON_SLOW_CYCLES 4
78 * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't
79 * allow wait (__GFP_RECLAIM) for NOSLEEP page allocations. Use
80 * __GFP_NOWARN, to suppress page allocation failure warnings.
82 #define VMW_PAGE_ALLOC_NOSLEEP (__GFP_HIGHMEM|__GFP_NOWARN)
85 * Use GFP_HIGHUSER when executing in a separate kernel thread
86 * context and allocation can sleep. This is less stressful to
87 * the guest memory system, since it allows the thread to block
88 * while memory is reclaimed, and won't take pages from emergency
91 #define VMW_PAGE_ALLOC_CANSLEEP (GFP_HIGHUSER)
93 /* Maximum number of refused pages we accumulate during inflation cycle */
94 #define VMW_BALLOON_MAX_REFUSED 16
97 * Hypervisor communication port definitions.
99 #define VMW_BALLOON_HV_PORT 0x5670
100 #define VMW_BALLOON_HV_MAGIC 0x456c6d6f
101 #define VMW_BALLOON_GUEST_ID 1 /* Linux */
103 enum vmwballoon_capabilities {
105 * Bit 0 is reserved and not associated to any capability.
107 VMW_BALLOON_BASIC_CMDS = (1 << 1),
108 VMW_BALLOON_BATCHED_CMDS = (1 << 2),
109 VMW_BALLOON_BATCHED_2M_CMDS = (1 << 3),
110 VMW_BALLOON_SIGNALLED_WAKEUP_CMD = (1 << 4),
113 #define VMW_BALLOON_CAPABILITIES (VMW_BALLOON_BASIC_CMDS \
114 | VMW_BALLOON_BATCHED_CMDS \
115 | VMW_BALLOON_BATCHED_2M_CMDS \
116 | VMW_BALLOON_SIGNALLED_WAKEUP_CMD)
118 #define VMW_BALLOON_2M_SHIFT (9)
119 #define VMW_BALLOON_NUM_PAGE_SIZES (2)
122 * Backdoor commands availability:
124 * START, GET_TARGET and GUEST_ID are always available,
126 * VMW_BALLOON_BASIC_CMDS:
127 * LOCK and UNLOCK commands,
128 * VMW_BALLOON_BATCHED_CMDS:
129 * BATCHED_LOCK and BATCHED_UNLOCK commands.
130 * VMW BALLOON_BATCHED_2M_CMDS:
131 * BATCHED_2M_LOCK and BATCHED_2M_UNLOCK commands,
132 * VMW VMW_BALLOON_SIGNALLED_WAKEUP_CMD:
133 * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command.
135 #define VMW_BALLOON_CMD_START 0
136 #define VMW_BALLOON_CMD_GET_TARGET 1
137 #define VMW_BALLOON_CMD_LOCK 2
138 #define VMW_BALLOON_CMD_UNLOCK 3
139 #define VMW_BALLOON_CMD_GUEST_ID 4
140 #define VMW_BALLOON_CMD_BATCHED_LOCK 6
141 #define VMW_BALLOON_CMD_BATCHED_UNLOCK 7
142 #define VMW_BALLOON_CMD_BATCHED_2M_LOCK 8
143 #define VMW_BALLOON_CMD_BATCHED_2M_UNLOCK 9
144 #define VMW_BALLOON_CMD_VMCI_DOORBELL_SET 10
148 #define VMW_BALLOON_SUCCESS 0
149 #define VMW_BALLOON_FAILURE -1
150 #define VMW_BALLOON_ERROR_CMD_INVALID 1
151 #define VMW_BALLOON_ERROR_PPN_INVALID 2
152 #define VMW_BALLOON_ERROR_PPN_LOCKED 3
153 #define VMW_BALLOON_ERROR_PPN_UNLOCKED 4
154 #define VMW_BALLOON_ERROR_PPN_PINNED 5
155 #define VMW_BALLOON_ERROR_PPN_NOTNEEDED 6
156 #define VMW_BALLOON_ERROR_RESET 7
157 #define VMW_BALLOON_ERROR_BUSY 8
159 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES (0x03000000)
161 /* Batch page description */
164 * Layout of a page in the batch page:
166 * +-------------+----------+--------+
168 * | Page number | Reserved | Status |
170 * +-------------+----------+--------+
173 * The reserved field should be set to 0.
175 #define VMW_BALLOON_BATCH_MAX_PAGES (PAGE_SIZE / sizeof(u64))
176 #define VMW_BALLOON_BATCH_STATUS_MASK ((1UL << 5) - 1)
177 #define VMW_BALLOON_BATCH_PAGE_MASK (~((1UL << PAGE_SHIFT) - 1))
179 struct vmballoon_batch_page {
180 u64 pages[VMW_BALLOON_BATCH_MAX_PAGES];
183 static u64 vmballoon_batch_get_pa(struct vmballoon_batch_page *batch, int idx)
185 return batch->pages[idx] & VMW_BALLOON_BATCH_PAGE_MASK;
188 static int vmballoon_batch_get_status(struct vmballoon_batch_page *batch,
191 return (int)(batch->pages[idx] & VMW_BALLOON_BATCH_STATUS_MASK);
194 static void vmballoon_batch_set_pa(struct vmballoon_batch_page *batch, int idx,
197 batch->pages[idx] = pa;
201 #define VMWARE_BALLOON_CMD(cmd, arg1, arg2, result) \
203 unsigned long __status, __dummy1, __dummy2, __dummy3; \
204 __asm__ __volatile__ ("inl %%dx" : \
210 "0"(VMW_BALLOON_HV_MAGIC), \
211 "1"(VMW_BALLOON_CMD_##cmd), \
212 "2"(VMW_BALLOON_HV_PORT), \
216 if (VMW_BALLOON_CMD_##cmd == VMW_BALLOON_CMD_START) \
222 #ifdef CONFIG_DEBUG_FS
223 struct vmballoon_stats {
225 unsigned int doorbell;
227 /* allocation statistics */
228 unsigned int alloc[VMW_BALLOON_NUM_PAGE_SIZES];
229 unsigned int alloc_fail[VMW_BALLOON_NUM_PAGE_SIZES];
230 unsigned int sleep_alloc;
231 unsigned int sleep_alloc_fail;
232 unsigned int refused_alloc[VMW_BALLOON_NUM_PAGE_SIZES];
233 unsigned int refused_free[VMW_BALLOON_NUM_PAGE_SIZES];
234 unsigned int free[VMW_BALLOON_NUM_PAGE_SIZES];
236 /* monitor operations */
237 unsigned int lock[VMW_BALLOON_NUM_PAGE_SIZES];
238 unsigned int lock_fail[VMW_BALLOON_NUM_PAGE_SIZES];
239 unsigned int unlock[VMW_BALLOON_NUM_PAGE_SIZES];
240 unsigned int unlock_fail[VMW_BALLOON_NUM_PAGE_SIZES];
242 unsigned int target_fail;
244 unsigned int start_fail;
245 unsigned int guest_type;
246 unsigned int guest_type_fail;
247 unsigned int doorbell_set;
248 unsigned int doorbell_unset;
251 #define STATS_INC(stat) (stat)++
253 #define STATS_INC(stat)
258 struct vmballoon_ops {
259 void (*add_page)(struct vmballoon *b, int idx, struct page *p);
260 int (*lock)(struct vmballoon *b, unsigned int num_pages,
261 bool is_2m_pages, unsigned int *target);
262 int (*unlock)(struct vmballoon *b, unsigned int num_pages,
263 bool is_2m_pages, unsigned int *target);
266 struct vmballoon_page_size {
267 /* list of reserved physical pages */
268 struct list_head pages;
270 /* transient list of non-balloonable pages */
271 struct list_head refused_pages;
272 unsigned int n_refused_pages;
276 struct vmballoon_page_size page_sizes[VMW_BALLOON_NUM_PAGE_SIZES];
278 /* supported page sizes. 1 == 4k pages only, 2 == 4k and 2m pages */
279 unsigned supported_page_sizes;
281 /* balloon size in pages */
288 /* adjustment rates (pages per second) */
289 unsigned int rate_alloc;
291 /* slowdown page allocations for next few cycles */
292 unsigned int slow_allocation_cycles;
294 unsigned long capabilities;
296 struct vmballoon_batch_page *batch_page;
297 unsigned int batch_max_pages;
300 const struct vmballoon_ops *ops;
302 #ifdef CONFIG_DEBUG_FS
304 struct vmballoon_stats stats;
306 /* debugfs file exporting statistics */
307 struct dentry *dbg_entry;
310 struct sysinfo sysinfo;
312 struct delayed_work dwork;
314 struct vmci_handle vmci_doorbell;
317 static struct vmballoon balloon;
320 * Send "start" command to the host, communicating supported version
323 static bool vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
325 unsigned long status, capabilities, dummy = 0;
328 STATS_INC(b->stats.start);
330 status = VMWARE_BALLOON_CMD(START, req_caps, dummy, capabilities);
333 case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
334 b->capabilities = capabilities;
337 case VMW_BALLOON_SUCCESS:
338 b->capabilities = VMW_BALLOON_BASIC_CMDS;
346 * 2MB pages are only supported with batching. If batching is for some
347 * reason disabled, do not use 2MB pages, since otherwise the legacy
348 * mechanism is used with 2MB pages, causing a failure.
350 if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
351 (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
352 b->supported_page_sizes = 2;
354 b->supported_page_sizes = 1;
357 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
358 STATS_INC(b->stats.start_fail);
363 static bool vmballoon_check_status(struct vmballoon *b, unsigned long status)
366 case VMW_BALLOON_SUCCESS:
369 case VMW_BALLOON_ERROR_RESET:
370 b->reset_required = true;
379 * Communicate guest type to the host so that it can adjust ballooning
380 * algorithm to the one most appropriate for the guest. This command
381 * is normally issued after sending "start" command and is part of
382 * standard reset sequence.
384 static bool vmballoon_send_guest_id(struct vmballoon *b)
386 unsigned long status, dummy = 0;
388 status = VMWARE_BALLOON_CMD(GUEST_ID, VMW_BALLOON_GUEST_ID, dummy,
391 STATS_INC(b->stats.guest_type);
393 if (vmballoon_check_status(b, status))
396 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
397 STATS_INC(b->stats.guest_type_fail);
401 static u16 vmballoon_page_size(bool is_2m_page)
404 return 1 << VMW_BALLOON_2M_SHIFT;
410 * Retrieve desired balloon size from the host.
412 static bool vmballoon_send_get_target(struct vmballoon *b, u32 *new_target)
414 unsigned long status;
415 unsigned long target;
417 unsigned long dummy = 0;
421 * si_meminfo() is cheap. Moreover, we want to provide dynamic
422 * max balloon size later. So let us call si_meminfo() every
425 si_meminfo(&b->sysinfo);
426 limit = b->sysinfo.totalram;
428 /* Ensure limit fits in 32-bits */
429 limit32 = (u32)limit;
430 if (limit != limit32)
434 STATS_INC(b->stats.target);
436 status = VMWARE_BALLOON_CMD(GET_TARGET, limit, dummy, target);
437 if (vmballoon_check_status(b, status)) {
438 *new_target = target;
442 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
443 STATS_INC(b->stats.target_fail);
448 * Notify the host about allocated page so that host can use it without
449 * fear that guest will need it. Host may reject some pages, we need to
450 * check the return value and maybe submit a different page.
452 static int vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn,
453 unsigned int *hv_status, unsigned int *target)
455 unsigned long status, dummy = 0;
462 STATS_INC(b->stats.lock[false]);
464 *hv_status = status = VMWARE_BALLOON_CMD(LOCK, pfn, dummy, *target);
465 if (vmballoon_check_status(b, status))
468 pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
469 STATS_INC(b->stats.lock_fail[false]);
473 static int vmballoon_send_batched_lock(struct vmballoon *b,
474 unsigned int num_pages, bool is_2m_pages, unsigned int *target)
476 unsigned long status;
477 unsigned long pfn = PHYS_PFN(virt_to_phys(b->batch_page));
479 STATS_INC(b->stats.lock[is_2m_pages]);
482 status = VMWARE_BALLOON_CMD(BATCHED_2M_LOCK, pfn, num_pages,
485 status = VMWARE_BALLOON_CMD(BATCHED_LOCK, pfn, num_pages,
488 if (vmballoon_check_status(b, status))
491 pr_debug("%s - batch ppn %lx, hv returns %ld\n", __func__, pfn, status);
492 STATS_INC(b->stats.lock_fail[is_2m_pages]);
497 * Notify the host that guest intends to release given page back into
498 * the pool of available (to the guest) pages.
500 static bool vmballoon_send_unlock_page(struct vmballoon *b, unsigned long pfn,
501 unsigned int *target)
503 unsigned long status, dummy = 0;
510 STATS_INC(b->stats.unlock[false]);
512 status = VMWARE_BALLOON_CMD(UNLOCK, pfn, dummy, *target);
513 if (vmballoon_check_status(b, status))
516 pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
517 STATS_INC(b->stats.unlock_fail[false]);
521 static bool vmballoon_send_batched_unlock(struct vmballoon *b,
522 unsigned int num_pages, bool is_2m_pages, unsigned int *target)
524 unsigned long status;
525 unsigned long pfn = PHYS_PFN(virt_to_phys(b->batch_page));
527 STATS_INC(b->stats.unlock[is_2m_pages]);
530 status = VMWARE_BALLOON_CMD(BATCHED_2M_UNLOCK, pfn, num_pages,
533 status = VMWARE_BALLOON_CMD(BATCHED_UNLOCK, pfn, num_pages,
536 if (vmballoon_check_status(b, status))
539 pr_debug("%s - batch ppn %lx, hv returns %ld\n", __func__, pfn, status);
540 STATS_INC(b->stats.unlock_fail[is_2m_pages]);
544 static struct page *vmballoon_alloc_page(gfp_t flags, bool is_2m_page)
547 return alloc_pages(flags, VMW_BALLOON_2M_SHIFT);
549 return alloc_page(flags);
552 static void vmballoon_free_page(struct page *page, bool is_2m_page)
555 __free_pages(page, VMW_BALLOON_2M_SHIFT);
561 * Quickly release all pages allocated for the balloon. This function is
562 * called when host decides to "reset" balloon for one reason or another.
563 * Unlike normal "deflate" we do not (shall not) notify host of the pages
566 static void vmballoon_pop(struct vmballoon *b)
568 struct page *page, *next;
569 unsigned is_2m_pages;
571 for (is_2m_pages = 0; is_2m_pages < VMW_BALLOON_NUM_PAGE_SIZES;
573 struct vmballoon_page_size *page_size =
574 &b->page_sizes[is_2m_pages];
575 u16 size_per_page = vmballoon_page_size(is_2m_pages);
577 list_for_each_entry_safe(page, next, &page_size->pages, lru) {
578 list_del(&page->lru);
579 vmballoon_free_page(page, is_2m_pages);
580 STATS_INC(b->stats.free[is_2m_pages]);
581 b->size -= size_per_page;
586 /* Clearing the batch_page unconditionally has no adverse effect */
587 free_page((unsigned long)b->batch_page);
588 b->batch_page = NULL;
592 * Notify the host of a ballooned page. If host rejects the page put it on the
593 * refuse list, those refused page are then released at the end of the
596 static int vmballoon_lock_page(struct vmballoon *b, unsigned int num_pages,
597 bool is_2m_pages, unsigned int *target)
599 int locked, hv_status;
600 struct page *page = b->page;
601 struct vmballoon_page_size *page_size = &b->page_sizes[false];
603 /* is_2m_pages can never happen as 2m pages support implies batching */
605 locked = vmballoon_send_lock_page(b, page_to_pfn(page), &hv_status,
608 STATS_INC(b->stats.refused_alloc[false]);
610 if (locked == -EIO &&
611 (hv_status == VMW_BALLOON_ERROR_RESET ||
612 hv_status == VMW_BALLOON_ERROR_PPN_NOTNEEDED)) {
613 vmballoon_free_page(page, false);
618 * Place page on the list of non-balloonable pages
619 * and retry allocation, unless we already accumulated
620 * too many of them, in which case take a breather.
622 if (page_size->n_refused_pages < VMW_BALLOON_MAX_REFUSED) {
623 page_size->n_refused_pages++;
624 list_add(&page->lru, &page_size->refused_pages);
626 vmballoon_free_page(page, false);
631 /* track allocated page */
632 list_add(&page->lru, &page_size->pages);
634 /* update balloon size */
640 static int vmballoon_lock_batched_page(struct vmballoon *b,
641 unsigned int num_pages, bool is_2m_pages, unsigned int *target)
644 u16 size_per_page = vmballoon_page_size(is_2m_pages);
646 locked = vmballoon_send_batched_lock(b, num_pages, is_2m_pages,
649 for (i = 0; i < num_pages; i++) {
650 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
651 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
653 vmballoon_free_page(p, is_2m_pages);
659 for (i = 0; i < num_pages; i++) {
660 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
661 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
662 struct vmballoon_page_size *page_size =
663 &b->page_sizes[is_2m_pages];
665 locked = vmballoon_batch_get_status(b->batch_page, i);
668 case VMW_BALLOON_SUCCESS:
669 list_add(&p->lru, &page_size->pages);
670 b->size += size_per_page;
672 case VMW_BALLOON_ERROR_PPN_PINNED:
673 case VMW_BALLOON_ERROR_PPN_INVALID:
674 if (page_size->n_refused_pages
675 < VMW_BALLOON_MAX_REFUSED) {
676 list_add(&p->lru, &page_size->refused_pages);
677 page_size->n_refused_pages++;
681 case VMW_BALLOON_ERROR_RESET:
682 case VMW_BALLOON_ERROR_PPN_NOTNEEDED:
683 vmballoon_free_page(p, is_2m_pages);
686 /* This should never happen */
695 * Release the page allocated for the balloon. Note that we first notify
696 * the host so it can make sure the page will be available for the guest
699 static int vmballoon_unlock_page(struct vmballoon *b, unsigned int num_pages,
700 bool is_2m_pages, unsigned int *target)
702 struct page *page = b->page;
703 struct vmballoon_page_size *page_size = &b->page_sizes[false];
705 /* is_2m_pages can never happen as 2m pages support implies batching */
707 if (!vmballoon_send_unlock_page(b, page_to_pfn(page), target)) {
708 list_add(&page->lru, &page_size->pages);
712 /* deallocate page */
713 vmballoon_free_page(page, false);
714 STATS_INC(b->stats.free[false]);
716 /* update balloon size */
722 static int vmballoon_unlock_batched_page(struct vmballoon *b,
723 unsigned int num_pages, bool is_2m_pages,
724 unsigned int *target)
726 int locked, i, ret = 0;
728 u16 size_per_page = vmballoon_page_size(is_2m_pages);
730 hv_success = vmballoon_send_batched_unlock(b, num_pages, is_2m_pages,
735 for (i = 0; i < num_pages; i++) {
736 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
737 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
738 struct vmballoon_page_size *page_size =
739 &b->page_sizes[is_2m_pages];
741 locked = vmballoon_batch_get_status(b->batch_page, i);
742 if (!hv_success || locked != VMW_BALLOON_SUCCESS) {
744 * That page wasn't successfully unlocked by the
745 * hypervisor, re-add it to the list of pages owned by
746 * the balloon driver.
748 list_add(&p->lru, &page_size->pages);
750 /* deallocate page */
751 vmballoon_free_page(p, is_2m_pages);
752 STATS_INC(b->stats.free[is_2m_pages]);
754 /* update balloon size */
755 b->size -= size_per_page;
763 * Release pages that were allocated while attempting to inflate the
764 * balloon but were refused by the host for one reason or another.
766 static void vmballoon_release_refused_pages(struct vmballoon *b,
769 struct page *page, *next;
770 struct vmballoon_page_size *page_size =
771 &b->page_sizes[is_2m_pages];
773 list_for_each_entry_safe(page, next, &page_size->refused_pages, lru) {
774 list_del(&page->lru);
775 vmballoon_free_page(page, is_2m_pages);
776 STATS_INC(b->stats.refused_free[is_2m_pages]);
779 page_size->n_refused_pages = 0;
782 static void vmballoon_add_page(struct vmballoon *b, int idx, struct page *p)
787 static void vmballoon_add_batched_page(struct vmballoon *b, int idx,
790 vmballoon_batch_set_pa(b->batch_page, idx,
791 (u64)page_to_pfn(p) << PAGE_SHIFT);
795 * Inflate the balloon towards its target size. Note that we try to limit
796 * the rate of allocation to make sure we are not choking the rest of the
799 static void vmballoon_inflate(struct vmballoon *b)
802 unsigned int allocations = 0;
803 unsigned int num_pages = 0;
805 gfp_t flags = VMW_PAGE_ALLOC_NOSLEEP;
808 pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
811 * First try NOSLEEP page allocations to inflate balloon.
813 * If we do not throttle nosleep allocations, we can drain all
814 * free pages in the guest quickly (if the balloon target is high).
815 * As a side-effect, draining free pages helps to inform (force)
816 * the guest to start swapping if balloon target is not met yet,
817 * which is a desired behavior. However, balloon driver can consume
818 * all available CPU cycles if too many pages are allocated in a
819 * second. Therefore, we throttle nosleep allocations even when
820 * the guest is not under memory pressure. OTOH, if we have already
821 * predicted that the guest is under memory pressure, then we
822 * slowdown page allocations considerably.
826 * Start with no sleep allocation rate which may be higher
827 * than sleeping allocation rate.
829 if (b->slow_allocation_cycles) {
830 rate = b->rate_alloc;
835 b->supported_page_sizes == VMW_BALLOON_NUM_PAGE_SIZES;
838 pr_debug("%s - goal: %d, no-sleep rate: %u, sleep rate: %d\n",
839 __func__, b->target - b->size, rate, b->rate_alloc);
841 while (!b->reset_required &&
842 b->size + num_pages * vmballoon_page_size(is_2m_pages)
846 if (flags == VMW_PAGE_ALLOC_NOSLEEP)
847 STATS_INC(b->stats.alloc[is_2m_pages]);
849 STATS_INC(b->stats.sleep_alloc);
851 page = vmballoon_alloc_page(flags, is_2m_pages);
853 STATS_INC(b->stats.alloc_fail[is_2m_pages]);
856 b->ops->lock(b, num_pages, true, &b->target);
859 * ignore errors from locking as we now switch
860 * to 4k pages and we might get different
869 if (flags == VMW_PAGE_ALLOC_CANSLEEP) {
871 * CANSLEEP page allocation failed, so guest
872 * is under severe memory pressure. Quickly
873 * decrease allocation rate.
875 b->rate_alloc = max(b->rate_alloc / 2,
876 VMW_BALLOON_RATE_ALLOC_MIN);
877 STATS_INC(b->stats.sleep_alloc_fail);
882 * NOSLEEP page allocation failed, so the guest is
883 * under memory pressure. Let us slow down page
884 * allocations for next few cycles so that the guest
885 * gets out of memory pressure. Also, if we already
886 * allocated b->rate_alloc pages, let's pause,
887 * otherwise switch to sleeping allocations.
889 b->slow_allocation_cycles = VMW_BALLOON_SLOW_CYCLES;
891 if (allocations >= b->rate_alloc)
894 flags = VMW_PAGE_ALLOC_CANSLEEP;
895 /* Lower rate for sleeping allocations. */
896 rate = b->rate_alloc;
900 b->ops->add_page(b, num_pages++, page);
901 if (num_pages == b->batch_max_pages) {
902 error = b->ops->lock(b, num_pages, is_2m_pages,
911 if (allocations >= rate) {
912 /* We allocated enough pages, let's take a break. */
918 b->ops->lock(b, num_pages, is_2m_pages, &b->target);
921 * We reached our goal without failures so try increasing
924 if (error == 0 && allocations >= b->rate_alloc) {
925 unsigned int mult = allocations / b->rate_alloc;
928 min(b->rate_alloc + mult * VMW_BALLOON_RATE_ALLOC_INC,
929 VMW_BALLOON_RATE_ALLOC_MAX);
932 vmballoon_release_refused_pages(b, true);
933 vmballoon_release_refused_pages(b, false);
937 * Decrease the size of the balloon allowing guest to use more memory.
939 static void vmballoon_deflate(struct vmballoon *b)
941 unsigned is_2m_pages;
943 pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
945 /* free pages to reach target */
946 for (is_2m_pages = 0; is_2m_pages < b->supported_page_sizes;
948 struct page *page, *next;
949 unsigned int num_pages = 0;
950 struct vmballoon_page_size *page_size =
951 &b->page_sizes[is_2m_pages];
953 list_for_each_entry_safe(page, next, &page_size->pages, lru) {
954 if (b->reset_required ||
957 * vmballoon_page_size(is_2m_pages)
958 < b->target + vmballoon_page_size(true)))
961 list_del(&page->lru);
962 b->ops->add_page(b, num_pages++, page);
964 if (num_pages == b->batch_max_pages) {
967 error = b->ops->unlock(b, num_pages,
968 is_2m_pages, &b->target);
978 b->ops->unlock(b, num_pages, is_2m_pages, &b->target);
982 static const struct vmballoon_ops vmballoon_basic_ops = {
983 .add_page = vmballoon_add_page,
984 .lock = vmballoon_lock_page,
985 .unlock = vmballoon_unlock_page
988 static const struct vmballoon_ops vmballoon_batched_ops = {
989 .add_page = vmballoon_add_batched_page,
990 .lock = vmballoon_lock_batched_page,
991 .unlock = vmballoon_unlock_batched_page
994 static bool vmballoon_init_batching(struct vmballoon *b)
998 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1002 b->batch_page = page_address(page);
1007 * Receive notification and resize balloon
1009 static void vmballoon_doorbell(void *client_data)
1011 struct vmballoon *b = client_data;
1013 STATS_INC(b->stats.doorbell);
1015 mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1019 * Clean up vmci doorbell
1021 static void vmballoon_vmci_cleanup(struct vmballoon *b)
1025 VMWARE_BALLOON_CMD(VMCI_DOORBELL_SET, VMCI_INVALID_ID,
1026 VMCI_INVALID_ID, error);
1027 STATS_INC(b->stats.doorbell_unset);
1029 if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1030 vmci_doorbell_destroy(b->vmci_doorbell);
1031 b->vmci_doorbell = VMCI_INVALID_HANDLE;
1036 * Initialize vmci doorbell, to get notified as soon as balloon changes
1038 static int vmballoon_vmci_init(struct vmballoon *b)
1040 unsigned long error, dummy;
1042 if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1045 error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1046 VMCI_PRIVILEGE_FLAG_RESTRICTED,
1047 vmballoon_doorbell, b);
1049 if (error != VMCI_SUCCESS)
1052 error = VMWARE_BALLOON_CMD(VMCI_DOORBELL_SET, b->vmci_doorbell.context,
1053 b->vmci_doorbell.resource, dummy);
1055 STATS_INC(b->stats.doorbell_set);
1057 if (error != VMW_BALLOON_SUCCESS)
1062 vmballoon_vmci_cleanup(b);
1067 * Perform standard reset sequence by popping the balloon (in case it
1068 * is not empty) and then restarting protocol. This operation normally
1069 * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1071 static void vmballoon_reset(struct vmballoon *b)
1075 vmballoon_vmci_cleanup(b);
1077 /* free all pages, skipping monitor unlock */
1080 if (!vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1083 if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1084 b->ops = &vmballoon_batched_ops;
1085 b->batch_max_pages = VMW_BALLOON_BATCH_MAX_PAGES;
1086 if (!vmballoon_init_batching(b)) {
1088 * We failed to initialize batching, inform the monitor
1089 * about it by sending a null capability.
1091 * The guest will retry in one second.
1093 vmballoon_send_start(b, 0);
1096 } else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1097 b->ops = &vmballoon_basic_ops;
1098 b->batch_max_pages = 1;
1101 b->reset_required = false;
1103 error = vmballoon_vmci_init(b);
1105 pr_err("failed to initialize vmci doorbell\n");
1107 if (!vmballoon_send_guest_id(b))
1108 pr_err("failed to send guest ID to the host\n");
1112 * Balloon work function: reset protocol, if needed, get the new size and
1113 * adjust balloon as needed. Repeat in 1 sec.
1115 static void vmballoon_work(struct work_struct *work)
1117 struct delayed_work *dwork = to_delayed_work(work);
1118 struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1119 unsigned int target;
1121 STATS_INC(b->stats.timer);
1123 if (b->reset_required)
1126 if (b->slow_allocation_cycles > 0)
1127 b->slow_allocation_cycles--;
1129 if (!b->reset_required && vmballoon_send_get_target(b, &target)) {
1130 /* update target, adjust size */
1133 if (b->size < target)
1134 vmballoon_inflate(b);
1135 else if (target == 0 ||
1136 b->size > target + vmballoon_page_size(true))
1137 vmballoon_deflate(b);
1141 * We are using a freezable workqueue so that balloon operations are
1142 * stopped while the system transitions to/from sleep/hibernation.
1144 queue_delayed_work(system_freezable_wq,
1145 dwork, round_jiffies_relative(HZ));
1151 #ifdef CONFIG_DEBUG_FS
1153 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1155 struct vmballoon *b = f->private;
1156 struct vmballoon_stats *stats = &b->stats;
1158 /* format capabilities info */
1160 "balloon capabilities: %#4x\n"
1161 "used capabilities: %#4lx\n"
1162 "is resetting: %c\n",
1163 VMW_BALLOON_CAPABILITIES, b->capabilities,
1164 b->reset_required ? 'y' : 'n');
1166 /* format size info */
1168 "target: %8d pages\n"
1169 "current: %8d pages\n",
1170 b->target, b->size);
1172 /* format rate info */
1174 "rateSleepAlloc: %8d pages/sec\n",
1181 "start: %8u (%4u failed)\n"
1182 "guestType: %8u (%4u failed)\n"
1183 "2m-lock: %8u (%4u failed)\n"
1184 "lock: %8u (%4u failed)\n"
1185 "2m-unlock: %8u (%4u failed)\n"
1186 "unlock: %8u (%4u failed)\n"
1187 "target: %8u (%4u failed)\n"
1188 "prim2mAlloc: %8u (%4u failed)\n"
1189 "primNoSleepAlloc: %8u (%4u failed)\n"
1190 "primCanSleepAlloc: %8u (%4u failed)\n"
1197 "doorbellSet: %8u\n"
1198 "doorbellUnset: %8u\n",
1201 stats->start, stats->start_fail,
1202 stats->guest_type, stats->guest_type_fail,
1203 stats->lock[true], stats->lock_fail[true],
1204 stats->lock[false], stats->lock_fail[false],
1205 stats->unlock[true], stats->unlock_fail[true],
1206 stats->unlock[false], stats->unlock_fail[false],
1207 stats->target, stats->target_fail,
1208 stats->alloc[true], stats->alloc_fail[true],
1209 stats->alloc[false], stats->alloc_fail[false],
1210 stats->sleep_alloc, stats->sleep_alloc_fail,
1213 stats->refused_alloc[true], stats->refused_alloc[false],
1214 stats->refused_free[true], stats->refused_free[false],
1215 stats->doorbell_set, stats->doorbell_unset);
1220 static int vmballoon_debug_open(struct inode *inode, struct file *file)
1222 return single_open(file, vmballoon_debug_show, inode->i_private);
1225 static const struct file_operations vmballoon_debug_fops = {
1226 .owner = THIS_MODULE,
1227 .open = vmballoon_debug_open,
1229 .llseek = seq_lseek,
1230 .release = single_release,
1233 static int __init vmballoon_debugfs_init(struct vmballoon *b)
1237 b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1238 &vmballoon_debug_fops);
1239 if (IS_ERR(b->dbg_entry)) {
1240 error = PTR_ERR(b->dbg_entry);
1241 pr_err("failed to create debugfs entry, error: %d\n", error);
1248 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1250 debugfs_remove(b->dbg_entry);
1255 static inline int vmballoon_debugfs_init(struct vmballoon *b)
1260 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1264 #endif /* CONFIG_DEBUG_FS */
1266 static int __init vmballoon_init(void)
1269 unsigned is_2m_pages;
1271 * Check if we are running on VMware's hypervisor and bail out
1274 if (x86_hyper_type != X86_HYPER_VMWARE)
1277 for (is_2m_pages = 0; is_2m_pages < VMW_BALLOON_NUM_PAGE_SIZES;
1279 INIT_LIST_HEAD(&balloon.page_sizes[is_2m_pages].pages);
1280 INIT_LIST_HEAD(&balloon.page_sizes[is_2m_pages].refused_pages);
1283 /* initialize rates */
1284 balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX;
1286 INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1288 error = vmballoon_debugfs_init(&balloon);
1292 balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1293 balloon.batch_page = NULL;
1294 balloon.page = NULL;
1295 balloon.reset_required = true;
1297 queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1303 * Using late_initcall() instead of module_init() allows the balloon to use the
1304 * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1305 * VMCI is probed only after the balloon is initialized. If the balloon is used
1306 * as a module, late_initcall() is equivalent to module_init().
1308 late_initcall(vmballoon_init);
1310 static void __exit vmballoon_exit(void)
1312 vmballoon_vmci_cleanup(&balloon);
1313 cancel_delayed_work_sync(&balloon.dwork);
1315 vmballoon_debugfs_exit(&balloon);
1318 * Deallocate all reserved memory, and reset connection with monitor.
1319 * Reset connection before deallocating memory to avoid potential for
1320 * additional spurious resets from guest touching deallocated pages.
1322 vmballoon_send_start(&balloon, 0);
1323 vmballoon_pop(&balloon);
1325 module_exit(vmballoon_exit);