xref: /freebsd/sys/dev/xen/balloon/balloon.c (revision 1829d5da5355930d5cfa8ec8add8ff47dc0bebab)
1 /******************************************************************************
2  * balloon.c
3  *
4  * Xen balloon driver - enables returning/claiming memory to/from Xen.
5  *
6  * Copyright (c) 2003, B Dragovic
7  * Copyright (c) 2003-2004, M Williamson, K Fraser
8  * Copyright (c) 2005 Dan M. Smith, IBM Corporation
9  *
10  * This file may be distributed separately from the Linux kernel, or
11  * incorporated into other software packages, subject to the following license:
12  *
13  * Permission is hereby granted, free of charge, to any person obtaining a copy
14  * of this source file (the "Software"), to deal in the Software without
15  * restriction, including without limitation the rights to use, copy, modify,
16  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
17  * and to permit persons to whom the Software is furnished to do so, subject to
18  * the following conditions:
19  *
20  * The above copyright notice and this permission notice shall be included in
21  * all copies or substantial portions of the Software.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
26  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
28  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
29  * IN THE SOFTWARE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/lock.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/malloc.h>
40 #include <sys/mutex.h>
41 #include <sys/sysctl.h>
42 
43 #include <machine/xen/xen-os.h>
44 #include <machine/xen/xenfunc.h>
45 #include <machine/xen/xenvar.h>
46 #include <xen/hypervisor.h>
47 #include <xen/xenbus/xenbusvar.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_page.h>
51 
52 MALLOC_DEFINE(M_BALLOON, "Balloon", "Xen Balloon Driver");
53 
54 struct mtx balloon_mutex;
55 
56 /*
57  * Protects atomic reservation decrease/increase against concurrent increases.
58  * Also protects non-atomic updates of current_pages and driver_pages, and
59  * balloon lists.
60  */
61 struct mtx balloon_lock;
62 
63 /* We increase/decrease in batches which fit in a page */
64 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
65 #define ARRAY_SIZE(A)	(sizeof(A) / sizeof(A[0]))
66 
67 struct balloon_stats {
68 	/* We aim for 'current allocation' == 'target allocation'. */
69 	unsigned long current_pages;
70 	unsigned long target_pages;
71 	/* We may hit the hard limit in Xen. If we do then we remember it. */
72 	unsigned long hard_limit;
73 	/*
74 	 * Drivers may alter the memory reservation independently, but they
75 	 * must inform the balloon driver so we avoid hitting the hard limit.
76 	 */
77 	unsigned long driver_pages;
78 	/* Number of pages in high- and low-memory balloons. */
79 	unsigned long balloon_low;
80 	unsigned long balloon_high;
81 };
82 
83 static struct balloon_stats balloon_stats;
84 #define bs balloon_stats
85 
86 SYSCTL_DECL(_dev_xen);
87 SYSCTL_NODE(_dev_xen, OID_AUTO, balloon, CTLFLAG_RD, NULL, "Balloon");
88 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, current, CTLFLAG_RD,
89     &bs.current_pages, 0, "Current allocation");
90 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, target, CTLFLAG_RD,
91     &bs.target_pages, 0, "Target allocation");
92 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, driver_pages, CTLFLAG_RD,
93     &bs.driver_pages, 0, "Driver pages");
94 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, hard_limit, CTLFLAG_RD,
95     &bs.hard_limit, 0, "Xen hard limit");
96 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, low_mem, CTLFLAG_RD,
97     &bs.balloon_low, 0, "Low-mem balloon");
98 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, high_mem, CTLFLAG_RD,
99     &bs.balloon_high, 0, "High-mem balloon");
100 
101 struct balloon_entry {
102 	vm_page_t page;
103 	STAILQ_ENTRY(balloon_entry) list;
104 };
105 
106 /* List of ballooned pages, threaded through the mem_map array. */
107 static STAILQ_HEAD(,balloon_entry) ballooned_pages;
108 
109 /* Main work function, always executed in process context. */
110 static void balloon_process(void *unused);
111 
112 #define IPRINTK(fmt, args...) \
113 	printk(KERN_INFO "xen_mem: " fmt, ##args)
114 #define WPRINTK(fmt, args...) \
115 	printk(KERN_WARNING "xen_mem: " fmt, ##args)
116 
117 /* balloon_append: add the given page to the balloon. */
118 static void
119 balloon_append(vm_page_t page)
120 {
121 	struct balloon_entry *entry;
122 
123 	entry = malloc(sizeof(struct balloon_entry), M_BALLOON, M_WAITOK);
124 	entry->page = page;
125 	STAILQ_INSERT_HEAD(&ballooned_pages, entry, list);
126 	bs.balloon_low++;
127 }
128 
129 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
130 static vm_page_t
131 balloon_retrieve(void)
132 {
133 	vm_page_t page;
134 	struct balloon_entry *entry;
135 
136 	if (STAILQ_EMPTY(&ballooned_pages))
137 		return NULL;
138 
139 	entry = STAILQ_FIRST(&ballooned_pages);
140 	STAILQ_REMOVE_HEAD(&ballooned_pages, list);
141 
142 	page = entry->page;
143 	free(entry, M_DEVBUF);
144 
145 	bs.balloon_low--;
146 
147 	return page;
148 }
149 
150 static void
151 balloon_alarm(void *unused)
152 {
153 	wakeup(balloon_process);
154 }
155 
156 static unsigned long
157 current_target(void)
158 {
159 	unsigned long target = min(bs.target_pages, bs.hard_limit);
160 	if (target > (bs.current_pages + bs.balloon_low + bs.balloon_high))
161 		target = bs.current_pages + bs.balloon_low + bs.balloon_high;
162 	return target;
163 }
164 
165 static unsigned long
166 minimum_target(void)
167 {
168 #ifdef XENHVM
169 #define max_pfn physmem
170 #endif
171 	unsigned long min_pages, curr_pages = current_target();
172 
173 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
174 	/* Simple continuous piecewiese linear function:
175 	 *  max MiB -> min MiB	gradient
176 	 *       0	   0
177 	 *      16	  16
178 	 *      32	  24
179 	 *     128	  72	(1/2)
180 	 *     512 	 168	(1/4)
181 	 *    2048	 360	(1/8)
182 	 *    8192	 552	(1/32)
183 	 *   32768	1320
184 	 *  131072	4392
185 	 */
186 	if (max_pfn < MB2PAGES(128))
187 		min_pages = MB2PAGES(8) + (max_pfn >> 1);
188 	else if (max_pfn < MB2PAGES(512))
189 		min_pages = MB2PAGES(40) + (max_pfn >> 2);
190 	else if (max_pfn < MB2PAGES(2048))
191 		min_pages = MB2PAGES(104) + (max_pfn >> 3);
192 	else
193 		min_pages = MB2PAGES(296) + (max_pfn >> 5);
194 #undef MB2PAGES
195 
196 	/* Don't enforce growth */
197 	return min(min_pages, curr_pages);
198 #ifndef CONFIG_XEN
199 #undef max_pfn
200 #endif
201 }
202 
203 static int
204 increase_reservation(unsigned long nr_pages)
205 {
206 	unsigned long  pfn, i;
207 	struct balloon_entry *entry;
208 	vm_page_t      page;
209 	long           rc;
210 	struct xen_memory_reservation reservation = {
211 		.address_bits = 0,
212 		.extent_order = 0,
213 		.domid        = DOMID_SELF
214 	};
215 
216 	if (nr_pages > ARRAY_SIZE(frame_list))
217 		nr_pages = ARRAY_SIZE(frame_list);
218 
219 	mtx_lock(&balloon_lock);
220 
221 	for (entry = STAILQ_FIRST(&ballooned_pages), i = 0;
222 	     i < nr_pages; i++, entry = STAILQ_NEXT(entry, list)) {
223 		KASSERT(entry, ("ballooned_pages list corrupt"));
224 		page = entry->page;
225 		frame_list[i] = (VM_PAGE_TO_PHYS(page) >> PAGE_SHIFT);
226 	}
227 
228 	set_xen_guest_handle(reservation.extent_start, frame_list);
229 	reservation.nr_extents   = nr_pages;
230 	rc = HYPERVISOR_memory_op(
231 		XENMEM_populate_physmap, &reservation);
232 	if (rc < nr_pages) {
233 		if (rc > 0) {
234 			int ret;
235 
236 			/* We hit the Xen hard limit: reprobe. */
237 			reservation.nr_extents = rc;
238 			ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
239 					&reservation);
240 			KASSERT(ret == rc, ("HYPERVISOR_memory_op failed"));
241 		}
242 		if (rc >= 0)
243 			bs.hard_limit = (bs.current_pages + rc -
244 					 bs.driver_pages);
245 		goto out;
246 	}
247 
248 	for (i = 0; i < nr_pages; i++) {
249 		page = balloon_retrieve();
250 		KASSERT(page, ("balloon_retrieve failed"));
251 
252 		pfn = (VM_PAGE_TO_PHYS(page) >> PAGE_SHIFT);
253 		KASSERT((xen_feature(XENFEAT_auto_translated_physmap) ||
254 			!phys_to_machine_mapping_valid(pfn)),
255 		    ("auto translated physmap but mapping is valid"));
256 
257 		set_phys_to_machine(pfn, frame_list[i]);
258 
259 #ifndef XENHVM
260 		/* Link back into the page tables if not highmem. */
261 		if (pfn < max_low_pfn) {
262 			int ret;
263 			ret = HYPERVISOR_update_va_mapping(
264 				(unsigned long)__va(pfn << PAGE_SHIFT),
265 				pfn_pte_ma(frame_list[i], PAGE_KERNEL),
266 				0);
267 			PASSING(ret == 0,
268 			    ("HYPERVISOR_update_va_mapping failed"));
269 		}
270 #endif
271 
272 		/* Relinquish the page back to the allocator. */
273 		vm_page_unwire(page, 0);
274 		vm_page_free(page);
275 	}
276 
277 	bs.current_pages += nr_pages;
278 	//totalram_pages = bs.current_pages;
279 
280  out:
281 	mtx_unlock(&balloon_lock);
282 
283 	return 0;
284 }
285 
286 static int
287 decrease_reservation(unsigned long nr_pages)
288 {
289 	unsigned long  pfn, i;
290 	vm_page_t      page;
291 	int            need_sleep = 0;
292 	int ret;
293 	struct xen_memory_reservation reservation = {
294 		.address_bits = 0,
295 		.extent_order = 0,
296 		.domid        = DOMID_SELF
297 	};
298 
299 	if (nr_pages > ARRAY_SIZE(frame_list))
300 		nr_pages = ARRAY_SIZE(frame_list);
301 
302 	for (i = 0; i < nr_pages; i++) {
303 		int color = 0;
304 		if ((page = vm_page_alloc(NULL, color++,
305 			    VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
306 			    VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
307 			nr_pages = i;
308 			need_sleep = 1;
309 			break;
310 		}
311 
312 		pfn = (VM_PAGE_TO_PHYS(page) >> PAGE_SHIFT);
313 		frame_list[i] = PFNTOMFN(pfn);
314 
315 #if 0
316 		if (!PageHighMem(page)) {
317 			v = phys_to_virt(pfn << PAGE_SHIFT);
318 			scrub_pages(v, 1);
319 #ifdef CONFIG_XEN
320 			ret = HYPERVISOR_update_va_mapping(
321 				(unsigned long)v, __pte_ma(0), 0);
322 			BUG_ON(ret);
323 #endif
324 		}
325 #endif
326 #ifdef CONFIG_XEN_SCRUB_PAGES
327 		else {
328 			v = kmap(page);
329 			scrub_pages(v, 1);
330 			kunmap(page);
331 		}
332 #endif
333 	}
334 
335 #ifdef CONFIG_XEN
336 	/* Ensure that ballooned highmem pages don't have kmaps. */
337 	kmap_flush_unused();
338 	flush_tlb_all();
339 #endif
340 
341 	mtx_lock(&balloon_lock);
342 
343 	/* No more mappings: invalidate P2M and add to balloon. */
344 	for (i = 0; i < nr_pages; i++) {
345 		pfn = MFNTOPFN(frame_list[i]);
346 		set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
347 		balloon_append(PHYS_TO_VM_PAGE(pfn << PAGE_SHIFT));
348 	}
349 
350 	set_xen_guest_handle(reservation.extent_start, frame_list);
351 	reservation.nr_extents   = nr_pages;
352 	ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
353 	KASSERT(ret == nr_pages, ("HYPERVISOR_memory_op failed"));
354 
355 	bs.current_pages -= nr_pages;
356 	//totalram_pages = bs.current_pages;
357 
358 	mtx_unlock(&balloon_lock);
359 
360 	return (need_sleep);
361 }
362 
363 /*
364  * We avoid multiple worker processes conflicting via the balloon mutex.
365  * We may of course race updates of the target counts (which are protected
366  * by the balloon lock), or with changes to the Xen hard limit, but we will
367  * recover from these in time.
368  */
369 static void
370 balloon_process(void *unused)
371 {
372 	int need_sleep = 0;
373 	long credit;
374 
375 	mtx_lock(&balloon_mutex);
376 	for (;;) {
377 		do {
378 			credit = current_target() - bs.current_pages;
379 			if (credit > 0)
380 				need_sleep = (increase_reservation(credit) != 0);
381 			if (credit < 0)
382 				need_sleep = (decrease_reservation(-credit) != 0);
383 
384 		} while ((credit != 0) && !need_sleep);
385 
386 		/* Schedule more work if there is some still to be done. */
387 		if (current_target() != bs.current_pages)
388 			timeout(balloon_alarm, NULL, ticks + hz);
389 
390 		msleep(balloon_process, &balloon_mutex, 0, "balloon", -1);
391 	}
392 	mtx_unlock(&balloon_mutex);
393 }
394 
395 /* Resets the Xen limit, sets new target, and kicks off processing. */
396 static void
397 set_new_target(unsigned long target)
398 {
399 	/* No need for lock. Not read-modify-write updates. */
400 	bs.hard_limit   = ~0UL;
401 	bs.target_pages = max(target, minimum_target());
402 	wakeup(balloon_process);
403 }
404 
405 static struct xenbus_watch target_watch =
406 {
407 	.node = "memory/target"
408 };
409 
410 /* React to a change in the target key */
411 static void
412 watch_target(struct xenbus_watch *watch,
413 	     const char **vec, unsigned int len)
414 {
415 	unsigned long long new_target;
416 	int err;
417 
418 	err = xenbus_scanf(XBT_NIL, "memory", "target", NULL,
419 	    "%llu", &new_target);
420 	if (err) {
421 		/* This is ok (for domain0 at least) - so just return */
422 		return;
423 	}
424 
425 	/* The given memory/target value is in KiB, so it needs converting to
426 	   pages.  PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
427 	*/
428 	set_new_target(new_target >> (PAGE_SHIFT - 10));
429 
430 }
431 
432 static void
433 balloon_init_watcher(void *arg)
434 {
435 	int err;
436 
437 	err = register_xenbus_watch(&target_watch);
438 	if (err)
439 		printf("Failed to set balloon watcher\n");
440 
441 }
442 SYSINIT(balloon_init_watcher, SI_SUB_PSEUDO, SI_ORDER_ANY,
443     balloon_init_watcher, NULL);
444 
445 static void
446 balloon_init(void *arg)
447 {
448 #ifndef XENHVM
449 	vm_page_t page;
450 #endif
451 
452 	if (!is_running_on_xen())
453 		return;
454 
455 	mtx_init(&balloon_lock, "balloon_lock", NULL, MTX_DEF);
456 	mtx_init(&balloon_mutex, "balloon_mutex", NULL, MTX_DEF);
457 
458 #ifndef XENHVM
459 	bs.current_pages = min(xen_start_info->nr_pages, max_pfn);
460 #else
461 	bs.current_pages = physmem;
462 #endif
463 	bs.target_pages  = bs.current_pages;
464 	bs.balloon_low   = 0;
465 	bs.balloon_high  = 0;
466 	bs.driver_pages  = 0UL;
467 	bs.hard_limit    = ~0UL;
468 
469 	kproc_create(balloon_process, NULL, NULL, 0, 0, "balloon");
470 //	init_timer(&balloon_timer);
471 //	balloon_timer.data = 0;
472 //	balloon_timer.function = balloon_alarm;
473 
474 #ifndef XENHVM
475 	/* Initialise the balloon with excess memory space. */
476 	for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
477 		page = PHYS_TO_VM_PAGE(pfn << PAGE_SHIFT);
478 		balloon_append(page);
479 	}
480 #endif
481 
482 	target_watch.callback = watch_target;
483 
484 	return;
485 }
486 SYSINIT(balloon_init, SI_SUB_PSEUDO, SI_ORDER_ANY, balloon_init, NULL);
487 
488 void balloon_update_driver_allowance(long delta);
489 
490 void
491 balloon_update_driver_allowance(long delta)
492 {
493 	mtx_lock(&balloon_lock);
494 	bs.driver_pages += delta;
495 	mtx_unlock(&balloon_lock);
496 }
497 
498 #if 0
499 static int dealloc_pte_fn(
500 	pte_t *pte, struct page *pte_page, unsigned long addr, void *data)
501 {
502 	unsigned long mfn = pte_mfn(*pte);
503 	int ret;
504 	struct xen_memory_reservation reservation = {
505 		.extent_start = &mfn,
506 		.nr_extents   = 1,
507 		.extent_order = 0,
508 		.domid        = DOMID_SELF
509 	};
510 	set_pte_at(&init_mm, addr, pte, __pte_ma(0));
511 	set_phys_to_machine(__pa(addr) >> PAGE_SHIFT, INVALID_P2M_ENTRY);
512 	ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
513 	KASSERT(ret == 1, ("HYPERVISOR_memory_op failed"));
514 	return 0;
515 }
516 
517 #endif
518 
519 #if 0
520 vm_page_t
521 balloon_alloc_empty_page_range(unsigned long nr_pages)
522 {
523 	vm_page_t pages;
524 	int i, rc;
525 	unsigned long *mfn_list;
526 	struct xen_memory_reservation reservation = {
527 		.address_bits = 0,
528 		.extent_order = 0,
529 		.domid        = DOMID_SELF
530 	};
531 
532 	pages = vm_page_alloc_contig(nr_pages, 0, -1, 4, 4)
533 	if (pages == NULL)
534 		return NULL;
535 
536 	mfn_list = malloc(nr_pages*sizeof(unsigned long), M_DEVBUF, M_WAITOK);
537 
538 	for (i = 0; i < nr_pages; i++) {
539 		mfn_list[i] = PFNTOMFN(VM_PAGE_TO_PHYS(pages[i]) >> PAGE_SHIFT);
540 		PFNTOMFN(i) = INVALID_P2M_ENTRY;
541 		reservation.extent_start = mfn_list;
542 		reservation.nr_extents = nr_pages;
543 		rc = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
544 		    &reservation);
545 		KASSERT(rc == nr_pages, ("HYPERVISOR_memory_op failed"));
546 	}
547 
548 	current_pages -= nr_pages;
549 
550 	wakeup(balloon_process);
551 
552 	return pages;
553 }
554 
555 void
556 balloon_dealloc_empty_page_range(vm_page_t page, unsigned long nr_pages)
557 {
558 	unsigned long i;
559 
560 	for (i = 0; i < nr_pages; i++)
561 		balloon_append(page + i);
562 
563 	wakeup(balloon_process);
564 }
565 #endif
566