xref: /freebsd/sys/dev/xen/balloon/balloon.c (revision d48760ffe9eaac809f8d31fb1b4e6b13ae52d59a)
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/param.h>
33 #include <sys/lock.h>
34 #include <sys/kernel.h>
35 #include <sys/kthread.h>
36 #include <sys/malloc.h>
37 #include <sys/mutex.h>
38 #include <sys/sysctl.h>
39 #include <sys/module.h>
40 
41 #include <vm/vm.h>
42 #include <vm/vm_page.h>
43 
44 #include <xen/xen-os.h>
45 #include <xen/hypervisor.h>
46 #include <xen/features.h>
47 #include <xen/xenstore/xenstorevar.h>
48 
49 static MALLOC_DEFINE(M_BALLOON, "Balloon", "Xen Balloon Driver");
50 
51 /* Convert from KB (as fetched from xenstore) to number of PAGES */
52 #define KB_TO_PAGE_SHIFT	(PAGE_SHIFT - 10)
53 
54 struct mtx balloon_mutex;
55 
56 /* We increase/decrease in batches which fit in a page */
57 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
58 
59 struct balloon_stats {
60 	/* We aim for 'current allocation' == 'target allocation'. */
61 	unsigned long current_pages;
62 	unsigned long target_pages;
63 	/* We may hit the hard limit in Xen. If we do then we remember it. */
64 	unsigned long hard_limit;
65 	/*
66 	 * Drivers may alter the memory reservation independently, but they
67 	 * must inform the balloon driver so we avoid hitting the hard limit.
68 	 */
69 	unsigned long driver_pages;
70 	/* Number of pages in high- and low-memory balloons. */
71 	unsigned long balloon_low;
72 	unsigned long balloon_high;
73 };
74 
75 static struct balloon_stats balloon_stats;
76 #define bs balloon_stats
77 
78 SYSCTL_DECL(_dev_xen);
79 static SYSCTL_NODE(_dev_xen, OID_AUTO, balloon,
80     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
81     "Balloon");
82 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, current, CTLFLAG_RD,
83     &bs.current_pages, 0, "Current allocation");
84 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, target, CTLFLAG_RD,
85     &bs.target_pages, 0, "Target allocation");
86 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, driver_pages, CTLFLAG_RD,
87     &bs.driver_pages, 0, "Driver pages");
88 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, hard_limit, CTLFLAG_RD,
89     &bs.hard_limit, 0, "Xen hard limit");
90 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, low_mem, CTLFLAG_RD,
91     &bs.balloon_low, 0, "Low-mem balloon");
92 SYSCTL_ULONG(_dev_xen_balloon, OID_AUTO, high_mem, CTLFLAG_RD,
93     &bs.balloon_high, 0, "High-mem balloon");
94 
95 /* List of ballooned pages, threaded through the mem_map array. */
96 static TAILQ_HEAD(,vm_page) ballooned_pages;
97 
98 /* Main work function, always executed in process context. */
99 static void balloon_process(void *unused);
100 
101 #define IPRINTK(fmt, args...) \
102 	printk(KERN_INFO "xen_mem: " fmt, ##args)
103 #define WPRINTK(fmt, args...) \
104 	printk(KERN_WARNING "xen_mem: " fmt, ##args)
105 
106 static unsigned long
current_target(void)107 current_target(void)
108 {
109 	unsigned long target = min(bs.target_pages, bs.hard_limit);
110 	if (target > (bs.current_pages + bs.balloon_low + bs.balloon_high))
111 		target = bs.current_pages + bs.balloon_low + bs.balloon_high;
112 	return (target);
113 }
114 
115 static unsigned long
minimum_target(void)116 minimum_target(void)
117 {
118 	unsigned long min_pages, curr_pages = current_target();
119 
120 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
121 	/*
122 	 * Simple continuous piecewiese linear function:
123 	 *  max MiB -> min MiB	gradient
124 	 *       0	   0
125 	 *      16	  16
126 	 *      32	  24
127 	 *     128	  72	(1/2)
128 	 *     512 	 168	(1/4)
129 	 *    2048	 360	(1/8)
130 	 *    8192	 552	(1/32)
131 	 *   32768	1320
132 	 *  131072	4392
133 	 */
134 	if (realmem < MB2PAGES(128))
135 		min_pages = MB2PAGES(8) + (realmem >> 1);
136 	else if (realmem < MB2PAGES(512))
137 		min_pages = MB2PAGES(40) + (realmem >> 2);
138 	else if (realmem < MB2PAGES(2048))
139 		min_pages = MB2PAGES(104) + (realmem >> 3);
140 	else
141 		min_pages = MB2PAGES(296) + (realmem >> 5);
142 #undef MB2PAGES
143 
144 	/* Don't enforce growth */
145 	return (min(min_pages, curr_pages));
146 }
147 
148 static int
increase_reservation(unsigned long nr_pages)149 increase_reservation(unsigned long nr_pages)
150 {
151 	unsigned long  i;
152 	vm_page_t      page;
153 	long           rc;
154 	struct xen_memory_reservation reservation = {
155 		.extent_order = 0,
156 		.domid        = DOMID_SELF
157 	};
158 
159 	mtx_assert(&balloon_mutex, MA_OWNED);
160 
161 	if (nr_pages > nitems(frame_list))
162 		nr_pages = nitems(frame_list);
163 
164 	for (page = TAILQ_FIRST(&ballooned_pages), i = 0;
165 	    i < nr_pages; i++, page = TAILQ_NEXT(page, plinks.q)) {
166 		KASSERT(page != NULL, ("ballooned_pages list corrupt"));
167 		frame_list[i] = (VM_PAGE_TO_PHYS(page) >> PAGE_SHIFT);
168 	}
169 
170 	set_xen_guest_handle(reservation.extent_start, frame_list);
171 	reservation.nr_extents   = nr_pages;
172 	rc = HYPERVISOR_memory_op(
173 		XENMEM_populate_physmap, &reservation);
174 	if (rc < nr_pages) {
175 		if (rc > 0) {
176 			int ret __diagused;
177 
178 			/* We hit the Xen hard limit: reprobe. */
179 			reservation.nr_extents = rc;
180 			ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
181 					&reservation);
182 			KASSERT(ret == rc, ("HYPERVISOR_memory_op failed"));
183 		}
184 		if (rc >= 0)
185 			bs.hard_limit = (bs.current_pages + rc -
186 					 bs.driver_pages);
187 		goto out;
188 	}
189 
190 	for (i = 0; i < nr_pages; i++) {
191 		page = TAILQ_FIRST(&ballooned_pages);
192 		KASSERT(page != NULL, ("Unable to get ballooned page"));
193 		TAILQ_REMOVE(&ballooned_pages, page, plinks.q);
194 		bs.balloon_low--;
195 
196 		KASSERT(xen_feature(XENFEAT_auto_translated_physmap),
197 		    ("auto translated physmap but mapping is valid"));
198 
199 		vm_page_free(page);
200 	}
201 
202 	bs.current_pages += nr_pages;
203 
204  out:
205 	return (0);
206 }
207 
208 static int
decrease_reservation(unsigned long nr_pages)209 decrease_reservation(unsigned long nr_pages)
210 {
211 	unsigned long  i;
212 	vm_page_t      page;
213 	int            need_sleep = 0;
214 	int ret __diagused;
215 	struct xen_memory_reservation reservation = {
216 		.extent_order = 0,
217 		.domid        = DOMID_SELF
218 	};
219 
220 	mtx_assert(&balloon_mutex, MA_OWNED);
221 
222 	if (nr_pages > nitems(frame_list))
223 		nr_pages = nitems(frame_list);
224 
225 	for (i = 0; i < nr_pages; i++) {
226 		/*
227 		 * Zero the page, or else we might be leaking important data to
228 		 * other domains on the same host. Xen doesn't scrub ballooned
229 		 * out memory pages, the guest is in charge of making sure that
230 		 * no information is leaked.
231 		 */
232 		if ((page = vm_page_alloc_noobj(VM_ALLOC_ZERO)) == NULL) {
233 			nr_pages = i;
234 			need_sleep = 1;
235 			break;
236 		}
237 
238 		frame_list[i] = (VM_PAGE_TO_PHYS(page) >> PAGE_SHIFT);
239 
240 		TAILQ_INSERT_HEAD(&ballooned_pages, page, plinks.q);
241 		bs.balloon_low++;
242 	}
243 
244 	set_xen_guest_handle(reservation.extent_start, frame_list);
245 	reservation.nr_extents   = nr_pages;
246 	ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
247 	KASSERT(ret == nr_pages, ("HYPERVISOR_memory_op failed"));
248 
249 	bs.current_pages -= nr_pages;
250 
251 	return (need_sleep);
252 }
253 
254 /*
255  * We avoid multiple worker processes conflicting via the balloon mutex.
256  * We may of course race updates of the target counts (which are protected
257  * by the balloon lock), or with changes to the Xen hard limit, but we will
258  * recover from these in time.
259  */
260 static void
balloon_process(void * unused)261 balloon_process(void *unused)
262 {
263 	int need_sleep = 0;
264 	long credit;
265 
266 	mtx_lock(&balloon_mutex);
267 	for (;;) {
268 		int sleep_time;
269 
270 		do {
271 			credit = current_target() - bs.current_pages;
272 			if (credit > 0)
273 				need_sleep = (increase_reservation(credit) != 0);
274 			if (credit < 0)
275 				need_sleep = (decrease_reservation(-credit) != 0);
276 
277 		} while ((credit != 0) && !need_sleep);
278 
279 		/* Schedule more work if there is some still to be done. */
280 		if (current_target() != bs.current_pages)
281 			sleep_time = hz;
282 		else
283 			sleep_time = 0;
284 
285 		msleep(balloon_process, &balloon_mutex, 0, "balloon",
286 		       sleep_time);
287 	}
288 	mtx_unlock(&balloon_mutex);
289 }
290 
291 /* Resets the Xen limit, sets new target, and kicks off processing. */
292 static void
set_new_target(unsigned long target)293 set_new_target(unsigned long target)
294 {
295 	/* No need for lock. Not read-modify-write updates. */
296 	bs.hard_limit   = ~0UL;
297 	bs.target_pages = max(target, minimum_target());
298 	wakeup(balloon_process);
299 }
300 
301 static struct xs_watch target_watch =
302 {
303 	.node = "memory/target",
304 	.max_pending = 1,
305 };
306 
307 /* React to a change in the target key */
308 static void
watch_target(struct xs_watch * watch,const char ** vec,unsigned int len)309 watch_target(struct xs_watch *watch,
310 	     const char **vec, unsigned int len)
311 {
312 	unsigned long long new_target;
313 	int err;
314 
315 	err = xs_scanf(XST_NIL, "memory", "target", NULL,
316 	    "%llu", &new_target);
317 	if (err) {
318 		/* This is ok (for domain0 at least) - so just return */
319 		return;
320 	}
321 
322 	/*
323 	 * The given memory/target value is in KiB, so it needs converting to
324 	 * pages.  PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
325 	 */
326 	set_new_target(new_target >> KB_TO_PAGE_SHIFT);
327 }
328 
329 /*------------------ Private Device Attachment Functions  --------------------*/
330 /**
331  * \brief Identify instances of this device type in the system.
332  *
333  * \param driver  The driver performing this identify action.
334  * \param parent  The NewBus parent device for any devices this method adds.
335  */
336 static void
xenballoon_identify(driver_t * driver,device_t parent)337 xenballoon_identify(driver_t *driver, device_t parent)
338 {
339 	/*
340 	 * A single device instance for our driver is always present
341 	 * in a system operating under Xen.
342 	 */
343 	BUS_ADD_CHILD(parent, 0, driver->name, 0);
344 }
345 
346 /**
347  * \brief Probe for the existence of the Xen Balloon device
348  *
349  * \param dev  NewBus device_t for this Xen control instance.
350  *
351  * \return  Always returns 0 indicating success.
352  */
353 static int
xenballoon_probe(device_t dev)354 xenballoon_probe(device_t dev)
355 {
356 
357 	device_set_desc(dev, "Xen Balloon Device");
358 	return (0);
359 }
360 
361 /**
362  * \brief Attach the Xen Balloon device.
363  *
364  * \param dev  NewBus device_t for this Xen control instance.
365  *
366  * \return  On success, 0. Otherwise an errno value indicating the
367  *          type of failure.
368  */
369 static int
xenballoon_attach(device_t dev)370 xenballoon_attach(device_t dev)
371 {
372 	int err;
373 
374 	mtx_init(&balloon_mutex, "balloon_mutex", NULL, MTX_DEF);
375 
376 	bs.current_pages = realmem;
377 	bs.target_pages  = bs.current_pages;
378 	bs.balloon_low   = 0;
379 	bs.balloon_high  = 0;
380 	bs.driver_pages  = 0UL;
381 	bs.hard_limit    = ~0UL;
382 
383 	kproc_create(balloon_process, NULL, NULL, 0, 0, "balloon");
384 
385 	target_watch.callback = watch_target;
386 
387 	err = xs_register_watch(&target_watch);
388 	if (err)
389 		device_printf(dev,
390 		    "xenballon: failed to set balloon watcher\n");
391 
392 	return (err);
393 }
394 
395 /*-------------------- Private Device Attachment Data  -----------------------*/
396 static device_method_t xenballoon_methods[] = {
397 	/* Device interface */
398 	DEVMETHOD(device_identify,	xenballoon_identify),
399 	DEVMETHOD(device_probe,         xenballoon_probe),
400 	DEVMETHOD(device_attach,        xenballoon_attach),
401 
402 	DEVMETHOD_END
403 };
404 
405 DEFINE_CLASS_0(xenballoon, xenballoon_driver, xenballoon_methods, 0);
406 
407 DRIVER_MODULE(xenballoon, xenstore, xenballoon_driver, NULL, NULL);
408