1 | /* $Id: VBoxServiceBalloon.cpp 82968 2020-02-04 10:35:17Z vboxsync $ */
|
---|
2 | /** @file
|
---|
3 | * VBoxService - Memory Ballooning.
|
---|
4 | */
|
---|
5 |
|
---|
6 | /*
|
---|
7 | * Copyright (C) 2006-2020 Oracle Corporation
|
---|
8 | *
|
---|
9 | * This file is part of VirtualBox Open Source Edition (OSE), as
|
---|
10 | * available from http://www.virtualbox.org. This file is free software;
|
---|
11 | * you can redistribute it and/or modify it under the terms of the GNU
|
---|
12 | * General Public License (GPL) as published by the Free Software
|
---|
13 | * Foundation, in version 2 as it comes in the "COPYING" file of the
|
---|
14 | * VirtualBox OSE distribution. VirtualBox OSE is distributed in the
|
---|
15 | * hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
|
---|
16 | */
|
---|
17 |
|
---|
18 |
|
---|
19 | /** @page pg_vgsvc_memballoon VBoxService - Memory Ballooning
|
---|
20 | *
|
---|
21 | * The Memory Ballooning subservice works with VBoxGuest, PGM and GMM to
|
---|
22 | * dynamically reallocate memory between VMs.
|
---|
23 | *
|
---|
24 | * Memory ballooning is typically used to deal with overcomitting memory on the
|
---|
25 | * host. It allowes you to borrow memory from one or more VMs and make it
|
---|
26 | * available to others. In theory it could also be used to make memory
|
---|
27 | * available to the host system, however memory fragmentation typically makes
|
---|
28 | * that difficult.
|
---|
29 | *
|
---|
30 | * The memory ballooning subservices talks to PGM, GMM and Main via the VMMDev.
|
---|
31 | * It polls for change requests at an interval and executes them when they
|
---|
32 | * arrive. There are two ways we implement the actual ballooning, either
|
---|
33 | * VBoxGuest allocates kernel memory and donates it to the host, or this service
|
---|
34 | * allocates process memory which VBoxGuest then locks down and donates to the
|
---|
35 | * host. While we prefer the former method it is not practicable on all OS and
|
---|
36 | * we have to use the latter.
|
---|
37 | *
|
---|
38 | */
|
---|
39 |
|
---|
40 |
|
---|
41 | /*********************************************************************************************************************************
|
---|
42 | * Header Files *
|
---|
43 | *********************************************************************************************************************************/
|
---|
44 | #include <iprt/assert.h>
|
---|
45 | #include <iprt/mem.h>
|
---|
46 | #include <iprt/stream.h>
|
---|
47 | #include <iprt/string.h>
|
---|
48 | #include <iprt/semaphore.h>
|
---|
49 | #include <iprt/system.h>
|
---|
50 | #include <iprt/thread.h>
|
---|
51 | #include <iprt/time.h>
|
---|
52 | #include <VBox/err.h>
|
---|
53 | #include <VBox/VBoxGuestLib.h>
|
---|
54 | #include "VBoxServiceInternal.h"
|
---|
55 | #include "VBoxServiceUtils.h"
|
---|
56 |
|
---|
57 | #ifdef RT_OS_LINUX
|
---|
58 | # include <iprt/param.h>
|
---|
59 | # include <sys/mman.h>
|
---|
60 | # ifndef MADV_DONTFORK
|
---|
61 | # define MADV_DONTFORK 10
|
---|
62 | # endif
|
---|
63 | #endif
|
---|
64 |
|
---|
65 |
|
---|
66 |
|
---|
67 | /*********************************************************************************************************************************
|
---|
68 | * Global Variables *
|
---|
69 | *********************************************************************************************************************************/
|
---|
70 | /** The balloon size. */
|
---|
71 | static uint32_t g_cMemBalloonChunks = 0;
|
---|
72 |
|
---|
73 | /** The semaphore we're blocking on. */
|
---|
74 | static RTSEMEVENTMULTI g_MemBalloonEvent = NIL_RTSEMEVENTMULTI;
|
---|
75 |
|
---|
76 | /** The array holding the R3 pointers of the balloon. */
|
---|
77 | static void **g_pavBalloon = NULL;
|
---|
78 |
|
---|
79 | #ifdef RT_OS_LINUX
|
---|
80 | /** True = madvise(MADV_DONTFORK) works, false otherwise. */
|
---|
81 | static bool g_fSysMadviseWorks;
|
---|
82 | #endif
|
---|
83 |
|
---|
84 |
|
---|
85 | /**
|
---|
86 | * Check whether madvise() works.
|
---|
87 | */
|
---|
88 | static void vgsvcBalloonInitMadvise(void)
|
---|
89 | {
|
---|
90 | #ifdef RT_OS_LINUX
|
---|
91 | void *pv = (void*)mmap(NULL, PAGE_SIZE, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
|
---|
92 | if (pv != MAP_FAILED)
|
---|
93 | {
|
---|
94 | g_fSysMadviseWorks = madvise(pv, PAGE_SIZE, MADV_DONTFORK) == 0;
|
---|
95 | munmap(pv, PAGE_SIZE);
|
---|
96 | }
|
---|
97 | #endif
|
---|
98 | }
|
---|
99 |
|
---|
100 |
|
---|
101 | /**
|
---|
102 | * Allocate a chunk of the balloon. Fulfil the prerequisite that we can lock this memory
|
---|
103 | * and protect it against fork() in R0. See also suplibOsPageAlloc().
|
---|
104 | */
|
---|
105 | static void *VGSvcBalloonAllocChunk(void)
|
---|
106 | {
|
---|
107 | size_t cb = VMMDEV_MEMORY_BALLOON_CHUNK_SIZE;
|
---|
108 | char *pu8;
|
---|
109 |
|
---|
110 | #ifdef RT_OS_LINUX
|
---|
111 | if (!g_fSysMadviseWorks)
|
---|
112 | cb += 2 * PAGE_SIZE;
|
---|
113 |
|
---|
114 | pu8 = (char*)mmap(NULL, cb, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
|
---|
115 | if (pu8 == MAP_FAILED)
|
---|
116 | return NULL;
|
---|
117 |
|
---|
118 | if (g_fSysMadviseWorks)
|
---|
119 | {
|
---|
120 | /*
|
---|
121 | * It is not fatal if we fail here but a forked child (e.g. the ALSA sound server)
|
---|
122 | * could crash. Linux < 2.6.16 does not implement madvise(MADV_DONTFORK) but the
|
---|
123 | * kernel seems to split bigger VMAs and that is all that we want -- later we set the
|
---|
124 | * VM_DONTCOPY attribute in supdrvOSLockMemOne().
|
---|
125 | */
|
---|
126 | madvise(pu8, cb, MADV_DONTFORK);
|
---|
127 | }
|
---|
128 | else
|
---|
129 | {
|
---|
130 | /*
|
---|
131 | * madvise(MADV_DONTFORK) is not available (most probably Linux 2.4). Enclose any
|
---|
132 | * mmapped region by two unmapped pages to guarantee that there is exactly one VM
|
---|
133 | * area struct of the very same size as the mmap area.
|
---|
134 | */
|
---|
135 | RTMemProtect(pu8, PAGE_SIZE, RTMEM_PROT_NONE);
|
---|
136 | RTMemProtect(pu8 + cb - PAGE_SIZE, PAGE_SIZE, RTMEM_PROT_NONE);
|
---|
137 | pu8 += PAGE_SIZE;
|
---|
138 | }
|
---|
139 |
|
---|
140 | #else
|
---|
141 |
|
---|
142 | pu8 = (char*)RTMemPageAlloc(cb);
|
---|
143 | if (!pu8)
|
---|
144 | return pu8;
|
---|
145 |
|
---|
146 | #endif
|
---|
147 |
|
---|
148 | memset(pu8, 0, VMMDEV_MEMORY_BALLOON_CHUNK_SIZE);
|
---|
149 | return pu8;
|
---|
150 | }
|
---|
151 |
|
---|
152 |
|
---|
153 | /**
|
---|
154 | * Free an allocated chunk undoing VGSvcBalloonAllocChunk().
|
---|
155 | */
|
---|
156 | static void vgsvcBalloonFreeChunk(void *pv)
|
---|
157 | {
|
---|
158 | char *pu8 = (char*)pv;
|
---|
159 | size_t cb = VMMDEV_MEMORY_BALLOON_CHUNK_SIZE;
|
---|
160 |
|
---|
161 | #ifdef RT_OS_LINUX
|
---|
162 |
|
---|
163 | if (!g_fSysMadviseWorks)
|
---|
164 | {
|
---|
165 | cb += 2 * PAGE_SIZE;
|
---|
166 | pu8 -= PAGE_SIZE;
|
---|
167 | /* This is not really necessary */
|
---|
168 | RTMemProtect(pu8, PAGE_SIZE, RTMEM_PROT_READ | RTMEM_PROT_WRITE);
|
---|
169 | RTMemProtect(pu8 + cb - PAGE_SIZE, PAGE_SIZE, RTMEM_PROT_READ | RTMEM_PROT_WRITE);
|
---|
170 | }
|
---|
171 | munmap(pu8, cb);
|
---|
172 |
|
---|
173 | #else
|
---|
174 |
|
---|
175 | RTMemPageFree(pu8, cb);
|
---|
176 |
|
---|
177 | #endif
|
---|
178 | }
|
---|
179 |
|
---|
180 |
|
---|
181 | /**
|
---|
182 | * Adapt the R0 memory balloon by granting/reclaiming 1MB chunks to/from R0.
|
---|
183 | *
|
---|
184 | * returns IPRT status code.
|
---|
185 | * @param cNewChunks The new number of 1MB chunks in the balloon.
|
---|
186 | */
|
---|
187 | static int vgsvcBalloonSetUser(uint32_t cNewChunks)
|
---|
188 | {
|
---|
189 | if (cNewChunks == g_cMemBalloonChunks)
|
---|
190 | return VINF_SUCCESS;
|
---|
191 |
|
---|
192 | VGSvcVerbose(3, "vgsvcBalloonSetUser: cNewChunks=%u g_cMemBalloonChunks=%u\n", cNewChunks, g_cMemBalloonChunks);
|
---|
193 | int rc = VINF_SUCCESS;
|
---|
194 | if (cNewChunks > g_cMemBalloonChunks)
|
---|
195 | {
|
---|
196 | /* inflate */
|
---|
197 | g_pavBalloon = (void**)RTMemRealloc(g_pavBalloon, cNewChunks * sizeof(void*));
|
---|
198 | uint32_t i;
|
---|
199 | for (i = g_cMemBalloonChunks; i < cNewChunks; i++)
|
---|
200 | {
|
---|
201 | void *pv = VGSvcBalloonAllocChunk();
|
---|
202 | if (!pv)
|
---|
203 | break;
|
---|
204 | rc = VbglR3MemBalloonChange(pv, /* inflate=*/ true);
|
---|
205 | if (RT_SUCCESS(rc))
|
---|
206 | {
|
---|
207 | g_pavBalloon[i] = pv;
|
---|
208 | #ifndef RT_OS_SOLARIS
|
---|
209 | /*
|
---|
210 | * Protect against access by dangling pointers (ignore errors as it may fail).
|
---|
211 | * On Solaris it corrupts the address space leaving the process unkillable. This
|
---|
212 | * could perhaps be related to what the underlying segment driver does; currently
|
---|
213 | * just disable it.
|
---|
214 | */
|
---|
215 | RTMemProtect(pv, VMMDEV_MEMORY_BALLOON_CHUNK_SIZE, RTMEM_PROT_NONE);
|
---|
216 | #endif
|
---|
217 | g_cMemBalloonChunks++;
|
---|
218 | }
|
---|
219 | else
|
---|
220 | {
|
---|
221 | vgsvcBalloonFreeChunk(pv);
|
---|
222 | break;
|
---|
223 | }
|
---|
224 | }
|
---|
225 | VGSvcVerbose(3, "vgsvcBalloonSetUser: inflation complete. chunks=%u rc=%d\n", i, rc);
|
---|
226 | }
|
---|
227 | else
|
---|
228 | {
|
---|
229 | /* deflate */
|
---|
230 | uint32_t i;
|
---|
231 | for (i = g_cMemBalloonChunks; i-- > cNewChunks;)
|
---|
232 | {
|
---|
233 | void *pv = g_pavBalloon[i];
|
---|
234 | rc = VbglR3MemBalloonChange(pv, /* inflate=*/ false);
|
---|
235 | if (RT_SUCCESS(rc))
|
---|
236 | {
|
---|
237 | #ifndef RT_OS_SOLARIS
|
---|
238 | /* unprotect */
|
---|
239 | RTMemProtect(pv, VMMDEV_MEMORY_BALLOON_CHUNK_SIZE, RTMEM_PROT_READ | RTMEM_PROT_WRITE);
|
---|
240 | #endif
|
---|
241 | vgsvcBalloonFreeChunk(pv);
|
---|
242 | g_pavBalloon[i] = NULL;
|
---|
243 | g_cMemBalloonChunks--;
|
---|
244 | }
|
---|
245 | else
|
---|
246 | break;
|
---|
247 | VGSvcVerbose(3, "vgsvcBalloonSetUser: deflation complete. chunks=%u rc=%d\n", i, rc);
|
---|
248 | }
|
---|
249 | }
|
---|
250 |
|
---|
251 | return VINF_SUCCESS;
|
---|
252 | }
|
---|
253 |
|
---|
254 |
|
---|
255 | /**
|
---|
256 | * @interface_method_impl{VBOXSERVICE,pfnInit}
|
---|
257 | */
|
---|
258 | static DECLCALLBACK(int) vgsvcBalloonInit(void)
|
---|
259 | {
|
---|
260 | VGSvcVerbose(3, "vgsvcBalloonInit\n");
|
---|
261 |
|
---|
262 | int rc = RTSemEventMultiCreate(&g_MemBalloonEvent);
|
---|
263 | AssertRCReturn(rc, rc);
|
---|
264 |
|
---|
265 | vgsvcBalloonInitMadvise();
|
---|
266 |
|
---|
267 | g_cMemBalloonChunks = 0;
|
---|
268 | uint32_t cNewChunks = 0;
|
---|
269 | bool fHandleInR3;
|
---|
270 |
|
---|
271 | /* Check balloon size */
|
---|
272 | rc = VbglR3MemBalloonRefresh(&cNewChunks, &fHandleInR3);
|
---|
273 | if (RT_SUCCESS(rc))
|
---|
274 | {
|
---|
275 | VGSvcVerbose(3, "MemBalloon: New balloon size %d MB (%s memory)\n", cNewChunks, fHandleInR3 ? "R3" : "R0");
|
---|
276 | if (fHandleInR3)
|
---|
277 | rc = vgsvcBalloonSetUser(cNewChunks);
|
---|
278 | else
|
---|
279 | g_cMemBalloonChunks = cNewChunks;
|
---|
280 | }
|
---|
281 | if (RT_FAILURE(rc))
|
---|
282 | {
|
---|
283 | /* If the service was not found, we disable this service without
|
---|
284 | causing VBoxService to fail. */
|
---|
285 | if ( rc == VERR_NOT_IMPLEMENTED
|
---|
286 | #ifdef RT_OS_WINDOWS /** @todo r=bird: Windows kernel driver should return VERR_NOT_IMPLEMENTED,
|
---|
287 | * VERR_INVALID_PARAMETER has too many other uses. */
|
---|
288 | || rc == VERR_INVALID_PARAMETER
|
---|
289 | #endif
|
---|
290 | )
|
---|
291 | {
|
---|
292 | VGSvcVerbose(0, "MemBalloon: Memory ballooning support is not available\n");
|
---|
293 | rc = VERR_SERVICE_DISABLED;
|
---|
294 | }
|
---|
295 | else
|
---|
296 | {
|
---|
297 | VGSvcVerbose(3, "MemBalloon: VbglR3MemBalloonRefresh failed with %Rrc\n", rc);
|
---|
298 | rc = VERR_SERVICE_DISABLED; /** @todo Playing safe for now, figure out the exact status codes here. */
|
---|
299 | }
|
---|
300 | RTSemEventMultiDestroy(g_MemBalloonEvent);
|
---|
301 | g_MemBalloonEvent = NIL_RTSEMEVENTMULTI;
|
---|
302 | }
|
---|
303 |
|
---|
304 | return rc;
|
---|
305 | }
|
---|
306 |
|
---|
307 |
|
---|
308 | /**
|
---|
309 | * Query the size of the memory balloon, given as a page count.
|
---|
310 | *
|
---|
311 | * @returns Number of pages.
|
---|
312 | * @param cbPage The page size.
|
---|
313 | */
|
---|
314 | uint32_t VGSvcBalloonQueryPages(uint32_t cbPage)
|
---|
315 | {
|
---|
316 | Assert(cbPage > 0);
|
---|
317 | return g_cMemBalloonChunks * (VMMDEV_MEMORY_BALLOON_CHUNK_SIZE / cbPage);
|
---|
318 | }
|
---|
319 |
|
---|
320 |
|
---|
321 | /**
|
---|
322 | * @interface_method_impl{VBOXSERVICE,pfnWorker}
|
---|
323 | */
|
---|
324 | static DECLCALLBACK(int) vgsvcBalloonWorker(bool volatile *pfShutdown)
|
---|
325 | {
|
---|
326 | /* Start monitoring of the stat event change event. */
|
---|
327 | int rc = VbglR3CtlFilterMask(VMMDEV_EVENT_BALLOON_CHANGE_REQUEST, 0);
|
---|
328 | if (RT_FAILURE(rc))
|
---|
329 | {
|
---|
330 | VGSvcVerbose(3, "vgsvcBalloonInit: VbglR3CtlFilterMask failed with %Rrc\n", rc);
|
---|
331 | return rc;
|
---|
332 | }
|
---|
333 |
|
---|
334 | /*
|
---|
335 | * Tell the control thread that it can continue
|
---|
336 | * spawning services.
|
---|
337 | */
|
---|
338 | RTThreadUserSignal(RTThreadSelf());
|
---|
339 |
|
---|
340 | /*
|
---|
341 | * Now enter the loop retrieving runtime data continuously.
|
---|
342 | */
|
---|
343 | for (;;)
|
---|
344 | {
|
---|
345 | uint32_t fEvents = 0;
|
---|
346 |
|
---|
347 | /* Check if an update interval change is pending. */
|
---|
348 | rc = VbglR3WaitEvent(VMMDEV_EVENT_BALLOON_CHANGE_REQUEST, 0 /* no wait */, &fEvents);
|
---|
349 | if ( RT_SUCCESS(rc)
|
---|
350 | && (fEvents & VMMDEV_EVENT_BALLOON_CHANGE_REQUEST))
|
---|
351 | {
|
---|
352 | uint32_t cNewChunks;
|
---|
353 | bool fHandleInR3;
|
---|
354 | rc = VbglR3MemBalloonRefresh(&cNewChunks, &fHandleInR3);
|
---|
355 | if (RT_SUCCESS(rc))
|
---|
356 | {
|
---|
357 | VGSvcVerbose(3, "vgsvcBalloonInit: new balloon size %d MB (%s memory)\n", cNewChunks, fHandleInR3 ? "R3" : "R0");
|
---|
358 | if (fHandleInR3)
|
---|
359 | {
|
---|
360 | rc = vgsvcBalloonSetUser(cNewChunks);
|
---|
361 | if (RT_FAILURE(rc))
|
---|
362 | {
|
---|
363 | VGSvcVerbose(3, "vgsvcBalloonInit: failed to set balloon size %d MB (%s memory)\n",
|
---|
364 | cNewChunks, fHandleInR3 ? "R3" : "R0");
|
---|
365 | }
|
---|
366 | else
|
---|
367 | VGSvcVerbose(3, "vgsvcBalloonInit: successfully set requested balloon size %d.\n", cNewChunks);
|
---|
368 | }
|
---|
369 | else
|
---|
370 | g_cMemBalloonChunks = cNewChunks;
|
---|
371 | }
|
---|
372 | else
|
---|
373 | VGSvcVerbose(3, "vgsvcBalloonInit: VbglR3MemBalloonRefresh failed with %Rrc\n", rc);
|
---|
374 | }
|
---|
375 |
|
---|
376 | /*
|
---|
377 | * Block for a while.
|
---|
378 | *
|
---|
379 | * The event semaphore takes care of ignoring interruptions and it
|
---|
380 | * allows us to implement service wakeup later.
|
---|
381 | */
|
---|
382 | if (*pfShutdown)
|
---|
383 | break;
|
---|
384 | int rc2 = RTSemEventMultiWait(g_MemBalloonEvent, 5000);
|
---|
385 | if (*pfShutdown)
|
---|
386 | break;
|
---|
387 | if (rc2 != VERR_TIMEOUT && RT_FAILURE(rc2))
|
---|
388 | {
|
---|
389 | VGSvcError("vgsvcBalloonInit: RTSemEventMultiWait failed; rc2=%Rrc\n", rc2);
|
---|
390 | rc = rc2;
|
---|
391 | break;
|
---|
392 | }
|
---|
393 | }
|
---|
394 |
|
---|
395 | /* Cancel monitoring of the memory balloon change event. */
|
---|
396 | rc = VbglR3CtlFilterMask(0, VMMDEV_EVENT_BALLOON_CHANGE_REQUEST);
|
---|
397 | if (RT_FAILURE(rc))
|
---|
398 | VGSvcVerbose(3, "vgsvcBalloonInit: VbglR3CtlFilterMask failed with %Rrc\n", rc);
|
---|
399 |
|
---|
400 | VGSvcVerbose(3, "vgsvcBalloonInit: finished mem balloon change request thread\n");
|
---|
401 | return VINF_SUCCESS;
|
---|
402 | }
|
---|
403 |
|
---|
404 |
|
---|
405 | /**
|
---|
406 | * @interface_method_impl{VBOXSERVICE,pfnStop}
|
---|
407 | */
|
---|
408 | static DECLCALLBACK(void) vgsvcBalloonStop(void)
|
---|
409 | {
|
---|
410 | RTSemEventMultiSignal(g_MemBalloonEvent);
|
---|
411 | }
|
---|
412 |
|
---|
413 |
|
---|
414 | /**
|
---|
415 | * @interface_method_impl{VBOXSERVICE,pfnTerm}
|
---|
416 | */
|
---|
417 | static DECLCALLBACK(void) vgsvcBalloonTerm(void)
|
---|
418 | {
|
---|
419 | if (g_MemBalloonEvent != NIL_RTSEMEVENTMULTI)
|
---|
420 | {
|
---|
421 | RTSemEventMultiDestroy(g_MemBalloonEvent);
|
---|
422 | g_MemBalloonEvent = NIL_RTSEMEVENTMULTI;
|
---|
423 | }
|
---|
424 | }
|
---|
425 |
|
---|
426 |
|
---|
427 | /**
|
---|
428 | * The 'memballoon' service description.
|
---|
429 | */
|
---|
430 | VBOXSERVICE g_MemBalloon =
|
---|
431 | {
|
---|
432 | /* pszName. */
|
---|
433 | "memballoon",
|
---|
434 | /* pszDescription. */
|
---|
435 | "Memory Ballooning",
|
---|
436 | /* pszUsage. */
|
---|
437 | NULL,
|
---|
438 | /* pszOptions. */
|
---|
439 | NULL,
|
---|
440 | /* methods */
|
---|
441 | VGSvcDefaultPreInit,
|
---|
442 | VGSvcDefaultOption,
|
---|
443 | vgsvcBalloonInit,
|
---|
444 | vgsvcBalloonWorker,
|
---|
445 | vgsvcBalloonStop,
|
---|
446 | vgsvcBalloonTerm
|
---|
447 | };
|
---|