migration/rdma: Drop rdma_add_block() error handling
[qemu/armbru.git] / hw / xen / xen-mapcache.c
blobf7d974677d143da9541466b923cdb2539b1f892c
1 /*
2 * Copyright (C) 2011 Citrix Ltd.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
7 * Contributions after 2012-01-13 are licensed under the terms of the
8 * GNU GPL, version 2 or (at your option) any later version.
9 */
11 #include "qemu/osdep.h"
12 #include "qemu/units.h"
13 #include "qemu/error-report.h"
15 #include <sys/resource.h>
17 #include "hw/xen/xen_native.h"
18 #include "qemu/bitmap.h"
20 #include "sysemu/runstate.h"
21 #include "sysemu/xen-mapcache.h"
22 #include "trace.h"
25 //#define MAPCACHE_DEBUG
27 #ifdef MAPCACHE_DEBUG
28 # define DPRINTF(fmt, ...) do { \
29 fprintf(stderr, "xen_mapcache: " fmt, ## __VA_ARGS__); \
30 } while (0)
31 #else
32 # define DPRINTF(fmt, ...) do { } while (0)
33 #endif
35 #if HOST_LONG_BITS == 32
36 # define MCACHE_BUCKET_SHIFT 16
37 # define MCACHE_MAX_SIZE (1UL<<31) /* 2GB Cap */
38 #else
39 # define MCACHE_BUCKET_SHIFT 20
40 # define MCACHE_MAX_SIZE (1UL<<35) /* 32GB Cap */
41 #endif
42 #define MCACHE_BUCKET_SIZE (1UL << MCACHE_BUCKET_SHIFT)
44 /* This is the size of the virtual address space reserve to QEMU that will not
45 * be use by MapCache.
46 * From empirical tests I observed that qemu use 75MB more than the
47 * max_mcache_size.
49 #define NON_MCACHE_MEMORY_SIZE (80 * MiB)
51 typedef struct MapCacheEntry {
52 hwaddr paddr_index;
53 uint8_t *vaddr_base;
54 unsigned long *valid_mapping;
55 uint32_t lock;
56 #define XEN_MAPCACHE_ENTRY_DUMMY (1 << 0)
57 uint8_t flags;
58 hwaddr size;
59 struct MapCacheEntry *next;
60 } MapCacheEntry;
62 typedef struct MapCacheRev {
63 uint8_t *vaddr_req;
64 hwaddr paddr_index;
65 hwaddr size;
66 QTAILQ_ENTRY(MapCacheRev) next;
67 bool dma;
68 } MapCacheRev;
70 typedef struct MapCache {
71 MapCacheEntry *entry;
72 unsigned long nr_buckets;
73 QTAILQ_HEAD(, MapCacheRev) locked_entries;
75 /* For most cases (>99.9%), the page address is the same. */
76 MapCacheEntry *last_entry;
77 unsigned long max_mcache_size;
78 unsigned int mcache_bucket_shift;
80 phys_offset_to_gaddr_t phys_offset_to_gaddr;
81 QemuMutex lock;
82 void *opaque;
83 } MapCache;
85 static MapCache *mapcache;
87 static inline void mapcache_lock(void)
89 qemu_mutex_lock(&mapcache->lock);
92 static inline void mapcache_unlock(void)
94 qemu_mutex_unlock(&mapcache->lock);
97 static inline int test_bits(int nr, int size, const unsigned long *addr)
99 unsigned long res = find_next_zero_bit(addr, size + nr, nr);
100 if (res >= nr + size)
101 return 1;
102 else
103 return 0;
106 void xen_map_cache_init(phys_offset_to_gaddr_t f, void *opaque)
108 unsigned long size;
109 struct rlimit rlimit_as;
111 mapcache = g_new0(MapCache, 1);
113 mapcache->phys_offset_to_gaddr = f;
114 mapcache->opaque = opaque;
115 qemu_mutex_init(&mapcache->lock);
117 QTAILQ_INIT(&mapcache->locked_entries);
119 if (geteuid() == 0) {
120 rlimit_as.rlim_cur = RLIM_INFINITY;
121 rlimit_as.rlim_max = RLIM_INFINITY;
122 mapcache->max_mcache_size = MCACHE_MAX_SIZE;
123 } else {
124 getrlimit(RLIMIT_AS, &rlimit_as);
125 rlimit_as.rlim_cur = rlimit_as.rlim_max;
127 if (rlimit_as.rlim_max != RLIM_INFINITY) {
128 warn_report("QEMU's maximum size of virtual"
129 " memory is not infinity");
131 if (rlimit_as.rlim_max < MCACHE_MAX_SIZE + NON_MCACHE_MEMORY_SIZE) {
132 mapcache->max_mcache_size = rlimit_as.rlim_max -
133 NON_MCACHE_MEMORY_SIZE;
134 } else {
135 mapcache->max_mcache_size = MCACHE_MAX_SIZE;
139 setrlimit(RLIMIT_AS, &rlimit_as);
141 mapcache->nr_buckets =
142 (((mapcache->max_mcache_size >> XC_PAGE_SHIFT) +
143 (1UL << (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT)) - 1) >>
144 (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT));
146 size = mapcache->nr_buckets * sizeof (MapCacheEntry);
147 size = (size + XC_PAGE_SIZE - 1) & ~(XC_PAGE_SIZE - 1);
148 DPRINTF("%s, nr_buckets = %lx size %lu\n", __func__,
149 mapcache->nr_buckets, size);
150 mapcache->entry = g_malloc0(size);
153 static void xen_remap_bucket(MapCacheEntry *entry,
154 void *vaddr,
155 hwaddr size,
156 hwaddr address_index,
157 bool dummy)
159 uint8_t *vaddr_base;
160 xen_pfn_t *pfns;
161 int *err;
162 unsigned int i;
163 hwaddr nb_pfn = size >> XC_PAGE_SHIFT;
165 trace_xen_remap_bucket(address_index);
167 pfns = g_new0(xen_pfn_t, nb_pfn);
168 err = g_new0(int, nb_pfn);
170 if (entry->vaddr_base != NULL) {
171 if (!(entry->flags & XEN_MAPCACHE_ENTRY_DUMMY)) {
172 ram_block_notify_remove(entry->vaddr_base, entry->size,
173 entry->size);
177 * If an entry is being replaced by another mapping and we're using
178 * MAP_FIXED flag for it - there is possibility of a race for vaddr
179 * address with another thread doing an mmap call itself
180 * (see man 2 mmap). To avoid that we skip explicit unmapping here
181 * and allow the kernel to destroy the previous mappings by replacing
182 * them in mmap call later.
184 * Non-identical replacements are not allowed therefore.
186 assert(!vaddr || (entry->vaddr_base == vaddr && entry->size == size));
188 if (!vaddr && munmap(entry->vaddr_base, entry->size) != 0) {
189 perror("unmap fails");
190 exit(-1);
193 g_free(entry->valid_mapping);
194 entry->valid_mapping = NULL;
196 for (i = 0; i < nb_pfn; i++) {
197 pfns[i] = (address_index << (MCACHE_BUCKET_SHIFT-XC_PAGE_SHIFT)) + i;
201 * If the caller has requested the mapping at a specific address use
202 * MAP_FIXED to make sure it's honored.
204 if (!dummy) {
205 vaddr_base = xenforeignmemory_map2(xen_fmem, xen_domid, vaddr,
206 PROT_READ | PROT_WRITE,
207 vaddr ? MAP_FIXED : 0,
208 nb_pfn, pfns, err);
209 if (vaddr_base == NULL) {
210 perror("xenforeignmemory_map2");
211 exit(-1);
213 } else {
215 * We create dummy mappings where we are unable to create a foreign
216 * mapping immediately due to certain circumstances (i.e. on resume now)
218 vaddr_base = mmap(vaddr, size, PROT_READ | PROT_WRITE,
219 MAP_ANON | MAP_SHARED | (vaddr ? MAP_FIXED : 0),
220 -1, 0);
221 if (vaddr_base == MAP_FAILED) {
222 perror("mmap");
223 exit(-1);
227 if (!(entry->flags & XEN_MAPCACHE_ENTRY_DUMMY)) {
228 ram_block_notify_add(vaddr_base, size, size);
231 entry->vaddr_base = vaddr_base;
232 entry->paddr_index = address_index;
233 entry->size = size;
234 entry->valid_mapping = g_new0(unsigned long,
235 BITS_TO_LONGS(size >> XC_PAGE_SHIFT));
237 if (dummy) {
238 entry->flags |= XEN_MAPCACHE_ENTRY_DUMMY;
239 } else {
240 entry->flags &= ~(XEN_MAPCACHE_ENTRY_DUMMY);
243 bitmap_zero(entry->valid_mapping, nb_pfn);
244 for (i = 0; i < nb_pfn; i++) {
245 if (!err[i]) {
246 bitmap_set(entry->valid_mapping, i, 1);
250 g_free(pfns);
251 g_free(err);
254 static uint8_t *xen_map_cache_unlocked(hwaddr phys_addr, hwaddr size,
255 uint8_t lock, bool dma)
257 MapCacheEntry *entry, *pentry = NULL,
258 *free_entry = NULL, *free_pentry = NULL;
259 hwaddr address_index;
260 hwaddr address_offset;
261 hwaddr cache_size = size;
262 hwaddr test_bit_size;
263 bool translated G_GNUC_UNUSED = false;
264 bool dummy = false;
266 tryagain:
267 address_index = phys_addr >> MCACHE_BUCKET_SHIFT;
268 address_offset = phys_addr & (MCACHE_BUCKET_SIZE - 1);
270 trace_xen_map_cache(phys_addr);
272 /* test_bit_size is always a multiple of XC_PAGE_SIZE */
273 if (size) {
274 test_bit_size = size + (phys_addr & (XC_PAGE_SIZE - 1));
276 if (test_bit_size % XC_PAGE_SIZE) {
277 test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
279 } else {
280 test_bit_size = XC_PAGE_SIZE;
283 if (mapcache->last_entry != NULL &&
284 mapcache->last_entry->paddr_index == address_index &&
285 !lock && !size &&
286 test_bits(address_offset >> XC_PAGE_SHIFT,
287 test_bit_size >> XC_PAGE_SHIFT,
288 mapcache->last_entry->valid_mapping)) {
289 trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
290 return mapcache->last_entry->vaddr_base + address_offset;
293 /* size is always a multiple of MCACHE_BUCKET_SIZE */
294 if (size) {
295 cache_size = size + address_offset;
296 if (cache_size % MCACHE_BUCKET_SIZE) {
297 cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
299 } else {
300 cache_size = MCACHE_BUCKET_SIZE;
303 entry = &mapcache->entry[address_index % mapcache->nr_buckets];
305 while (entry && (lock || entry->lock) && entry->vaddr_base &&
306 (entry->paddr_index != address_index || entry->size != cache_size ||
307 !test_bits(address_offset >> XC_PAGE_SHIFT,
308 test_bit_size >> XC_PAGE_SHIFT,
309 entry->valid_mapping))) {
310 if (!free_entry && !entry->lock) {
311 free_entry = entry;
312 free_pentry = pentry;
314 pentry = entry;
315 entry = entry->next;
317 if (!entry && free_entry) {
318 entry = free_entry;
319 pentry = free_pentry;
321 if (!entry) {
322 entry = g_new0(MapCacheEntry, 1);
323 pentry->next = entry;
324 xen_remap_bucket(entry, NULL, cache_size, address_index, dummy);
325 } else if (!entry->lock) {
326 if (!entry->vaddr_base || entry->paddr_index != address_index ||
327 entry->size != cache_size ||
328 !test_bits(address_offset >> XC_PAGE_SHIFT,
329 test_bit_size >> XC_PAGE_SHIFT,
330 entry->valid_mapping)) {
331 xen_remap_bucket(entry, NULL, cache_size, address_index, dummy);
335 if(!test_bits(address_offset >> XC_PAGE_SHIFT,
336 test_bit_size >> XC_PAGE_SHIFT,
337 entry->valid_mapping)) {
338 mapcache->last_entry = NULL;
339 #ifdef XEN_COMPAT_PHYSMAP
340 if (!translated && mapcache->phys_offset_to_gaddr) {
341 phys_addr = mapcache->phys_offset_to_gaddr(phys_addr, size);
342 translated = true;
343 goto tryagain;
345 #endif
346 if (!dummy && runstate_check(RUN_STATE_INMIGRATE)) {
347 dummy = true;
348 goto tryagain;
350 trace_xen_map_cache_return(NULL);
351 return NULL;
354 mapcache->last_entry = entry;
355 if (lock) {
356 MapCacheRev *reventry = g_new0(MapCacheRev, 1);
357 entry->lock++;
358 if (entry->lock == 0) {
359 fprintf(stderr,
360 "mapcache entry lock overflow: "HWADDR_FMT_plx" -> %p\n",
361 entry->paddr_index, entry->vaddr_base);
362 abort();
364 reventry->dma = dma;
365 reventry->vaddr_req = mapcache->last_entry->vaddr_base + address_offset;
366 reventry->paddr_index = mapcache->last_entry->paddr_index;
367 reventry->size = entry->size;
368 QTAILQ_INSERT_HEAD(&mapcache->locked_entries, reventry, next);
371 trace_xen_map_cache_return(mapcache->last_entry->vaddr_base + address_offset);
372 return mapcache->last_entry->vaddr_base + address_offset;
375 uint8_t *xen_map_cache(hwaddr phys_addr, hwaddr size,
376 uint8_t lock, bool dma)
378 uint8_t *p;
380 mapcache_lock();
381 p = xen_map_cache_unlocked(phys_addr, size, lock, dma);
382 mapcache_unlock();
383 return p;
386 ram_addr_t xen_ram_addr_from_mapcache(void *ptr)
388 MapCacheEntry *entry = NULL;
389 MapCacheRev *reventry;
390 hwaddr paddr_index;
391 hwaddr size;
392 ram_addr_t raddr;
393 int found = 0;
395 mapcache_lock();
396 QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
397 if (reventry->vaddr_req == ptr) {
398 paddr_index = reventry->paddr_index;
399 size = reventry->size;
400 found = 1;
401 break;
404 if (!found) {
405 fprintf(stderr, "%s, could not find %p\n", __func__, ptr);
406 QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
407 DPRINTF(" "HWADDR_FMT_plx" -> %p is present\n", reventry->paddr_index,
408 reventry->vaddr_req);
410 abort();
411 return 0;
414 entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
415 while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
416 entry = entry->next;
418 if (!entry) {
419 DPRINTF("Trying to find address %p that is not in the mapcache!\n", ptr);
420 raddr = 0;
421 } else {
422 raddr = (reventry->paddr_index << MCACHE_BUCKET_SHIFT) +
423 ((unsigned long) ptr - (unsigned long) entry->vaddr_base);
425 mapcache_unlock();
426 return raddr;
429 static void xen_invalidate_map_cache_entry_unlocked(uint8_t *buffer)
431 MapCacheEntry *entry = NULL, *pentry = NULL;
432 MapCacheRev *reventry;
433 hwaddr paddr_index;
434 hwaddr size;
435 int found = 0;
437 QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
438 if (reventry->vaddr_req == buffer) {
439 paddr_index = reventry->paddr_index;
440 size = reventry->size;
441 found = 1;
442 break;
445 if (!found) {
446 DPRINTF("%s, could not find %p\n", __func__, buffer);
447 QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
448 DPRINTF(" "HWADDR_FMT_plx" -> %p is present\n", reventry->paddr_index, reventry->vaddr_req);
450 return;
452 QTAILQ_REMOVE(&mapcache->locked_entries, reventry, next);
453 g_free(reventry);
455 if (mapcache->last_entry != NULL &&
456 mapcache->last_entry->paddr_index == paddr_index) {
457 mapcache->last_entry = NULL;
460 entry = &mapcache->entry[paddr_index % mapcache->nr_buckets];
461 while (entry && (entry->paddr_index != paddr_index || entry->size != size)) {
462 pentry = entry;
463 entry = entry->next;
465 if (!entry) {
466 DPRINTF("Trying to unmap address %p that is not in the mapcache!\n", buffer);
467 return;
469 entry->lock--;
470 if (entry->lock > 0 || pentry == NULL) {
471 return;
474 pentry->next = entry->next;
475 ram_block_notify_remove(entry->vaddr_base, entry->size, entry->size);
476 if (munmap(entry->vaddr_base, entry->size) != 0) {
477 perror("unmap fails");
478 exit(-1);
480 g_free(entry->valid_mapping);
481 g_free(entry);
484 void xen_invalidate_map_cache_entry(uint8_t *buffer)
486 mapcache_lock();
487 xen_invalidate_map_cache_entry_unlocked(buffer);
488 mapcache_unlock();
491 void xen_invalidate_map_cache(void)
493 unsigned long i;
494 MapCacheRev *reventry;
496 /* Flush pending AIO before destroying the mapcache */
497 bdrv_drain_all();
499 mapcache_lock();
501 QTAILQ_FOREACH(reventry, &mapcache->locked_entries, next) {
502 if (!reventry->dma) {
503 continue;
505 fprintf(stderr, "Locked DMA mapping while invalidating mapcache!"
506 " "HWADDR_FMT_plx" -> %p is present\n",
507 reventry->paddr_index, reventry->vaddr_req);
510 for (i = 0; i < mapcache->nr_buckets; i++) {
511 MapCacheEntry *entry = &mapcache->entry[i];
513 if (entry->vaddr_base == NULL) {
514 continue;
516 if (entry->lock > 0) {
517 continue;
520 if (munmap(entry->vaddr_base, entry->size) != 0) {
521 perror("unmap fails");
522 exit(-1);
525 entry->paddr_index = 0;
526 entry->vaddr_base = NULL;
527 entry->size = 0;
528 g_free(entry->valid_mapping);
529 entry->valid_mapping = NULL;
532 mapcache->last_entry = NULL;
534 mapcache_unlock();
537 static uint8_t *xen_replace_cache_entry_unlocked(hwaddr old_phys_addr,
538 hwaddr new_phys_addr,
539 hwaddr size)
541 MapCacheEntry *entry;
542 hwaddr address_index, address_offset;
543 hwaddr test_bit_size, cache_size = size;
545 address_index = old_phys_addr >> MCACHE_BUCKET_SHIFT;
546 address_offset = old_phys_addr & (MCACHE_BUCKET_SIZE - 1);
548 assert(size);
549 /* test_bit_size is always a multiple of XC_PAGE_SIZE */
550 test_bit_size = size + (old_phys_addr & (XC_PAGE_SIZE - 1));
551 if (test_bit_size % XC_PAGE_SIZE) {
552 test_bit_size += XC_PAGE_SIZE - (test_bit_size % XC_PAGE_SIZE);
554 cache_size = size + address_offset;
555 if (cache_size % MCACHE_BUCKET_SIZE) {
556 cache_size += MCACHE_BUCKET_SIZE - (cache_size % MCACHE_BUCKET_SIZE);
559 entry = &mapcache->entry[address_index % mapcache->nr_buckets];
560 while (entry && !(entry->paddr_index == address_index &&
561 entry->size == cache_size)) {
562 entry = entry->next;
564 if (!entry) {
565 DPRINTF("Trying to update an entry for "HWADDR_FMT_plx \
566 "that is not in the mapcache!\n", old_phys_addr);
567 return NULL;
570 address_index = new_phys_addr >> MCACHE_BUCKET_SHIFT;
571 address_offset = new_phys_addr & (MCACHE_BUCKET_SIZE - 1);
573 fprintf(stderr, "Replacing a dummy mapcache entry for "HWADDR_FMT_plx \
574 " with "HWADDR_FMT_plx"\n", old_phys_addr, new_phys_addr);
576 xen_remap_bucket(entry, entry->vaddr_base,
577 cache_size, address_index, false);
578 if (!test_bits(address_offset >> XC_PAGE_SHIFT,
579 test_bit_size >> XC_PAGE_SHIFT,
580 entry->valid_mapping)) {
581 DPRINTF("Unable to update a mapcache entry for "HWADDR_FMT_plx"!\n",
582 old_phys_addr);
583 return NULL;
586 return entry->vaddr_base + address_offset;
589 uint8_t *xen_replace_cache_entry(hwaddr old_phys_addr,
590 hwaddr new_phys_addr,
591 hwaddr size)
593 uint8_t *p;
595 mapcache_lock();
596 p = xen_replace_cache_entry_unlocked(old_phys_addr, new_phys_addr, size);
597 mapcache_unlock();
598 return p;