xref: /openbmc/qemu/tcg/region.c (revision af531756)
1 /*
2  * Memory region management for Tiny Code Generator for QEMU
3  *
4  * Copyright (c) 2008 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include "qemu/units.h"
27 #include "qapi/error.h"
28 #include "exec/exec-all.h"
29 #include "tcg/tcg.h"
30 #include "tcg-internal.h"
31 
32 
33 struct tcg_region_tree {
34     QemuMutex lock;
35     GTree *tree;
36     /* padding to avoid false sharing is computed at run-time */
37 };
38 
39 /*
40  * We divide code_gen_buffer into equally-sized "regions" that TCG threads
41  * dynamically allocate from as demand dictates. Given appropriate region
42  * sizing, this minimizes flushes even when some TCG threads generate a lot
43  * more code than others.
44  */
45 struct tcg_region_state {
46     QemuMutex lock;
47 
48     /* fields set at init time */
49     void *start_aligned;
50     void *after_prologue;
51     size_t n;
52     size_t size; /* size of one region */
53     size_t stride; /* .size + guard size */
54     size_t total_size; /* size of entire buffer, >= n * stride */
55 
56     /* fields protected by the lock */
57     size_t current; /* current region index */
58     size_t agg_size_full; /* aggregate size of full regions */
59 };
60 
61 static struct tcg_region_state region;
62 
63 /*
64  * This is an array of struct tcg_region_tree's, with padding.
65  * We use void * to simplify the computation of region_trees[i]; each
66  * struct is found every tree_size bytes.
67  */
68 static void *region_trees;
69 static size_t tree_size;
70 
71 bool in_code_gen_buffer(const void *p)
72 {
73     /*
74      * Much like it is valid to have a pointer to the byte past the
75      * end of an array (so long as you don't dereference it), allow
76      * a pointer to the byte past the end of the code gen buffer.
77      */
78     return (size_t)(p - region.start_aligned) <= region.total_size;
79 }
80 
81 #ifdef CONFIG_DEBUG_TCG
82 const void *tcg_splitwx_to_rx(void *rw)
83 {
84     /* Pass NULL pointers unchanged. */
85     if (rw) {
86         g_assert(in_code_gen_buffer(rw));
87         rw += tcg_splitwx_diff;
88     }
89     return rw;
90 }
91 
92 void *tcg_splitwx_to_rw(const void *rx)
93 {
94     /* Pass NULL pointers unchanged. */
95     if (rx) {
96         rx -= tcg_splitwx_diff;
97         /* Assert that we end with a pointer in the rw region. */
98         g_assert(in_code_gen_buffer(rx));
99     }
100     return (void *)rx;
101 }
102 #endif /* CONFIG_DEBUG_TCG */
103 
104 /* compare a pointer @ptr and a tb_tc @s */
105 static int ptr_cmp_tb_tc(const void *ptr, const struct tb_tc *s)
106 {
107     if (ptr >= s->ptr + s->size) {
108         return 1;
109     } else if (ptr < s->ptr) {
110         return -1;
111     }
112     return 0;
113 }
114 
115 static gint tb_tc_cmp(gconstpointer ap, gconstpointer bp, gpointer userdata)
116 {
117     const struct tb_tc *a = ap;
118     const struct tb_tc *b = bp;
119 
120     /*
121      * When both sizes are set, we know this isn't a lookup.
122      * This is the most likely case: every TB must be inserted; lookups
123      * are a lot less frequent.
124      */
125     if (likely(a->size && b->size)) {
126         if (a->ptr > b->ptr) {
127             return 1;
128         } else if (a->ptr < b->ptr) {
129             return -1;
130         }
131         /* a->ptr == b->ptr should happen only on deletions */
132         g_assert(a->size == b->size);
133         return 0;
134     }
135     /*
136      * All lookups have either .size field set to 0.
137      * From the glib sources we see that @ap is always the lookup key. However
138      * the docs provide no guarantee, so we just mark this case as likely.
139      */
140     if (likely(a->size == 0)) {
141         return ptr_cmp_tb_tc(a->ptr, b);
142     }
143     return ptr_cmp_tb_tc(b->ptr, a);
144 }
145 
146 static void tb_destroy(gpointer value)
147 {
148     TranslationBlock *tb = value;
149     qemu_spin_destroy(&tb->jmp_lock);
150 }
151 
152 static void tcg_region_trees_init(void)
153 {
154     size_t i;
155 
156     tree_size = ROUND_UP(sizeof(struct tcg_region_tree), qemu_dcache_linesize);
157     region_trees = qemu_memalign(qemu_dcache_linesize, region.n * tree_size);
158     for (i = 0; i < region.n; i++) {
159         struct tcg_region_tree *rt = region_trees + i * tree_size;
160 
161         qemu_mutex_init(&rt->lock);
162         rt->tree = g_tree_new_full(tb_tc_cmp, NULL, NULL, tb_destroy);
163     }
164 }
165 
166 static struct tcg_region_tree *tc_ptr_to_region_tree(const void *p)
167 {
168     size_t region_idx;
169 
170     /*
171      * Like tcg_splitwx_to_rw, with no assert.  The pc may come from
172      * a signal handler over which the caller has no control.
173      */
174     if (!in_code_gen_buffer(p)) {
175         p -= tcg_splitwx_diff;
176         if (!in_code_gen_buffer(p)) {
177             return NULL;
178         }
179     }
180 
181     if (p < region.start_aligned) {
182         region_idx = 0;
183     } else {
184         ptrdiff_t offset = p - region.start_aligned;
185 
186         if (offset > region.stride * (region.n - 1)) {
187             region_idx = region.n - 1;
188         } else {
189             region_idx = offset / region.stride;
190         }
191     }
192     return region_trees + region_idx * tree_size;
193 }
194 
195 void tcg_tb_insert(TranslationBlock *tb)
196 {
197     struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr);
198 
199     g_assert(rt != NULL);
200     qemu_mutex_lock(&rt->lock);
201     g_tree_insert(rt->tree, &tb->tc, tb);
202     qemu_mutex_unlock(&rt->lock);
203 }
204 
205 void tcg_tb_remove(TranslationBlock *tb)
206 {
207     struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr);
208 
209     g_assert(rt != NULL);
210     qemu_mutex_lock(&rt->lock);
211     g_tree_remove(rt->tree, &tb->tc);
212     qemu_mutex_unlock(&rt->lock);
213 }
214 
215 /*
216  * Find the TB 'tb' such that
217  * tb->tc.ptr <= tc_ptr < tb->tc.ptr + tb->tc.size
218  * Return NULL if not found.
219  */
220 TranslationBlock *tcg_tb_lookup(uintptr_t tc_ptr)
221 {
222     struct tcg_region_tree *rt = tc_ptr_to_region_tree((void *)tc_ptr);
223     TranslationBlock *tb;
224     struct tb_tc s = { .ptr = (void *)tc_ptr };
225 
226     if (rt == NULL) {
227         return NULL;
228     }
229 
230     qemu_mutex_lock(&rt->lock);
231     tb = g_tree_lookup(rt->tree, &s);
232     qemu_mutex_unlock(&rt->lock);
233     return tb;
234 }
235 
236 static void tcg_region_tree_lock_all(void)
237 {
238     size_t i;
239 
240     for (i = 0; i < region.n; i++) {
241         struct tcg_region_tree *rt = region_trees + i * tree_size;
242 
243         qemu_mutex_lock(&rt->lock);
244     }
245 }
246 
247 static void tcg_region_tree_unlock_all(void)
248 {
249     size_t i;
250 
251     for (i = 0; i < region.n; i++) {
252         struct tcg_region_tree *rt = region_trees + i * tree_size;
253 
254         qemu_mutex_unlock(&rt->lock);
255     }
256 }
257 
258 void tcg_tb_foreach(GTraverseFunc func, gpointer user_data)
259 {
260     size_t i;
261 
262     tcg_region_tree_lock_all();
263     for (i = 0; i < region.n; i++) {
264         struct tcg_region_tree *rt = region_trees + i * tree_size;
265 
266         g_tree_foreach(rt->tree, func, user_data);
267     }
268     tcg_region_tree_unlock_all();
269 }
270 
271 size_t tcg_nb_tbs(void)
272 {
273     size_t nb_tbs = 0;
274     size_t i;
275 
276     tcg_region_tree_lock_all();
277     for (i = 0; i < region.n; i++) {
278         struct tcg_region_tree *rt = region_trees + i * tree_size;
279 
280         nb_tbs += g_tree_nnodes(rt->tree);
281     }
282     tcg_region_tree_unlock_all();
283     return nb_tbs;
284 }
285 
286 static void tcg_region_tree_reset_all(void)
287 {
288     size_t i;
289 
290     tcg_region_tree_lock_all();
291     for (i = 0; i < region.n; i++) {
292         struct tcg_region_tree *rt = region_trees + i * tree_size;
293 
294         /* Increment the refcount first so that destroy acts as a reset */
295         g_tree_ref(rt->tree);
296         g_tree_destroy(rt->tree);
297     }
298     tcg_region_tree_unlock_all();
299 }
300 
301 static void tcg_region_bounds(size_t curr_region, void **pstart, void **pend)
302 {
303     void *start, *end;
304 
305     start = region.start_aligned + curr_region * region.stride;
306     end = start + region.size;
307 
308     if (curr_region == 0) {
309         start = region.after_prologue;
310     }
311     /* The final region may have a few extra pages due to earlier rounding. */
312     if (curr_region == region.n - 1) {
313         end = region.start_aligned + region.total_size;
314     }
315 
316     *pstart = start;
317     *pend = end;
318 }
319 
320 static void tcg_region_assign(TCGContext *s, size_t curr_region)
321 {
322     void *start, *end;
323 
324     tcg_region_bounds(curr_region, &start, &end);
325 
326     s->code_gen_buffer = start;
327     s->code_gen_ptr = start;
328     s->code_gen_buffer_size = end - start;
329     s->code_gen_highwater = end - TCG_HIGHWATER;
330 }
331 
332 static bool tcg_region_alloc__locked(TCGContext *s)
333 {
334     if (region.current == region.n) {
335         return true;
336     }
337     tcg_region_assign(s, region.current);
338     region.current++;
339     return false;
340 }
341 
342 /*
343  * Request a new region once the one in use has filled up.
344  * Returns true on error.
345  */
346 bool tcg_region_alloc(TCGContext *s)
347 {
348     bool err;
349     /* read the region size now; alloc__locked will overwrite it on success */
350     size_t size_full = s->code_gen_buffer_size;
351 
352     qemu_mutex_lock(&region.lock);
353     err = tcg_region_alloc__locked(s);
354     if (!err) {
355         region.agg_size_full += size_full - TCG_HIGHWATER;
356     }
357     qemu_mutex_unlock(&region.lock);
358     return err;
359 }
360 
361 /*
362  * Perform a context's first region allocation.
363  * This function does _not_ increment region.agg_size_full.
364  */
365 static void tcg_region_initial_alloc__locked(TCGContext *s)
366 {
367     bool err = tcg_region_alloc__locked(s);
368     g_assert(!err);
369 }
370 
371 void tcg_region_initial_alloc(TCGContext *s)
372 {
373     qemu_mutex_lock(&region.lock);
374     tcg_region_initial_alloc__locked(s);
375     qemu_mutex_unlock(&region.lock);
376 }
377 
378 /* Call from a safe-work context */
379 void tcg_region_reset_all(void)
380 {
381     unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs);
382     unsigned int i;
383 
384     qemu_mutex_lock(&region.lock);
385     region.current = 0;
386     region.agg_size_full = 0;
387 
388     for (i = 0; i < n_ctxs; i++) {
389         TCGContext *s = qatomic_read(&tcg_ctxs[i]);
390         tcg_region_initial_alloc__locked(s);
391     }
392     qemu_mutex_unlock(&region.lock);
393 
394     tcg_region_tree_reset_all();
395 }
396 
397 static size_t tcg_n_regions(size_t tb_size, unsigned max_cpus)
398 {
399 #ifdef CONFIG_USER_ONLY
400     return 1;
401 #else
402     size_t n_regions;
403 
404     /*
405      * It is likely that some vCPUs will translate more code than others,
406      * so we first try to set more regions than max_cpus, with those regions
407      * being of reasonable size. If that's not possible we make do by evenly
408      * dividing the code_gen_buffer among the vCPUs.
409      */
410     /* Use a single region if all we have is one vCPU thread */
411     if (max_cpus == 1 || !qemu_tcg_mttcg_enabled()) {
412         return 1;
413     }
414 
415     /*
416      * Try to have more regions than max_cpus, with each region being >= 2 MB.
417      * If we can't, then just allocate one region per vCPU thread.
418      */
419     n_regions = tb_size / (2 * MiB);
420     if (n_regions <= max_cpus) {
421         return max_cpus;
422     }
423     return MIN(n_regions, max_cpus * 8);
424 #endif
425 }
426 
427 /*
428  * Minimum size of the code gen buffer.  This number is randomly chosen,
429  * but not so small that we can't have a fair number of TB's live.
430  *
431  * Maximum size, MAX_CODE_GEN_BUFFER_SIZE, is defined in tcg-target.h.
432  * Unless otherwise indicated, this is constrained by the range of
433  * direct branches on the host cpu, as used by the TCG implementation
434  * of goto_tb.
435  */
436 #define MIN_CODE_GEN_BUFFER_SIZE     (1 * MiB)
437 
438 #if TCG_TARGET_REG_BITS == 32
439 #define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (32 * MiB)
440 #ifdef CONFIG_USER_ONLY
441 /*
442  * For user mode on smaller 32 bit systems we may run into trouble
443  * allocating big chunks of data in the right place. On these systems
444  * we utilise a static code generation buffer directly in the binary.
445  */
446 #define USE_STATIC_CODE_GEN_BUFFER
447 #endif
448 #else /* TCG_TARGET_REG_BITS == 64 */
449 #ifdef CONFIG_USER_ONLY
450 /*
451  * As user-mode emulation typically means running multiple instances
452  * of the translator don't go too nuts with our default code gen
453  * buffer lest we make things too hard for the OS.
454  */
455 #define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (128 * MiB)
456 #else
457 /*
458  * We expect most system emulation to run one or two guests per host.
459  * Users running large scale system emulation may want to tweak their
460  * runtime setup via the tb-size control on the command line.
461  */
462 #define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (1 * GiB)
463 #endif
464 #endif
465 
466 #define DEFAULT_CODE_GEN_BUFFER_SIZE \
467   (DEFAULT_CODE_GEN_BUFFER_SIZE_1 < MAX_CODE_GEN_BUFFER_SIZE \
468    ? DEFAULT_CODE_GEN_BUFFER_SIZE_1 : MAX_CODE_GEN_BUFFER_SIZE)
469 
470 #ifdef USE_STATIC_CODE_GEN_BUFFER
471 static uint8_t static_code_gen_buffer[DEFAULT_CODE_GEN_BUFFER_SIZE]
472     __attribute__((aligned(CODE_GEN_ALIGN)));
473 
474 static int alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp)
475 {
476     void *buf, *end;
477     size_t size;
478 
479     if (splitwx > 0) {
480         error_setg(errp, "jit split-wx not supported");
481         return -1;
482     }
483 
484     /* page-align the beginning and end of the buffer */
485     buf = static_code_gen_buffer;
486     end = static_code_gen_buffer + sizeof(static_code_gen_buffer);
487     buf = QEMU_ALIGN_PTR_UP(buf, qemu_real_host_page_size);
488     end = QEMU_ALIGN_PTR_DOWN(end, qemu_real_host_page_size);
489 
490     size = end - buf;
491 
492     /* Honor a command-line option limiting the size of the buffer.  */
493     if (size > tb_size) {
494         size = QEMU_ALIGN_DOWN(tb_size, qemu_real_host_page_size);
495     }
496 
497     region.start_aligned = buf;
498     region.total_size = size;
499 
500     return PROT_READ | PROT_WRITE;
501 }
502 #elif defined(_WIN32)
503 static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp)
504 {
505     void *buf;
506 
507     if (splitwx > 0) {
508         error_setg(errp, "jit split-wx not supported");
509         return -1;
510     }
511 
512     buf = VirtualAlloc(NULL, size, MEM_RESERVE | MEM_COMMIT,
513                              PAGE_EXECUTE_READWRITE);
514     if (buf == NULL) {
515         error_setg_win32(errp, GetLastError(),
516                          "allocate %zu bytes for jit buffer", size);
517         return false;
518     }
519 
520     region.start_aligned = buf;
521     region.total_size = size;
522 
523     return PAGE_READ | PAGE_WRITE | PAGE_EXEC;
524 }
525 #else
526 static int alloc_code_gen_buffer_anon(size_t size, int prot,
527                                       int flags, Error **errp)
528 {
529     void *buf;
530 
531     buf = mmap(NULL, size, prot, flags, -1, 0);
532     if (buf == MAP_FAILED) {
533         error_setg_errno(errp, errno,
534                          "allocate %zu bytes for jit buffer", size);
535         return -1;
536     }
537 
538     region.start_aligned = buf;
539     region.total_size = size;
540     return prot;
541 }
542 
543 #ifndef CONFIG_TCG_INTERPRETER
544 #ifdef CONFIG_POSIX
545 #include "qemu/memfd.h"
546 
547 static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp)
548 {
549     void *buf_rw = NULL, *buf_rx = MAP_FAILED;
550     int fd = -1;
551 
552     buf_rw = qemu_memfd_alloc("tcg-jit", size, 0, &fd, errp);
553     if (buf_rw == NULL) {
554         goto fail;
555     }
556 
557     buf_rx = mmap(NULL, size, PROT_READ | PROT_EXEC, MAP_SHARED, fd, 0);
558     if (buf_rx == MAP_FAILED) {
559         goto fail_rx;
560     }
561 
562     close(fd);
563     region.start_aligned = buf_rw;
564     region.total_size = size;
565     tcg_splitwx_diff = buf_rx - buf_rw;
566 
567     return PROT_READ | PROT_WRITE;
568 
569  fail_rx:
570     error_setg_errno(errp, errno, "failed to map shared memory for execute");
571  fail:
572     if (buf_rx != MAP_FAILED) {
573         munmap(buf_rx, size);
574     }
575     if (buf_rw) {
576         munmap(buf_rw, size);
577     }
578     if (fd >= 0) {
579         close(fd);
580     }
581     return -1;
582 }
583 #endif /* CONFIG_POSIX */
584 
585 #ifdef CONFIG_DARWIN
586 #include <mach/mach.h>
587 
588 extern kern_return_t mach_vm_remap(vm_map_t target_task,
589                                    mach_vm_address_t *target_address,
590                                    mach_vm_size_t size,
591                                    mach_vm_offset_t mask,
592                                    int flags,
593                                    vm_map_t src_task,
594                                    mach_vm_address_t src_address,
595                                    boolean_t copy,
596                                    vm_prot_t *cur_protection,
597                                    vm_prot_t *max_protection,
598                                    vm_inherit_t inheritance);
599 
600 static int alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp)
601 {
602     kern_return_t ret;
603     mach_vm_address_t buf_rw, buf_rx;
604     vm_prot_t cur_prot, max_prot;
605 
606     /* Map the read-write portion via normal anon memory. */
607     if (!alloc_code_gen_buffer_anon(size, PROT_READ | PROT_WRITE,
608                                     MAP_PRIVATE | MAP_ANONYMOUS, errp)) {
609         return -1;
610     }
611 
612     buf_rw = (mach_vm_address_t)region.start_aligned;
613     buf_rx = 0;
614     ret = mach_vm_remap(mach_task_self(),
615                         &buf_rx,
616                         size,
617                         0,
618                         VM_FLAGS_ANYWHERE,
619                         mach_task_self(),
620                         buf_rw,
621                         false,
622                         &cur_prot,
623                         &max_prot,
624                         VM_INHERIT_NONE);
625     if (ret != KERN_SUCCESS) {
626         /* TODO: Convert "ret" to a human readable error message. */
627         error_setg(errp, "vm_remap for jit splitwx failed");
628         munmap((void *)buf_rw, size);
629         return -1;
630     }
631 
632     if (mprotect((void *)buf_rx, size, PROT_READ | PROT_EXEC) != 0) {
633         error_setg_errno(errp, errno, "mprotect for jit splitwx");
634         munmap((void *)buf_rx, size);
635         munmap((void *)buf_rw, size);
636         return -1;
637     }
638 
639     tcg_splitwx_diff = buf_rx - buf_rw;
640     return PROT_READ | PROT_WRITE;
641 }
642 #endif /* CONFIG_DARWIN */
643 #endif /* CONFIG_TCG_INTERPRETER */
644 
645 static int alloc_code_gen_buffer_splitwx(size_t size, Error **errp)
646 {
647 #ifndef CONFIG_TCG_INTERPRETER
648 # ifdef CONFIG_DARWIN
649     return alloc_code_gen_buffer_splitwx_vmremap(size, errp);
650 # endif
651 # ifdef CONFIG_POSIX
652     return alloc_code_gen_buffer_splitwx_memfd(size, errp);
653 # endif
654 #endif
655     error_setg(errp, "jit split-wx not supported");
656     return -1;
657 }
658 
659 static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp)
660 {
661     ERRP_GUARD();
662     int prot, flags;
663 
664     if (splitwx) {
665         prot = alloc_code_gen_buffer_splitwx(size, errp);
666         if (prot >= 0) {
667             return prot;
668         }
669         /*
670          * If splitwx force-on (1), fail;
671          * if splitwx default-on (-1), fall through to splitwx off.
672          */
673         if (splitwx > 0) {
674             return -1;
675         }
676         error_free_or_abort(errp);
677     }
678 
679     /*
680      * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect
681      * rejects a permission change from RWX -> NONE when reserving the
682      * guard pages later.  We can go the other way with the same number
683      * of syscalls, so always begin with PROT_NONE.
684      */
685     prot = PROT_NONE;
686     flags = MAP_PRIVATE | MAP_ANONYMOUS;
687 #ifdef CONFIG_DARWIN
688     /* Applicable to both iOS and macOS (Apple Silicon). */
689     if (!splitwx) {
690         flags |= MAP_JIT;
691     }
692 #endif
693 
694     return alloc_code_gen_buffer_anon(size, prot, flags, errp);
695 }
696 #endif /* USE_STATIC_CODE_GEN_BUFFER, WIN32, POSIX */
697 
698 /*
699  * Initializes region partitioning.
700  *
701  * Called at init time from the parent thread (i.e. the one calling
702  * tcg_context_init), after the target's TCG globals have been set.
703  *
704  * Region partitioning works by splitting code_gen_buffer into separate regions,
705  * and then assigning regions to TCG threads so that the threads can translate
706  * code in parallel without synchronization.
707  *
708  * In softmmu the number of TCG threads is bounded by max_cpus, so we use at
709  * least max_cpus regions in MTTCG. In !MTTCG we use a single region.
710  * Note that the TCG options from the command-line (i.e. -accel accel=tcg,[...])
711  * must have been parsed before calling this function, since it calls
712  * qemu_tcg_mttcg_enabled().
713  *
714  * In user-mode we use a single region.  Having multiple regions in user-mode
715  * is not supported, because the number of vCPU threads (recall that each thread
716  * spawned by the guest corresponds to a vCPU thread) is only bounded by the
717  * OS, and usually this number is huge (tens of thousands is not uncommon).
718  * Thus, given this large bound on the number of vCPU threads and the fact
719  * that code_gen_buffer is allocated at compile-time, we cannot guarantee
720  * that the availability of at least one region per vCPU thread.
721  *
722  * However, this user-mode limitation is unlikely to be a significant problem
723  * in practice. Multi-threaded guests share most if not all of their translated
724  * code, which makes parallel code generation less appealing than in softmmu.
725  */
726 void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus)
727 {
728     const size_t page_size = qemu_real_host_page_size;
729     size_t region_size;
730     int have_prot, need_prot;
731 
732     /* Size the buffer.  */
733     if (tb_size == 0) {
734         size_t phys_mem = qemu_get_host_physmem();
735         if (phys_mem == 0) {
736             tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE;
737         } else {
738             tb_size = QEMU_ALIGN_DOWN(phys_mem / 8, page_size);
739             tb_size = MIN(DEFAULT_CODE_GEN_BUFFER_SIZE, tb_size);
740         }
741     }
742     if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) {
743         tb_size = MIN_CODE_GEN_BUFFER_SIZE;
744     }
745     if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) {
746         tb_size = MAX_CODE_GEN_BUFFER_SIZE;
747     }
748 
749     have_prot = alloc_code_gen_buffer(tb_size, splitwx, &error_fatal);
750     assert(have_prot >= 0);
751 
752     /* Request large pages for the buffer and the splitwx.  */
753     qemu_madvise(region.start_aligned, region.total_size, QEMU_MADV_HUGEPAGE);
754     if (tcg_splitwx_diff) {
755         qemu_madvise(region.start_aligned + tcg_splitwx_diff,
756                      region.total_size, QEMU_MADV_HUGEPAGE);
757     }
758 
759     /*
760      * Make region_size a multiple of page_size, using aligned as the start.
761      * As a result of this we might end up with a few extra pages at the end of
762      * the buffer; we will assign those to the last region.
763      */
764     region.n = tcg_n_regions(tb_size, max_cpus);
765     region_size = tb_size / region.n;
766     region_size = QEMU_ALIGN_DOWN(region_size, page_size);
767 
768     /* A region must have at least 2 pages; one code, one guard */
769     g_assert(region_size >= 2 * page_size);
770     region.stride = region_size;
771 
772     /* Reserve space for guard pages. */
773     region.size = region_size - page_size;
774     region.total_size -= page_size;
775 
776     /*
777      * The first region will be smaller than the others, via the prologue,
778      * which has yet to be allocated.  For now, the first region begins at
779      * the page boundary.
780      */
781     region.after_prologue = region.start_aligned;
782 
783     /* init the region struct */
784     qemu_mutex_init(&region.lock);
785 
786     /*
787      * Set guard pages in the rw buffer, as that's the one into which
788      * buffer overruns could occur.  Do not set guard pages in the rx
789      * buffer -- let that one use hugepages throughout.
790      * Work with the page protections set up with the initial mapping.
791      */
792     need_prot = PAGE_READ | PAGE_WRITE;
793 #ifndef CONFIG_TCG_INTERPRETER
794     if (tcg_splitwx_diff == 0) {
795         need_prot |= PAGE_EXEC;
796     }
797 #endif
798     for (size_t i = 0, n = region.n; i < n; i++) {
799         void *start, *end;
800 
801         tcg_region_bounds(i, &start, &end);
802         if (have_prot != need_prot) {
803             int rc;
804 
805             if (need_prot == (PAGE_READ | PAGE_WRITE | PAGE_EXEC)) {
806                 rc = qemu_mprotect_rwx(start, end - start);
807             } else if (need_prot == (PAGE_READ | PAGE_WRITE)) {
808                 rc = qemu_mprotect_rw(start, end - start);
809             } else {
810                 g_assert_not_reached();
811             }
812             if (rc) {
813                 error_setg_errno(&error_fatal, errno,
814                                  "mprotect of jit buffer");
815             }
816         }
817         if (have_prot != 0) {
818             /* Guard pages are nice for bug detection but are not essential. */
819             (void)qemu_mprotect_none(end, page_size);
820         }
821     }
822 
823     tcg_region_trees_init();
824 
825     /*
826      * Leave the initial context initialized to the first region.
827      * This will be the context into which we generate the prologue.
828      * It is also the only context for CONFIG_USER_ONLY.
829      */
830     tcg_region_initial_alloc__locked(&tcg_init_ctx);
831 }
832 
833 void tcg_region_prologue_set(TCGContext *s)
834 {
835     /* Deduct the prologue from the first region.  */
836     g_assert(region.start_aligned == s->code_gen_buffer);
837     region.after_prologue = s->code_ptr;
838 
839     /* Recompute boundaries of the first region. */
840     tcg_region_assign(s, 0);
841 
842     /* Register the balance of the buffer with gdb. */
843     tcg_register_jit(tcg_splitwx_to_rx(region.after_prologue),
844                      region.start_aligned + region.total_size -
845                      region.after_prologue);
846 }
847 
848 /*
849  * Returns the size (in bytes) of all translated code (i.e. from all regions)
850  * currently in the cache.
851  * See also: tcg_code_capacity()
852  * Do not confuse with tcg_current_code_size(); that one applies to a single
853  * TCG context.
854  */
855 size_t tcg_code_size(void)
856 {
857     unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs);
858     unsigned int i;
859     size_t total;
860 
861     qemu_mutex_lock(&region.lock);
862     total = region.agg_size_full;
863     for (i = 0; i < n_ctxs; i++) {
864         const TCGContext *s = qatomic_read(&tcg_ctxs[i]);
865         size_t size;
866 
867         size = qatomic_read(&s->code_gen_ptr) - s->code_gen_buffer;
868         g_assert(size <= s->code_gen_buffer_size);
869         total += size;
870     }
871     qemu_mutex_unlock(&region.lock);
872     return total;
873 }
874 
875 /*
876  * Returns the code capacity (in bytes) of the entire cache, i.e. including all
877  * regions.
878  * See also: tcg_code_size()
879  */
880 size_t tcg_code_capacity(void)
881 {
882     size_t guard_size, capacity;
883 
884     /* no need for synchronization; these variables are set at init time */
885     guard_size = region.stride - region.size;
886     capacity = region.total_size;
887     capacity -= (region.n - 1) * guard_size;
888     capacity -= region.n * TCG_HIGHWATER;
889 
890     return capacity;
891 }
892