xref: /qemu/contrib/plugins/cflow.c (revision 513823e7521a09ed7ad1e32e6454bac3b2cbf52d)
1 /*
2  * Control Flow plugin
3  *
4  * This plugin will track changes to control flow and detect where
5  * instructions fault.
6  *
7  * Copyright (c) 2024 Linaro Ltd
8  *
9  * SPDX-License-Identifier: GPL-2.0-or-later
10  */
11 #include <glib.h>
12 #include <inttypes.h>
13 #include <stdio.h>
14 #include <stdlib.h>
15 #include <string.h>
16 #include <unistd.h>
17 
18 #include <qemu-plugin.h>
19 
20 QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION;
21 
22 typedef enum {
23     SORT_HOTTEST,  /* hottest branch insn */
24     SORT_EXCEPTION,    /* most early exits */
25     SORT_POPDEST,  /* most destinations (usually ret's) */
26 } ReportType;
27 
28 ReportType report = SORT_HOTTEST;
29 int topn = 10;
30 
31 typedef struct {
32     uint64_t daddr;
33     uint64_t dcount;
34 } DestData;
35 
36 /* A node is an address where we can go to multiple places */
37 typedef struct {
38     GMutex lock;
39     /* address of the branch point */
40     uint64_t addr;
41     /* array of DestData */
42     GArray *dests;
43     /* early exit/fault count */
44     uint64_t early_exit;
45     /* jump destination count */
46     uint64_t dest_count;
47     /* instruction data */
48     char *insn_disas;
49     /* symbol? */
50     const char *symbol;
51     /* times translated as last in block? */
52     int last_count;
53     /* times translated in the middle of block? */
54     int mid_count;
55 } NodeData;
56 
57 typedef enum {
58     /* last insn in block, expected flow control */
59     LAST_INSN = (1 << 0),
60     /* mid-block insn, can only be an exception */
61     EXCP_INSN = (1 << 1),
62     /* multiple disassembly, may have changed */
63     MULT_INSN = (1 << 2),
64 } InsnTypes;
65 
66 typedef struct {
67     /* address of the branch point */
68     uint64_t addr;
69     /* disassembly */
70     char *insn_disas;
71     /* symbol? */
72     const char *symbol;
73     /* types */
74     InsnTypes type_flag;
75 } InsnData;
76 
77 /* We use this to track the current execution state */
78 typedef struct {
79     /* address of current translated block */
80     uint64_t tb_pc;
81     /* address of end of block */
82     uint64_t end_block;
83     /* next pc after end of block */
84     uint64_t pc_after_block;
85     /* address of last executed PC */
86     uint64_t last_pc;
87 } VCPUScoreBoard;
88 
89 /* descriptors for accessing the above scoreboard */
90 static qemu_plugin_u64 tb_pc;
91 static qemu_plugin_u64 end_block;
92 static qemu_plugin_u64 pc_after_block;
93 static qemu_plugin_u64 last_pc;
94 
95 
96 static GMutex node_lock;
97 static GHashTable *nodes;
98 struct qemu_plugin_scoreboard *state;
99 
100 /* SORT_HOTTEST */
101 static gint hottest(gconstpointer a, gconstpointer b)
102 {
103     NodeData *na = (NodeData *) a;
104     NodeData *nb = (NodeData *) b;
105 
106     return na->dest_count > nb->dest_count ? -1 :
107         na->dest_count == nb->dest_count ? 0 : 1;
108 }
109 
110 static gint exception(gconstpointer a, gconstpointer b)
111 {
112     NodeData *na = (NodeData *) a;
113     NodeData *nb = (NodeData *) b;
114 
115     return na->early_exit > nb->early_exit ? -1 :
116         na->early_exit == nb->early_exit ? 0 : 1;
117 }
118 
119 static gint popular(gconstpointer a, gconstpointer b)
120 {
121     NodeData *na = (NodeData *) a;
122     NodeData *nb = (NodeData *) b;
123 
124     return na->dests->len > nb->dests->len ? -1 :
125         na->dests->len == nb->dests->len ? 0 : 1;
126 }
127 
128 /* Filter out non-branches - returns true to remove entry */
129 static gboolean filter_non_branches(gpointer key, gpointer value,
130                                     gpointer user_data)
131 {
132     NodeData *node = (NodeData *) value;
133 
134     return node->dest_count == 0;
135 }
136 
137 static void plugin_exit(qemu_plugin_id_t id, void *p)
138 {
139     g_autoptr(GString) result = g_string_new("collected ");
140     GList *data;
141     GCompareFunc sort = &hottest;
142     int i = 0;
143 
144     g_mutex_lock(&node_lock);
145     g_string_append_printf(result, "%d control flow nodes in the hash table\n",
146                            g_hash_table_size(nodes));
147 
148     /* remove all nodes that didn't branch */
149     g_hash_table_foreach_remove(nodes, filter_non_branches, NULL);
150 
151     data = g_hash_table_get_values(nodes);
152 
153     switch (report) {
154     case SORT_HOTTEST:
155         sort = &hottest;
156         break;
157     case SORT_EXCEPTION:
158         sort = &exception;
159         break;
160     case SORT_POPDEST:
161         sort = &popular;
162         break;
163     }
164 
165     data = g_list_sort(data, sort);
166 
167     for (GList *l = data;
168          l != NULL && i < topn;
169          l = l->next, i++) {
170         NodeData *n = l->data;
171         const char *type = n->mid_count ? "sync fault" : "branch";
172         g_string_append_printf(result, "  addr: 0x%"PRIx64 " %s: %s (%s)\n",
173                                n->addr, n->symbol, n->insn_disas, type);
174         if (n->early_exit) {
175             g_string_append_printf(result, "    early exits %"PRId64"\n",
176                                    n->early_exit);
177         }
178         g_string_append_printf(result, "    branches %"PRId64"\n",
179                                n->dest_count);
180         for (int j = 0; j < n->dests->len; j++) {
181             DestData *dd = &g_array_index(n->dests, DestData, j);
182             g_string_append_printf(result, "      to 0x%"PRIx64" (%"PRId64")\n",
183                                    dd->daddr, dd->dcount);
184         }
185     }
186 
187     qemu_plugin_outs(result->str);
188 
189     g_mutex_unlock(&node_lock);
190 }
191 
192 static void plugin_init(void)
193 {
194     g_mutex_init(&node_lock);
195     nodes = g_hash_table_new(g_int64_hash, g_int64_equal);
196     state = qemu_plugin_scoreboard_new(sizeof(VCPUScoreBoard));
197 
198     /* score board declarations */
199     tb_pc = qemu_plugin_scoreboard_u64_in_struct(state, VCPUScoreBoard, tb_pc);
200     end_block = qemu_plugin_scoreboard_u64_in_struct(state, VCPUScoreBoard,
201                                                      end_block);
202     pc_after_block = qemu_plugin_scoreboard_u64_in_struct(state, VCPUScoreBoard,
203                                                           pc_after_block);
204     last_pc = qemu_plugin_scoreboard_u64_in_struct(state, VCPUScoreBoard,
205                                                    last_pc);
206 }
207 
208 static NodeData *create_node(uint64_t addr)
209 {
210     NodeData *node = g_new0(NodeData, 1);
211     g_mutex_init(&node->lock);
212     node->addr = addr;
213     node->dests = g_array_new(true, true, sizeof(DestData));
214     return node;
215 }
216 
217 static NodeData *fetch_node(uint64_t addr, bool create_if_not_found)
218 {
219     NodeData *node = NULL;
220 
221     g_mutex_lock(&node_lock);
222     node = (NodeData *) g_hash_table_lookup(nodes, &addr);
223     if (!node && create_if_not_found) {
224         node = create_node(addr);
225         g_hash_table_insert(nodes, &node->addr, node);
226     }
227     g_mutex_unlock(&node_lock);
228     return node;
229 }
230 
231 /*
232  * Called when we detect a non-linear execution (pc !=
233  * pc_after_block). This could be due to a fault causing some sort of
234  * exit exception (if last_pc != block_end) or just a taken branch.
235  */
236 static void vcpu_tb_branched_exec(unsigned int cpu_index, void *udata)
237 {
238     uint64_t lpc = qemu_plugin_u64_get(last_pc, cpu_index);
239     uint64_t ebpc = qemu_plugin_u64_get(end_block, cpu_index);
240     uint64_t npc = qemu_plugin_u64_get(pc_after_block, cpu_index);
241     uint64_t pc = qemu_plugin_u64_get(tb_pc, cpu_index);
242 
243     /* return early for address 0 */
244     if (!lpc) {
245         return;
246     }
247 
248     NodeData *node = fetch_node(lpc, true);
249     DestData *data = NULL;
250     bool early_exit = (lpc != ebpc);
251     GArray *dests;
252 
253     /* the condition should never hit */
254     g_assert(pc != npc);
255 
256     g_mutex_lock(&node->lock);
257 
258     if (early_exit) {
259         fprintf(stderr, "%s: pc=%"PRIx64", epbc=%"PRIx64
260                 " npc=%"PRIx64", lpc=%"PRIx64"\n",
261                 __func__, pc, ebpc, npc, lpc);
262         node->early_exit++;
263         if (!node->mid_count) {
264             /* count now as we've only just allocated */
265             node->mid_count++;
266         }
267     }
268 
269     dests = node->dests;
270     for (int i = 0; i < dests->len; i++) {
271         if (g_array_index(dests, DestData, i).daddr == pc) {
272             data = &g_array_index(dests, DestData, i);
273         }
274     }
275 
276     /* we've never seen this before, allocate a new entry */
277     if (!data) {
278         DestData new_entry = { .daddr = pc };
279         g_array_append_val(dests, new_entry);
280         data = &g_array_index(dests, DestData, dests->len - 1);
281         g_assert(data->daddr == pc);
282     }
283 
284     data->dcount++;
285     node->dest_count++;
286 
287     g_mutex_unlock(&node->lock);
288 }
289 
290 /*
291  * At the start of each block we need to resolve two things:
292  *
293  *  - is last_pc == block_end, if not we had an early exit
294  *  - is start of block last_pc + insn width, if not we jumped
295  *
296  * Once those are dealt with we can instrument the rest of the
297  * instructions for their execution.
298  *
299  */
300 static void vcpu_tb_trans(qemu_plugin_id_t id, struct qemu_plugin_tb *tb)
301 {
302     uint64_t pc = qemu_plugin_tb_vaddr(tb);
303     size_t insns = qemu_plugin_tb_n_insns(tb);
304     struct qemu_plugin_insn *first_insn = qemu_plugin_tb_get_insn(tb, 0);
305     struct qemu_plugin_insn *last_insn = qemu_plugin_tb_get_insn(tb, insns - 1);
306 
307     /*
308      * check if we are executing linearly after the last block. We can
309      * handle both early block exits and normal branches in the
310      * callback if we hit it.
311      */
312     qemu_plugin_register_vcpu_tb_exec_inline_per_vcpu(
313         tb, QEMU_PLUGIN_INLINE_STORE_U64, tb_pc, pc);
314     qemu_plugin_register_vcpu_tb_exec_cond_cb(
315         tb, vcpu_tb_branched_exec, QEMU_PLUGIN_CB_NO_REGS,
316         QEMU_PLUGIN_COND_NE, pc_after_block, pc, NULL);
317 
318     /*
319      * Now we can set start/end for this block so the next block can
320      * check where we are at. Do this on the first instruction and not
321      * the TB so we don't get mixed up with above.
322      */
323     qemu_plugin_register_vcpu_insn_exec_inline_per_vcpu(first_insn,
324                                                       QEMU_PLUGIN_INLINE_STORE_U64,
325                                                       end_block, qemu_plugin_insn_vaddr(last_insn));
326     qemu_plugin_register_vcpu_insn_exec_inline_per_vcpu(first_insn,
327                                                       QEMU_PLUGIN_INLINE_STORE_U64,
328                                                       pc_after_block,
329                                                       qemu_plugin_insn_vaddr(last_insn) +
330                                                       qemu_plugin_insn_size(last_insn));
331 
332     for (int idx = 0; idx < qemu_plugin_tb_n_insns(tb); ++idx) {
333         struct qemu_plugin_insn *insn = qemu_plugin_tb_get_insn(tb, idx);
334         uint64_t ipc = qemu_plugin_insn_vaddr(insn);
335         /*
336          * If this is a potential branch point check if we could grab
337          * the disassembly for it. If it is the last instruction
338          * always create an entry.
339          */
340         NodeData *node = fetch_node(ipc, last_insn);
341         if (node) {
342             g_mutex_lock(&node->lock);
343             if (!node->insn_disas) {
344                 node->insn_disas = qemu_plugin_insn_disas(insn);
345             }
346             if (!node->symbol) {
347                 node->symbol = qemu_plugin_insn_symbol(insn);
348             }
349             if (last_insn == insn) {
350                 node->last_count++;
351             } else {
352                 node->mid_count++;
353             }
354             g_mutex_unlock(&node->lock);
355         }
356 
357         /* Store the PC of what we are about to execute */
358         qemu_plugin_register_vcpu_insn_exec_inline_per_vcpu(insn,
359                                                             QEMU_PLUGIN_INLINE_STORE_U64,
360                                                             last_pc, ipc);
361     }
362 }
363 
364 QEMU_PLUGIN_EXPORT
365 int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
366                         int argc, char **argv)
367 {
368     for (int i = 0; i < argc; i++) {
369         char *opt = argv[i];
370         g_auto(GStrv) tokens = g_strsplit(opt, "=", 2);
371         if (g_strcmp0(tokens[0], "sort") == 0) {
372             if (g_strcmp0(tokens[1], "hottest") == 0) {
373                 report = SORT_HOTTEST;
374             } else if (g_strcmp0(tokens[1], "early") == 0) {
375                 report = SORT_EXCEPTION;
376             } else if (g_strcmp0(tokens[1], "exceptions") == 0) {
377                 report = SORT_POPDEST;
378             } else {
379                 fprintf(stderr, "failed to parse: %s\n", tokens[1]);
380                 return -1;
381             }
382         } else {
383             fprintf(stderr, "option parsing failed: %s\n", opt);
384             return -1;
385         }
386     }
387 
388     plugin_init();
389 
390     qemu_plugin_register_vcpu_tb_trans_cb(id, vcpu_tb_trans);
391     qemu_plugin_register_atexit_cb(id, plugin_exit, NULL);
392     return 0;
393 }
394