1.. 2 Copyright (C) 2017, Emilio G. Cota <cota@braap.org> 3 Copyright (c) 2019, Linaro Limited 4 Written by Emilio Cota and Alex Bennée 5 6================ 7QEMU TCG Plugins 8================ 9 10QEMU TCG plugins provide a way for users to run experiments taking 11advantage of the total system control emulation can have over a guest. 12It provides a mechanism for plugins to subscribe to events during 13translation and execution and optionally callback into the plugin 14during these events. TCG plugins are unable to change the system state 15only monitor it passively. However they can do this down to an 16individual instruction granularity including potentially subscribing 17to all load and store operations. 18 19API Stability 20============= 21 22This is a new feature for QEMU and it does allow people to develop 23out-of-tree plugins that can be dynamically linked into a running QEMU 24process. However the project reserves the right to change or break the 25API should it need to do so. The best way to avoid this is to submit 26your plugin upstream so they can be updated if/when the API changes. 27 28API versioning 29-------------- 30 31All plugins need to declare a symbol which exports the plugin API 32version they were built against. This can be done simply by:: 33 34 QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION; 35 36The core code will refuse to load a plugin that doesn't export a 37`qemu_plugin_version` symbol or if plugin version is outside of QEMU's 38supported range of API versions. 39 40Additionally the `qemu_info_t` structure which is passed to the 41`qemu_plugin_install` method of a plugin will detail the minimum and 42current API versions supported by QEMU. The API version will be 43incremented if new APIs are added. The minimum API version will be 44incremented if existing APIs are changed or removed. 45 46Exposure of QEMU internals 47-------------------------- 48 49The plugin architecture actively avoids leaking implementation details 50about how QEMU's translation works to the plugins. While there are 51conceptions such as translation time and translation blocks the 52details are opaque to plugins. The plugin is able to query select 53details of instructions and system configuration only through the 54exported *qemu_plugin* functions. 55 56Query Handle Lifetime 57--------------------- 58 59Each callback provides an opaque anonymous information handle which 60can usually be further queried to find out information about a 61translation, instruction or operation. The handles themselves are only 62valid during the lifetime of the callback so it is important that any 63information that is needed is extracted during the callback and saved 64by the plugin. 65 66Usage 67===== 68 69The QEMU binary needs to be compiled for plugin support:: 70 71 configure --enable-plugins 72 73Once built a program can be run with multiple plugins loaded each with 74their own arguments:: 75 76 $QEMU $OTHER_QEMU_ARGS \ 77 -plugin tests/plugin/libhowvec.so,arg=inline,arg=hint \ 78 -plugin tests/plugin/libhotblocks.so 79 80Arguments are plugin specific and can be used to modify their 81behaviour. In this case the howvec plugin is being asked to use inline 82ops to count and break down the hint instructions by type. 83 84Plugin Life cycle 85================= 86 87First the plugin is loaded and the public qemu_plugin_install function 88is called. The plugin will then register callbacks for various plugin 89events. Generally plugins will register a handler for the *atexit* 90if they want to dump a summary of collected information once the 91program/system has finished running. 92 93When a registered event occurs the plugin callback is invoked. The 94callbacks may provide additional information. In the case of a 95translation event the plugin has an option to enumerate the 96instructions in a block of instructions and optionally register 97callbacks to some or all instructions when they are executed. 98 99There is also a facility to add an inline event where code to 100increment a counter can be directly inlined with the translation. 101Currently only a simple increment is supported. This is not atomic so 102can miss counts. If you want absolute precision you should use a 103callback which can then ensure atomicity itself. 104 105Finally when QEMU exits all the registered *atexit* callbacks are 106invoked. 107 108Internals 109========= 110 111Locking 112------- 113 114We have to ensure we cannot deadlock, particularly under MTTCG. For 115this we acquire a lock when called from plugin code. We also keep the 116list of callbacks under RCU so that we do not have to hold the lock 117when calling the callbacks. This is also for performance, since some 118callbacks (e.g. memory access callbacks) might be called very 119frequently. 120 121 * A consequence of this is that we keep our own list of CPUs, so that 122 we do not have to worry about locking order wrt cpu_list_lock. 123 * Use a recursive lock, since we can get registration calls from 124 callbacks. 125 126As a result registering/unregistering callbacks is "slow", since it 127takes a lock. But this is very infrequent; we want performance when 128calling (or not calling) callbacks, not when registering them. Using 129RCU is great for this. 130 131We support the uninstallation of a plugin at any time (e.g. from 132plugin callbacks). This allows plugins to remove themselves if they no 133longer want to instrument the code. This operation is asynchronous 134which means callbacks may still occur after the uninstall operation is 135requested. The plugin isn't completely uninstalled until the safe work 136has executed while all vCPUs are quiescent. 137 138Example Plugins 139=============== 140 141There are a number of plugins included with QEMU and you are 142encouraged to contribute your own plugins plugins upstream. There is a 143`contrib/plugins` directory where they can go. 144 145- tests/plugins 146 147These are some basic plugins that are used to test and exercise the 148API during the `make check-tcg` target. 149 150- contrib/plugins/hotblocks.c 151 152The hotblocks plugin allows you to examine the where hot paths of 153execution are in your program. Once the program has finished you will 154get a sorted list of blocks reporting the starting PC, translation 155count, number of instructions and execution count. This will work best 156with linux-user execution as system emulation tends to generate 157re-translations as blocks from different programs get swapped in and 158out of system memory. 159 160If your program is single-threaded you can use the `inline` option for 161slightly faster (but not thread safe) counters. 162 163Example:: 164 165 ./aarch64-linux-user/qemu-aarch64 \ 166 -plugin contrib/plugins/libhotblocks.so -d plugin \ 167 ./tests/tcg/aarch64-linux-user/sha1 168 SHA1=15dd99a1991e0b3826fede3deffc1feba42278e6 169 collected 903 entries in the hash table 170 pc, tcount, icount, ecount 171 0x0000000041ed10, 1, 5, 66087 172 0x000000004002b0, 1, 4, 66087 173 ... 174 175- contrib/plugins/hotpages.c 176 177Similar to hotblocks but this time tracks memory accesses:: 178 179 ./aarch64-linux-user/qemu-aarch64 \ 180 -plugin contrib/plugins/libhotpages.so -d plugin \ 181 ./tests/tcg/aarch64-linux-user/sha1 182 SHA1=15dd99a1991e0b3826fede3deffc1feba42278e6 183 Addr, RCPUs, Reads, WCPUs, Writes 184 0x000055007fe000, 0x0001, 31747952, 0x0001, 8835161 185 0x000055007ff000, 0x0001, 29001054, 0x0001, 8780625 186 0x00005500800000, 0x0001, 687465, 0x0001, 335857 187 0x0000000048b000, 0x0001, 130594, 0x0001, 355 188 0x0000000048a000, 0x0001, 1826, 0x0001, 11 189 190- contrib/plugins/howvec.c 191 192This is an instruction classifier so can be used to count different 193types of instructions. It has a number of options to refine which get 194counted. You can give an argument for a class of instructions to break 195it down fully, so for example to see all the system registers 196accesses:: 197 198 ./aarch64-softmmu/qemu-system-aarch64 $(QEMU_ARGS) \ 199 -append "root=/dev/sda2 systemd.unit=benchmark.service" \ 200 -smp 4 -plugin ./contrib/plugins/libhowvec.so,arg=sreg -d plugin 201 202which will lead to a sorted list after the class breakdown:: 203 204 Instruction Classes: 205 Class: UDEF not counted 206 Class: SVE (68 hits) 207 Class: PCrel addr (47789483 hits) 208 Class: Add/Sub (imm) (192817388 hits) 209 Class: Logical (imm) (93852565 hits) 210 Class: Move Wide (imm) (76398116 hits) 211 Class: Bitfield (44706084 hits) 212 Class: Extract (5499257 hits) 213 Class: Cond Branch (imm) (147202932 hits) 214 Class: Exception Gen (193581 hits) 215 Class: NOP not counted 216 Class: Hints (6652291 hits) 217 Class: Barriers (8001661 hits) 218 Class: PSTATE (1801695 hits) 219 Class: System Insn (6385349 hits) 220 Class: System Reg counted individually 221 Class: Branch (reg) (69497127 hits) 222 Class: Branch (imm) (84393665 hits) 223 Class: Cmp & Branch (110929659 hits) 224 Class: Tst & Branch (44681442 hits) 225 Class: AdvSimd ldstmult (736 hits) 226 Class: ldst excl (9098783 hits) 227 Class: Load Reg (lit) (87189424 hits) 228 Class: ldst noalloc pair (3264433 hits) 229 Class: ldst pair (412526434 hits) 230 Class: ldst reg (imm) (314734576 hits) 231 Class: Loads & Stores (2117774 hits) 232 Class: Data Proc Reg (223519077 hits) 233 Class: Scalar FP (31657954 hits) 234 Individual Instructions: 235 Instr: mrs x0, sp_el0 (2682661 hits) (op=0xd5384100/ System Reg) 236 Instr: mrs x1, tpidr_el2 (1789339 hits) (op=0xd53cd041/ System Reg) 237 Instr: mrs x2, tpidr_el2 (1513494 hits) (op=0xd53cd042/ System Reg) 238 Instr: mrs x0, tpidr_el2 (1490823 hits) (op=0xd53cd040/ System Reg) 239 Instr: mrs x1, sp_el0 (933793 hits) (op=0xd5384101/ System Reg) 240 Instr: mrs x2, sp_el0 (699516 hits) (op=0xd5384102/ System Reg) 241 Instr: mrs x4, tpidr_el2 (528437 hits) (op=0xd53cd044/ System Reg) 242 Instr: mrs x30, ttbr1_el1 (480776 hits) (op=0xd538203e/ System Reg) 243 Instr: msr ttbr1_el1, x30 (480713 hits) (op=0xd518203e/ System Reg) 244 Instr: msr vbar_el1, x30 (480671 hits) (op=0xd518c01e/ System Reg) 245 ... 246 247To find the argument shorthand for the class you need to examine the 248source code of the plugin at the moment, specifically the `*opt` 249argument in the InsnClassExecCount tables. 250 251- contrib/plugins/lockstep.c 252 253This is a debugging tool for developers who want to find out when and 254where execution diverges after a subtle change to TCG code generation. 255It is not an exact science and results are likely to be mixed once 256asynchronous events are introduced. While the use of -icount can 257introduce determinism to the execution flow it doesn't always follow 258the translation sequence will be exactly the same. Typically this is 259caused by a timer firing to service the GUI causing a block to end 260early. However in some cases it has proved to be useful in pointing 261people at roughly where execution diverges. The only argument you need 262for the plugin is a path for the socket the two instances will 263communicate over:: 264 265 266 ./sparc-softmmu/qemu-system-sparc -monitor none -parallel none \ 267 -net none -M SS-20 -m 256 -kernel day11/zImage.elf \ 268 -plugin ./contrib/plugins/liblockstep.so,arg=lockstep-sparc.sock \ 269 -d plugin,nochain 270 271which will eventually report:: 272 273 qemu-system-sparc: warning: nic lance.0 has no peer 274 @ 0x000000ffd06678 vs 0x000000ffd001e0 (2/1 since last) 275 @ 0x000000ffd07d9c vs 0x000000ffd06678 (3/1 since last) 276 Δ insn_count @ 0x000000ffd07d9c (809900609) vs 0x000000ffd06678 (809900612) 277 previously @ 0x000000ffd06678/10 (809900609 insns) 278 previously @ 0x000000ffd001e0/4 (809900599 insns) 279 previously @ 0x000000ffd080ac/2 (809900595 insns) 280 previously @ 0x000000ffd08098/5 (809900593 insns) 281 previously @ 0x000000ffd080c0/1 (809900588 insns) 282 283- contrib/plugins/hwprofile 284 285The hwprofile tool can only be used with system emulation and allows 286the user to see what hardware is accessed how often. It has a number of options: 287 288 * arg=read or arg=write 289 290 By default the plugin tracks both reads and writes. You can use one 291 of these options to limit the tracking to just one class of accesses. 292 293 * arg=source 294 295 Will include a detailed break down of what the guest PC that made the 296 access was. Not compatible with arg=pattern. Example output:: 297 298 cirrus-low-memory @ 0xfffffd00000a0000 299 pc:fffffc0000005cdc, 1, 256 300 pc:fffffc0000005ce8, 1, 256 301 pc:fffffc0000005cec, 1, 256 302 303 * arg=pattern 304 305 Instead break down the accesses based on the offset into the HW 306 region. This can be useful for seeing the most used registers of a 307 device. Example output:: 308 309 pci0-conf @ 0xfffffd01fe000000 310 off:00000004, 1, 1 311 off:00000010, 1, 3 312 off:00000014, 1, 3 313 off:00000018, 1, 2 314 off:0000001c, 1, 2 315 off:00000020, 1, 2 316 ... 317