xref: /qemu/docs/devel/multi-thread-tcg.rst (revision d884e272)
1..
2  Copyright (c) 2015-2020 Linaro Ltd.
3
4  This work is licensed under the terms of the GNU GPL, version 2 or
5  later. See the COPYING file in the top-level directory.
6
7==================
8Multi-threaded TCG
9==================
10
11This document outlines the design for multi-threaded TCG (a.k.a MTTCG)
12system-mode emulation. user-mode emulation has always mirrored the
13thread structure of the translated executable although some of the
14changes done for MTTCG system emulation have improved the stability of
15linux-user emulation.
16
17The original system-mode TCG implementation was single threaded and
18dealt with multiple CPUs with simple round-robin scheduling. This
19simplified a lot of things but became increasingly limited as systems
20being emulated gained additional cores and per-core performance gains
21for host systems started to level off.
22
23vCPU Scheduling
24===============
25
26We introduce a new running mode where each vCPU will run on its own
27user-space thread. This is enabled by default for all FE/BE
28combinations where the host memory model is able to accommodate the
29guest (TCG_GUEST_DEFAULT_MO & ~TCG_TARGET_DEFAULT_MO is zero) and the
30guest has had the required work done to support this safely
31(TARGET_SUPPORTS_MTTCG).
32
33System emulation will fall back to the original round robin approach
34if:
35
36* forced by --accel tcg,thread=single
37* enabling --icount mode
38* 64 bit guests on 32 bit hosts (TCG_OVERSIZED_GUEST)
39
40In the general case of running translated code there should be no
41inter-vCPU dependencies and all vCPUs should be able to run at full
42speed. Synchronisation will only be required while accessing internal
43shared data structures or when the emulated architecture requires a
44coherent representation of the emulated machine state.
45
46Shared Data Structures
47======================
48
49Main Run Loop
50-------------
51
52Even when there is no code being generated there are a number of
53structures associated with the hot-path through the main run-loop.
54These are associated with looking up the next translation block to
55execute. These include:
56
57    tb_jmp_cache (per-vCPU, cache of recent jumps)
58    tb_ctx.htable (global hash table, phys address->tb lookup)
59
60As TB linking only occurs when blocks are in the same page this code
61is critical to performance as looking up the next TB to execute is the
62most common reason to exit the generated code.
63
64DESIGN REQUIREMENT: Make access to lookup structures safe with
65multiple reader/writer threads. Minimise any lock contention to do it.
66
67The hot-path avoids using locks where possible. The tb_jmp_cache is
68updated with atomic accesses to ensure consistent results. The fall
69back QHT based hash table is also designed for lockless lookups. Locks
70are only taken when code generation is required or TranslationBlocks
71have their block-to-block jumps patched.
72
73Global TCG State
74----------------
75
76User-mode emulation
77~~~~~~~~~~~~~~~~~~~
78
79We need to protect the entire code generation cycle including any post
80generation patching of the translated code. This also implies a shared
81translation buffer which contains code running on all cores. Any
82execution path that comes to the main run loop will need to hold a
83mutex for code generation. This also includes times when we need flush
84code or entries from any shared lookups/caches. Structures held on a
85per-vCPU basis won't need locking unless other vCPUs will need to
86modify them.
87
88DESIGN REQUIREMENT: Add locking around all code generation and TB
89patching.
90
91(Current solution)
92
93Code generation is serialised with mmap_lock().
94
95!User-mode emulation
96~~~~~~~~~~~~~~~~~~~~
97
98Each vCPU has its own TCG context and associated TCG region, thereby
99requiring no locking during translation.
100
101Translation Blocks
102------------------
103
104Currently the whole system shares a single code generation buffer
105which when full will force a flush of all translations and start from
106scratch again. Some operations also force a full flush of translations
107including:
108
109  - debugging operations (breakpoint insertion/removal)
110  - some CPU helper functions
111  - linux-user spawning its first thread
112  - operations related to TCG Plugins
113
114This is done with the async_safe_run_on_cpu() mechanism to ensure all
115vCPUs are quiescent when changes are being made to shared global
116structures.
117
118More granular translation invalidation events are typically due
119to a change of the state of a physical page:
120
121  - code modification (self modify code, patching code)
122  - page changes (new page mapping in linux-user mode)
123
124While setting the invalid flag in a TranslationBlock will stop it
125being used when looked up in the hot-path there are a number of other
126book-keeping structures that need to be safely cleared.
127
128Any TranslationBlocks which have been patched to jump directly to the
129now invalid blocks need the jump patches reversing so they will return
130to the C code.
131
132There are a number of look-up caches that need to be properly updated
133including the:
134
135  - jump lookup cache
136  - the physical-to-tb lookup hash table
137  - the global page table
138
139The global page table (l1_map) which provides a multi-level look-up
140for PageDesc structures which contain pointers to the start of a
141linked list of all Translation Blocks in that page (see page_next).
142
143Both the jump patching and the page cache involve linked lists that
144the invalidated TranslationBlock needs to be removed from.
145
146DESIGN REQUIREMENT: Safely handle invalidation of TBs
147                      - safely patch/revert direct jumps
148                      - remove central PageDesc lookup entries
149                      - ensure lookup caches/hashes are safely updated
150
151(Current solution)
152
153The direct jump themselves are updated atomically by the TCG
154tb_set_jmp_target() code. Modification to the linked lists that allow
155searching for linked pages are done under the protection of tb->jmp_lock,
156where tb is the destination block of a jump. Each origin block keeps a
157pointer to its destinations so that the appropriate lock can be acquired before
158iterating over a jump list.
159
160The global page table is a lockless radix tree; cmpxchg is used
161to atomically insert new elements.
162
163The lookup caches are updated atomically and the lookup hash uses QHT
164which is designed for concurrent safe lookup.
165
166Parallel code generation is supported. QHT is used at insertion time
167as the synchronization point across threads, thereby ensuring that we only
168keep track of a single TranslationBlock for each guest code block.
169
170Memory maps and TLBs
171--------------------
172
173The memory handling code is fairly critical to the speed of memory
174access in the emulated system. The SoftMMU code is designed so the
175hot-path can be handled entirely within translated code. This is
176handled with a per-vCPU TLB structure which once populated will allow
177a series of accesses to the page to occur without exiting the
178translated code. It is possible to set flags in the TLB address which
179will ensure the slow-path is taken for each access. This can be done
180to support:
181
182  - Memory regions (dividing up access to PIO, MMIO and RAM)
183  - Dirty page tracking (for code gen, SMC detection, migration and display)
184  - Virtual TLB (for translating guest address->real address)
185
186When the TLB tables are updated by a vCPU thread other than their own
187we need to ensure it is done in a safe way so no inconsistent state is
188seen by the vCPU thread.
189
190Some operations require updating a number of vCPUs TLBs at the same
191time in a synchronised manner.
192
193DESIGN REQUIREMENTS:
194
195  - TLB Flush All/Page
196    - can be across-vCPUs
197    - cross vCPU TLB flush may need other vCPU brought to halt
198    - change may need to be visible to the calling vCPU immediately
199  - TLB Flag Update
200    - usually cross-vCPU
201    - want change to be visible as soon as possible
202  - TLB Update (update a CPUTLBEntry, via tlb_set_page_with_attrs)
203    - This is a per-vCPU table - by definition can't race
204    - updated by its own thread when the slow-path is forced
205
206(Current solution)
207
208We have updated cputlb.c to defer operations when a cross-vCPU
209operation with async_run_on_cpu() which ensures each vCPU sees a
210coherent state when it next runs its work (in a few instructions
211time).
212
213A new set up operations (tlb_flush_*_all_cpus) take an additional flag
214which when set will force synchronisation by setting the source vCPUs
215work as "safe work" and exiting the cpu run loop. This ensure by the
216time execution restarts all flush operations have completed.
217
218TLB flag updates are all done atomically and are also protected by the
219corresponding page lock.
220
221(Known limitation)
222
223Not really a limitation but the wait mechanism is overly strict for
224some architectures which only need flushes completed by a barrier
225instruction. This could be a future optimisation.
226
227Emulated hardware state
228-----------------------
229
230Currently thanks to KVM work any access to IO memory is automatically protected
231by the BQL (Big QEMU Lock). Any IO region that doesn't use the BQL is expected
232to do its own locking.
233
234However IO memory isn't the only way emulated hardware state can be
235modified. Some architectures have model specific registers that
236trigger hardware emulation features. Generally any translation helper
237that needs to update more than a single vCPUs of state should take the
238BQL.
239
240As the BQL, or global iothread mutex is shared across the system we
241push the use of the lock as far down into the TCG code as possible to
242minimise contention.
243
244(Current solution)
245
246MMIO access automatically serialises hardware emulation by way of the
247BQL. Currently Arm targets serialise all ARM_CP_IO register accesses
248and also defer the reset/startup of vCPUs to the vCPU context by way
249of async_run_on_cpu().
250
251Updates to interrupt state are also protected by the BQL as they can
252often be cross vCPU.
253
254Memory Consistency
255==================
256
257Between emulated guests and host systems there are a range of memory
258consistency models. Even emulating weakly ordered systems on strongly
259ordered hosts needs to ensure things like store-after-load re-ordering
260can be prevented when the guest wants to.
261
262Memory Barriers
263---------------
264
265Barriers (sometimes known as fences) provide a mechanism for software
266to enforce a particular ordering of memory operations from the point
267of view of external observers (e.g. another processor core). They can
268apply to any memory operations as well as just loads or stores.
269
270The Linux kernel has an excellent `write-up
271<https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/plain/Documentation/memory-barriers.txt>`_
272on the various forms of memory barrier and the guarantees they can
273provide.
274
275Barriers are often wrapped around synchronisation primitives to
276provide explicit memory ordering semantics. However they can be used
277by themselves to provide safe lockless access by ensuring for example
278a change to a signal flag will only be visible once the changes to
279payload are.
280
281DESIGN REQUIREMENT: Add a new tcg_memory_barrier op
282
283This would enforce a strong load/store ordering so all loads/stores
284complete at the memory barrier. On single-core non-SMP strongly
285ordered backends this could become a NOP.
286
287Aside from explicit standalone memory barrier instructions there are
288also implicit memory ordering semantics which comes with each guest
289memory access instruction. For example all x86 load/stores come with
290fairly strong guarantees of sequential consistency whereas Arm has
291special variants of load/store instructions that imply acquire/release
292semantics.
293
294In the case of a strongly ordered guest architecture being emulated on
295a weakly ordered host the scope for a heavy performance impact is
296quite high.
297
298DESIGN REQUIREMENTS: Be efficient with use of memory barriers
299       - host systems with stronger implied guarantees can skip some barriers
300       - merge consecutive barriers to the strongest one
301
302(Current solution)
303
304The system currently has a tcg_gen_mb() which will add memory barrier
305operations if code generation is being done in a parallel context. The
306tcg_optimize() function attempts to merge barriers up to their
307strongest form before any load/store operations. The solution was
308originally developed and tested for linux-user based systems. All
309backends have been converted to emit fences when required. So far the
310following front-ends have been updated to emit fences when required:
311
312    - target-i386
313    - target-arm
314    - target-aarch64
315    - target-alpha
316    - target-mips
317
318Memory Control and Maintenance
319------------------------------
320
321This includes a class of instructions for controlling system cache
322behaviour. While QEMU doesn't model cache behaviour these instructions
323are often seen when code modification has taken place to ensure the
324changes take effect.
325
326Synchronisation Primitives
327--------------------------
328
329There are two broad types of synchronisation primitives found in
330modern ISAs: atomic instructions and exclusive regions.
331
332The first type offer a simple atomic instruction which will guarantee
333some sort of test and conditional store will be truly atomic w.r.t.
334other cores sharing access to the memory. The classic example is the
335x86 cmpxchg instruction.
336
337The second type offer a pair of load/store instructions which offer a
338guarantee that a region of memory has not been touched between the
339load and store instructions. An example of this is Arm's ldrex/strex
340pair where the strex instruction will return a flag indicating a
341successful store only if no other CPU has accessed the memory region
342since the ldrex.
343
344Traditionally TCG has generated a series of operations that work
345because they are within the context of a single translation block so
346will have completed before another CPU is scheduled. However with
347the ability to have multiple threads running to emulate multiple CPUs
348we will need to explicitly expose these semantics.
349
350DESIGN REQUIREMENTS:
351  - Support classic atomic instructions
352  - Support load/store exclusive (or load link/store conditional) pairs
353  - Generic enough infrastructure to support all guest architectures
354CURRENT OPEN QUESTIONS:
355  - How problematic is the ABA problem in general?
356
357(Current solution)
358
359The TCG provides a number of atomic helpers (tcg_gen_atomic_*) which
360can be used directly or combined to emulate other instructions like
361Arm's ldrex/strex instructions. While they are susceptible to the ABA
362problem so far common guests have not implemented patterns where
363this may be a problem - typically presenting a locking ABI which
364assumes cmpxchg like semantics.
365
366The code also includes a fall-back for cases where multi-threaded TCG
367ops can't work (e.g. guest atomic width > host atomic width). In this
368case an EXCP_ATOMIC exit occurs and the instruction is emulated with
369an exclusive lock which ensures all emulation is serialised.
370
371While the atomic helpers look good enough for now there may be a need
372to look at solutions that can more closely model the guest
373architectures semantics.
374