xref: /qemu/docs/devel/multi-thread-tcg.rst (revision 2abf0da2)
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
113This is done with the async_safe_run_on_cpu() mechanism to ensure all
114vCPUs are quiescent when changes are being made to shared global
115structures.
116
117More granular translation invalidation events are typically due
118to a change of the state of a physical page:
119
120  - code modification (self modify code, patching code)
121  - page changes (new page mapping in linux-user mode)
122
123While setting the invalid flag in a TranslationBlock will stop it
124being used when looked up in the hot-path there are a number of other
125book-keeping structures that need to be safely cleared.
126
127Any TranslationBlocks which have been patched to jump directly to the
128now invalid blocks need the jump patches reversing so they will return
129to the C code.
130
131There are a number of look-up caches that need to be properly updated
132including the:
133
134  - jump lookup cache
135  - the physical-to-tb lookup hash table
136  - the global page table
137
138The global page table (l1_map) which provides a multi-level look-up
139for PageDesc structures which contain pointers to the start of a
140linked list of all Translation Blocks in that page (see page_next).
141
142Both the jump patching and the page cache involve linked lists that
143the invalidated TranslationBlock needs to be removed from.
144
145DESIGN REQUIREMENT: Safely handle invalidation of TBs
146                      - safely patch/revert direct jumps
147                      - remove central PageDesc lookup entries
148                      - ensure lookup caches/hashes are safely updated
149
150(Current solution)
151
152The direct jump themselves are updated atomically by the TCG
153tb_set_jmp_target() code. Modification to the linked lists that allow
154searching for linked pages are done under the protection of tb->jmp_lock,
155where tb is the destination block of a jump. Each origin block keeps a
156pointer to its destinations so that the appropriate lock can be acquired before
157iterating over a jump list.
158
159The global page table is a lockless radix tree; cmpxchg is used
160to atomically insert new elements.
161
162The lookup caches are updated atomically and the lookup hash uses QHT
163which is designed for concurrent safe lookup.
164
165Parallel code generation is supported. QHT is used at insertion time
166as the synchronization point across threads, thereby ensuring that we only
167keep track of a single TranslationBlock for each guest code block.
168
169Memory maps and TLBs
170--------------------
171
172The memory handling code is fairly critical to the speed of memory
173access in the emulated system. The SoftMMU code is designed so the
174hot-path can be handled entirely within translated code. This is
175handled with a per-vCPU TLB structure which once populated will allow
176a series of accesses to the page to occur without exiting the
177translated code. It is possible to set flags in the TLB address which
178will ensure the slow-path is taken for each access. This can be done
179to support:
180
181  - Memory regions (dividing up access to PIO, MMIO and RAM)
182  - Dirty page tracking (for code gen, SMC detection, migration and display)
183  - Virtual TLB (for translating guest address->real address)
184
185When the TLB tables are updated by a vCPU thread other than their own
186we need to ensure it is done in a safe way so no inconsistent state is
187seen by the vCPU thread.
188
189Some operations require updating a number of vCPUs TLBs at the same
190time in a synchronised manner.
191
192DESIGN REQUIREMENTS:
193
194  - TLB Flush All/Page
195    - can be across-vCPUs
196    - cross vCPU TLB flush may need other vCPU brought to halt
197    - change may need to be visible to the calling vCPU immediately
198  - TLB Flag Update
199    - usually cross-vCPU
200    - want change to be visible as soon as possible
201  - TLB Update (update a CPUTLBEntry, via tlb_set_page_with_attrs)
202    - This is a per-vCPU table - by definition can't race
203    - updated by its own thread when the slow-path is forced
204
205(Current solution)
206
207We have updated cputlb.c to defer operations when a cross-vCPU
208operation with async_run_on_cpu() which ensures each vCPU sees a
209coherent state when it next runs its work (in a few instructions
210time).
211
212A new set up operations (tlb_flush_*_all_cpus) take an additional flag
213which when set will force synchronisation by setting the source vCPUs
214work as "safe work" and exiting the cpu run loop. This ensure by the
215time execution restarts all flush operations have completed.
216
217TLB flag updates are all done atomically and are also protected by the
218corresponding page lock.
219
220(Known limitation)
221
222Not really a limitation but the wait mechanism is overly strict for
223some architectures which only need flushes completed by a barrier
224instruction. This could be a future optimisation.
225
226Emulated hardware state
227-----------------------
228
229Currently thanks to KVM work any access to IO memory is automatically protected
230by the BQL (Big QEMU Lock). Any IO region that doesn't use the BQL is expected
231to do its own locking.
232
233However IO memory isn't the only way emulated hardware state can be
234modified. Some architectures have model specific registers that
235trigger hardware emulation features. Generally any translation helper
236that needs to update more than a single vCPUs of state should take the
237BQL.
238
239As the BQL, or global iothread mutex is shared across the system we
240push the use of the lock as far down into the TCG code as possible to
241minimise contention.
242
243(Current solution)
244
245MMIO access automatically serialises hardware emulation by way of the
246BQL. Currently Arm targets serialise all ARM_CP_IO register accesses
247and also defer the reset/startup of vCPUs to the vCPU context by way
248of async_run_on_cpu().
249
250Updates to interrupt state are also protected by the BQL as they can
251often be cross vCPU.
252
253Memory Consistency
254==================
255
256Between emulated guests and host systems there are a range of memory
257consistency models. Even emulating weakly ordered systems on strongly
258ordered hosts needs to ensure things like store-after-load re-ordering
259can be prevented when the guest wants to.
260
261Memory Barriers
262---------------
263
264Barriers (sometimes known as fences) provide a mechanism for software
265to enforce a particular ordering of memory operations from the point
266of view of external observers (e.g. another processor core). They can
267apply to any memory operations as well as just loads or stores.
268
269The Linux kernel has an excellent `write-up
270<https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/plain/Documentation/memory-barriers.txt>`_
271on the various forms of memory barrier and the guarantees they can
272provide.
273
274Barriers are often wrapped around synchronisation primitives to
275provide explicit memory ordering semantics. However they can be used
276by themselves to provide safe lockless access by ensuring for example
277a change to a signal flag will only be visible once the changes to
278payload are.
279
280DESIGN REQUIREMENT: Add a new tcg_memory_barrier op
281
282This would enforce a strong load/store ordering so all loads/stores
283complete at the memory barrier. On single-core non-SMP strongly
284ordered backends this could become a NOP.
285
286Aside from explicit standalone memory barrier instructions there are
287also implicit memory ordering semantics which comes with each guest
288memory access instruction. For example all x86 load/stores come with
289fairly strong guarantees of sequential consistency whereas Arm has
290special variants of load/store instructions that imply acquire/release
291semantics.
292
293In the case of a strongly ordered guest architecture being emulated on
294a weakly ordered host the scope for a heavy performance impact is
295quite high.
296
297DESIGN REQUIREMENTS: Be efficient with use of memory barriers
298       - host systems with stronger implied guarantees can skip some barriers
299       - merge consecutive barriers to the strongest one
300
301(Current solution)
302
303The system currently has a tcg_gen_mb() which will add memory barrier
304operations if code generation is being done in a parallel context. The
305tcg_optimize() function attempts to merge barriers up to their
306strongest form before any load/store operations. The solution was
307originally developed and tested for linux-user based systems. All
308backends have been converted to emit fences when required. So far the
309following front-ends have been updated to emit fences when required:
310
311    - target-i386
312    - target-arm
313    - target-aarch64
314    - target-alpha
315    - target-mips
316
317Memory Control and Maintenance
318------------------------------
319
320This includes a class of instructions for controlling system cache
321behaviour. While QEMU doesn't model cache behaviour these instructions
322are often seen when code modification has taken place to ensure the
323changes take effect.
324
325Synchronisation Primitives
326--------------------------
327
328There are two broad types of synchronisation primitives found in
329modern ISAs: atomic instructions and exclusive regions.
330
331The first type offer a simple atomic instruction which will guarantee
332some sort of test and conditional store will be truly atomic w.r.t.
333other cores sharing access to the memory. The classic example is the
334x86 cmpxchg instruction.
335
336The second type offer a pair of load/store instructions which offer a
337guarantee that a region of memory has not been touched between the
338load and store instructions. An example of this is Arm's ldrex/strex
339pair where the strex instruction will return a flag indicating a
340successful store only if no other CPU has accessed the memory region
341since the ldrex.
342
343Traditionally TCG has generated a series of operations that work
344because they are within the context of a single translation block so
345will have completed before another CPU is scheduled. However with
346the ability to have multiple threads running to emulate multiple CPUs
347we will need to explicitly expose these semantics.
348
349DESIGN REQUIREMENTS:
350  - Support classic atomic instructions
351  - Support load/store exclusive (or load link/store conditional) pairs
352  - Generic enough infrastructure to support all guest architectures
353CURRENT OPEN QUESTIONS:
354  - How problematic is the ABA problem in general?
355
356(Current solution)
357
358The TCG provides a number of atomic helpers (tcg_gen_atomic_*) which
359can be used directly or combined to emulate other instructions like
360Arm's ldrex/strex instructions. While they are susceptible to the ABA
361problem so far common guests have not implemented patterns where
362this may be a problem - typically presenting a locking ABI which
363assumes cmpxchg like semantics.
364
365The code also includes a fall-back for cases where multi-threaded TCG
366ops can't work (e.g. guest atomic width > host atomic width). In this
367case an EXCP_ATOMIC exit occurs and the instruction is emulated with
368an exclusive lock which ensures all emulation is serialised.
369
370While the atomic helpers look good enough for now there may be a need
371to look at solutions that can more closely model the guest
372architectures semantics.
373