ae68fdaba9
10 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Emilio G. Cota
|
0ac20318ce |
tcg: remove tb_lock
Use mmap_lock in user-mode to protect TCG state and the page descriptors.
In !user-mode, each vCPU has its own TCG state, so no locks needed.
Per-page locks are used to protect the page descriptors.
Per-TB locks are used in both modes to protect TB jumps.
Some notes:
- tb_lock is removed from notdirty_mem_write by passing a
locked page_collection to tb_invalidate_phys_page_fast.
- tcg_tb_lookup/remove/insert/etc have their own internal lock(s),
so there is no need to further serialize access to them.
- do_tb_flush is run in a safe async context, meaning no other
vCPU threads are running. Therefore acquiring mmap_lock there
is just to please tools such as thread sanitizer.
- Not visible in the diff, but tb_invalidate_phys_page already
has an assert_memory_lock.
- cpu_io_recompile is !user-only, so no mmap_lock there.
- Added mmap_unlock()'s before all siglongjmp's that could
be called in user-mode while mmap_lock is held.
+ Added an assert for !have_mmap_lock() after returning from
the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic.
Performance numbers before/after:
Host: AMD Opteron(tm) Processor 6376
ubuntu 17.04 ppc64 bootup+shutdown time
700 +-+--+----+------+------------+-----------+------------*--+-+
| + + + + + *B |
| before ***B*** ** * |
|tb lock removal ###D### *** |
600 +-+ *** +-+
| ** # |
| *B* #D |
| *** * ## |
500 +-+ *** ### +-+
| * *** ### |
| *B* # ## |
| ** * #D# |
400 +-+ ** ## +-+
| ** ### |
| ** ## |
| ** # ## |
300 +-+ * B* #D# +-+
| B *** ### |
| * ** #### |
| * *** ### |
200 +-+ B *B #D# +-+
| #B* * ## # |
| #* ## |
| + D##D# + + + + |
100 +-+--+----+------+------------+-----------+------------+--+-+
1 8 16 Guest CPUs 48 64
png: https://imgur.com/HwmBHXe
debian jessie aarch64 bootup+shutdown time
90 +-+--+-----+-----+------------+------------+------------+--+-+
| + + + + + + |
| before ***B*** B |
80 +tb lock removal ###D### **D +-+
| **### |
| **## |
70 +-+ ** # +-+
| ** ## |
| ** # |
60 +-+ *B ## +-+
| ** ## |
| *** #D |
50 +-+ *** ## +-+
| * ** ### |
| **B* ### |
40 +-+ **** # ## +-+
| **** #D# |
| ***B** ### |
30 +-+ B***B** #### +-+
| B * * # ### |
| B ###D# |
20 +-+ D ##D## +-+
| D# |
| + + + + + + |
10 +-+--+-----+-----+------------+------------+------------+--+-+
1 8 16 Guest CPUs 48 64
png: https://imgur.com/iGpGFtv
The gains are high for 4-8 CPUs. Beyond that point, however, unrelated
lock contention significantly hurts scalability.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
||
|
Emilio G. Cota
|
128ed2278c |
tcg: move tb_ctx.tb_phys_invalidate_count to tcg_ctx
Thereby making it per-TCGContext. Once we remove tb_lock, this will avoid an atomic increment every time a TB is invalidated. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> |
||
|
Emilio G. Cota
|
be2cdc5e35 |
tcg: track TBs with per-region BST's
This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> |
||
|
Emilio G. Cota
|
44ded3d048 |
tcg: take tb_ctx out of TCGContext
Groundwork for supporting multiple TCG contexts. Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> |
||
|
Emilio G. Cota
|
2ac01d6daf |
translate-all: use a binary search tree to track TBs in TBContext
This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> |
||
|
Emilio G. Cota
|
6e3b2bfd6a |
tcg: allocate TB structs before the corresponding translated code
Allocating an arbitrarily-sized array of tbs results in either (a) a lot of memory wasted or (b) unnecessary flushes of the code cache when we run out of TB structs in the array. An obvious solution would be to just malloc a TB struct when needed, and keep the TB array as an array of pointers (recall that tb_find_pc() needs the TB array to run in O(log n)). Perhaps a better solution, which is implemented in this patch, is to allocate TB's right before the translated code they describe. This results in some memory waste due to padding to have code and TBs in separate cache lines--for instance, I measured 4.7% of padding in the used portion of code_gen_buffer when booting aarch64 Linux on a host with 64-byte cache lines. However, it can allow for optimizations in some host architectures, since TCG backends could safely assume that the TB and the corresponding translated code are very close to each other in memory. See this message by rth for a detailed explanation: https://lists.gnu.org/archive/html/qemu-devel/2017-03/msg05172.html Subject: Re: GSoC 2017 Proposal: TCG performance enhancements Message-ID: <1e67644b-4b30-887e-d329-1848e94c9484@twiddle.net> Suggested-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Pranith Kumar <bobby.prani@gmail.com> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1496790745-314-3-git-send-email-cota@braap.org> [rth: Simplify the arithmetic in tcg_tb_alloc] Signed-off-by: Richard Henderson <rth@twiddle.net> |
||
|
Sergey Fedorov
|
3359baad36 |
tcg: Make tb_flush() thread safe
Use async_safe_run_on_cpu() to make tb_flush() thread safe. This is possible now that code generation does not happen in the middle of execution. It can happen that multiple threads schedule a safe work to flush the translation buffer. To keep statistics and debugging output sane, always check if the translation buffer has already been flushed. Signed-off-by: Sergey Fedorov <serge.fdrv@gmail.com> Signed-off-by: Sergey Fedorov <sergey.fedorov@linaro.org> [AJB: minor re-base fixes] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Message-Id: <1470158864-17651-13-git-send-email-alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
||
|
Markus Armbruster
|
2a6a4076e1 |
Clean up ill-advised or unusual header guards
Cleaned up with scripts/clean-header-guards.pl. Signed-off-by: Markus Armbruster <armbru@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> |
||
|
Emilio G. Cota
|
909eaac9bb |
tb hash: track translated blocks with qht
Having a fixed-size hash table for keeping track of all translation blocks
is suboptimal: some workloads are just too big or too small to get maximum
performance from the hash table. The MRU promotion policy helps improve
performance when the hash table is a little undersized, but it cannot
make up for severely undersized hash tables.
Furthermore, frequent MRU promotions result in writes that are a scalability
bottleneck. For scalability, lookups should only perform reads, not writes.
This is not a big deal for now, but it will become one once MTTCG matures.
The appended fixes these issues by using qht as the implementation of
the TB hash table. This solution is superior to other alternatives considered,
namely:
- master: implementation in QEMU before this patchset
- xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU.
- xxhash-rcu: fixed buckets + xxhash + RCU list + MRU.
MRU is implemented here by adding an intermediate struct
that contains the u32 hash and a pointer to the TB; this
allows us, on an MRU promotion, to copy said struct (that is not
at the head), and put this new copy at the head. After a grace
period, the original non-head struct can be eliminated, and
after another grace period, freed.
- qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize +
no MRU for lookups; MRU for inserts.
The appended solution is the following:
- qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize +
no MRU for lookups; MRU for inserts.
The plots below compare the considered solutions. The Y axis shows the
boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis
sweeps the number of buckets (or initial number of buckets for qht-autoresize).
The plots in PNG format (and with errorbars) can be seen here:
http://imgur.com/a/Awgnq
Each test runs 5 times, and the entire QEMU process is pinned to a
single core for repeatability of results.
Host: Intel Xeon E5-2690
28 ++------------+-------------+-------------+-------------+------------++
A***** + + + master **A*** +
27 ++ * xxhash ##B###++
| A******A****** xxhash-rcu $$C$$$ |
26 C$$ A******A****** qht-fixed-nomru*%%D%%%++
D%%$$ A******A******A*qht-dyn-mru A*E****A
25 ++ %%$$ qht-dyn-nomru &&F&&&++
B#####% |
24 ++ #C$$$$$ ++
| B### $ |
| ## C$$$$$$ |
23 ++ # C$$$$$$ ++
| B###### C$$$$$$ %%%D
22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C
| D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E
21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B
+ E@@@ F&&& + E@ + F&&& + +
20 ++------------+-------------+-------------+-------------+------------++
14 16 18 20 22 24
log2 number of buckets
Host: Intel i7-4790K
14.5 ++------------+------------+-------------+------------+------------++
A** + + + master **A*** +
14 ++ ** xxhash ##B###++
13.5 ++ ** xxhash-rcu $$C$$$++
| qht-fixed-nomru %%D%%% |
13 ++ A****** qht-dyn-mru @@E@@@++
| A*****A******A****** qht-dyn-nomru &&F&&& |
12.5 C$$ A******A******A*****A****** ***A
12 ++ $$ A*** ++
D%%% $$ |
11.5 ++ %% ++
B### %C$$$$$$ |
11 ++ ## D%%%%% C$$$$$ ++
| # % C$$$$$$ |
10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C
10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B
+ F&& D%%%%%%B######B######B#####B###@@@D%%% +
9.5 ++------------+------------+-------------+------------+------------++
14 16 18 20 22 24
log2 number of buckets
Note that the original point before this patch series is X=15 for "master";
the little sensitivity to the increased number of buckets is due to the
poor hashing function in master.
xxhash-rcu has significant overhead due to the constant churn of allocating
and deallocating intermediate structs for implementing MRU. An alternative
would be do consider failed lookups as "maybe not there", and then
acquire the external lock (tb_lock in this case) to really confirm that
there was indeed a failed lookup. This, however, would not be enough
to implement dynamic resizing--this is more complex: see
"Resizable, Scalable, Concurrent Hash Tables via Relativistic
Programming" by Triplett, McKenney and Walpole. This solution was
discarded due to the very coarse RCU read critical sections that we have
in MTTCG; resizing requires waiting for readers after every pointer update,
and resizes require many pointer updates, so this would quickly become
prohibitive.
qht-fixed-nomru shows that MRU promotion is advisable for undersized
hash tables.
However, qht-dyn-mru shows that MRU promotion is not important if the
hash table is properly sized: there is virtually no difference in
performance between qht-dyn-nomru and qht-dyn-mru.
Before this patch, we're at X=15 on "xxhash"; after this patch, we're at
X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we
can achieve with optimum sizing of the hash table, while keeping the hash
table scalable for readers.
The improvement we get before and after this patch for booting debian jessie
with arm-softmmu is:
- Intel Xeon E5-2690: 10.5% less time
- Intel i7-4790K: 5.2% less time
We could get this same improvement _for this particular workload_ by
statically increasing the size of the hash table. But this would hurt
workloads that do not need a large hash table. The dynamic (upward)
resizing allows us to start small and enlarge the hash table as needed.
A quick note on downsizing: the table is resized back to 2**15 buckets
on every tb_flush; this makes sense because it is not guaranteed that the
table will reach the same number of TBs later on (e.g. most bootup code is
thrown away after boot); it makes sense to grow the hash table as
more code blocks are translated. This also avoids the complication of
having to build downsizing hysteresis logic into qht.
Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org>
Signed-off-by: Richard Henderson <rth@twiddle.net>
|
||
|
Paolo Bonzini
|
00f6da6a1a |
exec: extract exec/tb-context.h
TCG backends do not need most of exec-all.h; extract what they actually need to a separate file or move it directly to tcg.h. The next patch will stop including exec-all.h from everywhere. Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |