qemu/cpu-exec.c

662 lines
21 KiB
C
Raw Normal View History

/*
* emulator main execution loop
*
* Copyright (c) 2003-2005 Fabrice Bellard
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2 of the License, or (at your option) any later version.
*
* This library is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with this library; if not, see <http://www.gnu.org/licenses/>.
*/
#include "qemu/osdep.h"
#include "cpu.h"
#include "trace.h"
#include "disas/disas.h"
#include "exec/exec-all.h"
#include "tcg.h"
#include "qemu/atomic.h"
#include "sysemu/qtest.h"
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
#include "qemu/timer.h"
#include "exec/address-spaces.h"
#include "qemu/rcu.h"
#include "exec/tb-hash.h"
#include "exec/log.h"
#if defined(TARGET_I386) && !defined(CONFIG_USER_ONLY)
#include "hw/i386/apic.h"
#endif
#include "sysemu/replay.h"
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
/* -icount align implementation. */
typedef struct SyncClocks {
int64_t diff_clk;
int64_t last_cpu_icount;
int64_t realtime_clock;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
} SyncClocks;
#if !defined(CONFIG_USER_ONLY)
/* Allow the guest to have a max 3ms advance.
* The difference between the 2 clocks could therefore
* oscillate around 0.
*/
#define VM_CLOCK_ADVANCE 3000000
#define THRESHOLD_REDUCE 1.5
#define MAX_DELAY_PRINT_RATE 2000000000LL
#define MAX_NB_PRINTS 100
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
static void align_clocks(SyncClocks *sc, const CPUState *cpu)
{
int64_t cpu_icount;
if (!icount_align_option) {
return;
}
cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
sc->diff_clk += cpu_icount_to_ns(sc->last_cpu_icount - cpu_icount);
sc->last_cpu_icount = cpu_icount;
if (sc->diff_clk > VM_CLOCK_ADVANCE) {
#ifndef _WIN32
struct timespec sleep_delay, rem_delay;
sleep_delay.tv_sec = sc->diff_clk / 1000000000LL;
sleep_delay.tv_nsec = sc->diff_clk % 1000000000LL;
if (nanosleep(&sleep_delay, &rem_delay) < 0) {
sc->diff_clk = rem_delay.tv_sec * 1000000000LL + rem_delay.tv_nsec;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
} else {
sc->diff_clk = 0;
}
#else
Sleep(sc->diff_clk / SCALE_MS);
sc->diff_clk = 0;
#endif
}
}
static void print_delay(const SyncClocks *sc)
{
static float threshold_delay;
static int64_t last_realtime_clock;
static int nb_prints;
if (icount_align_option &&
sc->realtime_clock - last_realtime_clock >= MAX_DELAY_PRINT_RATE &&
nb_prints < MAX_NB_PRINTS) {
if ((-sc->diff_clk / (float)1000000000LL > threshold_delay) ||
(-sc->diff_clk / (float)1000000000LL <
(threshold_delay - THRESHOLD_REDUCE))) {
threshold_delay = (-sc->diff_clk / 1000000000LL) + 1;
printf("Warning: The guest is now late by %.1f to %.1f seconds\n",
threshold_delay - 1,
threshold_delay);
nb_prints++;
last_realtime_clock = sc->realtime_clock;
}
}
}
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
static void init_delay_params(SyncClocks *sc,
const CPUState *cpu)
{
if (!icount_align_option) {
return;
}
sc->realtime_clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
sc->diff_clk = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) - sc->realtime_clock;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
sc->last_cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
if (sc->diff_clk < max_delay) {
max_delay = sc->diff_clk;
}
if (sc->diff_clk > max_advance) {
max_advance = sc->diff_clk;
}
/* Print every 2s max if the guest is late. We limit the number
of printed messages to NB_PRINT_MAX(currently 100) */
print_delay(sc);
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
}
#else
static void align_clocks(SyncClocks *sc, const CPUState *cpu)
{
}
static void init_delay_params(SyncClocks *sc, const CPUState *cpu)
{
}
#endif /* CONFIG USER ONLY */
/* Execute a TB, and fix up the CPU state afterwards if necessary */
static inline tcg_target_ulong cpu_tb_exec(CPUState *cpu, TranslationBlock *itb)
{
CPUArchState *env = cpu->env_ptr;
uintptr_t ret;
TranslationBlock *last_tb;
int tb_exit;
uint8_t *tb_ptr = itb->tc_ptr;
qemu_log_mask_and_addr(CPU_LOG_EXEC, itb->pc,
"Trace %p [" TARGET_FMT_lx "] %s\n",
itb->tc_ptr, itb->pc, lookup_symbol(itb->pc));
#if defined(DEBUG_DISAS)
if (qemu_loglevel_mask(CPU_LOG_TB_CPU)
&& qemu_log_in_addr_range(itb->pc)) {
#if defined(TARGET_I386)
log_cpu_state(cpu, CPU_DUMP_CCOP);
#else
log_cpu_state(cpu, 0);
#endif
}
#endif /* DEBUG_DISAS */
cpu->can_do_io = !use_icount;
ret = tcg_qemu_tb_exec(env, tb_ptr);
cpu->can_do_io = 1;
last_tb = (TranslationBlock *)(ret & ~TB_EXIT_MASK);
tb_exit = ret & TB_EXIT_MASK;
trace_exec_tb_exit(last_tb, tb_exit);
if (tb_exit > TB_EXIT_IDX1) {
/* We didn't start executing this TB (eg because the instruction
* counter hit zero); we must restore the guest PC to the address
* of the start of the TB.
*/
CPUClass *cc = CPU_GET_CLASS(cpu);
qemu_log_mask_and_addr(CPU_LOG_EXEC, last_tb->pc,
"Stopped execution of TB chain before %p ["
TARGET_FMT_lx "] %s\n",
last_tb->tc_ptr, last_tb->pc,
lookup_symbol(last_tb->pc));
if (cc->synchronize_from_tb) {
cc->synchronize_from_tb(cpu, last_tb);
} else {
assert(cc->set_pc);
cc->set_pc(cpu, last_tb->pc);
}
}
if (tb_exit == TB_EXIT_REQUESTED) {
/* We were asked to stop executing TBs (probably a pending
* interrupt. We've now stopped, so clear the flag.
*/
atomic_set(&cpu->tcg_exit_req, 0);
}
return ret;
}
#ifndef CONFIG_USER_ONLY
/* Execute the code without caching the generated code. An interpreter
could be used if available. */
static void cpu_exec_nocache(CPUState *cpu, int max_cycles,
TranslationBlock *orig_tb, bool ignore_icount)
{
TranslationBlock *tb;
/* Should never happen.
We only end up here when an existing TB is too long. */
if (max_cycles > CF_COUNT_MASK)
max_cycles = CF_COUNT_MASK;
tb = tb_gen_code(cpu, orig_tb->pc, orig_tb->cs_base, orig_tb->flags,
max_cycles | CF_NOCACHE
| (ignore_icount ? CF_IGNORE_ICOUNT : 0));
tb->orig_tb = orig_tb;
/* execute the generated code */
trace_exec_tb_nocache(tb, tb->pc);
cpu_tb_exec(cpu, tb);
tb_phys_invalidate(tb, -1);
tb_free(tb);
}
#endif
static void cpu_exec_step(CPUState *cpu)
{
CPUArchState *env = (CPUArchState *)cpu->env_ptr;
TranslationBlock *tb;
target_ulong cs_base, pc;
uint32_t flags;
cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
tb = tb_gen_code(cpu, pc, cs_base, flags,
1 | CF_NOCACHE | CF_IGNORE_ICOUNT);
tb->orig_tb = NULL;
/* execute the generated code */
trace_exec_tb_nocache(tb, pc);
cpu_tb_exec(cpu, tb);
tb_phys_invalidate(tb, -1);
tb_free(tb);
}
void cpu_exec_step_atomic(CPUState *cpu)
{
start_exclusive();
/* Since we got here, we know that parallel_cpus must be true. */
parallel_cpus = false;
cpu_exec_step(cpu);
parallel_cpus = true;
end_exclusive();
}
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>
2016-06-08 21:55:32 +03:00
struct tb_desc {
target_ulong pc;
target_ulong cs_base;
CPUArchState *env;
tb_page_addr_t phys_page1;
uint32_t flags;
};
static bool tb_cmp(const void *p, const void *d)
{
const TranslationBlock *tb = p;
const struct tb_desc *desc = d;
if (tb->pc == desc->pc &&
tb->page_addr[0] == desc->phys_page1 &&
tb->cs_base == desc->cs_base &&
tb->flags == desc->flags &&
!atomic_read(&tb->invalid)) {
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>
2016-06-08 21:55:32 +03:00
/* check next page if needed */
if (tb->page_addr[1] == -1) {
return true;
} else {
tb_page_addr_t phys_page2;
target_ulong virt_page2;
virt_page2 = (desc->pc & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE;
phys_page2 = get_page_addr_code(desc->env, virt_page2);
if (tb->page_addr[1] == phys_page2) {
return true;
}
}
}
return false;
}
static TranslationBlock *tb_htable_lookup(CPUState *cpu,
target_ulong pc,
target_ulong cs_base,
uint32_t flags)
{
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>
2016-06-08 21:55:32 +03:00
tb_page_addr_t phys_pc;
struct tb_desc desc;
tb hash: hash phys_pc, pc, and flags with xxhash For some workloads such as arm bootup, tb_phys_hash is performance-critical. The is due to the high frequency of accesses to the hash table, originated by (frequent) TLB flushes that wipe out the cpu-private tb_jmp_cache's. More info: https://lists.nongnu.org/archive/html/qemu-devel/2016-03/msg05098.html To dig further into this I modified an arm image booting debian jessie to immediately shut down after boot. Analysis revealed that quite a bit of time is unnecessarily spent in tb_phys_hash: the cause is poor hashing that results in very uneven loading of chains in the hash table's buckets; the longest observed chain had ~550 elements. The appended addresses this with two changes: 1) Use xxhash as the hash table's hash function. xxhash is a fast, high-quality hashing function. 2) Feed the hashing function with not just tb_phys, but also pc and flags. This improves performance over using just tb_phys for hashing, since that resulted in some hash buckets having many TB's, while others getting very few; with these changes, the longest observed chain on a single hash bucket is brought down from ~550 to ~40. Tests show that the other element checked for in tb_find_physical, cs_base, is always a match when tb_phys+pc+flags are a match, so hashing cs_base is wasteful. It could be that this is an ARM-only thing, though. UPDATE: On Tue, Apr 05, 2016 at 08:41:43 -0700, Richard Henderson wrote: > The cs_base field is only used by i386 (in 16-bit modes), and sparc (for a TB > consisting of only a delay slot). > It may well still turn out to be reasonable to ignore cs_base for hashing. BTW, after this change the hash table should not be called "tb_hash_phys" anymore; this is addressed later in this series. This change gives consistent bootup time improvements. I tested two host machines: - Intel Xeon E5-2690: 11.6% less time - Intel i7-4790K: 19.2% less time Increasing the number of hash buckets yields further improvements. However, using a larger, fixed number of buckets can degrade performance for other workloads that do not translate as many blocks (600K+ for debian-jessie arm bootup). This is dealt with later in this series. Reviewed-by: Sergey Fedorov <sergey.fedorov@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-8-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 21:55:25 +03:00
uint32_t h;
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>
2016-06-08 21:55:32 +03:00
desc.env = (CPUArchState *)cpu->env_ptr;
desc.cs_base = cs_base;
desc.flags = flags;
desc.pc = pc;
phys_pc = get_page_addr_code(desc.env, pc);
desc.phys_page1 = phys_pc & TARGET_PAGE_MASK;
tb hash: hash phys_pc, pc, and flags with xxhash For some workloads such as arm bootup, tb_phys_hash is performance-critical. The is due to the high frequency of accesses to the hash table, originated by (frequent) TLB flushes that wipe out the cpu-private tb_jmp_cache's. More info: https://lists.nongnu.org/archive/html/qemu-devel/2016-03/msg05098.html To dig further into this I modified an arm image booting debian jessie to immediately shut down after boot. Analysis revealed that quite a bit of time is unnecessarily spent in tb_phys_hash: the cause is poor hashing that results in very uneven loading of chains in the hash table's buckets; the longest observed chain had ~550 elements. The appended addresses this with two changes: 1) Use xxhash as the hash table's hash function. xxhash is a fast, high-quality hashing function. 2) Feed the hashing function with not just tb_phys, but also pc and flags. This improves performance over using just tb_phys for hashing, since that resulted in some hash buckets having many TB's, while others getting very few; with these changes, the longest observed chain on a single hash bucket is brought down from ~550 to ~40. Tests show that the other element checked for in tb_find_physical, cs_base, is always a match when tb_phys+pc+flags are a match, so hashing cs_base is wasteful. It could be that this is an ARM-only thing, though. UPDATE: On Tue, Apr 05, 2016 at 08:41:43 -0700, Richard Henderson wrote: > The cs_base field is only used by i386 (in 16-bit modes), and sparc (for a TB > consisting of only a delay slot). > It may well still turn out to be reasonable to ignore cs_base for hashing. BTW, after this change the hash table should not be called "tb_hash_phys" anymore; this is addressed later in this series. This change gives consistent bootup time improvements. I tested two host machines: - Intel Xeon E5-2690: 11.6% less time - Intel i7-4790K: 19.2% less time Increasing the number of hash buckets yields further improvements. However, using a larger, fixed number of buckets can degrade performance for other workloads that do not translate as many blocks (600K+ for debian-jessie arm bootup). This is dealt with later in this series. Reviewed-by: Sergey Fedorov <sergey.fedorov@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-8-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 21:55:25 +03:00
h = tb_hash_func(phys_pc, pc, flags);
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>
2016-06-08 21:55:32 +03:00
return qht_lookup(&tcg_ctx.tb_ctx.htable, tb_cmp, &desc, h);
}
static inline TranslationBlock *tb_find(CPUState *cpu,
TranslationBlock *last_tb,
int tb_exit)
{
CPUArchState *env = (CPUArchState *)cpu->env_ptr;
TranslationBlock *tb;
target_ulong cs_base, pc;
uint32_t flags;
bool have_tb_lock = false;
/* we record a subset of the CPU state. It will
always be the same before a given translated block
is executed. */
cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
tb = atomic_rcu_read(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)]);
if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base ||
tb->flags != flags)) {
tb = tb_htable_lookup(cpu, pc, cs_base, flags);
if (!tb) {
/* mmap_lock is needed by tb_gen_code, and mmap_lock must be
* taken outside tb_lock. As system emulation is currently
* single threaded the locks are NOPs.
*/
mmap_lock();
tb_lock();
have_tb_lock = true;
/* There's a chance that our desired tb has been translated while
* taking the locks so we check again inside the lock.
*/
tb = tb_htable_lookup(cpu, pc, cs_base, flags);
if (!tb) {
/* if no translated code available, then translate it now */
tb = tb_gen_code(cpu, pc, cs_base, flags, 0);
}
mmap_unlock();
}
/* We add the TB in the virtual pc hash table for the fast lookup */
atomic_set(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)], tb);
}
#ifndef CONFIG_USER_ONLY
/* We don't take care of direct jumps when address mapping changes in
* system emulation. So it's not safe to make a direct jump to a TB
* spanning two pages because the mapping for the second page can change.
*/
if (tb->page_addr[1] != -1) {
last_tb = NULL;
}
#endif
/* See if we can patch the calling TB. */
if (last_tb && !qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
if (!have_tb_lock) {
tb_lock();
have_tb_lock = true;
}
if (!tb->invalid) {
tb_add_jump(last_tb, tb_exit, tb);
}
}
if (have_tb_lock) {
tb_unlock();
}
return tb;
}
static inline bool cpu_handle_halt(CPUState *cpu)
{
if (cpu->halted) {
#if defined(TARGET_I386) && !defined(CONFIG_USER_ONLY)
if ((cpu->interrupt_request & CPU_INTERRUPT_POLL)
&& replay_interrupt()) {
X86CPU *x86_cpu = X86_CPU(cpu);
apic_poll_irq(x86_cpu->apic_state);
cpu_reset_interrupt(cpu, CPU_INTERRUPT_POLL);
}
#endif
if (!cpu_has_work(cpu)) {
current_cpu = NULL;
return true;
}
cpu->halted = 0;
}
return false;
}
static inline void cpu_handle_debug_exception(CPUState *cpu)
{
CPUClass *cc = CPU_GET_CLASS(cpu);
CPUWatchpoint *wp;
if (!cpu->watchpoint_hit) {
QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
wp->flags &= ~BP_WATCHPOINT_HIT;
}
}
cc->debug_excp_handler(cpu);
}
static inline bool cpu_handle_exception(CPUState *cpu, int *ret)
{
if (cpu->exception_index >= 0) {
if (cpu->exception_index >= EXCP_INTERRUPT) {
/* exit request from the cpu execution loop */
*ret = cpu->exception_index;
if (*ret == EXCP_DEBUG) {
cpu_handle_debug_exception(cpu);
}
cpu->exception_index = -1;
return true;
} else {
#if defined(CONFIG_USER_ONLY)
/* if user mode only, we simulate a fake exception
which will be handled outside the cpu execution
loop */
#if defined(TARGET_I386)
CPUClass *cc = CPU_GET_CLASS(cpu);
cc->do_interrupt(cpu);
#endif
*ret = cpu->exception_index;
cpu->exception_index = -1;
return true;
#else
if (replay_exception()) {
CPUClass *cc = CPU_GET_CLASS(cpu);
cc->do_interrupt(cpu);
cpu->exception_index = -1;
} else if (!replay_has_interrupt()) {
/* give a chance to iothread in replay mode */
*ret = EXCP_INTERRUPT;
return true;
}
#endif
}
#ifndef CONFIG_USER_ONLY
} else if (replay_has_exception()
&& cpu->icount_decr.u16.low + cpu->icount_extra == 0) {
/* try to cause an exception pending in the log */
cpu_exec_nocache(cpu, 1, tb_find(cpu, NULL, 0), true);
*ret = -1;
return true;
#endif
}
return false;
}
static inline void cpu_handle_interrupt(CPUState *cpu,
TranslationBlock **last_tb)
{
CPUClass *cc = CPU_GET_CLASS(cpu);
int interrupt_request = cpu->interrupt_request;
if (unlikely(interrupt_request)) {
if (unlikely(cpu->singlestep_enabled & SSTEP_NOIRQ)) {
/* Mask out external interrupts for this step. */
interrupt_request &= ~CPU_INTERRUPT_SSTEP_MASK;
}
if (interrupt_request & CPU_INTERRUPT_DEBUG) {
cpu->interrupt_request &= ~CPU_INTERRUPT_DEBUG;
cpu->exception_index = EXCP_DEBUG;
cpu_loop_exit(cpu);
}
if (replay_mode == REPLAY_MODE_PLAY && !replay_has_interrupt()) {
/* Do nothing */
} else if (interrupt_request & CPU_INTERRUPT_HALT) {
replay_interrupt();
cpu->interrupt_request &= ~CPU_INTERRUPT_HALT;
cpu->halted = 1;
cpu->exception_index = EXCP_HLT;
cpu_loop_exit(cpu);
}
#if defined(TARGET_I386)
else if (interrupt_request & CPU_INTERRUPT_INIT) {
X86CPU *x86_cpu = X86_CPU(cpu);
CPUArchState *env = &x86_cpu->env;
replay_interrupt();
cpu_svm_check_intercept_param(env, SVM_EXIT_INIT, 0);
do_cpu_init(x86_cpu);
cpu->exception_index = EXCP_HALTED;
cpu_loop_exit(cpu);
}
#else
else if (interrupt_request & CPU_INTERRUPT_RESET) {
replay_interrupt();
cpu_reset(cpu);
cpu_loop_exit(cpu);
}
#endif
/* The target hook has 3 exit conditions:
False when the interrupt isn't processed,
True when it is, and we should restart on a new TB,
and via longjmp via cpu_loop_exit. */
else {
replay_interrupt();
if (cc->cpu_exec_interrupt(cpu, interrupt_request)) {
*last_tb = NULL;
}
/* The target hook may have updated the 'cpu->interrupt_request';
* reload the 'interrupt_request' value */
interrupt_request = cpu->interrupt_request;
}
if (interrupt_request & CPU_INTERRUPT_EXITTB) {
cpu->interrupt_request &= ~CPU_INTERRUPT_EXITTB;
/* ensure that no TB jump will be modified as
the program flow was changed */
*last_tb = NULL;
}
}
if (unlikely(atomic_read(&cpu->exit_request) || replay_has_interrupt())) {
atomic_set(&cpu->exit_request, 0);
cpu->exception_index = EXCP_INTERRUPT;
cpu_loop_exit(cpu);
}
}
static inline void cpu_loop_exec_tb(CPUState *cpu, TranslationBlock *tb,
TranslationBlock **last_tb, int *tb_exit,
SyncClocks *sc)
{
uintptr_t ret;
if (unlikely(atomic_read(&cpu->exit_request))) {
return;
}
trace_exec_tb(tb, tb->pc);
ret = cpu_tb_exec(cpu, tb);
*last_tb = (TranslationBlock *)(ret & ~TB_EXIT_MASK);
*tb_exit = ret & TB_EXIT_MASK;
switch (*tb_exit) {
case TB_EXIT_REQUESTED:
/* Something asked us to stop executing
* chained TBs; just continue round the main
* loop. Whatever requested the exit will also
* have set something else (eg exit_request or
* interrupt_request) which we will handle
* next time around the loop. But we need to
* ensure the tcg_exit_req read in generated code
* comes before the next read of cpu->exit_request
* or cpu->interrupt_request.
*/
smp_rmb();
*last_tb = NULL;
break;
case TB_EXIT_ICOUNT_EXPIRED:
{
/* Instruction counter expired. */
#ifdef CONFIG_USER_ONLY
abort();
#else
int insns_left = cpu->icount_decr.u32;
if (cpu->icount_extra && insns_left >= 0) {
/* Refill decrementer and continue execution. */
cpu->icount_extra += insns_left;
insns_left = MIN(0xffff, cpu->icount_extra);
cpu->icount_extra -= insns_left;
cpu->icount_decr.u16.low = insns_left;
} else {
if (insns_left > 0) {
/* Execute remaining instructions. */
cpu_exec_nocache(cpu, insns_left, *last_tb, false);
align_clocks(sc, cpu);
}
cpu->exception_index = EXCP_INTERRUPT;
*last_tb = NULL;
cpu_loop_exit(cpu);
}
break;
#endif
}
default:
break;
}
}
/* main execution loop */
int cpu_exec(CPUState *cpu)
{
CPUClass *cc = CPU_GET_CLASS(cpu);
int ret;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
SyncClocks sc;
/* replay_interrupt may need current_cpu */
current_cpu = cpu;
if (cpu_handle_halt(cpu)) {
return EXCP_HALTED;
}
atomic_mb_set(&tcg_current_cpu, cpu);
rcu_read_lock();
if (unlikely(atomic_mb_read(&exit_request))) {
cpu->exit_request = 1;
}
cc->cpu_exec_enter(cpu);
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
/* Calculate difference between guest clock and host clock.
* This delay includes the delay of the last cycle, so
* what we have to do is sleep until it is 0. As for the
* advance/delay we gain here, we try to fix it next time.
*/
init_delay_params(&sc, cpu);
for(;;) {
/* prepare setjmp context for exception handling */
if (sigsetjmp(cpu->jmp_env, 0) == 0) {
TranslationBlock *tb, *last_tb = NULL;
int tb_exit = 0;
/* if an exception is pending, we execute it here */
if (cpu_handle_exception(cpu, &ret)) {
break;
}
for(;;) {
cpu_handle_interrupt(cpu, &last_tb);
tb = tb_find(cpu, last_tb, tb_exit);
cpu_loop_exec_tb(cpu, tb, &last_tb, &tb_exit, &sc);
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 13:56:31 +04:00
/* Try to align the host and virtual clocks
if the guest is in advance */
align_clocks(&sc, cpu);
} /* for(;;) */
} else {
#if defined(__clang__) || !QEMU_GNUC_PREREQ(4, 6)
/* Some compilers wrongly smash all local variables after
* siglongjmp. There were bug reports for gcc 4.5.0 and clang.
* Reload essential local variables here for those compilers.
* Newer versions of gcc would complain about this code (-Wclobbered). */
cpu = current_cpu;
cc = CPU_GET_CLASS(cpu);
#else /* buggy compiler */
/* Assert that the compiler does not smash local variables. */
g_assert(cpu == current_cpu);
g_assert(cc == CPU_GET_CLASS(cpu));
#endif /* buggy compiler */
cpu->can_do_io = 1;
tb_lock_reset();
}
} /* for(;;) */
cc->cpu_exec_exit(cpu);
rcu_read_unlock();
/* fail safe : never use current_cpu outside cpu_exec() */
current_cpu = NULL;
/* Does not need atomic_mb_set because a spurious wakeup is okay. */
atomic_set(&tcg_current_cpu, NULL);
return ret;
}