v15->v16: - Remove the lfsr patch and use linear probing as lfsr is not really necessary in most cases. - Move the paravirt PV_CALLEE_SAVE_REGS_THUNK code to an asm header. - Add a patch to collect PV qspinlock statistics which also supersedes the PV lock hash debug patch. - Add PV qspinlock performance numbers. v14->v15: - Incorporate PeterZ's v15 qspinlock patch and improve upon the PV qspinlock code by dynamically allocating the hash table as well as some other performance optimization. - Simplified the Xen PV qspinlock code as suggested by David Vrabel <david.vrabel@xxxxxxxxxx>. - Add benchmarking data for 3.19 kernel to compare the performance of a spinlock heavy test with and without the qspinlock patch under different cpufreq drivers and scaling governors. v13->v14: - Patches 1 & 2: Add queue_spin_unlock_wait() to accommodate commit 78bff1c86 from Oleg Nesterov. - Fix the system hang problem when using PV qspinlock in an over-committed guest due to a racing condition in the pv_set_head_in_tail() function. - Increase the MAYHALT_THRESHOLD from 10 to 1024. - Change kick_cpu into a regular function pointer instead of a callee-saved function. - Change lock statistics code to use separate bits for different statistics. v12->v13: - Change patch 9 to generate separate versions of the queue_spin_lock_slowpath functions for bare metal and PV guest. This reduces the performance impact of the PV code on bare metal systems. v11->v12: - Based on PeterZ's version of the qspinlock patch (https://lkml.org/lkml/2014/6/15/63). - Incorporated many of the review comments from Konrad Wilk and Paolo Bonzini. - The pvqspinlock code is largely from my previous version with PeterZ's way of going from queue tail to head and his idea of using callee saved calls to KVM and XEN codes. v10->v11: - Use a simple test-and-set unfair lock to simplify the code, but performance may suffer a bit for large guest with many CPUs. - Take out Raghavendra KT's test results as the unfair lock changes may render some of his results invalid. - Add PV support without increasing the size of the core queue node structure. - Other minor changes to address some of the feedback comments. v9->v10: - Make some minor changes to qspinlock.c to accommodate review feedback. - Change author to PeterZ for 2 of the patches. - Include Raghavendra KT's test results in patch 18. v8->v9: - Integrate PeterZ's version of the queue spinlock patch with some modification: http://lkml.kernel.org/r/20140310154236.038181843@xxxxxxxxxxxxx - Break the more complex patches into smaller ones to ease review effort. - Fix a racing condition in the PV qspinlock code. v7->v8: - Remove one unneeded atomic operation from the slowpath, thus improving performance. - Simplify some of the codes and add more comments. - Test for X86_FEATURE_HYPERVISOR CPU feature bit to enable/disable unfair lock. - Reduce unfair lock slowpath lock stealing frequency depending on its distance from the queue head. - Add performance data for IvyBridge-EX CPU. v6->v7: - Remove an atomic operation from the 2-task contending code - Shorten the names of some macros - Make the queue waiter to attempt to steal lock when unfair lock is enabled. - Remove lock holder kick from the PV code and fix a race condition - Run the unfair lock & PV code on overcommitted KVM guests to collect performance data. v5->v6: - Change the optimized 2-task contending code to make it fairer at the expense of a bit of performance. - Add a patch to support unfair queue spinlock for Xen. - Modify the PV qspinlock code to follow what was done in the PV ticketlock. - Add performance data for the unfair lock as well as the PV support code. v4->v5: - Move the optimized 2-task contending code to the generic file to enable more architectures to use it without code duplication. - Address some of the style-related comments by PeterZ. - Allow the use of unfair queue spinlock in a real para-virtualized execution environment. - Add para-virtualization support to the qspinlock code by ensuring that the lock holder and queue head stay alive as much as possible. v3->v4: - Remove debugging code and fix a configuration error - Simplify the qspinlock structure and streamline the code to make it perform a bit better - Add an x86 version of asm/qspinlock.h for holding x86 specific optimization. - Add an optimized x86 code path for 2 contending tasks to improve low contention performance. v2->v3: - Simplify the code by using numerous mode only without an unfair option. - Use the latest smp_load_acquire()/smp_store_release() barriers. - Move the queue spinlock code to kernel/locking. - Make the use of queue spinlock the default for x86-64 without user configuration. - Additional performance tuning. v1->v2: - Add some more comments to document what the code does. - Add a numerous CPU mode to support >= 16K CPUs - Add a configuration option to allow lock stealing which can further improve performance in many cases. - Enable wakeup of queue head CPU at unlock time for non-numerous CPU mode. This patch set has 3 different sections: 1) Patches 1-6: Introduces a queue-based spinlock implementation that can replace the default ticket spinlock without increasing the size of the spinlock data structure. As a result, critical kernel data structures that embed spinlock won't increase in size and break data alignments. 2) Patch 7: Enables the use of unfair lock in a virtual guest. This can resolve some of the locking related performance issues due to halting the waiting CPUs after spinning for a certain amount of time. The unlock code will detect the a sleeping waiter and wake it up. This is essentially the same logic as the PV ticketlock code. 3) Patches 8-14: Add paravirtualization spinlock support. This is needed as paravirt spinlock was enabled by default on distribution like RHEL 7. The queue spinlock has slightly better performance than the ticket spinlock in uncontended case. Its performance can be much better with moderate to heavy contention. This patch has the potential of however, will not be able to support TSX. The queue spinlock is especially suitable for NUMA machines with at least 2 sockets. Though even at the 2-socket level, there can be significant speedup depending on the workload. The purpose of this patch set is not to solve any particular spinlock contention problems. Those need to be solved by refactoring the code to make more efficient use of the lock or finer granularity ones. The main purpose is to make the lock contention problems more tolerable until someone can spend the time and effort to fix them. The PV portion of the patch series is geared toward performance optimization for bare metal. However, the qspinlock performance in virtual guest should still be comparable with ticket spinlock at low load and much better at high load. Native qspinlock patch performance ---------------------------------- In term of the performance benefit of this patch, I ran the high_systime workload (which does a lot of fork() and exit()) at various load levels (500, 1000, 1500 and 2000 users) on a 4-socket IvyBridge-EX bare-metal system (60 cores, 120 threads) with intel_pstate driver and performance scaling governor. The JPM (jobs/minutes) and execution time results were as follows: Kernel JPM Execution Time ------ --- -------------- At 500 users: 3.19 118857.14 26.25s 3.19-qspinlock 134889.75 23.13s % change +13.5% -11.9% At 1000 users: 3.19 204255.32 30.55s 3.19-qspinlock 239631.34 26.04s % change +17.3% -14.8% At 1500 users: 3.19 177272.73 52.80s 3.19-qspinlock 326132.40 28.70s % change +84.0% -45.6% At 2000 users: 3.19 196690.31 63.45s 3.19-qspinlock 341730.56 36.52s % change +73.7% -42.4% It turns out that this workload was causing quite a lot of spinlock contention in the vanilla 3.19 kernel. The performance advantage of this patch increases with heavier loads. With the powersave governor, the JPM data were as follows: Users 3.19 3.19-qspinlock % Change ----- ---- -------------- -------- 500 112635.38 132596.69 +17.7% 1000 171240.40 240369.80 +40.4% 1500 130507.53 324436.74 +148.6% 2000 175972.93 341637.01 +94.1% With the qspinlock patch, there wasn't too much difference in performance between the 2 scaling governors. Without this patch, the powersave governor was much slower than the performance governor. By disabling the intel_pstate driver and use acpi_cpufreq instead, the benchmark performance (JPM) at 1000 users level for the performance and ondemand governors were: Governor 3.19 3.19-qspinlock % Change -------- ---- -------------- -------- performance 124949.94 219950.65 +76.0% ondemand 4838.90 206690.96 +4171% The performance was just horrible when there was significant spinlock contention with the ondemand governor. There was also significant run-to-run variation. A second run of the same benchmark gave a result of 22115 JPMs. With the qspinlock patch, however, the performance was much more stable under different cpufreq drivers and governors. That is not the case with the default ticket spinlock implementation. The %CPU times spent on spinlock contention (from perf) with the performance governor and the intel_pstate driver were: Kernel Function 3.19 kernel 3.19-qspinlock kernel --------------- ----------- --------------------- At 500 users: _raw_spin_lock* 28.23% 2.25% queue_spin_lock_slowpath N/A 4.05% At 1000 users: _raw_spin_lock* 23.21% 2.25% queue_spin_lock_slowpath N/A 4.42% At 1500 users: _raw_spin_lock* 29.07% 2.24% queue_spin_lock_slowpath N/A 4.49% At 2000 users: _raw_spin_lock* 29.15% 2.26% queue_spin_lock_slowpath N/A 4.82% The top spinlock related entries in the perf profile for the 3.19 kernel at 1000 users were: 7.40% reaim [kernel.kallsyms] [k] _raw_spin_lock_irqsave |--58.96%-- rwsem_wake |--20.02%-- release_pages |--15.88%-- pagevec_lru_move_fn |--1.53%-- get_page_from_freelist |--0.78%-- __wake_up |--0.55%-- try_to_wake_up --2.28%-- [...] 3.13% reaim [kernel.kallsyms] [k] _raw_spin_lock |--37.55%-- free_one_page |--17.47%-- __cache_free_alien |--4.95%-- __rcu_process_callbacks |--2.93%-- __pte_alloc |--2.68%-- __drain_alien_cache |--2.56%-- ext4_do_update_inode |--2.54%-- try_to_wake_up |--2.46%-- pgd_free |--2.32%-- cache_alloc_refill |--2.32%-- pgd_alloc |--2.32%-- free_pcppages_bulk |--1.88%-- do_wp_page |--1.77%-- handle_pte_fault |--1.58%-- do_anonymous_page |--1.56%-- rmqueue_bulk.clone.0 |--1.35%-- copy_pte_range |--1.25%-- zap_pte_range |--1.13%-- cache_flusharray |--0.88%-- __pmd_alloc |--0.70%-- wake_up_new_task |--0.66%-- __pud_alloc |--0.59%-- ext4_discard_preallocations --6.53%-- [...] With the qspinlock patch, the perf profile at 1000 users was: 3.25% reaim [kernel.kallsyms] [k] queue_spin_lock_slowpath |--62.00%-- _raw_spin_lock_irqsave | |--77.49%-- rwsem_wake | |--11.99%-- release_pages | |--4.34%-- pagevec_lru_move_fn | |--1.93%-- get_page_from_freelist | |--1.90%-- prepare_to_wait_exclusive | |--1.29%-- __wake_up | |--0.74%-- finish_wait |--11.63%-- _raw_spin_lock | |--31.11%-- try_to_wake_up | |--18.48%-- free_one_page | |--13.97%-- __cache_free_alien | |--7.77%-- free_pcppages_bulk | |--7.12%-- __drain_alien_cache | |--6.17%-- rmqueue_bulk.clone.0 | |--4.17%-- __rcu_process_callbacks | |--2.22%-- cache_alloc_refill | |--2.15%-- wake_up_new_task | |--1.80%-- ext4_do_update_inode | |--1.52%-- cache_flusharray | |--0.89%-- __mutex_unlock_slowpath | |--0.64%-- ttwu_queue |--11.19%-- _raw_spin_lock_irq | |--98.95%-- rwsem_down_write_failed | |--0.93%-- __schedule |--7.91%-- queue_read_lock_slowpath | _raw_read_lock | |--96.79%-- do_wait | |--2.44%-- do_prlimit | chrdev_open | do_dentry_open | vfs_open | do_last | path_openat | do_filp_open | do_sys_open | sys_open | system_call | __GI___libc_open |--7.05%-- queue_write_lock_slowpath | _raw_write_lock_irq | |--35.36%-- release_task | |--32.76%-- copy_process | do_exit | do_group_exit | sys_exit_group | system_call --0.22%-- [...] This demonstrates the benefit of this patch for those applications that run on multi-socket machines and can cause significant spinlock contentions in the kernel. I also got report that the queue spinlock patch can improve the performance of an I/O and interrupt intensive stress test with a lot of spinlock contention on a 2-socket system by up to 20%. Daniel Blueman of NumaScale had tested the v14 qspinlock patch on an older 512-core NumaConnect system for testing with 3.19.1: $ src/reaim -c data/reaim.config -f data/workfile.compute -i 16 -e 256 Num Parent Child Child Jobs per Jobs/min/ Std_dev Std_dev JTI Forked Time SysTime UTime Minute Child Time Percent 1 2.15 0.36 1.19 2818.60 2818.60 0.00 0.00 100 17 6.29 42.41 22.43 16378.38 963.43 0.27 4.43 95 33 56.24 1611.77 47.04 3555.83 107.75 1.58 2.88 97 49 120.88 5576.87 63.54 2456.49 50.13 1.75 1.47 98 65 199.17 12328.42 81.27 1977.71 30.43 2.42 1.23 98 81 270.89 20943.99 103.99 1812.03 22.37 3.85 1.44 98 97 315.31 29211.63 121.30 1864.26 19.22 4.48 1.44 98 113 397.68 42766.15 144.48 1721.94 15.24 7.04 1.80 98 129 520.74 63442.27 162.03 1501.21 11.64 10.24 1.99 98 ... Patched with qspinlocks v14: Num Parent Child Child Jobs per Jobs/min/ Std_dev Std_dev JTI Forked Time SysTime UTime Minute Child Time Percent 1 1.98 0.36 1.21 3060.61 3060.61 0.00 0.00 100 17 5.60 25.41 22.39 18396.43 1082.14 0.16 3.00 97 33 12.67 153.64 49.17 15783.74 478.30 0.50 4.16 95 49 21.15 365.55 76.61 14039.72 286.52 0.83 4.19 95 65 27.57 556.87 105.96 14287.27 219.80 0.75 2.80 97 81 33.07 712.92 127.85 14843.06 183.25 1.16 3.69 96 97 42.81 1009.02 161.87 13730.90 141.56 1.75 4.27 95 113 49.13 1166.04 185.10 13938.12 123.35 2.11 4.53 95 129 56.62 1262.07 231.92 13806.78 107.03 2.67 4.99 95 145 66.41 1420.29 250.24 13231.44 91.25 2.95 4.70 95 179 90.15 2237.78 325.18 12032.61 67.22 4.23 4.94 95 193 85.41 1865.81 326.64 13693.71 70.95 4.29 5.32 94 220 93.58 2298.77 376.46 14246.63 64.76 4.97 5.65 94 Max Jobs per Minute 18396.43 The compute workload is mostly userspace code with some synchronization between working threads (HPC type workload). With large user count, the patched kernel has close to 8X the performance of an unpatched kernel. Paravirt qspinlock patch performance ------------------------------------ A linux kernel build workload (make -j <n>) was run on KVM and Xen guests with and without the qspinlock patch. The host machine was an 8-socket Westmere-EX server with 4 active cores/socket (8x4). For the KVM test, the VM configurations were: 1) 32 NUMA-pinned vCPUs (8x4) 2) 32 vCPUs (no pinning or NUMA) 3) 60 vCPUs (overcommitted) The kernel build times for different 4.0-rc6 based kernels were: Kernel VM1 VM2 VM3 ------ --- --- --- PV ticket lock 3m49.7s 11m45.6s 15m59.3s PV qspinlock 3m50.4s 5m51.6s 15m33.6s For VM1, both the qspinlock and ticket lock have essentially the same performance. VM2 illustrated the performance benefit of qspinlock versus ticket spinlock which got reduced in VM3 due to the overhead of constant vCPUs halting and kicking. For the Xen test, the VM configurations were: 1) 32 vCPUs (no pinning or NUMA) 2) 64 vCPUs 2) 128 vCPUs The kernel build times for different 4.0 based kernels were: Kernel VM1 VM2 VM3 ------ --- --- --- PV ticket lock 7m27.3s 16m14.1s 18m26.9s PV qspinlock 5m51.6s 12m02.6s 14m19.2s Here, the PV qspinlock was faster than the PV ticket lock. David Vrabel (1): pvqspinlock, x86: Enable PV qspinlock for Xen Peter Zijlstra (Intel) (4): qspinlock: Add pending bit qspinlock: Optimize for smaller NR_CPUS qspinlock: Revert to test-and-set on hypervisors pvqspinlock, x86: Implement the paravirt qspinlock call patching Waiman Long (9): qspinlock: A simple generic 4-byte queue spinlock qspinlock, x86: Enable x86-64 to use queue spinlock qspinlock: Extract out code snippets for the next patch qspinlock: Use a simple write to grab the lock pvqspinlock: Implement simple paravirt support for the qspinlock pvqspinlock, x86: Enable PV qspinlock for KVM pvqspinlock: Only kick CPU at unlock time pvqspinlock: Improve slowpath performance by avoiding cmpxchg pvqspinlock: Collect slowpath lock statistics arch/x86/Kconfig | 3 +- arch/x86/include/asm/paravirt.h | 29 ++- arch/x86/include/asm/paravirt_types.h | 10 + arch/x86/include/asm/qspinlock.h | 57 ++++ arch/x86/include/asm/qspinlock_paravirt.h | 6 + arch/x86/include/asm/spinlock.h | 5 + arch/x86/include/asm/spinlock_types.h | 4 + arch/x86/kernel/kvm.c | 43 +++ arch/x86/kernel/paravirt-spinlocks.c | 24 ++- arch/x86/kernel/paravirt_patch_32.c | 22 +- arch/x86/kernel/paravirt_patch_64.c | 22 +- arch/x86/xen/spinlock.c | 64 ++++- include/asm-generic/qspinlock.h | 139 +++++++++ include/asm-generic/qspinlock_types.h | 79 +++++ kernel/Kconfig.locks | 7 + kernel/locking/Makefile | 1 + kernel/locking/mcs_spinlock.h | 1 + kernel/locking/qspinlock.c | 473 ++++++++++++++++++++++++++++ kernel/locking/qspinlock_paravirt.h | 478 +++++++++++++++++++++++++++++ 19 files changed, 1452 insertions(+), 15 deletions(-) create mode 100644 arch/x86/include/asm/qspinlock.h create mode 100644 arch/x86/include/asm/qspinlock_paravirt.h create mode 100644 include/asm-generic/qspinlock.h create mode 100644 include/asm-generic/qspinlock_types.h create mode 100644 kernel/locking/qspinlock.c create mode 100644 kernel/locking/qspinlock_paravirt.h _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/virtualization
