- Linuxarm - mailweb.openeuler.org

Re: [PATCH net v4 1/2] net: sched: fix packet stuck problem for lockless qdisc
by Michal Kubecek 21 Apr '21

21 Apr '21

On Wed, Apr 21, 2021 at 04:21:54PM +0800, Yunsheng Lin wrote: > > I tried using below shell to simulate your testcase: > > #!/bin/sh > > for((i=0; i<20; i++)) > do > taskset -c 0-31 netperf -t TCP_STREAM -H 192.168.100.2 -l 30 -- -m 1048576 > done > > And I got a quite stable result: 9413~9416 (10^6bits/sec) for 10G netdev. Perhaps try it without the taskset, in my test, there was only one connection. > > > > https://github.com/mkubecek/nperf > > > > It is still raw and a lot of features are missing but it can be already > > used for multithreaded TCP_STREAM and TCP_RR tests. In particular, the > > output above was with > > > > nperf -H 172.17.1.1 -l 30 -i 20 --exact -t TCP_STREAM -M 1 > > I tried your nperf too, unfortunately it does not seem to work on my > system(arm64), which exits very quickly and output the blow result: > > root@(none):/home# nperf -H 192.168.100.2 -l 30 -i 20 --exact -t TCP_STREAM -M 1 > server: 192.168.100.2, port 12543 > iterations: 20, threads: 1, test length: 30 > test: TCP_STREAM, message size: 1048576 > > 1 4.0 B/s, avg 4.0 B/s, mdev 0.0 B/s ( 0.0%) [...] Did you start nperfd on the other side? (It plays a role similar to netserver for netperf.) Few days ago I noticed that there is something wrong with error handling in case of failed connection but didn't get to fixing it yet. I'll try running some tests also on other architectures, including arm64 and s390x (to catch potential endinanity issues). Michal

1 0

Re: [PATCH V5] app/test-pmd: support cleanup txq mbufs command
by Ferruh Yigit 21 Apr '21

21 Apr '21

On 4/21/2021 9:09 AM, Lijun Ou wrote: > From: Chengwen Feng <fengchengwen(a)huawei.com> > > This patch supports cleanup txq mbufs command: > port cleanup (port_id) txq (queue_id) (free_cnt) > > Signed-off-by: Chengwen Feng <fengchengwen(a)huawei.com> > Signed-off-by: Lijun Ou <oulijun(a)huawei.com> > --- > V4->V5: > - rewrite patch title > - define the new cmd. > - Fix the comments given by Ferruh.yigit > > V3->V4: > - revert the V3 scheme. > > V2->V3: > - The command implementation is changed so that the queuestate does > not depend on the command execution. > > V1->V2: > - use Tx instead of TX > - add note in doc > --- > app/test-pmd/cmdline.c | 85 +++++++++++++++++++++++++++++ > doc/guides/rel_notes/release_21_05.rst | 2 + > doc/guides/testpmd_app_ug/testpmd_funcs.rst | 9 +++ > 3 files changed, 96 insertions(+) > Can you please update 'cmd_help_long_parsed' for the help string, this was in v4 and seems dropped during change.

1 0

Re: [Intel-wired-lan] [PATCH V2 net] ice: Re-organizes reqstd/avail {R, T}XQ check/code for efficiency+readability
by Paul Menzel 21 Apr '21

21 Apr '21

Dear Salil, Thank you very much for your patch. In the git commit message summary, could you please use imperative mood [1]? > Re-organize reqstd/avail {R, T}XQ check/code for efficiency+readability It’s a bit long though. Maybe: > Avoid unnecessary assignment with user specified {R,T}XQs Am 14.04.21 um 00:44 schrieb Salil Mehta: > If user has explicitly requested the number of {R,T}XQs, then it is > unnecessary to get the count of already available {R,T}XQs from the > PF avail_{r,t}xqs bitmap. This value will get overridden by user specified > value in any case. > > This patch does minor re-organization of the code for improving the flow > and readabiltiy. This scope of improvement was found during the review of readabil*it*y > the ICE driver code. > > FYI, I could not test this change due to unavailability of the hardware. > It would be helpful if somebody can test this patch and provide Tested-by > Tag. Many thanks! This should go outside the commit message (below the --- for example). > Fixes: 87324e747fde ("ice: Implement ethtool ops for channels") Did you check the behavior before is actually a bug? Or is it just for the detection heuristic for commits to be applied to the stable series? > Cc: intel-wired-lan(a)lists.osuosl.org > Cc: Jeff Kirsher <jeffrey.t.kirsher(a)intel.com> > Signed-off-by: Salil Mehta <salil.mehta(a)huawei.com> > -- > Change V1->V2 > (*) Fixed the comments from Anthony Nguyen(Intel) > Link: https://lkml.org/lkml/2021/4/12/1997 > --- > drivers/net/ethernet/intel/ice/ice_lib.c | 14 ++++++++------ > 1 file changed, 8 insertions(+), 6 deletions(-) > > diff --git a/drivers/net/ethernet/intel/ice/ice_lib.c b/drivers/net/ethernet/intel/ice/ice_lib.c > index d13c7fc8fb0a..d77133d6baa7 100644 > --- a/drivers/net/ethernet/intel/ice/ice_lib.c > +++ b/drivers/net/ethernet/intel/ice/ice_lib.c > @@ -161,12 +161,13 @@ static void ice_vsi_set_num_qs(struct ice_vsi *vsi, u16 vf_id) > > switch (vsi->type) { > case ICE_VSI_PF: > - vsi->alloc_txq = min3(pf->num_lan_msix, > - ice_get_avail_txq_count(pf), > - (u16)num_online_cpus()); > if (vsi->req_txq) { > vsi->alloc_txq = vsi->req_txq; > vsi->num_txq = vsi->req_txq; > + } else { > + vsi->alloc_txq = min3(pf->num_lan_msix, > + ice_get_avail_txq_count(pf), > + (u16)num_online_cpus()); > } I am curious, did you check the compiler actually creates different code, or did it notice the inefficiency by itself and optimized it already? > > pf->num_lan_tx = vsi->alloc_txq; > @@ -175,12 +176,13 @@ static void ice_vsi_set_num_qs(struct ice_vsi *vsi, u16 vf_id) > if (!test_bit(ICE_FLAG_RSS_ENA, pf->flags)) { > vsi->alloc_rxq = 1; > } else { > - vsi->alloc_rxq = min3(pf->num_lan_msix, > - ice_get_avail_rxq_count(pf), > - (u16)num_online_cpus()); > if (vsi->req_rxq) { > vsi->alloc_rxq = vsi->req_rxq; > vsi->num_rxq = vsi->req_rxq; > + } else { > + vsi->alloc_rxq = min3(pf->num_lan_msix, > + ice_get_avail_rxq_count(pf), > + (u16)num_online_cpus()); > } > } > Kind regards, Paul

2 3

Re: [PATCH net v4 1/2] net: sched: fix packet stuck problem for lockless qdisc
by Michal Kubecek 21 Apr '21

21 Apr '21

On Wed, Apr 21, 2021 at 09:52:40AM +0800, Yunsheng Lin wrote: > On 2021/4/21 4:34, Michal Kubecek wrote: > > However, I noticed something disturbing in the results of a simple > > 1-thread TCP_STREAM test (client sends data through a TCP connection to > > server using long writes, we measure the amount of data received by the > > server): > > > > server: 172.17.1.1, port 12543 > > iterations: 20, threads: 1, test length: 30 > > test: TCP_STREAM, message size: 1048576 > > > > 1 927403548.4 B/s, avg 927403548.4 B/s, mdev 0.0 B/s ( 0.0%) > > 2 1176317172.1 B/s, avg 1051860360.2 B/s, mdev 124456811.8 B/s ( 11.8%), confid. +/- 1581348251.3 B/s (150.3%) > > 3 927335837.8 B/s, avg 1010352186.1 B/s, mdev 117354970.3 B/s ( 11.6%), confid. +/- 357073677.2 B/s ( 35.3%) > > 4 1176728045.1 B/s, avg 1051946150.8 B/s, mdev 124576544.7 B/s ( 11.8%), confid. +/- 228863127.8 B/s ( 21.8%) > > 5 1176788216.3 B/s, avg 1076914563.9 B/s, mdev 122102985.3 B/s ( 11.3%), confid. +/- 169478943.5 B/s ( 15.7%) > > 6 1158167055.1 B/s, avg 1090456645.8 B/s, mdev 115504209.5 B/s ( 10.6%), confid. +/- 132805140.8 B/s ( 12.2%) > > 7 1176243474.4 B/s, avg 1102711907.0 B/s, mdev 111069717.1 B/s ( 10.1%), confid. +/- 110956822.2 B/s ( 10.1%) > > 8 1176771142.8 B/s, avg 1111969311.5 B/s, mdev 106744173.5 B/s ( 9.6%), confid. +/- 95417120.0 B/s ( 8.6%) > > 9 1176206364.6 B/s, avg 1119106761.8 B/s, mdev 102644185.2 B/s ( 9.2%), confid. +/- 83685200.5 B/s ( 7.5%) > > 10 1175888409.4 B/s, avg 1124784926.6 B/s, mdev 98855550.5 B/s ( 8.8%), confid. +/- 74537085.1 B/s ( 6.6%) > > 11 1176541407.6 B/s, avg 1129490061.2 B/s, mdev 95422224.8 B/s ( 8.4%), confid. +/- 67230249.7 B/s ( 6.0%) > > 12 934185352.8 B/s, avg 1113214668.9 B/s, mdev 106114984.5 B/s ( 9.5%), confid. +/- 70420712.5 B/s ( 6.3%) > > 13 1176550558.1 B/s, avg 1118086660.3 B/s, mdev 103339448.9 B/s ( 9.2%), confid. +/- 65002902.4 B/s ( 5.8%) > > 14 1176521808.8 B/s, avg 1122260599.5 B/s, mdev 100711151.3 B/s ( 9.0%), confid. +/- 60333655.0 B/s ( 5.4%) > > 15 1176744840.8 B/s, avg 1125892882.3 B/s, mdev 98240838.2 B/s ( 8.7%), confid. +/- 56319052.3 B/s ( 5.0%) > > 16 1176593778.5 B/s, avg 1129061688.3 B/s, mdev 95909740.8 B/s ( 8.5%), confid. +/- 52771633.5 B/s ( 4.7%) > > 17 1176583967.4 B/s, avg 1131857116.5 B/s, mdev 93715582.2 B/s ( 8.3%), confid. +/- 49669258.6 B/s ( 4.4%) > > 18 1176853301.8 B/s, avg 1134356904.5 B/s, mdev 91656530.2 B/s ( 8.1%), confid. +/- 46905244.8 B/s ( 4.1%) > > 19 1176592845.7 B/s, avg 1136579848.8 B/s, mdev 89709043.8 B/s ( 7.9%), confid. +/- 44424855.9 B/s ( 3.9%) > > 20 1176608117.3 B/s, avg 1138581262.2 B/s, mdev 87871692.6 B/s ( 7.7%), confid. +/- 42193098.5 B/s ( 3.7%) > > all avg 1138581262.2 B/s, mdev 87871692.6 B/s ( 7.7%), confid. +/- 42193098.5 B/s ( 3.7%) > > > > Each line shows result of one 30 second long test and average, mean > > deviation and 99% confidence interval half width through the iterations > > so far. While 17 iteration results are essentially the wire speed minus > > TCP overhead, iterations 1, 3 and 12 are more than 20% lower. As results > > of the same test on unpatched 5.12-rc7 are much more consistent (the > > lowest iteration result through the whole test was 1175939718.3 and the > > mean deviation only 276889.1 B/s), it doesn't seeem to be just a random > > fluctuation. > > I think I need to relearn the statistial math to understand the above > "99% confidence interval half width ":) An easy way to understand it is that if the last column shows 42 MB/s, it means that with 99% confidence (probability), the measured average is within 42 MB/s off the actual one. > But the problem do not seems related too much with "99% confidence > interval half width ", but with "mean deviation"? Mean deviation is a symptom here. What worries me is that most results show the same value (corresponding to fully saturated line) with very little variation, in some iterations (1, 3 and 12 here) we can suddenly see much lower value (by ~2.5 GB/s, i.e. 20-25%). And as each iteration runs the connection for 30 seconds, it cannot be just some short glitch. I managed to get tcpdump captures yesterday but they are huge even with "-s 128" (client ~5.6 GB, server ~9.0 GB) so that working with them is rather slow so I did not find anything interesting yet. > I tried using netperf, which seems only show throughput of 9415.06 > (10^6bits/sec) using 10G netdev. which tool did you used to show the > above number? 9415.06 * 10^6 b/s is 1176.9 * 10^6 B/s so it's about the same as the numbers above (the good ones, that is). As this was part of a longer test trying different thread counts from 1 to 128, I was using another utility I started writing recently: https://github.com/mkubecek/nperf It is still raw and a lot of features are missing but it can be already used for multithreaded TCP_STREAM and TCP_RR tests. In particular, the output above was with nperf -H 172.17.1.1 -l 30 -i 20 --exact -t TCP_STREAM -M 1 The results are with 1 thread so that they should be also reproducible with netperf too. But it needs to be repeated enough times, when I wanted to get the packet captures, I did 40 iterations and only two of them showed lower result. Michal

1 0

[RFC PATCH v5 0/4] scheduler: expose the topology of clusters and add cluster scheduler
by Barry Song 21 Apr '21

21 Apr '21

ARM64 server chip Kunpeng 920 has 6 or 8 clusters in each NUMA node, and each cluster has 4 cpus. All clusters share L3 cache data while each cluster has local L3 tag. On the other hand, each cluster will share some internal system bus. This means cache is much more affine inside one cluster than across clusters. +-----------------------------------+ +---------+ | +------+ +------+ +---------------------------+ | | | CPU0 | | cpu1 | | +-----------+ | | | +------+ +------+ | | | | | | +----+ L3 | | | | +------+ +------+ cluster | | tag | | | | | CPU2 | | CPU3 | | | | | | | +------+ +------+ | +-----------+ | | | | | | +-----------------------------------+ | | +-----------------------------------+ | | | +------+ +------+ +--------------------------+ | | | | | | | +-----------+ | | | +------+ +------+ | | | | | | | | L3 | | | | +------+ +------+ +----+ tag | | | | | | | | | | | | | | +------+ +------+ | +-----------+ | | | | | | +-----------------------------------+ | L3 | | data | +-----------------------------------+ | | | +------+ +------+ | +-----------+ | | | | | | | | | | | | | +------+ +------+ +----+ L3 | | | | | | tag | | | | +------+ +------+ | | | | | | | | | | ++ +-----------+ | | | +------+ +------+ |---------------------------+ | +-----------------------------------| | | +-----------------------------------| | | | +------+ +------+ +---------------------------+ | | | | | | | +-----------+ | | | +------+ +------+ | | | | | | +----+ L3 | | | | +------+ +------+ | | tag | | | | | | | | | | | | | | +------+ +------+ | +-----------+ | | | | | | +-----------------------------------+ | | +-----------------------------------+ | | | +------+ +------+ +--------------------------+ | | | | | | | +-----------+ | | | +------+ +------+ | | | | | There is a similar need for clustering in x86. Some x86 cores could share L2 caches that is similar to the cluster in Kupeng 920 (e.g. on Jacobsville there are 6 clusters of 4 Atom cores, each cluster sharing a separate L2, and 24 cores sharing L3). Having a sched_domain for clusters will bring two aspects of improvement: 1. spreading unrelated tasks among clusters, which decreases the contention of resources and improve the throughput. unrelated tasks might be put randomly without cluster sched_domain: +-------------------+ +-----------------+ | +----+ +----+ | | | | |task| |task| | | | | |1 | |2 | | | | | +----+ +----+ | | | | | | | | cluster1 | | cluster2 | +-------------------+ +-----------------+ but with cluster sched_domain, they are likely to spread due to LB: +-------------------+ +-----------------+ | +----+ | | +----+ | | |task| | | |task| | | |1 | | | |2 | | | +----+ | | +----+ | | | | | | cluster1 | | cluster2 | +-------------------+ +-----------------+ 2. gathering related tasks within a cluster, which improves the cache affinity of tasks talking with each other. Without cluster sched_domain, related tasks might be put randomly. In case task1-8 have relationship as below: Task1 wakes up task4 Task2 wakes up task5 Task3 wakes up task6 Task4 wakes up task7 With the tuning of select_idle_cpu() to scan local cluster first, those tasks might get a chance to be gathered like: +---------------------------+ +----------------------+ | +----+ +-----+ | | +----+ +-----+ | | |task| |task | | | |task| |task | | | |1 | | 4 | | | |2 | |5 | | | +----+ +-----+ | | +----+ +-----+ | | | | | | cluster1 | | cluster2 | | | | | | | | | | +-----+ +------+ | | +-----+ +------+ | | |task | | task | | | |task | |task | | | |3 | | 6 | | | |4 | |8 | | | +-----+ +------+ | | +-----+ +------+ | +---------------------------+ +----------------------+ Otherwise, the result might be: +---------------------------+ +----------------------+ | +----+ +-----+ | | +----+ +-----+ | | |task| |task | | | |task| |task | | | |1 | | 2 | | | |5 | |6 | | | +----+ +-----+ | | +----+ +-----+ | | | | | | cluster1 | | cluster2 | | | | | | | | | | +-----+ +------+ | | +-----+ +------+ | | |task | | task | | | |task | |task | | | |3 | | 4 | | | |7 | |8 | | | +-----+ +------+ | | +-----+ +------+ | +---------------------------+ +----------------------+ -v5: * split "add scheduler level for clusters" into two patches to evaluate the impact of spreading and gathering separately; * add a tracepoint of select_idle_cpu for debug purpose; add bcc script in commit log; * add cluster_id = -1 in reset_cpu_topology() * rebased to tip/sched/core -v4: * rebased to tip/sched/core with the latest unified code of select_idle_cpu * added Tim's patch for x86 Jacobsville * also added benchmark data of spreading unrelated tasks * avoided the iteration of sched_domain by moving to static_key(addressing Vincent's comment * used acpi_cpu_id for acpi_find_processor_node(addressing Masa's comment) Barry Song (2): scheduler: add scheduler level for clusters scheduler: scan idle cpu in cluster before scanning the whole llc Jonathan Cameron (1): topology: Represent clusters of CPUs within a die Tim Chen (1): scheduler: Add cluster scheduler level for x86 Documentation/admin-guide/cputopology.rst | 26 +++++++++++-- arch/arm64/Kconfig | 7 ++++ arch/arm64/kernel/topology.c | 2 + arch/x86/Kconfig | 8 ++++ arch/x86/include/asm/smp.h | 7 ++++ arch/x86/include/asm/topology.h | 1 + arch/x86/kernel/cpu/cacheinfo.c | 1 + arch/x86/kernel/cpu/common.c | 3 ++ arch/x86/kernel/smpboot.c | 43 ++++++++++++++++++++- drivers/acpi/pptt.c | 63 +++++++++++++++++++++++++++++++ drivers/base/arch_topology.c | 15 ++++++++ drivers/base/topology.c | 10 +++++ include/linux/acpi.h | 5 +++ include/linux/arch_topology.h | 5 +++ include/linux/sched/cluster.h | 19 ++++++++++ include/linux/sched/topology.h | 7 ++++ include/linux/topology.h | 13 +++++++ include/trace/events/sched.h | 22 +++++++++++ kernel/sched/core.c | 20 ++++++++++ kernel/sched/fair.c | 36 +++++++++++++++++- kernel/sched/sched.h | 1 + kernel/sched/topology.c | 5 +++ 22 files changed, 313 insertions(+), 6 deletions(-) create mode 100644 include/linux/sched/cluster.h -- 1.8.3.1

5 14

Re: [PATCH] lib: bitmap: check the overflow of pagebuf
by Song Bao Hua (Barry Song) 21 Apr '21

21 Apr '21

> -----Original Message----- > From: tiantao (H) > Sent: Tuesday, April 20, 2021 11:49 PM > To: Song Bao Hua (Barry Song) <song.bao.hua(a)hisilicon.com>; Jonathan Cameron > <jonathan.cameron(a)huawei.com> > Cc: linuxarm(a)openeuler.org; tiantao (H) <tiantao6(a)hisilicon.com> > Subject: [PATCH] lib: bitmap: check the overflow of pagebuf > > Move the detection of pagebuf overflow from node_read_cpumap to the > more appropriate bitmap_print_to_pagebuf function. Other interfaces > call bitmap_print_to_pagebuf to print without pagebuf overflow due > to too many cpus. This is far from a good commit log to describe the background. > > Signed-off-by: Tian Tao <tiantao6(a)hisilicon.com> > --- > drivers/base/node.c | 3 --- > lib/bitmap.c | 3 +++ > 2 files changed, 3 insertions(+), 3 deletions(-) > > diff --git a/drivers/base/node.c b/drivers/base/node.c > index 2c36f61d..e24530c3 100644 > --- a/drivers/base/node.c > +++ b/drivers/base/node.c > @@ -33,9 +33,6 @@ static ssize_t node_read_cpumap(struct device *dev, bool list, > char *buf) > cpumask_var_t mask; > struct node *node_dev = to_node(dev); > > - /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */ > - BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1)); > - > if (!alloc_cpumask_var(&mask, GFP_KERNEL)) > return 0; > > diff --git a/lib/bitmap.c b/lib/bitmap.c > index 74ceb02..170fc48 100644 > --- a/lib/bitmap.c > +++ b/lib/bitmap.c > @@ -482,6 +482,9 @@ int bitmap_print_to_pagebuf(bool list, char *buf, const > unsigned long *maskp, > { > ptrdiff_t len = PAGE_SIZE - offset_in_page(buf); > > + /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */ > + BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1)); > + I don't think it is the right place. expecting cpumap_print_to_pagebuf. BTW, it is not 2008/04/07 now, so just remove it. > return list ? scnprintf(buf, len, "%*pbl\n", nmaskbits, maskp) : > scnprintf(buf, len, "%*pb\n", nmaskbits, maskp); > } > -- > 2.7.4 I would prefer we send two patches as a series "clarify and cleanup CPU and NUMA topology ABIs" with a cover letter with the below as 1/2. 2/2 would be the patch moving the place of cpu topology ABI doc. From b32c0c00a187d4fe4c49d54d30650b0cacb2c351 Mon Sep 17 00:00:00 2001 From: Tian Tao <tiantao6(a)hisilicon.com> Date: Wed, 21 Apr 2021 14:36:11 +1200 Subject: [PATCH 1/2] CPU, NUMA topology ABIs: clarify the overflow issue of sysfs pagebuf Both numa node and cpu use cpu bitmap like 3,ffffffff to expose hardware topology. When cpu number is large, the page buffer of sysfs will over- flow. This doesn't really happen nowadays as the maximum NR_CPUS is 8196 for X86_64 and 4096 for ARM64 since 8196 * 9 / 32 = 2305 which is still smaller than 4KB page size. So this BUILD_BUG_ON() is pretty much preventing future problems similar with Y2K. On the other hand, it should be more sensible to move the guard to common code which can protect both cpu and numa: /sys/devices/system/cpu/cpu0/topology/die_cpus etc. /sys/devices/system/node/node0/cpumap etc. Topology bitmaps shouldn't be larger than PAGE_SIZE as lstopo and numactl depend on them. But other ABIs exposing cpu lists are not really used by common applications, so this patch also marks those lists could be trimmed. Signed-off-by: Tian Tao <tiantao6(a)hisilicon.com> Signed-off-by: Barry Song <song.bao.hua(a)hisilicon.com> --- Documentation/ABI/stable/sysfs-devices-node | 5 ++++- Documentation/admin-guide/cputopology.rst | 15 +++++++++++++++ drivers/base/node.c | 3 --- include/linux/cpumask.h | 6 ++++++ 4 files changed, 25 insertions(+), 4 deletions(-) diff --git a/Documentation/ABI/stable/sysfs-devices-node b/Documentation/ABI/stable/sysfs-devices-node index 484fc04bcc25..9832a17b2b15 100644 --- a/Documentation/ABI/stable/sysfs-devices-node +++ b/Documentation/ABI/stable/sysfs-devices-node @@ -47,7 +47,10 @@ What: /sys/devices/system/node/nodeX/cpulist Date: October 2002 Contact: Linux Memory Management list <linux-mm(a)kvack.org> Description: - The CPUs associated to the node. + The CPUs associated to the node. The format is like 0-3, + 8-11, 12-13. The maximum size is PAGE_SIZE, so the tail + of the string will be trimmed while its size is larger + than PAGE_SIZE. What: /sys/devices/system/node/nodeX/meminfo Date: October 2002 diff --git a/Documentation/admin-guide/cputopology.rst b/Documentation/admin-guide/cputopology.rst index b90dafcc8237..8fac776a5ffa 100644 --- a/Documentation/admin-guide/cputopology.rst +++ b/Documentation/admin-guide/cputopology.rst @@ -44,6 +44,9 @@ core_cpus: core_cpus_list: human-readable list of CPUs within the same core. + The format is like 0-3, 8-11, 12-13. The maximum size is PAGE_SIZE, + so the tail of the string will be trimmed while its size is larger + than PAGE_SIZE. (deprecated name: "thread_siblings_list"); package_cpus: @@ -54,6 +57,9 @@ package_cpus: package_cpus_list: human-readable list of CPUs sharing the same physical_package_id. + The format is like 0-3, 8-11, 12-13. The maximum size is PAGE_SIZE, + so the tail of the string will be trimmed while its size is larger + than PAGE_SIZE. (deprecated name: "core_siblings_list") die_cpus: @@ -63,6 +69,9 @@ die_cpus: die_cpus_list: human-readable list of CPUs within the same die. + The format is like 0-3, 8-11, 12-13. The maximum size is PAGE_SIZE, + so the tail of the string will be trimmed while its size is larger + than PAGE_SIZE. book_siblings: @@ -73,6 +82,9 @@ book_siblings_list: human-readable list of cpuX's hardware threads within the same book_id. + The format is like 0-3, 8-11, 12-13. The maximum size is PAGE_SIZE, + so the tail of the string will be trimmed while its size is larger + than PAGE_SIZE. drawer_siblings: @@ -83,6 +95,9 @@ drawer_siblings_list: human-readable list of cpuX's hardware threads within the same drawer_id. + The format is like 0-3, 8-11, 12-13. The maximum size is PAGE_SIZE, + so the tail of the string will be trimmed while its size is larger + than PAGE_SIZE. Architecture-neutral, drivers/base/topology.c, exports these attributes. However, the book and drawer related sysfs files will only be created if diff --git a/drivers/base/node.c b/drivers/base/node.c index f449dbb2c746..50324d06bcd5 100644 --- a/drivers/base/node.c +++ b/drivers/base/node.c @@ -33,9 +33,6 @@ static ssize_t node_read_cpumap(struct device *dev, bool list, char *buf) cpumask_var_t mask; struct node *node_dev = to_node(dev); - /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */ - BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1)); - if (!alloc_cpumask_var(&mask, GFP_KERNEL)) return 0; diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h index 383684e30f12..81f145e0c742 100644 --- a/include/linux/cpumask.h +++ b/include/linux/cpumask.h @@ -12,6 +12,7 @@ #include <linux/bitmap.h> #include <linux/atomic.h> #include <linux/bug.h> +#include <asm/page.h> /* Don't assign or return these: may not be this big! */ typedef struct cpumask { DECLARE_BITMAP(bits, NR_CPUS); } cpumask_t; @@ -924,6 +925,11 @@ static inline const struct cpumask *get_cpu_mask(unsigned int cpu) static inline ssize_t cpumap_print_to_pagebuf(bool list, char *buf, const struct cpumask *mask) { + /* + * 32bits requires 9bytes: "ff,ffffffff", thus, too many CPUs will + * cause the overflow of sysfs pagebuf + */ + BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1)); return bitmap_print_to_pagebuf(list, buf, cpumask_bits(mask), nr_cpu_ids); } -- 2.25.1 Thanks Barry

1 0

Re: [RFC PATCH v5 4/4] scheduler: Add cluster scheduler level for x86
by Song Bao Hua (Barry Song) 21 Apr '21

21 Apr '21

> -----Original Message----- > From: Tim Chen [mailto:tim.c.chen@linux.intel.com] > Sent: Wednesday, April 21, 2021 6:32 AM > To: Song Bao Hua (Barry Song) <song.bao.hua(a)hisilicon.com>; > catalin.marinas(a)arm.com; will(a)kernel.org; rjw(a)rjwysocki.net; > vincent.guittot(a)linaro.org; bp(a)alien8.de; tglx(a)linutronix.de; > mingo(a)redhat.com; lenb(a)kernel.org; peterz(a)infradead.org; > dietmar.eggemann(a)arm.com; rostedt(a)goodmis.org; bsegall(a)google.com; > mgorman(a)suse.de > Cc: msys.mizuma(a)gmail.com; valentin.schneider(a)arm.com; > gregkh(a)linuxfoundation.org; Jonathan Cameron <jonathan.cameron(a)huawei.com>; > juri.lelli(a)redhat.com; mark.rutland(a)arm.com; sudeep.holla(a)arm.com; > aubrey.li(a)linux.intel.com; linux-arm-kernel(a)lists.infradead.org; > linux-kernel(a)vger.kernel.org; linux-acpi(a)vger.kernel.org; x86(a)kernel.org; > xuwei (O) <xuwei5(a)huawei.com>; Zengtao (B) <prime.zeng(a)hisilicon.com>; > guodong.xu(a)linaro.org; yangyicong <yangyicong(a)huawei.com>; Liguozhu (Kenneth) > <liguozhu(a)hisilicon.com>; linuxarm(a)openeuler.org; hpa(a)zytor.com > Subject: Re: [RFC PATCH v5 4/4] scheduler: Add cluster scheduler level for x86 > > > > On 3/23/21 4:21 PM, Song Bao Hua (Barry Song) wrote: > > >> > >> On 3/18/21 9:16 PM, Barry Song wrote: > >>> From: Tim Chen <tim.c.chen(a)linux.intel.com> > >>> > >>> There are x86 CPU architectures (e.g. Jacobsville) where L2 cahce > >>> is shared among a cluster of cores instead of being exclusive > >>> to one single core. > >>> > >>> To prevent oversubscription of L2 cache, load should be > >>> balanced between such L2 clusters, especially for tasks with > >>> no shared data. > >>> > >>> Also with cluster scheduling policy where tasks are woken up > >>> in the same L2 cluster, we will benefit from keeping tasks > >>> related to each other and likely sharing data in the same L2 > >>> cluster. > >>> > >>> Add CPU masks of CPUs sharing the L2 cache so we can build such > >>> L2 cluster scheduler domain. > >>> > >>> Signed-off-by: Tim Chen <tim.c.chen(a)linux.intel.com> > >>> Signed-off-by: Barry Song <song.bao.hua(a)hisilicon.com> > >> > >> > >> Barry, > >> > >> Can you also add this chunk to the patch. > >> Thanks. > > > > Sure, Tim, Thanks. I'll put that into patch 4/4 in v6. > > > > Barry, > > This chunk will also need to be added to return cluster id for x86. > Please add it in your next rev. Yes. Thanks. I'll put this in either RFC v7 or Patch v1. For spreading path, things are much easier, though packing path is quite tricky. But It seems RFC v6 has been quite close to what we want to achieve to pack related tasks by scanning cluster for tasks within same NUMA: https://lore.kernel.org/lkml/20210420001844.9116-1-song.bao.hua@hisilicon.c… If couples have been already in same LLC(numa), scanning clusters will gather them further. If they are running in different NUMA nodes, the original scanning LLC will move them to the same node, after that, scanning cluster might put them closer to each other. it seems it is kind of the two-level packing Dietmar has suggested. So perhaps we won't have RFC v7, I will probably send patch v1 afterwards. > > Thanks. > > Tim > > --- > > diff --git a/arch/x86/include/asm/topology.h > b/arch/x86/include/asm/topology.h > index 800fa48c9fcd..2548d824f103 100644 > --- a/arch/x86/include/asm/topology.h > +++ b/arch/x86/include/asm/topology.h > @@ -109,6 +109,7 @@ extern const struct cpumask *cpu_clustergroup_mask(int cpu); > #define topology_physical_package_id(cpu) (cpu_data(cpu).phys_proc_id) > #define topology_logical_die_id(cpu) (cpu_data(cpu).logical_die_id) > #define topology_die_id(cpu) (cpu_data(cpu).cpu_die_id) > +#define topology_cluster_id(cpu) (per_cpu(cpu_l2c_id, cpu)) > #define topology_core_id(cpu) (cpu_data(cpu).cpu_core_id) > > extern unsigned int __max_die_per_package; Thanks Barry

1 0

[PATCH net v4 0/2] fix packet stuck problem for lockless qdisc
by Yunsheng Lin 21 Apr '21

21 Apr '21

This patchset fixes the packet stuck problem mentioned in [1]. Patch 1: Add STATE_MISSED flag to fix packet stuck problem. Patch 2: Fix a tx_action rescheduling problem after STATE_MISSED flag is added in patch 1. V4: Change STATE_NEED_RESCHEDULE to STATE_MISSED and add patch 2. [1]. https://lkml.org/lkml/2019/10/9/42 Yunsheng Lin (2): net: sched: fix packet stuck problem for lockless qdisc net: sched: fix endless tx action reschedule during deactivation include/net/pkt_sched.h | 7 +------ include/net/sch_generic.h | 37 ++++++++++++++++++++++++++++++++++++- net/core/dev.c | 26 ++++++++++++++++++++++---- net/sched/sch_generic.c | 16 +++++++++++++++- 4 files changed, 74 insertions(+), 12 deletions(-) -- 2.7.4

2 6

Re: [RFC PATCH 2/3] vfio/hisilicon: register the driver to vfio
by Jason Gunthorpe 20 Apr '21

20 Apr '21

On Tue, Apr 20, 2021 at 09:28:46PM +0800, liulongfang wrote: > >> So, I still don't understand what the security risk you are talking about is, > >> and what do you think the security design should look like? > >> Can you elaborate on it? > > > > Each security domain must have its own PCI BDF. > > > The basic unit to perform the live migration function is the VF, and the basic > function of the VF is the business function of the device. If the live migration > function and the business function are completely separated, and they are bound > to their respective VFs, it will result in the ability to execute the business > in the guest A functional VF cannot perform live migration, and a VF with a live > migration function cannot perform business functions in the guest. > > In fact, the scenario that requires live migration is to migrate its business > functions to the VFs of other hosts when the VF's business functions are heavily > loaded. > This usage scenario itself requires that the VF needs to have these two functions > at the same time. Other vendors are managing, it is not an unsolvable problem. I think these patches can't advance in any form without somehow addressing the isolation issue. Jason

1 0

[PATCH net v3] net: sched: fix packet stuck problem for lockless qdisc
by Yunsheng Lin 20 Apr '21

20 Apr '21

Lockless qdisc has below concurrent problem: cpu0 cpu1 . . q->enqueue . . . qdisc_run_begin() . . . dequeue_skb() . . . sch_direct_xmit() . . . . q->enqueue . qdisc_run_begin() . return and do nothing . . qdisc_run_end() . cpu1 enqueue a skb without calling __qdisc_run() because cpu0 has not released the lock yet and spin_trylock() return false for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb enqueued by cpu1 when calling dequeue_skb() because cpu1 may enqueue the skb after cpu0 calling dequeue_skb() and before cpu0 calling qdisc_run_end(). Lockless qdisc has below another concurrent problem when tx_action is involved: cpu0(serving tx_action) cpu1 cpu2 . . . . q->enqueue . . qdisc_run_begin() . . dequeue_skb() . . . q->enqueue . . . . sch_direct_xmit() . . . qdisc_run_begin() . . return and do nothing . . . clear __QDISC_STATE_SCHED . . qdisc_run_begin() . . return and do nothing . . . . . . qdisc_run_end() . This patch fixes the above data race by: 1. Get the flag before doing spin_trylock(). 2. If the first spin_trylock() return false and the flag is not set before the first spin_trylock(), Set the flag and retry another spin_trylock() in case other CPU may not see the new flag after it releases the lock. 3. reschedule if the flags is set after the lock is released at the end of qdisc_run_end(). For tx_action case, the flags is also set when cpu1 is at the end if qdisc_run_end(), so tx_action will be rescheduled again to dequeue the skb enqueued by cpu2. Only clear the flag before retrying a dequeuing when dequeuing returns NULL in order to reduce the overhead of the above double spin_trylock() and __netif_schedule() calling. The performance impact of this patch, tested using pktgen and dummy netdev with pfifo_fast qdisc attached: threads without+this_patch with+this_patch delta 1 2.61Mpps 2.60Mpps -0.3% 2 3.97Mpps 3.82Mpps -3.7% 4 5.62Mpps 5.59Mpps -0.5% 8 2.78Mpps 2.77Mpps -0.3% 16 2.22Mpps 2.22Mpps -0.0% Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking") Signed-off-by: Yunsheng Lin <linyunsheng(a)huawei.com> --- V3: fix a compile error and a few comment typo, remove the __QDISC_STATE_DEACTIVATED checking, and update the performance data. V2: Avoid the overhead of fixing the data race as much as possible. --- include/net/sch_generic.h | 38 +++++++++++++++++++++++++++++++++++++- net/sched/sch_generic.c | 12 ++++++++++++ 2 files changed, 49 insertions(+), 1 deletion(-) diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..e3f46eb 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -36,6 +36,7 @@ struct qdisc_rate_table { enum qdisc_state_t { __QDISC_STATE_SCHED, __QDISC_STATE_DEACTIVATED, + __QDISC_STATE_NEED_RESCHEDULE, }; struct qdisc_size_table { @@ -159,8 +160,38 @@ static inline bool qdisc_is_empty(const struct Qdisc *qdisc) static inline bool qdisc_run_begin(struct Qdisc *qdisc) { if (qdisc->flags & TCQ_F_NOLOCK) { + bool dont_retry = test_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state); + + if (spin_trylock(&qdisc->seqlock)) + goto nolock_empty; + + /* If the flag is set before doing the spin_trylock() and + * the above spin_trylock() return false, it means other cpu + * holding the lock will do dequeuing for us, or it wil see + * the flag set after releasing lock and reschedule the + * net_tx_action() to do the dequeuing. + */ + if (dont_retry) + return false; + + /* We could do set_bit() before the first spin_trylock(), + * and avoid doing second spin_trylock() completely, then + * we could have multi cpus doing the set_bit(). Here use + * dont_retry to avoid doing the set_bit() and the second + * spin_trylock(), which has 5% performance improvement than + * doing the set_bit() before the first spin_trylock(). + */ + set_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state); + + /* Retry again in case other CPU may not see the new flag + * after it releases the lock at the end of qdisc_run_end(). + */ if (!spin_trylock(&qdisc->seqlock)) return false; + +nolock_empty: WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; @@ -176,8 +207,13 @@ static inline bool qdisc_run_begin(struct Qdisc *qdisc) static inline void qdisc_run_end(struct Qdisc *qdisc) { write_seqcount_end(&qdisc->running); - if (qdisc->flags & TCQ_F_NOLOCK) + if (qdisc->flags & TCQ_F_NOLOCK) { spin_unlock(&qdisc->seqlock); + + if (unlikely(test_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state))) + __netif_schedule(qdisc); + } } static inline bool qdisc_may_bulk(const struct Qdisc *qdisc) diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 44991ea..4953430 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -640,8 +640,10 @@ static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc) { struct pfifo_fast_priv *priv = qdisc_priv(qdisc); struct sk_buff *skb = NULL; + bool need_retry = true; int band; +retry: for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) { struct skb_array *q = band2list(priv, band); @@ -652,6 +654,16 @@ static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc) } if (likely(skb)) { qdisc_update_stats_at_dequeue(qdisc, skb); + } else if (need_retry && + test_and_clear_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state)) { + /* do another enqueuing after clearing the flag to + * avoid calling __netif_schedule(). + */ + smp_mb__after_atomic(); + need_retry = false; + + goto retry; } else { WRITE_ONCE(qdisc->empty, true); } -- 2.7.4

2 6

2024

2023

2022

2021

2020

Linuxarm