| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CAMOZA0KPPyCFO4XYKhRfrnpMV1CiOkOd0dX41yottKp0p3nYww@mail.gmail.com>
Date: Fri, 21 Nov 2025 15:33:58 +0100
From: Luigi Rizzo <lrizzo@...gle.com>
To: Thomas Gleixner <tglx@...utronix.de>
Cc: Marc Zyngier <maz@...nel.org>, Luigi Rizzo <rizzo.unipi@...il.com>,
Paolo Abeni <pabeni@...hat.com>, Andrew Morton <akpm@...ux-foundation.org>,
Sean Christopherson <seanjc@...gle.com>, linux-kernel@...r.kernel.org,
linux-arch@...r.kernel.org, Bjorn Helgaas <bhelgaas@...gle.com>,
Willem de Bruijn <willemb@...gle.com>
Subject: Re: [PATCH v2 3/8] genirq: soft_moderation: implement fixed moderation
On Fri, Nov 21, 2025 at 11:58 AM Luigi Rizzo <lrizzo@...gle.com> wrote:
>
> On Wed, Nov 19, 2025 at 3:43 PM Thomas Gleixner <tglx@...utronix.de> wrote:
> >
> ...
> >
> > ...
> > That said, I'm happy to answer _informed_ questions about semantics,
> > rules and requirements if they are not obvious from the code.
Some followup on the cost of mask_irq() which is rarely called without
moderation, but is hit hard with moderation:
For msix devices, mask_irq() eventually calls the following function
static inline void pci_msix_mask(struct msi_desc *desc)
{
desc->pci.msix_ctrl |= PCI_MSIX_ENTRY_CTRL_MASKBIT;
pci_msix_write_vector_ctrl(desc, desc->pci.msix_ctrl); //
eventually, writel()
/* Flush write to device */
readl(desc->pci.mask_base);
}
The final readl() is very expensive (serialized after the write, and takes
a full round trip through a busy PCIe bus). Its real purpose is to make
sure the function returns only after the hardware has seen the mask.
Except for a few critical places (setup/shutdown, others?) it does not
seem necessary to wait, because the irq layer knows how to drop stray
interrupts.
Here is an experiment removing the readl() where it is not critical,
during an experiment with busy NVME workload (~1M intr/s total)
The average goes from 1800ns down to 50ns, the long tails almost
completely disappear.
Data is collected with kstats https://lwn.net/Articles/813303/
Distribution of mask_irq() times in nanoseconds collected with kstats
over a 5s period, heavy workload.
F=/sys/kernel/debug/kstats/mask_irq
echo reset > $F; sleep 5; grep CPUS $ | awk '{t+=$6*$8;n+=$6; print}
END { print "avg ", t/n, t*1e-9}'
-------------------------------------------------------
ORIGINAL CODE
count 945 avg 680 p 0.000195 n 945
count 1863 avg 742 p 0.000580 n 2808
count 2941 avg 805 p 0.001188 n 5749
count 6058 avg 866 p 0.002441 n 11807
count 10090 avg 930 p 0.004527 n 21897
count 17273 avg 994 p 0.008099 n 39170
count 88252 avg 1099 p 0.026348 n 127422
count 227984 avg 1223 p 0.073490 n 355406
count 405964 avg 1348 p 0.157434 n 761370
count 530397 avg 1474 p 0.267109 n 1291767
count 606284 avg 1601 p 0.392475 n 1898051
count 658027 avg 1728 p 0.528541 n 2556078
count 643644 avg 1854 p 0.661632 n 3199722
count 515443 avg 1980 p 0.768215 n 3715165
count 546976 avg 2156 p 0.881317 n 4262141
count 243037 avg 2415 p 0.931572 n 4505178
count 125700 avg 2674 p 0.957564 n 4630878
count 73707 avg 2934 p 0.972805 n 4704585
count 48124 avg 3190 p 0.982756 n 4752709
count 30397 avg 3446 p 0.989041 n 4783106
count 19108 avg 3700 p 0.992993 n 4802214
count 11228 avg 3956 p 0.995314 n 4813442
count 11484 avg 4316 p 0.997689 n 4824926
count 5447 avg 4835 p 0.998815 n 4830373
count 2699 avg 5341 p 0.999373 n 4833072
count 1343 avg 5855 p 0.999651 n 4834415
count 668 avg 6369 p 0.999789 n 4835083
count 382 avg 6880 p 0.999868 n 4835465
count 207 avg 7406 p 0.999911 n 4835672
count 149 avg 7918 p 0.999942 n 4835821
count 159 avg 8671 p 0.999975 n 4835980
count 61 avg 9574 p 0.999987 n 4836041
count 22 avg 10758 p 0.999992 n 4836063
count 15 avg 11925 p 0.999995 n 4836078
count 9 avg 12750 p 0.999997 n 4836087
count 6 avg 13739 p 0.999998 n 4836093
count 2 avg 14737 p 0.999998 n 4836095
count 1 avg 15699 p 0.999999 n 4836096
count 3 avg 17951 p 0.999999 n 4836099
count 1 avg 18680 p 1.000000 n 4836100
avg 1833.11 8.86512
-------------------------------------------------------
AFTER REMOVING READL()
count 45 avg 13 p 0.000009 n 45
count 5792 avg 14 p 0.001204 n 5837
count 36628 avg 15 p 0.008761 n 42465
count 108293 avg 17 p 0.031103 n 150758
count 529586 avg 18 p 0.140365 n 680344
count 352816 avg 21 p 0.213156 n 1033160
count 1082719 avg 22 p 0.436538 n 2115879
count 480375 avg 24 p 0.535646 n 2596254
count 628931 avg 26 p 0.665404 n 3225185
count 275178 avg 28 p 0.722178 n 3500363
count 158508 avg 30 p 0.754880 n 3658871
count 100066 avg 33 p 0.775525 n 3758937
count 42338 avg 38 p 0.784260 n 3801275
count 75427 avg 41 p 0.799822 n 3876702
count 43099 avg 45 p 0.808714 n 3919801
count 30181 avg 49 p 0.814941 n 3949982
count 3757 avg 53 p 0.815716 n 3953739
count 2104 avg 57 p 0.816150 n 3955843
count 3520 avg 62 p 0.816876 n 3959363
count 28199 avg 69 p 0.822694 n 3987562
count 72163 avg 76 p 0.837583 n 4059725
count 109582 avg 83 p 0.860191 n 4169307
count 35839 avg 90 p 0.867585 n 4205146
count 6622 avg 100 p 0.868951 n 4211768
count 19393 avg 109 p 0.872952 n 4231161
count 51844 avg 116 p 0.883649 n 4283005
count 94042 avg 122 p 0.903051 n 4377047
count 28129 avg 133 p 0.908854 n 4405176
count 91010 avg 151 p 0.927631 n 4496186
count 73043 avg 164 p 0.942701 n 4569229
count 20705 avg 185 p 0.946973 n 4589934
count 59477 avg 199 p 0.959244 n 4649411
count 10842 avg 212 p 0.961481 n 4660253
count 16374 avg 233 p 0.964859 n 4676627
count 35394 avg 242 p 0.972161 n 4712021
count 36254 avg 277 p 0.979641 n 4748275
count 16982 avg 304 p 0.983145 n 4765257
count 20411 avg 327 p 0.987356 n 4785668
count 14352 avg 362 p 0.990317 n 4800020
count 10931 avg 399 p 0.992572 n 4810951
count 8384 avg 436 p 0.994302 n 4819335
count 3184 avg 467 p 0.994959 n 4822519
count 4923 avg 487 p 0.995974 n 4827442
count 7341 avg 538 p 0.997489 n 4834783
count 2957 avg 606 p 0.998099 n 4837740
count 3688 avg 666 p 0.998860 n 4841428
count 1807 avg 732 p 0.999233 n 4843235
count 813 avg 800 p 0.999400 n 4844048
count 699 avg 856 p 0.999545 n 4844747
count 293 avg 927 p 0.999605 n 4845040
count 203 avg 985 p 0.999647 n 4845243
count 390 avg 1086 p 0.999727 n 4845633
count 273 avg 1213 p 0.999784 n 4845906
count 162 avg 1336 p 0.999817 n 4846068
count 89 avg 1465 p 0.999835 n 4846157
count 50 avg 1600 p 0.999846 n 4846207
count 33 avg 1727 p 0.999853 n 4846240
count 26 avg 1850 p 0.999858 n 4846266
count 38 avg 1993 p 0.999866 n 4846304
count 89 avg 2185 p 0.999884 n 4846393
count 132 avg 2444 p 0.999911 n 4846525
count 134 avg 2680 p 0.999939 n 4846659
count 92 avg 2932 p 0.999958 n 4846751
count 65 avg 3182 p 0.999971 n 4846816
count 42 avg 3448 p 0.999980 n 4846858
count 29 avg 3713 p 0.999986 n 4846887
count 11 avg 3934 p 0.999988 n 4846898
count 18 avg 4305 p 0.999992 n 4846916
count 10 avg 4784 p 0.999994 n 4846926
count 7 avg 5264 p 0.999996 n 4846933
count 5 avg 5957 p 0.999997 n 4846938
count 1 avg 6603 p 0.999997 n 4846939
count 1 avg 7098 p 0.999997 n 4846940
count 3 avg 7433 p 0.999998 n 4846943
count 1 avg 7892 p 0.999998 n 4846944
count 2 avg 8955 p 0.999998 n 4846946
count 3 avg 9424 p 0.999999 n 4846949
count 1 avg 11194 p 0.999999 n 4846950
count 2 avg 11619 p 1.000000 n 4846952
avg 51.3519 0.2489
-------------------------------------------------------
cheers
Luigi
Powered by blists - more mailing lists