Re: [PATCH v10 00/15] Linux RISC-V AIA Support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:

> On Mon, Oct 23, 2023 at 12:32 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>>
>> Björn Töpel <bjorn@xxxxxxxxxx> writes:
>>
>> > Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>> >
>> >> On Fri, Oct 20, 2023 at 10:07 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>> >>>
>> >>> Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>> >>>
>> >>> > On Fri, Oct 20, 2023 at 8:10 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>> >>> >>
>> >>> >> Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>> >>> >>
>> >>> >> > On Fri, Oct 20, 2023 at 2:17 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>> >>> >> >>
>> >>> >> >> Thanks for the quick reply!
>> >>> >> >>
>> >>> >> >> Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>> >>> >> >>
>> >>> >> >> > On Thu, Oct 19, 2023 at 7:13 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>> >>> >> >> >>
>> >>> >> >> >> Hi Anup,
>> >>> >> >> >>
>> >>> >> >> >> Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>> >>> >> >> >>
>> >>> >> >> >> > The RISC-V AIA specification is ratified as-per the RISC-V international
>> >>> >> >> >> > process. The latest ratified AIA specifcation can be found at:
>> >>> >> >> >> > https://github.com/riscv/riscv-aia/releases/download/1.0/riscv-interrupts-1.0.pdf
>> >>> >> >> >> >
>> >>> >> >> >> > At a high-level, the AIA specification adds three things:
>> >>> >> >> >> > 1) AIA CSRs
>> >>> >> >> >> >    - Improved local interrupt support
>> >>> >> >> >> > 2) Incoming Message Signaled Interrupt Controller (IMSIC)
>> >>> >> >> >> >    - Per-HART MSI controller
>> >>> >> >> >> >    - Support MSI virtualization
>> >>> >> >> >> >    - Support IPI along with virtualization
>> >>> >> >> >> > 3) Advanced Platform-Level Interrupt Controller (APLIC)
>> >>> >> >> >> >    - Wired interrupt controller
>> >>> >> >> >> >    - In MSI-mode, converts wired interrupt into MSIs (i.e. MSI generator)
>> >>> >> >> >> >    - In Direct-mode, injects external interrupts directly into HARTs
>> >>> >> >> >>
>> >>> >> >> >> Thanks for working on the AIA support! I had a look at the series, and
>> >>> >> >> >> have some concerns about interrupt ID abstraction.
>> >>> >> >> >>
>> >>> >> >> >> A bit of background, for readers not familiar with the AIA details.
>> >>> >> >> >>
>> >>> >> >> >> IMSIC allows for 2047 unique MSI ("msi-irq") sources per hart, and
>> >>> >> >> >> each MSI is dedicated to a certain hart. The series takes the approach
>> >>> >> >> >> to say that there are, e.g., 2047 interrupts ("lnx-irq") globally.
>> >>> >> >> >> Each lnx-irq consists of #harts * msi-irq -- a slice -- and in the
>> >>> >> >> >> slice only *one* msi-irq is acutally used.
>> >>> >> >> >>
>> >>> >> >> >> This scheme makes affinity changes more robust, because the interrupt
>> >>> >> >> >> sources on "other" harts are pre-allocated. On the other hand it
>> >>> >> >> >> requires to propagate irq masking to other harts via IPIs (this is
>> >>> >> >> >> mostly done up setup/tear down). It's also wasteful, because msi-irqs
>> >>> >> >> >> are hogged, and cannot be used.
>> >>> >> >> >>
>> >>> >> >> >> Contemporary storage/networking drivers usually uses queues per core
>> >>> >> >> >> (or a sub-set of cores). The current scheme wastes a lot of msi-irqs.
>> >>> >> >> >> If we instead used a scheme where "msi-irq == lnx-irq", instead of
>> >>> >> >> >> "lnq-irq = {hart 0;msi-irq x , ... hart N;msi-irq x}", there would be
>> >>> >> >> >> a lot MSIs for other users. 1-1 vs 1-N. E.g., if a storage device
>> >>> >> >> >> would like to use 5 queues (5 cores) on a 128 core system, the current
>> >>> >> >> >> scheme would consume 5 * 128 MSIs, instead of just 5.
>> >>> >> >> >>
>> >>> >> >> >> On the plus side:
>> >>> >> >> >> * Changing interrupts affinity will never fail, because the interrupts
>> >>> >> >> >>   on each hart is pre-allocated.
>> >>> >> >> >>
>> >>> >> >> >> On the negative side:
>> >>> >> >> >> * Wasteful interrupt usage, and a system can potientially "run out" of
>> >>> >> >> >>   interrupts. Especially for many core systems.
>> >>> >> >> >> * Interrupt masking need to proagate to harts via IPIs (there's no
>> >>> >> >> >>   broadcast csr in IMSIC), and a more complex locking scheme IMSIC
>> >>> >> >> >>
>> >>> >> >> >> Summary:
>> >>> >> >> >> The current series caps the number of global interrupts to maximum
>> >>> >> >> >> 2047 MSIs for all cores (whole system). A better scheme, IMO, would be
>> >>> >> >> >> to expose 2047 * #harts unique MSIs.
>> >>> >> >> >>
>> >>> >> >> >> I think this could simplify/remove(?) the locking as well.
>> >>> >> >> >
>> >>> >> >> > Exposing 2047 * #harts unique MSIs has multiple issues:
>> >>> >> >> > 1) The irq_set_affinity() does not work for MSIs because each
>> >>> >> >> >      IRQ is not tied to a particular HART. This means we can't
>> >>> >> >> >      balance the IRQ processing load among HARTs.
>> >>> >> >>
>> >>> >> >> Yes, you can balance. In your code, each *active* MSI is still
>> >>> >> >> bound/active to a specific hard together with the affinity mask. In an
>> >>> >> >> 1-1 model you would still need to track the affinity mask, but the
>> >>> >> >> irq_set_affinity() would be different. It would try to allocate a new
>> >>> >> >> MSI from the target CPU, and then switch to having that MSI active.
>> >>> >> >>
>> >>> >> >> That's what x86 does AFAIU, which is also constrained by the # of
>> >>> >> >> available MSIs.
>> >>> >> >>
>> >>> >> >> The downside, as I pointed out, is that the set affinity action can
>> >>> >> >> fail for a certain target CPU.
>> >>> >> >
>> >>> >> > Yes, irq_set_affinity() can fail for the suggested approach plus for
>> >>> >> > RISC-V AIA, one HART does not have access to other HARTs
>> >>> >> > MSI enable/disable bits so the approach will also involve IPI.
>> >>> >>
>> >>> >> Correct, but the current series does a broadcast to all cores, where the
>> >>> >> 1-1 approach is at most an IPI to a single core.
>> >>> >>
>> >>> >> 128+c machines are getting more common, and you have devices that you
>> >>> >> bring up/down on a per-core basis. Broadcasting IPIs to all cores, when
>> >>> >> dealing with a per-core activity is a pretty noisy neighbor.
>> >>> >
>> >>> > Broadcast IPI in the current approach is only done upon MSI mask/unmask
>> >>> > operation. It is not done upon set_affinity() of interrupt handling.
>> >>>
>> >>> I'm aware. We're on the same page here.
>> >>>
>> >>> >>
>> >>> >> This could be fixed in the existing 1-n approach, by not require to sync
>> >>> >> the cores that are not handling the MSI in question. "Lazy disable"
>> >>> >
>> >>> > Incorrect. The approach you are suggesting involves an IPI upon every
>> >>> > irq_set_affinity(). This is because a HART can only enable it's own
>> >>> > MSI ID so when an IRQ is moved to from HART A to HART B with
>> >>> > a different ID X on HART B then we will need an IPI in irq_set_affinit()
>> >>> > to enable ID X on HART B.
>> >>>
>> >>> Yes, the 1-1 approach will require an IPI to one target cpu on affinity
>> >>> changes, and similar on mask/unmask.
>> >>>
>> >>> The 1-n approach, require no-IPI on affinity changes (nice!), but IPI
>> >>> broadcast to all cores on mask/unmask (not so nice).
>> >>>
>> >>> >> >> My concern is interrupts become a scarce resource with this
>> >>> >> >> implementation, but maybe my view is incorrect. I've seen bare-metal
>> >>> >> >> x86 systems (no VMs) with ~200 cores, and ~2000 interrupts, but maybe
>> >>> >> >> that is considered "a lot of interrupts".
>> >>> >> >>
>> >>> >> >> As long as we don't get into scenarios where we're running out of
>> >>> >> >> interrupts, due to the software design.
>> >>> >> >>
>> >>> >> >
>> >>> >> > The current approach is simpler and ensures irq_set_affinity
>> >>> >> > always works. The limit of max 2047 IDs is sufficient for many
>> >>> >> > systems (if not all).
>> >>> >>
>> >>> >> Let me give you another view. On a 128c system each core has ~16 unique
>> >>> >> interrupts for disposal. E.g. the Intel E800 NIC has more than 2048
>> >>> >> network queue pairs for each PF.
>> >>> >
>> >>> > Clearly, this example is a hypothetical and represents a poorly
>> >>> > designed platform.
>> >>> >
>> >>> > Having just 16 IDs per-Core is a very poor design choice. In fact, the
>> >>> > Server SoC spec mandates a minimum 255 IDs.
>> >>>
>> >>> You are misreading. A 128c system with 2047 MSIs per-core, will only
>> >>> have 16 *per-core unique* (2047/128) interrupts with the current series.
>> >>>
>> >>> I'm not saying that each IMSIC has 16 IDs, I'm saying that in a 128c
>> >>> system with the maximum amount of MSIs possible in the spec, you'll end
>> >>> up with 16 *unique* interrupts per core.
>> >>
>> >> -ENOPARSE
>> >>
>> >> I don't see how this applies to the current approach because we treat
>> >> MSI ID space as global across cores so if a system has 2047 MSIs
>> >> per-core then we have 2047 MSIs across all cores.
>> >
>> > Ok, I'll try again! :-)
>> >
>> > Let's assume that each core in the 128c system has some per-core
>> > resources, say a two NIC queue pairs, and a storage queue pair. This
>> > will consume, e.g., 2*2 + 2 (6) MSI sources from the global namespace.
>> >
>> > If each core does this it'll be 6*128 MSI sources of the global
>> > namespace.
>> >
>> > The maximum number of "privates" MSI sources a core can utilize is 16.
>> >
>> > I'm trying (it's does seem to go that well ;-)) to point out that it's
>> > only 16 unique sources per core. For, say, a 256 core system it would be
>> > 8. 2047 MSI sources in a system is not much.
>> >
>> > Say that I want to spin up 24 NIC queues with one MSI each on each core
>> > on my 128c system. That's not possible with this series, while with an
>> > 1-1 system it wouldn't be an issue.
>> >
>> > Clearer, or still weird?
>> >
>> >>
>> >>>
>> >>> > Regarding NICs which support a large number of queues, the driver
>> >>> > will typically enable only one queue per-core and set the affinity to
>> >>> > separate cores. We have user-space data plane applications based
>> >>> > on DPDK which are capable of using a large number of NIC queues
>> >>> > but these applications are polling based and don't use MSIs.
>> >>>
>> >>> That's one sample point, and clearly not the only one. There are *many*
>> >>> different usage models. Just because you *assign* MSI, doesn't mean they
>> >>> are firing all the time.
>> >>>
>> >>> I can show you a couple of networking setups where this is clearly not
>> >>> enough. Each core has a large number of QoS queues, and each queue would
>> >>> very much like to have a dedicated MSI.
>> >>>
>> >>> >> > When we encounter a system requiring a large number of MSIs,
>> >>> >> > we can either:
>> >>> >> > 1) Extend the AIA spec to support greater than 2047 IDs
>> >>> >> > 2) Re-think the approach in the IMSIC driver
>> >>> >> >
>> >>> >> > The choice between #1 and #2 above depends on the
>> >>> >> > guarantees we want for irq_set_affinity().
>> >>> >>
>> >>> >> The irq_set_affinity() behavior is better with this series, but I think
>> >>> >> the other downsides: number of available interrupt sources, and IPI
>> >>> >> broadcast are worse.
>> >>> >
>> >>> > The IPI overhead in the approach you are suggesting will be
>> >>> > even bad compared to the IPI overhead of the current approach
>> >>> > because we will end-up doing IPI upon every irq_set_affinity()
>> >>> > in the suggested approach compared to doing IPI upon every
>> >>> > mask/unmask in the current approach.
>> >>>
>> >>> Again, very workload dependent.
>> >>>
>> >>> This series does IPI broadcast on masking/unmasking, which means that
>> >>> cores that don't care get interrupted because, say, a network queue-pair
>> >>> is setup on another core.
>> >>>
>> >>> Some workloads never change the irq affinity.
>> >>
>> >> There are various events which irq affinity such as irq balance,
>> >> CPU hotplug, system suspend, etc.
>> >>
>> >> Also, the 1-1 approach does IPI upon set_affinity, mask and
>> >> unmask whereas the 1-n approach does IPI only upon mask
>> >> and unmask.
>> >
>> > An important distinction; When you say IPI on mask/unmask it is a
>> > broadcast IPI to *all* cores, which is pretty instrusive.
>> >
>> > The 1-1 variant does an IPI to a *one* target core.
>> >
>> >>> I'm just pointing out that there are pro/cons with both variants.
>> >>>
>> >>> > The biggest advantage of the current approach is a reliable
>> >>> > irq_set_affinity() which is a very valuable thing to have.
>> >>>
>> >>> ...and I'm arguing that we're paying a big price for that.
>> >>>
>> >>> > ARM systems easily support a large number of LPIs per-core.
>> >>> > For example, GIC-700 supports 56000 LPIs per-core.
>> >>> > (Refer, https://developer.arm.com/documentation/101516/0300/About-the-GIC-700/Features)
>> >>>
>> >>> Yeah, but this is not the GIC. This is something that looks more like
>> >>> the x86 world. We'll be stuck with a lot of implementations with AIA 1.0
>> >>> spec, and many cores.
>> >>
>> >> Well, RISC-V AIA is neigher ARM GIG not x86 APIC. All I am saying
>> >> is that there are systems with large number per-core interrupt IDs
>> >> for handling MSIs.
>> >
>> > Yes, and while that is nice, it's not what IMSIC is.
>>
>> Some follow-ups, after thinking more about it more over the weekend.
>>
>> * Do one really need an IPI for irq_set_affinity() for the 1-1 model?
>>   Why touch the enable/disable bits when moving interrupts?
>
> In the 1-1 model, the ID on the current HART and target HART upon
> irq_set_affinity will be different so we can't leave the unused ID on
> current HART enabled because it can lead to spurious interrupts
> when the ID on current HART is re-used for some other device.

Hmm, is this really an actual problem, or a theoretical one? The
implementation need to track what's in-use, so can we ever get into this
situation?

Somewhat related; I had a similar question for imsic_pci_{un,}mask_irq()
-- why not only do the the default mask operation (only
pci_msi_{un,}mask_irq()), but instead propagate to the IMSIC
mask/unmask?

> There is also a possibility of receiving an interrupt while the ID was
> moved to a new target HART in-which case we have to detect and
> re-trigger interrupt on the new target HART. In fact, x86 APLIC does
> an IPI to take care of this case.

This case I get, and the implementation can track that both are in use.
It's the spurious one that I'm dubious of (don't get).

>>
>> * In my book the IMSIC looks very much like the x86 LAPIC, which also
>>   has few interrupts (IMSIC <2048, LAPIC 256). The IRQ matrix allocator
>>   [1], and a scheme similar to LAPIC [2] would be a good fit. This is
>>   the 1-1 model, but more sophisticated than what I've been describing
>>   (e.g. properly handling mangaged/regular irqs). As a bonus we would
>>   get the IRQ matrix debugfs/tracepoint support.
>>
>
> Yes, I have been evaluating the 1-1 model for the past few days. I also
> have a working implementation with a simple per-CPU bitmap based
> allocator which handles both legacy MSI (block of 1,2,4,8,16, or 32 IDs)
> and MSI-X.
>
> The irq matrix allocator needs to be improved for handling legacy MSI
> so initially I will post a v11 series which works for me and converging
> with irq matrix allocator can be future work.

What's missing/needs to be improved for legacy MSI (legacy MSI ==
!MSI-X, right?) in the matrix allocator?


Björn





[Index of Archives]     [Device Tree Compilter]     [Device Tree Spec]     [Linux Driver Backports]     [Video for Linux]     [Linux USB Devel]     [Linux PCI Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Yosemite Backpacking]


  Powered by Linux