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

[Björn Töpel]

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.


Now, back to the weekend for real! ;-) (https://xkcd.com/386/)
Björn