From: Yong-Xuan Wang <yongxuan.wang@sifive.com>
To: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
Cc: qemu-devel@nongnu.org, qemu-riscv@nongnu.org,
greentime.hu@sifive.com, vincent.chen@sifive.com,
frank.chang@sifive.com, jim.shu@sifive.com,
Palmer Dabbelt <palmer@dabbelt.com>,
Alistair Francis <alistair.francis@wdc.com>,
Weiwei Li <liwei1518@gmail.com>,
Liu Zhiwei <zhiwei_liu@linux.alibaba.com>
Subject: Re: [PATCH 1/4] hw/riscv/virt: KVM AIA refinement
Date: Wed, 19 Feb 2025 19:25:08 +0800 [thread overview]
Message-ID: <CAMWQL2j32880zSM2ZesnAYReJiU2BkN7dwfGQDjAbde5t3uLUQ@mail.gmail.com> (raw)
In-Reply-To: <37cbfff3-9571-47e5-b955-41a85c7ef730@ventanamicro.com>
Hi Daniel,
On Tue, Feb 18, 2025 at 3:24 AM Daniel Henrique Barboza
<dbarboza@ventanamicro.com> wrote:
>
>
>
> On 2/17/25 5:19 AM, Yong-Xuan Wang wrote:
> > KVM AIA is only needed to be set when the virt machine use the AIA MSI.
> > So we can move the KVM AIA configuration into virt_create_aia() to reduce
> > the condition checking.
> >
> > Signed-off-by: Yong-Xuan Wang <yongxuan.wang@sifive.com>
> > ---
>
> Unfortunately this doesn't work.
>
> The reason is that kvm_riscv_aia_create(), as it is now, is called only once
> during virt_machine_init() and it's already handling initialization for each socket:
>
>
> for (socket = 0; socket < socket_count; socket++) {
> socket_imsic_base = imsic_base + socket * (1U << group_shift);
> hart_count = riscv_socket_hart_count(machine, socket);
> base_hart = riscv_socket_first_hartid(machine, socket);
>
> if (max_hart_per_socket < hart_count) {
> max_hart_per_socket = hart_count;
> }
>
> for (i = 0; i < hart_count; i++) {
> imsic_addr = socket_imsic_base + i * IMSIC_HART_SIZE(guest_bits);
> ret = kvm_device_access(aia_fd, KVM_DEV_RISCV_AIA_GRP_ADDR,
> KVM_DEV_RISCV_AIA_ADDR_IMSIC(i + base_hart),
> &imsic_addr, true, NULL);
> if (ret < 0) {
> error_report("KVM AIA: failed to set the IMSIC address for hart %d", i);
> exit(1);
> }
> }
> }
>
> After this change, kvm_riscv_aia_create() is being called once for each socket since it's
> now being called inside virt_create_aia(). And this will cause errors when running qemu-kvm
> with more than one socket:
>
> ./qemu-system-riscv64 \
> -machine virt,accel=kvm,aia=aplic-imsic -m 2G \
> -object memory-backend-ram,size=1G,id=m0 \
> -object memory-backend-ram,size=1G,id=m1 \
> -smp 2,sockets=2,cores=1,threads=1 \
> -numa node,memdev=m0,cpus=0,nodeid=0 \
> -numa node,memdev=m1,cpus=1,nodeid=1 \
> (...)
> qemu-system-riscv64: KVM AIA: failed to set the IMSIC address for hart 0
>
Oh I forgot to test the NUMA config. Sorry.
>
> To make this patch work we would need changes in kvm_riscv_aia_create() to handle just the
> current socket. The loop I mentioned above is one place, and there's another place where
> we set group_bits and group_shift if socket_count > 1.
>
Also we need to find a place to initialize the in-kernel AIA after
setting up all the IMSICs among sockets. This would make things more
complicated. I will remove this patch in the next version. Thank you!
Regards,
Yong-Xuan
> To be honest I'm not sure if all these extra required changes are worth the simplification
> this patch is proposing.
>
>
> Thanks,
>
> Daniel
>
>
>
>
>
> > hw/riscv/virt.c | 79 +++++++++++++++++++++++--------------------------
> > 1 file changed, 37 insertions(+), 42 deletions(-)
> >
> > diff --git a/hw/riscv/virt.c b/hw/riscv/virt.c
> > index dae46f4733cd..a52117ef71ee 100644
> > --- a/hw/riscv/virt.c
> > +++ b/hw/riscv/virt.c
> > @@ -58,14 +58,6 @@
> > #include "qapi/qapi-visit-common.h"
> > #include "hw/virtio/virtio-iommu.h"
> >
> > -/* KVM AIA only supports APLIC MSI. APLIC Wired is always emulated by QEMU. */
> > -static bool virt_use_kvm_aia_aplic_imsic(RISCVVirtAIAType aia_type)
> > -{
> > - bool msimode = aia_type == VIRT_AIA_TYPE_APLIC_IMSIC;
> > -
> > - return riscv_is_kvm_aia_aplic_imsic(msimode);
> > -}
> > -
> > static bool virt_use_emulated_aplic(RISCVVirtAIAType aia_type)
> > {
> > bool msimode = aia_type == VIRT_AIA_TYPE_APLIC_IMSIC;
> > @@ -1298,10 +1290,12 @@ static DeviceState *virt_create_plic(const MemMapEntry *memmap, int socket,
> > return ret;
> > }
> >
> > -static DeviceState *virt_create_aia(RISCVVirtAIAType aia_type, int aia_guests,
> > +static DeviceState *virt_create_aia(RISCVVirtState *s,
> > const MemMapEntry *memmap, int socket,
> > int base_hartid, int hart_count)
> > {
> > + RISCVVirtAIAType aia_type = s->aia_type;
> > + int aia_guests = s->aia_guests;
> > int i;
> > hwaddr addr = 0;
> > uint32_t guest_bits;
> > @@ -1309,6 +1303,28 @@ static DeviceState *virt_create_aia(RISCVVirtAIAType aia_type, int aia_guests,
> > DeviceState *aplic_m = NULL;
> > bool msimode = aia_type == VIRT_AIA_TYPE_APLIC_IMSIC;
> >
> > + if (!kvm_enabled()) {
> > + /* Per-socket M-level APLIC */
> > + aplic_m = riscv_aplic_create(memmap[VIRT_APLIC_M].base +
> > + socket * memmap[VIRT_APLIC_M].size,
> > + memmap[VIRT_APLIC_M].size,
> > + (msimode) ? 0 : base_hartid,
> > + (msimode) ? 0 : hart_count,
> > + VIRT_IRQCHIP_NUM_SOURCES,
> > + VIRT_IRQCHIP_NUM_PRIO_BITS,
> > + msimode, true, NULL);
> > + }
> > +
> > + /* Per-socket S-level APLIC */
> > + aplic_s = riscv_aplic_create(memmap[VIRT_APLIC_S].base +
> > + socket * memmap[VIRT_APLIC_S].size,
> > + memmap[VIRT_APLIC_S].size,
> > + (msimode) ? 0 : base_hartid,
> > + (msimode) ? 0 : hart_count,
> > + VIRT_IRQCHIP_NUM_SOURCES,
> > + VIRT_IRQCHIP_NUM_PRIO_BITS,
> > + msimode, false, aplic_m);
> > +
> > if (msimode) {
> > if (!kvm_enabled()) {
> > /* Per-socket M-level IMSICs */
> > @@ -1329,32 +1345,20 @@ static DeviceState *virt_create_aia(RISCVVirtAIAType aia_type, int aia_guests,
> > base_hartid + i, false, 1 + aia_guests,
> > VIRT_IRQCHIP_NUM_MSIS);
> > }
> > - }
> >
> > - if (!kvm_enabled()) {
> > - /* Per-socket M-level APLIC */
> > - aplic_m = riscv_aplic_create(memmap[VIRT_APLIC_M].base +
> > - socket * memmap[VIRT_APLIC_M].size,
> > - memmap[VIRT_APLIC_M].size,
> > - (msimode) ? 0 : base_hartid,
> > - (msimode) ? 0 : hart_count,
> > - VIRT_IRQCHIP_NUM_SOURCES,
> > - VIRT_IRQCHIP_NUM_PRIO_BITS,
> > - msimode, true, NULL);
> > - }
> >
> > - /* Per-socket S-level APLIC */
> > - aplic_s = riscv_aplic_create(memmap[VIRT_APLIC_S].base +
> > - socket * memmap[VIRT_APLIC_S].size,
> > - memmap[VIRT_APLIC_S].size,
> > - (msimode) ? 0 : base_hartid,
> > - (msimode) ? 0 : hart_count,
> > + if (kvm_irqchip_in_kernel()) {
> > + kvm_riscv_aia_create(MACHINE(s), IMSIC_MMIO_GROUP_MIN_SHIFT,
> > VIRT_IRQCHIP_NUM_SOURCES,
> > - VIRT_IRQCHIP_NUM_PRIO_BITS,
> > - msimode, false, aplic_m);
> > + VIRT_IRQCHIP_NUM_MSIS,
> > + memmap[VIRT_APLIC_S].base,
> > + memmap[VIRT_IMSIC_S].base,
> > + aia_guests);
> > + }
> >
> > - if (kvm_enabled() && msimode) {
> > - riscv_aplic_set_kvm_msicfgaddr(RISCV_APLIC(aplic_s), addr);
> > + if (kvm_enabled()) {
> > + riscv_aplic_set_kvm_msicfgaddr(RISCV_APLIC(aplic_s), addr);
> > + }
> > }
> >
> > return kvm_enabled() ? aplic_s : aplic_m;
> > @@ -1621,9 +1625,8 @@ static void virt_machine_init(MachineState *machine)
> > s->irqchip[i] = virt_create_plic(memmap, i,
> > base_hartid, hart_count);
> > } else {
> > - s->irqchip[i] = virt_create_aia(s->aia_type, s->aia_guests,
> > - memmap, i, base_hartid,
> > - hart_count);
> > + s->irqchip[i] = virt_create_aia(s, memmap, i,
> > + base_hartid, hart_count);
> > }
> >
> > /* Try to use different IRQCHIP instance based device type */
> > @@ -1641,14 +1644,6 @@ static void virt_machine_init(MachineState *machine)
> > }
> > }
> >
> > - if (kvm_enabled() && virt_use_kvm_aia_aplic_imsic(s->aia_type)) {
> > - kvm_riscv_aia_create(machine, IMSIC_MMIO_GROUP_MIN_SHIFT,
> > - VIRT_IRQCHIP_NUM_SOURCES, VIRT_IRQCHIP_NUM_MSIS,
> > - memmap[VIRT_APLIC_S].base,
> > - memmap[VIRT_IMSIC_S].base,
> > - s->aia_guests);
> > - }
> > -
> > if (riscv_is_32bit(&s->soc[0])) {
> > #if HOST_LONG_BITS == 64
> > /* limit RAM size in a 32-bit system */
>
next prev parent reply other threads:[~2025-02-19 11:26 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-02-17 8:19 [PATCH 0/4] riscv: AIA: refinement for KVM acceleration Yong-Xuan Wang
2025-02-17 8:19 ` [PATCH 1/4] hw/riscv/virt: KVM AIA refinement Yong-Xuan Wang
2025-02-17 19:24 ` Daniel Henrique Barboza
2025-02-19 11:25 ` Yong-Xuan Wang [this message]
2025-02-17 8:19 ` [PATCH 2/4] hw/intc/imsic: refine the IMSIC realize Yong-Xuan Wang
2025-02-17 8:19 ` [PATCH 3/4] hw/intc/aplic: refine the APLIC realize Yong-Xuan Wang
2025-02-17 8:19 ` [PATCH 4/4] hw/intc/aplic: refine kvm_msicfgaddr Yong-Xuan Wang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAMWQL2j32880zSM2ZesnAYReJiU2BkN7dwfGQDjAbde5t3uLUQ@mail.gmail.com \
--to=yongxuan.wang@sifive.com \
--cc=alistair.francis@wdc.com \
--cc=dbarboza@ventanamicro.com \
--cc=frank.chang@sifive.com \
--cc=greentime.hu@sifive.com \
--cc=jim.shu@sifive.com \
--cc=liwei1518@gmail.com \
--cc=palmer@dabbelt.com \
--cc=qemu-devel@nongnu.org \
--cc=qemu-riscv@nongnu.org \
--cc=vincent.chen@sifive.com \
--cc=zhiwei_liu@linux.alibaba.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).