On 12/3/2025 11:39 PM, Alejandro Jimenez wrote:
Hi alejandro,
Hi Sairaj,
On 11/18/25 3:24 AM, Sairaj Kodilkar wrote:
AMD IOMMU uses MMIO registers 0x170-0x180 to generate the interrupts
when guest enables xt support through control register [IntCapXTEn]. The
"the interrupts" is a bit vague, considering that it is only a
specific type of interrupts that originate from the IOMMU itself.
I think the quote from the spec is pretty clear and we might want to
use just that:
"
When MMIO 0x18[IntCapXTEn]=1, interrupts originating from the IOMMU
itself are sent based on the programming in XT IOMMU Interrupt Control
Registers in MMIO 0x170-0x180 instead of the programming in the
IOMMU's MSI capability registers.
"
I'd prefer if this was documented in the code comments, but I would go
a step further and mention the specific interrupts, e.g.:
The mapping to MMIO offsets and interrupts controlled by each is as
follows:
XT IOMMU General Interrupt Control Register (0x170): Event Log
exception interrupts and Completion wait interrupts
XT IOMMU PPR Interrupt Control Register (0x178): PPR Log exception
interrupts
XT IOMMU GA Log Interrupt Control Register (0x180): GA Log exception
interrupts (irrelevant in vIOMMU case)
I think we can avoid mentioning other two registers (PPR and GAlog) as
we do not plan to support these
features as of now.
guest programs these registers with appropriate vector and destination
ID instead of writing to PCI MSI capability.
Current AMD vIOMMU is capable of generating interrupts only through PCI
MSI capability and does not care about xt mode. Because of this AMD
vIOMMU cannot generate event log interrupts when the guest has enabled
xt mode.
At first I thought that statement was not correct. If that were the
case, I was wondering why we don't currently have issues with
Completion Wait, since they are also controlled by MMIO 0x170 offset.
But the Linux driver doesn't rely on the completion interrupt AFAICT,
it just sets the completion store bit and monitors the semaphore to
detect the completion. We might not be so lucky with other OSs, so
good catch.
Maybe this patch qualifies for the fixes tag ?
Introduce a new flag "intcapxten" which is set when guest writes control
And similarly to the xten case in PATCH 2, we also need to migrate
this new field. It can be added to the your proposed vmstate_xt.
Right
register [IntCapXTEn] (bit 51) and use vector and destination field in
the XT MMIO register (0x170) to support XT mode.
Signed-off-by: Sairaj Kodilkar <[email protected]>
---
hw/i386/amd_iommu.c | 51 ++++++++++++++++++++++++++++++++++++++------
hw/i386/amd_iommu.h | 3 +++
hw/i386/trace-events | 1 +
3 files changed, 49 insertions(+), 6 deletions(-)
diff --git a/hw/i386/amd_iommu.c b/hw/i386/amd_iommu.c
index 7f08fc31111a..c6bc13c93679 100644
--- a/hw/i386/amd_iommu.c
+++ b/hw/i386/amd_iommu.c
@@ -200,18 +200,52 @@ static void amdvi_assign_andq(AMDVIState *s,
hwaddr addr, uint64_t val)
amdvi_writeq_raw(s, addr, amdvi_readq(s, addr) & val);
}
+union mmio_xt_intr {
+ uint64_t val;
+ struct {
+ uint64_t rsvd_1:2,
+ destination_mode:1,
+ rsvd_2:5,
+ destination_lo:24,
+ vector:8,
+ delivery_mode:1,
+ rsvd_3:15,
+ destination_hi:8;
+ };
+};
We should define this union mmio_xt_intr in amd_iommu.h header, right
after struct irte_ga.
Right.
+
+static void amdvi_build_xt_msi_msg(AMDVIState *s, MSIMessage *msg)
+{
+ union mmio_xt_intr xt_reg;
+ struct X86IOMMUIrq irq;
+
+ xt_reg.val = amdvi_readq(s, AMDVI_MMIO_XT_GEN_INTR);
+
+ irq.vector = xt_reg.vector;
+ irq.delivery_mode = xt_reg.delivery_mode;
+ irq.dest_mode = xt_reg.destination_mode;
+ irq.dest = (xt_reg.destination_hi << 24) | xt_reg.destination_lo;
+ irq.trigger_mode = 0;
+ irq.redir_hint = 0;
+
Based on my reading from the MSI details, hardcoding redir_hint=0
results in the dest_mode field essentially being a nop, since the
messages will be delivered in physical mode i.e. only to the APIC ID
in the dest field.
From Intel's SDM Vol3, 12.11.1 Message Address Register Format:
• If RH is 0, then the DM bit is ignored and the message is sent ahead
independent of whether the physical or logical destination mode is used.
This statement sounds ambiguious. My understanding is that the RH field
is used to determine
the "lowest priority processor" among all possible cpus that can receive
the interrupt. This means
that in case of physical destination mode, it should forward the
interrupt to the cpu with matching
destination mode but in case of logical destination mode it should
forward the interrupt to the
cpu with lowest priority.
I confirmed this by looking at kernel KVM function -->
kvm_irq_delivery_to_apic(). So I think the
what SDM is trying to tell by "DM bit is ignored is that" is that it
will not try to find the cpu
with lowest priority interrupt but rather forward it to the APIC
hardware directly (i.e. to all the
logical processors in case of logical destination mode).
So we don't need to add a error report here.
I am not sure if the current implementations of AMD IOMMU drivers ever
use the logical mode, but I am thinking we should at least catch that
case with a warning e.g.
if (xt_reg.destination_mode) {
error_report_once(...
}
A follow up question, since you chose to explicitly set redir_hint to
0, and something that bothers me about the remap functions is that
they set:
irq->redir_hint = irte.lo.fields_remap.rq_eoi;
where AFAICT, redir_hint and rq_eoi are semantically different and
control unrelated behaviors. So I've been wondering why these
assignments are done. No need to answer this specifically, but if you
have a better understanding of it please let me know.
Currently I don't have idea about this. The specification does not
provide any relation between
rq_eoi and redir_hint.
+ x86_iommu_irq_to_msi_message(&irq, msg);
+}
+
static void amdvi_generate_msi_interrupt(AMDVIState *s)
{
MSIMessage msg = {};
- MemTxAttrs attrs = {
- .requester_id = pci_requester_id(&s->pci->dev)
- };
- if (msi_enabled(&s->pci->dev)) {
+ if (s->intcapxten) {
+ trace_amdvi_generate_msi_interrupt("XT GEN");
+ amdvi_build_xt_msi_msg(s, &msg);
+ } else if (msi_enabled(&s->pci->dev)) {
+ trace_amdvi_generate_msi_interrupt("MSI");
msg = msi_get_message(&s->pci->dev, 0);
- address_space_stl_le(&address_space_memory, msg.address,
msg.data,
- attrs, NULL);
+ } else {
+ trace_amdvi_generate_msi_interrupt("NO MSI");
+ return;
}
+ apic_get_class(NULL)->send_msi(&msg);
This is great. The method of writing to the address space directly
still confuses me, using the APIC helper for MSI delivery seems to be
the appropriate way.
}
static uint32_t get_next_eventlog_entry(AMDVIState *s)
@@ -1490,6 +1524,7 @@ static inline void amdvi_mmio_get_name(hwaddr
addr,
MMIO_REG_TO_STRING(AMDVI_MMIO_PPR_BASE, name)
MMIO_REG_TO_STRING(AMDVI_MMIO_PPR_HEAD, name)
MMIO_REG_TO_STRING(AMDVI_MMIO_PPR_TAIL, name)
+ MMIO_REG_TO_STRING(AMDVI_MMIO_XT_GEN_INTR, name)
default:
name = "UNHANDLED";
}
@@ -1549,6 +1584,7 @@ static void
amdvi_handle_control_write(AMDVIState *s)
AMDVI_MMIO_CONTROL_CMDBUFLEN);
s->ga_enabled = !!(control & AMDVI_MMIO_CONTROL_GAEN);
s->xten = !!(control & AMDVI_MMIO_CONTROL_XTEN) && s->xtsup;
+ s->intcapxten = !!(control & AMDVI_MMIO_CONTROL_INTCAPXTEN) &&
s->xtsup;
I think that s->intcapxten should check for s->xten instead of
s->xtsup, that way we create only one dependency on the input
parameter, and the rest flows logically from what the guest driver
configures.
I did think about it quite a bit, the specs does not provide any dependency
between intcapxten and xtsup, that's why I avoided.
On that topic (I should have mentioned this on patch 2), I think that
it might be reasonable to make s->xten also conditional on whether
s->ga_enabled is true. The spec just says they should both be set:
"In systems with x2APIC enabled, software must set MMIO 0x18[XTEn]=1
and MMIO 0x18[GAEn]=1.This enables the use of the 128-bit IRTE format
with 32-bit destination field. Even if Guest Virtual APIC
will not be used, software must set MMIO 0x18[GAEn]=1.
"
my understanding is that we can't support X2APIC mode without 128-bit
IRTE format (which is controlled by ga_enabled), so it makes sense to
block X2APIC (i.e. xten) unless ga_enabled is already set.
Yep I agree with this one, we can enforce ga_enabled but then we cannot
have
s->intcapxten = !!(control & AMDVI_MMIO_CONTROL_INTCAPXTEN) && s->xten;
because it'll create a dependency between s->intcapxten and s->ga_enabled
So in the end we'd have something like:
s->ga_enabled = !!(control & AMDVI_MMIO_CONTROL_GAEN);
s->xten = !!(control & AMDVI_MMIO_CONTROL_XTEN) &&
s->xtsup && s->ga_enabled; // this should go in PATCH 2
s->intcapxten = !!(control & AMDVI_MMIO_CONTROL_INTCAPXTEN) && s->xten;
What do you think?
PS: just a small formatting suggestion: Please keep the charater per
line within
~72 so that its easier to read on the email client
Thanks
Sairaj