On Fri, Mar 11, 2016 at 01:15:20PM +1100, Alexey Kardashevskiy wrote:

On 03/10/2016 04:18 PM, David Gibson wrote:

quoted

On Wed, Mar 09, 2016 at 07:46:47PM +1100, Alexey Kardashevskiy wrote:

quoted

On 03/08/2016 10:08 PM, David Gibson wrote:

quoted

On Mon, Mar 07, 2016 at 02:41:16PM +1100, Alexey Kardashevskiy wrote:

quoted

This allows the host kernel to handle H_PUT_TCE, H_PUT_TCE_INDIRECT
and H_STUFF_TCE requests targeted an IOMMU TCE table used for VFIO
without passing them to user space which saves time on switching
to user space and back.

Both real and virtual modes are supported. The kernel tries to
handle a TCE request in the real mode, if fails it passes the request
to the virtual mode to complete the operation. If it a virtual mode
handler fails, the request is passed to user space; this is not expected
to happen ever though.

Well... not expect to happen with a qemu which uses this.  Presumably
it will fall back to userspace routinely if you have an old qemu that
doesn't add the liobn mappings.

Ah. Ok, thanks, I'll add this to the commit log.

Ok.

quoted

quoted

quoted

The first user of this is VFIO on POWER. Trampolines to the VFIO external
user API functions are required for this patch.

I'm not sure what you mean by "trampoline" here.

For example, look at kvm_vfio_group_get_external_user. It calls
symbol_get(vfio_group_get_external_user) and then calls a function via the
returned pointer.

Is there a better word for this?

Uh.. probably although I don't immediately know what.  "Trampoline"
usually refers to code on the stack used for bouncing places, which
isn't what this resembles.

"Dynamic wrapper"?

Sure, that'll do.

quoted

quoted

quoted

quoted

This uses a VFIO KVM device to associate a logical bus number (LIOBN)
with an VFIO IOMMU group fd and enable in-kernel handling of map/unmap
requests.

Group fd?  Or container fd?  The group fd wouldn't make a lot of
sense.

Group. KVM has no idea about containers.

That's not going to fly.  Having a liobn registered against just one
group in a container makes no sense at all.  Conceptually, if not
physically, the container shares a single set of TCE tables.  If
handling that means teaching KVM the concept of containers, then so be
it.

Btw, I'm not sure yet if extending the existing vfio kvm device to
make the vfio<->kvm linkages makes sense.  I think the reason some x86
machines need that is quite different from how we're using it for
Power.  I haven't got a clear enough picture yet to be sure either
way.

The other option that would seem likely to me would be a "bind VFIO
container" ioctl() on the fd associated with a kernel accelerated TCE table.

Oh, I just noticed this response. I need to digest it. Looks like this is
going to take other 2 years to upstream...

quoted

quoted

quoted

quoted

To make use of the feature, the user space has to create a guest view
of the TCE table via KVM_CAP_SPAPR_TCE/KVM_CAP_SPAPR_TCE_64 and
then associate a LIOBN with this table via VFIO KVM device,
a KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE_LIOBN property (which is added in
the next patch).

Tests show that this patch increases transmission speed from 220MB/s
to 750..1020MB/s on 10Gb network (Chelsea CXGB3 10Gb ethernet card).

Is that with or without DDW (i.e. with or without a 64-bit DMA window)?

Without DDW, I should have mentioned this. The patch is from the times when
there was no DDW :(

Ok.

quoted

quoted

quoted

Signed-off-by: Alexey Kardashevskiy <redacted>
---
 arch/powerpc/kvm/book3s_64_vio.c    | 184 +++++++++++++++++++++++++++++++++++
 arch/powerpc/kvm/book3s_64_vio_hv.c | 186 ++++++++++++++++++++++++++++++++++++
 2 files changed, 370 insertions(+)

diff --git a/arch/powerpc/kvm/book3s_64_vio.c b/arch/powerpc/kvm/book3s_64_vio.c
index 7965fc7..9417d12 100644
--- a/arch/powerpc/kvm/book3s_64_vio.c
+++ b/arch/powerpc/kvm/book3s_64_vio.c

@@ -33,6 +33,7 @@
 #include <asm/kvm_ppc.h>
 #include <asm/kvm_book3s.h>
 #include <asm/mmu-hash64.h>
+#include <asm/mmu_context.h>
 #include <asm/hvcall.h>
 #include <asm/synch.h>
 #include <asm/ppc-opcode.h>

@@ -317,11 +318,161 @@ fail:
 	return ret;
 }

+static long kvmppc_tce_iommu_mapped_dec(struct iommu_table *tbl,
+		unsigned long entry)
+{
+	struct mm_iommu_table_group_mem_t *mem = NULL;
+	const unsigned long pgsize = 1ULL << tbl->it_page_shift;
+	unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
+
+	if (!pua)
+		return H_HARDWARE;
+
+	mem = mm_iommu_lookup(*pua, pgsize);
+	if (!mem)
+		return H_HARDWARE;
+
+	mm_iommu_mapped_dec(mem);
+
+	*pua = 0;
+
+	return H_SUCCESS;
+}
+
+static long kvmppc_tce_iommu_unmap(struct iommu_table *tbl,
+		unsigned long entry)
+{
+	enum dma_data_direction dir = DMA_NONE;
+	unsigned long hpa = 0;
+
+	if (iommu_tce_xchg(tbl, entry, &hpa, &dir))
+		return H_HARDWARE;
+
+	if (dir == DMA_NONE)
+		return H_SUCCESS;
+
+	return kvmppc_tce_iommu_mapped_dec(tbl, entry);
+}
+
+long kvmppc_tce_iommu_map(struct kvm *kvm, struct iommu_table *tbl,
+		unsigned long entry, unsigned long gpa,
+		enum dma_data_direction dir)
+{
+	long ret;
+	unsigned long hpa, ua, *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
+	struct mm_iommu_table_group_mem_t *mem;
+
+	if (!pua)
+		return H_HARDWARE;

H_HARDWARE?  Or H_PARAMETER?  This essentially means the guest has
supplied a bad physical address, doesn't it?

Well, may be. I'll change. If it not H_TOO_HARD, it does not make any
difference after all :)

quoted

quoted

+	if (kvmppc_gpa_to_ua(kvm, gpa, &ua, NULL))
+		return H_HARDWARE;
+
+	mem = mm_iommu_lookup(ua, 1ULL << tbl->it_page_shift);
+	if (!mem)
+		return H_HARDWARE;
+
+	if (mm_iommu_ua_to_hpa(mem, ua, &hpa))
+		return H_HARDWARE;
+
+	if (mm_iommu_mapped_inc(mem))
+		return H_HARDWARE;
+
+	ret = iommu_tce_xchg(tbl, entry, &hpa, &dir);
+	if (ret) {
+		mm_iommu_mapped_dec(mem);
+		return H_TOO_HARD;
+	}
+
+	if (dir != DMA_NONE)
+		kvmppc_tce_iommu_mapped_dec(tbl, entry);
+
+	*pua = ua;

IIUC this means you have a copy of the UA for every group attached to
the TCE table, but they'll all be the same. Any way to avoid that
duplication?

It is for every container, not a group. On P8, I allow multiple groups to go
to the same container, that means that a container has one or two
iommu_table, and each iommu_table has this "ua" list but since tables are
different (window size, page size, content), these "ua" arrays are also
different.

Erm.. but h_put_tce iterates h_put_tce_iommu through all the groups
attached to the stt, and each one seems to update pua.

Or is that what the if (kg->tbl == tbltmp) continue; is supposed to
avoid?  In which case what ensures that the stt->groups list is
ordered by tbl pointer?

Nothing. In the normal case (POWER8 IODA2) all groups on the same liobn have
the same iommu_table, so the first group's one gets updated, other do not
but it is ok as they use the same table.

Right, which is another indication that group is the wrong concept to
use here.

In a bad case (POWER7 IODA1, multiple containers per LIOBN) the same @ua can
be updated more than once. Well, not a huge loss.

Ugh.. this really seems to be based on knowing the specific cases we
have in practice, rather than writing code that's correct based on ly
on the properties that the objects are defined to have.  The latter
approach will make for much more robust and extensible code.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson