Re: [PATCH kernel v9 29/32] vfio: powerpc/spapr: Register memory and define IOMMU v2

On 05/01/2015 03:23 PM, David Gibson wrote:

On Fri, May 01, 2015 at 02:35:23PM +1000, Alexey Kardashevskiy wrote:

quoted

On 04/30/2015 04:55 PM, David Gibson wrote:

quoted

On Sat, Apr 25, 2015 at 10:14:53PM +1000, Alexey Kardashevskiy wrote:

quoted

The existing implementation accounts the whole DMA window in
the locked_vm counter. This is going to be worse with multiple
containers and huge DMA windows. Also, real-time accounting would requite
additional tracking of accounted pages due to the page size difference -
IOMMU uses 4K pages and system uses 4K or 64K pages.

Another issue is that actual pages pinning/unpinning happens on every
DMA map/unmap request. This does not affect the performance much now as
we spend way too much time now on switching context between
guest/userspace/host but this will start to matter when we add in-kernel
DMA map/unmap acceleration.

This introduces a new IOMMU type for SPAPR - VFIO_SPAPR_TCE_v2_IOMMU.
New IOMMU deprecates VFIO_IOMMU_ENABLE/VFIO_IOMMU_DISABLE and introduces
2 new ioctls to register/unregister DMA memory -
VFIO_IOMMU_SPAPR_REGISTER_MEMORY and VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY -
which receive user space address and size of a memory region which
needs to be pinned/unpinned and counted in locked_vm.
New IOMMU splits physical pages pinning and TCE table update into 2 different
operations. It requires 1) guest pages to be registered first 2) consequent
map/unmap requests to work only with pre-registered memory.
For the default single window case this means that the entire guest
(instead of 2GB) needs to be pinned before using VFIO.
When a huge DMA window is added, no additional pinning will be
required, otherwise it would be guest RAM + 2GB.

The new memory registration ioctls are not supported by
VFIO_SPAPR_TCE_IOMMU. Dynamic DMA window and in-kernel acceleration
will require memory to be preregistered in order to work.

The accounting is done per the user process.

This advertises v2 SPAPR TCE IOMMU and restricts what the userspace
can do with v1 or v2 IOMMUs.

Signed-off-by: Alexey Kardashevskiy <redacted>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <redacted>
---
Changes:
v9:
* s/tce_get_hva_cached/tce_iommu_use_page_v2/

v7:
* now memory is registered per mm (i.e. process)
* moved memory registration code to powerpc/mmu
* merged "vfio: powerpc/spapr: Define v2 IOMMU" into this
* limited new ioctls to v2 IOMMU
* updated doc
* unsupported ioclts return -ENOTTY instead of -EPERM

v6:
* tce_get_hva_cached() returns hva via a pointer

v4:
* updated docs
* s/kzmalloc/vzalloc/
* in tce_pin_pages()/tce_unpin_pages() removed @vaddr, @size and
replaced offset with index
* renamed vfio_iommu_type_register_memory to vfio_iommu_spapr_register_memory
and removed duplicating vfio_iommu_spapr_register_memory
---
  Documentation/vfio.txt              |  23 ++++
  drivers/vfio/vfio_iommu_spapr_tce.c | 230 +++++++++++++++++++++++++++++++++++-
  include/uapi/linux/vfio.h           |  27 +++++
  3 files changed, 274 insertions(+), 6 deletions(-)

diff --git a/Documentation/vfio.txt b/Documentation/vfio.txt
index 96978ec..94328c8 100644
--- a/Documentation/vfio.txt
+++ b/Documentation/vfio.txt

@@ -427,6 +427,29 @@ The code flow from the example above should be slightly changed:

  	....

+5) There is v2 of SPAPR TCE IOMMU. It deprecates VFIO_IOMMU_ENABLE/
+VFIO_IOMMU_DISABLE and implements 2 new ioctls:
+VFIO_IOMMU_SPAPR_REGISTER_MEMORY and VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY
+(which are unsupported in v1 IOMMU).

A summary of the semantic differeces between v1 and v2 would be nice.
At this point it's not really clear to me if there's a case for
creating v2, or if this could just be done by adding (optional)
functionality to v1.

v1: memory preregistration is not supported; explicit enable/disable ioctls
are required

v2: memory preregistration is required; explicit enable/disable are
prohibited (as they are not needed).

Mixing these in one IOMMU type caused a lot of problems like should I
increment locked_vm by the 32bit window size on enable() or not; what do I
do about pages pinning when map/map (check if it is from registered memory
and do not pin?).

Having 2 IOMMU models makes everything a lot simpler.

Ok.  Would it simplify it further if you made v2 only usable on IODA2
hardware?

Very little. V2 addresses memory pinning issue which is handled the same 
way on ioda2 and older hardware, including KVM acceleration. Whether enable 
DDW or not - this is handled just fine via extra properties in the GET_INFO 
ioctl().

IODA2 and others are different in handling multiple groups per container 
but this does not require changes to userspace API.

And remember, the only machine I can use 100% of time is POWER7/P5IOC2 so 
it is really useful if at least some bits of the patchset can be tested 
there; if it was a bit less different from IODA2, I would have even 
implemented DDW there too :)

quoted

quoted

quoted

+PPC64 paravirtualized guests generate a lot of map/unmap requests,
+and the handling of those includes pinning/unpinning pages and updating
+mm::locked_vm counter to make sure we do not exceed the rlimit.
+The v2 IOMMU splits accounting and pinning into separate operations:
+
+- VFIO_IOMMU_SPAPR_REGISTER_MEMORY/VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY ioctls
+receive a user space address and size of the block to be pinned.
+Bisecting is not supported and VFIO_IOMMU_UNREGISTER_MEMORY is expected to
+be called with the exact address and size used for registering
+the memory block. The userspace is not expected to call these often.
+The ranges are stored in a linked list in a VFIO container.
+
+- VFIO_IOMMU_MAP_DMA/VFIO_IOMMU_UNMAP_DMA ioctls only update the actual
+IOMMU table and do not do pinning; instead these check that the userspace
+address is from pre-registered range.
+
+This separation helps in optimizing DMA for guests.
+
  -------------------------------------------------------------------------------

  [1] VFIO was originally an acronym for "Virtual Function I/O" in its

diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c
index 892a584..4cfc2c1 100644
--- a/drivers/vfio/vfio_iommu_spapr_tce.c
+++ b/drivers/vfio/vfio_iommu_spapr_tce.c

So, from things you said at other points, I thought the idea was that
this registration stuff could also be used on non-Power IOMMUs.  Did I
misunderstand, or is that a possibility for the future?

I never said a thing about non-PPC :) I seriously doubt any other arch has
this hypervisor interface with H_PUT_TCE (may be s390? :) ); for others
there is no profit from memory preregistration as they (at least x86) do map
the entire guest before it starts which essentially is that preregistration.


btw later we may want to implement simple IOMMU v3 which will do pinning +
locked_vm when mapping as x86 does, for http://dpdk.org/ - these things do
not really have to bother with preregistration (even if it just a single
additional ioctl).

quoted

quoted

@@ -21,6 +21,7 @@
  #include <linux/vfio.h>
  #include <asm/iommu.h>
  #include <asm/tce.h>
+#include <asm/mmu_context.h>

  #define DRIVER_VERSION  "0.1"
  #define DRIVER_AUTHOR   "aik@ozlabs.ru"

@@ -91,8 +92,58 @@ struct tce_container {
  	struct iommu_group *grp;
  	bool enabled;
  	unsigned long locked_pages;
+	bool v2;
  };

+static long tce_unregister_pages(struct tce_container *container,
+		__u64 vaddr, __u64 size)
+{
+	long ret;
+	struct mm_iommu_table_group_mem_t *mem;
+
+	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
+		return -EINVAL;
+
+	mem = mm_iommu_get(vaddr, size >> PAGE_SHIFT);
+	if (!mem)
+		return -EINVAL;
+
+	ret = mm_iommu_put(mem); /* undo kref_get() from mm_iommu_get() */
+	if (!ret)
+		ret = mm_iommu_put(mem);
+
+	return ret;
+}
+
+static long tce_register_pages(struct tce_container *container,
+		__u64 vaddr, __u64 size)
+{
+	long ret = 0;
+	struct mm_iommu_table_group_mem_t *mem;
+	unsigned long entries = size >> PAGE_SHIFT;
+
+	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
+			((vaddr + size) < vaddr))
+		return -EINVAL;
+
+	mem = mm_iommu_get(vaddr, entries);
+	if (!mem) {
+		ret = try_increment_locked_vm(entries);
+		if (ret)
+			return ret;
+
+		ret = mm_iommu_alloc(vaddr, entries, &mem);
+		if (ret) {
+			decrement_locked_vm(entries);
+			return ret;
+		}
+	}
+
+	container->enabled = true;
+
+	return 0;
+}

So requiring that registered regions get unregistered with exactly the
same addr/length is reasonable.  I'm a bit less convinced that
disallowing overlaps is a good idea.  What if two libraries in the
same process are trying to use VFIO - they may not know if the regions
they try to register are overlapping.

Sorry, I do not understand. A library allocates RAM. A library is expected
to do register it via additional ioctl, that's it. Another library allocates
another chunk of memory and it won't overlap and the registered areas won't
either.

So the case I'm thinking is where the library does VFIO using a buffer
passed into it from the program at large.  Another library does the
same.

The main program, unaware of the VFIO shenanigans passes different
parts of the same page to the 2 libraries.

This is somewhat similar to the case of the horribly, horribly broken
semantics of POSIX file range locks (it's both hard to implement and
dangerous in the multi-library case similar to above).

Ok. I'll implement x86-alike V3 SPAPR TCE IOMMU for these people, later :)

V2 addresses issues caused by H_PUT_TCE + DDW RTAS interfaces.

quoted

quoted

quoted

  static bool tce_page_is_contained(struct page *page, unsigned page_shift)
  {
  	/*

@@ -205,7 +256,7 @@ static void *tce_iommu_open(unsigned long arg)
  {
  	struct tce_container *container;

-	if (arg != VFIO_SPAPR_TCE_IOMMU) {
+	if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
  		pr_err("tce_vfio: Wrong IOMMU type\n");
  		return ERR_PTR(-EINVAL);
  	}

@@ -215,6 +266,7 @@ static void *tce_iommu_open(unsigned long arg)
  		return ERR_PTR(-ENOMEM);

  	mutex_init(&container->lock);
+	container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;

  	return container;
  }

@@ -243,6 +295,47 @@ static void tce_iommu_unuse_page(struct tce_container *container,
  	put_page(page);
  }

+static int tce_iommu_use_page_v2(unsigned long tce, unsigned long size,
+		unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)

You suggested s/tce_get_hpa/tce_iommu_use_page/ but in this particular patch
it is confusing as tce_iommu_unuse_page_v2() calls it to find corresponding
mm_iommu_table_group_mem_t by the userspace address address of a page being
stopped used.

tce_iommu_use_page (without v2) does use the page but this one I'll rename
back to tce_iommu_ua_to_hpa_v2(), is that ok?

Sorry, I couldn't follow this comment.

For V1 IOMMU, I used to have:
tce_get_hpa() - this converted UA to linear address and did gup();
tce_iommu_unuse_page() - this did put_page().

You suggested (*) to rename the first one to tce_use_page() which makes sense.

V2 introduces its own versions of use/unuse but these use preregistered 
memory and do not do gup()/put_page(). I named them:
tce_get_hpa_cached()
tce_iommu_unuse_page_v2()

then, replaying your comment (*) on V2 IOMMU, I renamed 
tce_get_hpa_cached() to tce_iommu_use_page_v2(). And I do not like the 
result now (in the chunk below). I'll rename it to 
tce_iommu_ua_to_hpa_v2(), will it be ok?

quoted

quoted

quoted

+{
+	long ret = 0;
+	struct mm_iommu_table_group_mem_t *mem;
+
+	mem = mm_iommu_lookup(tce, size);
+	if (!mem)
+		return -EINVAL;
+
+	ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
+	if (ret)
+		return -EINVAL;
+
+	*pmem = mem;
+
+	return 0;
+}
+
+static void tce_iommu_unuse_page_v2(struct iommu_table *tbl,
+		unsigned long entry)
+{
+	struct mm_iommu_table_group_mem_t *mem = NULL;
+	int ret;
+	unsigned long hpa = 0;
+	unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
+
+	if (!pua || !current || !current->mm)
+		return;
+
+	ret = tce_iommu_use_page_v2(*pua, IOMMU_PAGE_SIZE(tbl),
+			&hpa, &mem);
+	if (ret)
+		pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
+				__func__, *pua, entry, ret);
+	if (mem)
+		mm_iommu_mapped_update(mem, false);
+
+	*pua = 0;
+}
+

-- 
Alexey