16 Sep, 2016
2 commits
-
With our DMA ops enabled for PCI devices, we should avoid allocating
IOVAs which a host bridge might misinterpret as peer-to-peer DMA and
lead to faults, corruption or other badness. To be safe, punch out holes
for all of the relevant host bridge's windows when initialising a DMA
domain for a PCI device.CC: Marek Szyprowski
CC: Inki Dae
Reported-by: Lorenzo Pieralisi
Signed-off-by: Robin Murphy
Signed-off-by: Will Deacon -
When an MSI doorbell is located downstream of an IOMMU, attaching
devices to a DMA ops domain and switching on translation leads to a rude
shock when their attempt to write to the physical address returned by
the irqchip driver faults (or worse, writes into some already-mapped
buffer) and no interrupt is forthcoming.Address this by adding a hook for relevant irqchip drivers to call from
their compose_msi_msg() callback, to swizzle the physical address with
an appropriatly-mapped IOVA for any device attached to one of our DMA
ops domains.Acked-by: Thomas Gleixner
Acked-by: Marc Zyngier
Signed-off-by: Robin Murphy
Signed-off-by: Will Deacon
10 Aug, 2016
1 commit
-
Where a device driver has set a 64-bit DMA mask to indicate the absence
of addressing limitations, we still need to ensure that we don't
allocate IOVAs beyond the actual input size of the IOMMU. The reported
aperture is the most reliable way we have of inferring that input
address size, so use that to enforce a hard upper limit where available.Fixes: 0db2e5d18f76 ("iommu: Implement common IOMMU ops for DMA mapping")
Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel
09 Aug, 2016
1 commit
-
Due to the limitations of having to wait until we see a device's DMA
restrictions before we know how we want an IOVA domain initialised,
there is a window for error if a DMA ops domain is allocated but later
freed without ever being used. In that case, init_iova_domain() was
never called, so calling put_iova_domain() from iommu_put_dma_cookie()
ends up trying to take an uninitialised lock and crashing.Make things robust by skipping the call unless the IOVA domain actually
has been initialised, as we probably should have done from the start.Fixes: 0db2e5d18f76 ("iommu: Implement common IOMMU ops for DMA mapping")
Cc: stable@vger.kernel.org
Reported-by: Nate Watterson
Reviewed-by: Nate Watterson
Tested-by: Nate Watterson
Reviewed-by: Eric Auger
Tested-by: Eric Auger
Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel
04 Aug, 2016
1 commit
-
The dma-mapping core and the implementations do not change the DMA
attributes passed by pointer. Thus the pointer can point to const data.
However the attributes do not have to be a bitfield. Instead unsigned
long will do fine:1. This is just simpler. Both in terms of reading the code and setting
attributes. Instead of initializing local attributes on the stack
and passing pointer to it to dma_set_attr(), just set the bits.2. It brings safeness and checking for const correctness because the
attributes are passed by value.Semantic patches for this change (at least most of them):
virtual patch
virtual context@r@
identifier f, attrs;@@
f(...,
- struct dma_attrs *attrs
+ unsigned long attrs
, ...)
{
...
}@@
identifier r.f;
@@
f(...,
- NULL
+ 0
)and
// Options: --all-includes
virtual patch
virtual context@r@
identifier f, attrs;
type t;@@
t f(..., struct dma_attrs *attrs);@@
identifier r.f;
@@
f(...,
- NULL
+ 0
)Link: http://lkml.kernel.org/r/1468399300-5399-2-git-send-email-k.kozlowski@samsung.com
Signed-off-by: Krzysztof Kozlowski
Acked-by: Vineet Gupta
Acked-by: Robin Murphy
Acked-by: Hans-Christian Noren Egtvedt
Acked-by: Mark Salter [c6x]
Acked-by: Jesper Nilsson [cris]
Acked-by: Daniel Vetter [drm]
Reviewed-by: Bart Van Assche
Acked-by: Joerg Roedel [iommu]
Acked-by: Fabien Dessenne [bdisp]
Reviewed-by: Marek Szyprowski [vb2-core]
Acked-by: David Vrabel [xen]
Acked-by: Konrad Rzeszutek Wilk [xen swiotlb]
Acked-by: Joerg Roedel [iommu]
Acked-by: Richard Kuo [hexagon]
Acked-by: Geert Uytterhoeven [m68k]
Acked-by: Gerald Schaefer [s390]
Acked-by: Bjorn Andersson
Acked-by: Hans-Christian Noren Egtvedt [avr32]
Acked-by: Vineet Gupta [arc]
Acked-by: Robin Murphy [arm64 and dma-iommu]
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
09 May, 2016
3 commits
-
Now that we know exactly which page sizes our caller wants to use in the
given domain, we can restrict higher-order allocation attempts to just
those sizes, if any, and avoid wasting any time or effort on other sizes
which offer no benefit. In the same vein, this also lets us accommodate
a minimum order greater than 0 for special cases.Signed-off-by: Robin Murphy
Acked-by: Will Deacon
Tested-by: Yong Wu
Signed-off-by: Joerg Roedel -
Many IOMMUs support multiple page table formats, meaning that any given
domain may only support a subset of the hardware page sizes presented in
iommu_ops->pgsize_bitmap. There are also certain use-cases where the
creator of a domain may want to control which page sizes are used, for
example to force the use of hugepage mappings to reduce pagetable walk
depth.To this end, add a per-domain pgsize_bitmap to represent the subset of
page sizes actually in use, to make it possible for domains with
different requirements to coexist.Signed-off-by: Will Deacon
[rm: hijacked and rebased original patch with new commit message]
Signed-off-by: Robin Murphy
Acked-by: Will Deacon
Signed-off-by: Joerg Roedel -
Stop wasting IOVA space by over-aligning scatterlist segments for a
theoretical worst-case segment boundary mask, and instead take the real
limits into account to merge consecutive segments wherever appropriate,
so our callers can benefit from getting back nicely simplified lists.This also represents the last piece of functionality wanted by users of
the current arch/arm implementation, thus brings us a small step closer
to converting that over to the common code.Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel
05 Apr, 2016
1 commit
-
With the change to stashing just the IOVA-page-aligned remainder of the
CPU-page offset rather than the whole thing, the failure path in
__invalidate_sg() also needs tweaking to account for that in the case of
differing page sizes where the two offsets may not be equivalent.
Similarly in __finalise_sg(), lest the architecture-specific wrappers
later get the wrong address for cache maintenance on sync or unmap.Fixes: 164afb1d85b8 ("iommu/dma: Use correct offset in map_sg")
Reported-by: Magnus Damm
Signed-off-by: Robin Murphy
Cc: stable@ver.kernel.org # v4.4+
Signed-off-by: Joerg Roedel
07 Jan, 2016
1 commit
-
When mapping a non-page-aligned scatterlist entry, we copy the original
offset to the output DMA address before aligning it to hand off to
iommu_map_sg(), then later adding the IOVA page address portion to get
the final mapped address. However, when the IOVA page size is smaller
than the CPU page size, it is the offset within the IOVA page we want,
not that within the CPU page, which can easily be larger than an IOVA
page and thus result in an incorrect final address.Fix the bug by taking only the IOVA-aligned part of the offset as the
basis of the DMA address, not the whole thing.Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel
29 Dec, 2015
2 commits
-
Doug reports that the equivalent page allocator on 32-bit ARM exhibits
particularly pathalogical behaviour under memory pressure when
fragmentation is high, where allocating a 4MB buffer takes tens of
seconds and the number of calls to alloc_pages() is over 9000![1]We can drastically improve that situation without losing the other
benefits of high-order allocations when they would succeed, by assuming
memory pressure is relatively constant over the course of an allocation,
and not retrying allocations at orders we know to have failed before.
This way, the best-case behaviour remains unchanged, and in the worst
case we should see at most a dozen or so (MAX_ORDER - 1) failed attempts
before falling back to single pages for the remainder of the buffer.[1]:http://lists.infradead.org/pipermail/linux-arm-kernel/2015-December/394660.html
Reported-by: Douglas Anderson
Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel -
dma-iommu.c was naughtily relying on an implicit transitive #include of
linux/vmalloc.h, which is apparently not present on some architectures.
Add that, plus a couple more headers for other functions which are used
similarly.Reported-by: kbuild test robot
Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel
15 Oct, 2015
1 commit
-
Taking inspiration from the existing arch/arm code, break out some
generic functions to interface the DMA-API to the IOMMU-API. This will
do the bulk of the heavy lifting for IOMMU-backed dma-mapping.Since associating an IOVA allocator with an IOMMU domain is a fairly
common need, rather than introduce yet another private structure just to
do this for ourselves, extend the top-level struct iommu_domain with the
notion. A simple opaque cookie allows reuse by other IOMMU API users
with their various different incompatible allocator types.Signed-off-by: Robin Murphy
Signed-off-by: Joerg Roedel