A huge update this time, but a lot of that is just consolidating or
removing code:
- provide a common DMA_MAPPING_ERROR definition and avoid indirect
calls for dma_map_* error checking
- use direct calls for the DMA direct mapping case, avoiding huge
retpoline overhead for high performance workloads
- merge the swiotlb dma_map_ops into dma-direct
- provide a generic remapping DMA consistent allocator for architectures
that have devices that perform DMA that is not cache coherent. Based
on the existing arm64 implementation and also used for csky now.
- improve the dma-debug infrastructure, including dynamic allocation
of entries (Robin Murphy)
- default to providing chaining scatterlist everywhere, with opt-outs
for the few architectures (alpha, parisc, most arm32 variants) that
can't cope with it
- misc sparc32 dma-related cleanups
- remove the dma_mark_clean arch hook used by swiotlb on ia64 and
replace it with the generic noncoherent infrastructure
- fix the return type of dma_set_max_seg_size (Niklas Söderlund)
- move the dummy dma ops for not DMA capable devices from arm64 to
common code (Robin Murphy)
- ensure dma_alloc_coherent returns zeroed memory to avoid kernel data
leaks through userspace. We already did this for most common
architectures, but this ensures we do it everywhere.
dma_zalloc_coherent has been deprecated and can hopefully be
removed after -rc1 with a coccinelle script.
-----BEGIN PGP SIGNATURE-----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=in72
-----END PGP SIGNATURE-----
Merge tag 'dma-mapping-4.21' of git://git.infradead.org/users/hch/dma-mapping
Pull DMA mapping updates from Christoph Hellwig:
"A huge update this time, but a lot of that is just consolidating or
removing code:
- provide a common DMA_MAPPING_ERROR definition and avoid indirect
calls for dma_map_* error checking
- use direct calls for the DMA direct mapping case, avoiding huge
retpoline overhead for high performance workloads
- merge the swiotlb dma_map_ops into dma-direct
- provide a generic remapping DMA consistent allocator for
architectures that have devices that perform DMA that is not cache
coherent. Based on the existing arm64 implementation and also used
for csky now.
- improve the dma-debug infrastructure, including dynamic allocation
of entries (Robin Murphy)
- default to providing chaining scatterlist everywhere, with opt-outs
for the few architectures (alpha, parisc, most arm32 variants) that
can't cope with it
- misc sparc32 dma-related cleanups
- remove the dma_mark_clean arch hook used by swiotlb on ia64 and
replace it with the generic noncoherent infrastructure
- fix the return type of dma_set_max_seg_size (Niklas Söderlund)
- move the dummy dma ops for not DMA capable devices from arm64 to
common code (Robin Murphy)
- ensure dma_alloc_coherent returns zeroed memory to avoid kernel
data leaks through userspace. We already did this for most common
architectures, but this ensures we do it everywhere.
dma_zalloc_coherent has been deprecated and can hopefully be
removed after -rc1 with a coccinelle script"
* tag 'dma-mapping-4.21' of git://git.infradead.org/users/hch/dma-mapping: (73 commits)
dma-mapping: fix inverted logic in dma_supported
dma-mapping: deprecate dma_zalloc_coherent
dma-mapping: zero memory returned from dma_alloc_*
sparc/iommu: fix ->map_sg return value
sparc/io-unit: fix ->map_sg return value
arm64: default to the direct mapping in get_arch_dma_ops
PCI: Remove unused attr variable in pci_dma_configure
ia64: only select ARCH_HAS_DMA_COHERENT_TO_PFN if swiotlb is enabled
dma-mapping: bypass indirect calls for dma-direct
vmd: use the proper dma_* APIs instead of direct methods calls
dma-direct: merge swiotlb_dma_ops into the dma_direct code
dma-direct: use dma_direct_map_page to implement dma_direct_map_sg
dma-direct: improve addressability error reporting
swiotlb: remove dma_mark_clean
swiotlb: remove SWIOTLB_MAP_ERROR
ACPI / scan: Refactor _CCA enforcement
dma-mapping: factor out dummy DMA ops
dma-mapping: always build the direct mapping code
dma-mapping: move dma_cache_sync out of line
dma-mapping: move various slow path functions out of line
...
The powernv platform registers IOMMU groups and adds devices to them
from the pci_controller_ops::setup_bridge() hook except one case when
virtual functions (SRIOV VFs) are added from a bus notifier.
The pseries platform registers IOMMU groups from
the pci_controller_ops::dma_bus_setup() hook and adds devices from
the pci_controller_ops::dma_dev_setup() hook. The very same bus notifier
used for powernv does not add devices for pseries though as
__of_scan_bus() adds devices first, then it does the bus/dev DMA setup.
Both platforms use iommu_add_device() which takes a device and expects
it to have a valid IOMMU table struct with an iommu_table_group pointer
which in turn points the iommu_group struct (which represents
an IOMMU group). Although the helper seems easy to use, it relies on
some pre-existing device configuration and associated data structures
which it does not really need.
This simplifies iommu_add_device() to take the table_group pointer
directly. Pseries already has a table_group pointer handy and the bus
notified is not used anyway. For powernv, this copies the existing bus
notifier, makes it work for powernv only which means an easy way of
getting to the table_group pointer. This was tested on VFs but should
also support physical PCI hotplug.
Since iommu_add_device() receives the table_group pointer directly,
pseries does not do TCE cache invalidation (the hypervisor does) nor
allow multiple groups per a VFIO container (in other words sharing
an IOMMU table between partitionable endpoints), this removes
iommu_table_group_link from pseries.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This new memory does not have page structs as it is not plugged to
the host so gup() will fail anyway.
This adds 2 helpers:
- mm_iommu_newdev() to preregister the "memory device" memory so
the rest of API can still be used;
- mm_iommu_is_devmem() to know if the physical address is one of thise
new regions which we must avoid unpinning of.
This adds @mm to tce_page_is_contained() and iommu_tce_xchg() to test
if the memory is device memory to avoid pfn_to_page().
This adds a check for device memory in mm_iommu_ua_mark_dirty_rm() which
does delayed pages dirtying.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The powerpc iommu code already returns (~(dma_addr_t)0x0) on mapping
failures, so we can switch over to returning DMA_MAPPING_ERROR and let
the core dma-mapping code handle the rest.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
The powernv platform maintains 2 TCE tables for VFIO - a hardware TCE
table and a table with userspace addresses. These tables are radix trees,
we allocate indirect levels when they are written to. Since
the memory allocation is problematic in real mode, we have 2 accessors
to the entries:
- for virtual mode: it allocates the memory and it is always expected
to return non-NULL;
- fr real mode: it does not allocate and can return NULL.
Also, DMA windows can span to up to 55 bits of the address space and since
we never have this much RAM, such windows are sparse. However currently
the SPAPR TCE IOMMU driver walks through all TCEs to unpin DMA memory.
Since we maintain a userspace addresses table for VFIO which is a mirror
of the hardware table, we can use it to know which parts of the DMA
window have not been mapped and skip these so does this patch.
The bare metal systems do not have this problem as they use a bypass mode
of a PHB which maps RAM directly.
This helps a lot with sparse DMA windows, reducing the shutdown time from
about 3 minutes per 1 billion TCEs to a few seconds for 32GB sparse guest.
Just skipping the last level seems to be good enough.
As non-allocating accessor is used now in virtual mode as well, rename it
from IOMMU_TABLE_USERSPACE_ENTRY_RM (real mode) to _RO (read only).
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
At the moment the real mode handler of H_PUT_TCE calls iommu_tce_xchg_rm()
which in turn reads the old TCE and if it was a valid entry, marks
the physical page dirty if it was mapped for writing. Since it is in
real mode, realmode_pfn_to_page() is used instead of pfn_to_page()
to get the page struct. However SetPageDirty() itself reads the compound
page head and returns a virtual address for the head page struct and
setting dirty bit for that kills the system.
This adds additional dirty bit tracking into the MM/IOMMU API for use
in the real mode. Note that this does not change how VFIO and
KVM (in virtual mode) set this bit. The KVM (real mode) changes include:
- use the lowest bit of the cached host phys address to carry
the dirty bit;
- mark pages dirty when they are unpinned which happens when
the preregistered memory is released which always happens in virtual
mode;
- add mm_iommu_ua_mark_dirty_rm() helper to set delayed dirty bit;
- change iommu_tce_xchg_rm() to take the kvm struct for the mm to use
in the new mm_iommu_ua_mark_dirty_rm() helper;
- move iommu_tce_xchg_rm() to book3s_64_vio_hv.c (which is the only
caller anyway) to reduce the real mode KVM and IOMMU knowledge
across different subsystems.
This removes realmode_pfn_to_page() as it is not used anymore.
While we at it, remove some EXPORT_SYMBOL_GPL() as that code is for
the real mode only and modules cannot call it anyway.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This patch moves ASM_CONST() and stringify_in_c() into
dedicated asm-const.h, then cleans all related inclusions.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
[mpe: asm-compat.h should include asm-const.h]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
At the moment we allocate the entire TCE table, twice (hardware part and
userspace translation cache). This normally works as we normally have
contigous memory and the guest will map entire RAM for 64bit DMA.
However if we have sparse RAM (one example is a memory device), then
we will allocate TCEs which will never be used as the guest only maps
actual memory for DMA. If it is a single level TCE table, there is nothing
we can really do but if it a multilevel table, we can skip allocating
TCEs we know we won't need.
This adds ability to allocate only first level, saving memory.
This changes iommu_table::free() to avoid allocating of an extra level;
iommu_table::set() will do this when needed.
This adds @alloc parameter to iommu_table::exchange() to tell the callback
if it can allocate an extra level; the flag is set to "false" for
the realmode KVM handlers of H_PUT_TCE hcalls and the callback returns
H_TOO_HARD.
This still requires the entire table to be counted in mm::locked_vm.
To be conservative, this only does on-demand allocation when
the usespace cache table is requested which is the case of VFIO.
The example math for a system replicating a powernv setup with NVLink2
in a guest:
16GB RAM mapped at 0x0
128GB GPU RAM window (16GB of actual RAM) mapped at 0x244000000000
the table to cover that all with 64K pages takes:
(((0x244000000000 + 0x2000000000) >> 16)*8)>>20 = 4556MB
If we allocate only necessary TCE levels, we will only need:
(((0x400000000 + 0x400000000) >> 16)*8)>>20 = 4MB (plus some for indirect
levels).
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We want to support sparse memory and therefore huge chunks of DMA windows
do not need to be mapped. If a DMA window big enough to require 2 or more
indirect levels, and a DMA window is used to map all RAM (which is
a default case for 64bit window), we can actually save some memory by
not allocation TCE for regions which we are not going to map anyway.
The hardware tables alreary support indirect levels but we also keep
host-physical-to-userspace translation array which is allocated by
vmalloc() and is a flat array which might use quite some memory.
This converts it_userspace from vmalloc'ed array to a multi level table.
As the format becomes platform dependend, this replaces the direct access
to it_usespace with a iommu_table_ops::useraddrptr hook which returns
a pointer to the userspace copy of a TCE; future extension will return
NULL if the level was not allocated.
This should not change non-KVM handling of TCE tables and it_userspace
will not be allocated for non-KVM tables.
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We are going to reuse multilevel TCE code for the userspace copy of
the TCE table and since it is big endian, let's make the copy big endian
too.
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Acked-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
DMA_ERROR_CODE is going to go away, so don't rely on it. Instead
define a ->mapping_error method for all IOMMU based dma operation
instances. The direct ops don't ever return an error and don't
need a ->mapping_error method.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
This reworks helpers for checking TCE update parameters in way they
can be used in KVM.
This should cause no behavioral change.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
So far iommu_table obejcts were only used in virtual mode and had
a single owner. We are going to change this by implementing in-kernel
acceleration of DMA mapping requests. The proposed acceleration
will handle requests in real mode and KVM will keep references to tables.
This adds a kref to iommu_table and defines new helpers to update it.
This replaces iommu_free_table() with iommu_tce_table_put() and makes
iommu_free_table() static. iommu_tce_table_get() is not used in this patch
but it will be in the following patch.
Since this touches prototypes, this also removes @node_name parameter as
it has never been really useful on powernv and carrying it for
the pseries platform code to iommu_free_table() seems to be quite
useless as well.
This should cause no behavioral change.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In real mode, TCE tables are invalidated using special
cache-inhibited store instructions which are not available in
virtual mode
This defines and implements exchange_rm() callback. This does not
define set_rm/clear_rm/flush_rm callbacks as there is no user for those -
exchange/exchange_rm are only to be used by KVM for VFIO.
The exchange_rm callback is defined for IODA1/IODA2 powernv platforms.
This replaces list_for_each_entry_rcu with its lockless version as
from now on pnv_pci_ioda2_tce_invalidate() can be called in
the real mode too.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
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 <k.kozlowski@samsung.com>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
Acked-by: Robin Murphy <robin.murphy@arm.com>
Acked-by: Hans-Christian Noren Egtvedt <egtvedt@samfundet.no>
Acked-by: Mark Salter <msalter@redhat.com> [c6x]
Acked-by: Jesper Nilsson <jesper.nilsson@axis.com> [cris]
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch> [drm]
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Acked-by: Joerg Roedel <jroedel@suse.de> [iommu]
Acked-by: Fabien Dessenne <fabien.dessenne@st.com> [bdisp]
Reviewed-by: Marek Szyprowski <m.szyprowski@samsung.com> [vb2-core]
Acked-by: David Vrabel <david.vrabel@citrix.com> [xen]
Acked-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> [xen swiotlb]
Acked-by: Joerg Roedel <jroedel@suse.de> [iommu]
Acked-by: Richard Kuo <rkuo@codeaurora.org> [hexagon]
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org> [m68k]
Acked-by: Gerald Schaefer <gerald.schaefer@de.ibm.com> [s390]
Acked-by: Bjorn Andersson <bjorn.andersson@linaro.org>
Acked-by: Hans-Christian Noren Egtvedt <egtvedt@samfundet.no> [avr32]
Acked-by: Vineet Gupta <vgupta@synopsys.com> [arc]
Acked-by: Robin Murphy <robin.murphy@arm.com> [arm64 and dma-iommu]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Instead of punching a hole in the linear mapping, just use normal
cachable memory, and apply the flush sequence documented in the
CPC625 (aka U3) user manual.
This allows us to remove quite a bit of code related to the early
allocation of the DART and the hole in the linear mapping. We can
also get rid of the copy of the DART for suspend/resume as the
original memory can just be saved/restored now, as long as we
properly sync the caches.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
[mpe: Integrate dart_init() fix to return ENODEV when DART disabled]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
To support "hybrid" DMA ops in a subsequent patch, we will need both
a direct DMA offset and an iommu pointer. Those are currently exclusive
(a union), so change them to be separate fields.
While there, also type iommu_table_base properly and make exist only
on CONFIG_PPC64 since it's not referenced on 32-bit at all.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This adds create/remove window ioctls to create and remove DMA windows.
sPAPR defines a Dynamic DMA windows capability which allows
para-virtualized guests to create additional DMA windows on a PCI bus.
The existing linux kernels use this new window to map the entire guest
memory and switch to the direct DMA operations saving time on map/unmap
requests which would normally happen in a big amounts.
This adds 2 ioctl handlers - VFIO_IOMMU_SPAPR_TCE_CREATE and
VFIO_IOMMU_SPAPR_TCE_REMOVE - to create and remove windows.
Up to 2 windows are supported now by the hardware and by this driver.
This changes VFIO_IOMMU_SPAPR_TCE_GET_INFO handler to return additional
information such as a number of supported windows and maximum number
levels of TCE tables.
DDW is added as a capability, not as a SPAPR TCE IOMMU v2 unique feature
as we still want to support v2 on platforms which cannot do DDW for
the sake of TCE acceleration in KVM (coming soon).
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
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.
In order to support memory pre-registration, we need a way to track
the use of every registered memory region and only allow unregistration
if a region is not in use anymore. So we need a way to tell from what
region the just cleared TCE was from.
This adds a userspace view of the TCE table into iommu_table struct.
It contains userspace address, one per TCE entry. The table is only
allocated when the ownership over an IOMMU group is taken which means
it is only used from outside of the powernv code (such as VFIO).
As v2 IOMMU supports IODA2 and pre-IODA2 IOMMUs (which do not support
DDW API), this creates a default DMA window for IODA2 for consistency.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This adds a way for the IOMMU user to know how much a new table will
use so it can be accounted in the locked_vm limit before allocation
happens.
This stores the allocated table size in pnv_pci_ioda2_get_table_size()
so the locked_vm counter can be updated correctly when a table is
being disposed.
This defines an iommu_table_group_ops callback to let VFIO know
how much memory will be locked if a table is created.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This extends iommu_table_group_ops by a set of callbacks to support
dynamic DMA windows management.
create_table() creates a TCE table with specific parameters.
it receives iommu_table_group to know nodeid in order to allocate
TCE table memory closer to the PHB. The exact format of allocated
multi-level table might be also specific to the PHB model (not
the case now though).
This callback calculated the DMA window offset on a PCI bus from @num
and stores it in a just created table.
set_window() sets the window at specified TVT index + @num on PHB.
unset_window() unsets the window from specified TVT.
This adds a free() callback to iommu_table_ops to free the memory
(potentially a tree of tables) allocated for the TCE table.
create_table() and free() are supposed to be called once per
VFIO container and set_window()/unset_window() are supposed to be
called for every group in a container.
This adds IOMMU capabilities to iommu_table_group such as default
32bit window parameters and others. This makes use of new values in
vfio_iommu_spapr_tce. IODA1/P5IOC2 do not support DDW so they do not
advertise pagemasks to the userspace.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
TCE tables might get too big in case of 4K IOMMU pages and DDW enabled
on huge guests (hundreds of GB of RAM) so the kernel might be unable to
allocate contiguous chunk of physical memory to store the TCE table.
To address this, POWER8 CPU (actually, IODA2) supports multi-level
TCE tables, up to 5 levels which splits the table into a tree of
smaller subtables.
This adds multi-level TCE tables support to
pnv_pci_ioda2_table_alloc_pages() and pnv_pci_ioda2_table_free_pages()
helpers.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
At the moment writing new TCE value to the IOMMU table fails with EBUSY
if there is a valid entry already. However PAPR specification allows
the guest to write new TCE value without clearing it first.
Another problem this patch is addressing is the use of pool locks for
external IOMMU users such as VFIO. The pool locks are to protect
DMA page allocator rather than entries and since the host kernel does
not control what pages are in use, there is no point in pool locks and
exchange()+put_page(oldtce) is sufficient to avoid possible races.
This adds an exchange() callback to iommu_table_ops which does the same
thing as set() plus it returns replaced TCE and DMA direction so
the caller can release the pages afterwards. The exchange() receives
a physical address unlike set() which receives linear mapping address;
and returns a physical address as the clear() does.
This implements exchange() for P5IOC2/IODA/IODA2. This adds a requirement
for a platform to have exchange() implemented in order to support VFIO.
This replaces iommu_tce_build() and iommu_clear_tce() with
a single iommu_tce_xchg().
This makes sure that TCE permission bits are not set in TCE passed to
IOMMU API as those are to be calculated by platform code from
DMA direction.
This moves SetPageDirty() to the IOMMU code to make it work for both
VFIO ioctl interface in in-kernel TCE acceleration (when it becomes
available later).
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This adds tce_iommu_take_ownership() and tce_iommu_release_ownership
which call in a loop iommu_take_ownership()/iommu_release_ownership()
for every table on the group. As there is just one now, no change in
behaviour is expected.
At the moment the iommu_table struct has a set_bypass() which enables/
disables DMA bypass on IODA2 PHB. This is exposed to POWERPC IOMMU code
which calls this callback when external IOMMU users such as VFIO are
about to get over a PHB.
The set_bypass() callback is not really an iommu_table function but
IOMMU/PE function. This introduces a iommu_table_group_ops struct and
adds take_ownership()/release_ownership() callbacks to it which are
called when an external user takes/releases control over the IOMMU.
This replaces set_bypass() with ownership callbacks as it is not
necessarily just bypass enabling, it can be something else/more
so let's give it more generic name.
The callbacks is implemented for IODA2 only. Other platforms (P5IOC2,
IODA1) will use the old iommu_take_ownership/iommu_release_ownership API.
The following patches will replace iommu_take_ownership/
iommu_release_ownership calls in IODA2 with full IOMMU table release/
create.
As we here and touching bypass control, this removes
pnv_pci_ioda2_setup_bypass_pe() as it does not do much
more compared to pnv_pci_ioda2_set_bypass. This moves tce_bypass_base
initialization to pnv_pci_ioda2_setup_dma_pe.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
So far one TCE table could only be used by one IOMMU group. However
IODA2 hardware allows programming the same TCE table address to
multiple PE allowing sharing tables.
This replaces a single pointer to a group in a iommu_table struct
with a linked list of groups which provides the way of invalidating
TCE cache for every PE when an actual TCE table is updated. This adds
pnv_pci_link_table_and_group() and pnv_pci_unlink_table_and_group()
helpers to manage the list. However without VFIO, it is still going
to be a single IOMMU group per iommu_table.
This changes iommu_add_device() to add a device to a first group
from the group list of a table as it is only called from the platform
init code or PCI bus notifier and at these moments there is only
one group per table.
This does not change TCE invalidation code to loop through all
attached groups in order to simplify this patch and because
it is not really needed in most cases. IODA2 is fixed in a later
patch.
This should cause no behavioural change.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Modern IBM POWERPC systems support multiple (currently two) TCE tables
per IOMMU group (a.k.a. PE). This adds a iommu_table_group container
for TCE tables. Right now just one table is supported.
This defines iommu_table_group struct which stores pointers to
iommu_group and iommu_table(s). This replaces iommu_table with
iommu_table_group where iommu_table was used to identify a group:
- iommu_register_group();
- iommudata of generic iommu_group;
This removes @data from iommu_table as it_table_group provides
same access to pnv_ioda_pe.
For IODA, instead of embedding iommu_table, the new iommu_table_group
keeps pointers to those. The iommu_table structs are allocated
dynamically.
For P5IOC2, both iommu_table_group and iommu_table are embedded into
PE struct. As there is no EEH and SRIOV support for P5IOC2,
iommu_free_table() should not be called on iommu_table struct pointers
so we can keep it embedded in pnv_phb::p5ioc2.
For pSeries, this replaces multiple calls of kzalloc_node() with a new
iommu_pseries_alloc_group() helper and stores the table group struct
pointer into the pci_dn struct. For release, a iommu_table_free_group()
helper is added.
This moves iommu_table struct allocation from SR-IOV code to
the generic DMA initialization code in pnv_pci_ioda_setup_dma_pe and
pnv_pci_ioda2_setup_dma_pe as this is where DMA is actually initialized.
This change is here because those lines had to be changed anyway.
This should cause no behavioural change.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This adds a iommu_table_ops struct and puts pointer to it into
the iommu_table struct. This moves tce_build/tce_free/tce_get/tce_flush
callbacks from ppc_md to the new struct where they really belong to.
This adds the requirement for @it_ops to be initialized before calling
iommu_init_table() to make sure that we do not leave any IOMMU table
with iommu_table_ops uninitialized. This is not a parameter of
iommu_init_table() though as there will be cases when iommu_init_table()
will not be called on TCE tables, for example - VFIO.
This does s/tce_build/set/, s/tce_free/clear/ and removes "tce_"
redundant prefixes.
This removes tce_xxx_rm handlers from ppc_md but does not add
them to iommu_table_ops as this will be done later if we decide to
support TCE hypercalls in real mode. This removes _vm callbacks as
only virtual mode is supported by now so this also removes @rm parameter.
For pSeries, this always uses tce_buildmulti_pSeriesLP/
tce_buildmulti_pSeriesLP. This changes multi callback to fall back to
tce_build_pSeriesLP/tce_free_pSeriesLP if FW_FEATURE_MULTITCE is not
present. The reason for this is we still have to support "multitce=off"
boot parameter in disable_multitce() and we do not want to walk through
all IOMMU tables in the system and replace "multi" callbacks with single
ones.
For powernv, this defines _ops per PHB type which are P5IOC2/IODA1/IODA2.
This makes the callbacks for them public. Later patches will extend
callbacks for IODA1/2.
No change in behaviour is expected.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Normally a bitmap from the iommu_table is used to track what TCE entry
is in use. Since we are going to use iommu_table without its locks and
do xchg() instead, it becomes essential not to put bits which are not
implied in the direction flag as the old TCE value (more precisely -
the permission bits) will be used to decide whether to put the page or not.
This adds iommu_direction_to_tce_perm() (its counterpart is there already)
and uses it for powernv's pnv_tce_build().
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This moves page pinning (get_user_pages_fast()/put_page()) code out of
the platform IOMMU code and puts it to VFIO IOMMU driver where it belongs
to as the platform code does not deal with page pinning.
This makes iommu_take_ownership()/iommu_release_ownership() deal with
the IOMMU table bitmap only.
This removes page unpinning from iommu_take_ownership() as the actual
TCE table might contain garbage and doing put_page() on it is undefined
behaviour.
Besides the last part, the rest of the patch is mechanical.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The set_iommu_table_base_and_group() name suggests that the function
sets table base and add a device to an IOMMU group.
The actual purpose for table base setting is to put some reference
into a device so later iommu_add_device() can get the IOMMU group
reference and the device to the group.
At the moment a group cannot be explicitly passed to iommu_add_device()
as we want it to work from the bus notifier, we can fix it later and
remove confusing calls of set_iommu_table_base().
This replaces set_iommu_table_base_and_group() with a couple of
set_iommu_table_base() + iommu_add_device() which makes reading the code
easier.
This adds few comments why set_iommu_table_base() and iommu_add_device()
are called where they are called.
For IODA1/2, this essentially removes iommu_add_device() call from
the pnv_pci_ioda_dma_dev_setup() as it will always fail at this particular
place:
- for physical PE, the device is already attached by iommu_add_device()
in pnv_pci_ioda_setup_dma_pe();
- for virtual PE, the sysfs entries are not ready to create all symlinks
so actual adding is happening in tce_iommu_bus_notifier.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Merge Richard's work to support SR-IOV on PowerNV. All generic PCI
patches acked by Bjorn.
Some minor conflicts with Daniel's pci_controller_ops work.
Conflicts:
arch/powerpc/include/asm/machdep.h
arch/powerpc/platforms/powernv/pci-ioda.c
If a pci_controller_ops struct is provided to iommu_init_early_dart,
populate that with the DMA setup ops, rather than ppc_md. If NULL is
provided, populate ppc_md as before.
This also patches the call sites for Maple and Power Mac to pass
NULL, so existing behaviour is preserved.
The benefit of making this optional is that it means we don't have
to change dart, Maple and Power Mac over to the controller_ops
system in one fell swoop.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Previously the iommu_table had the same lifetime as a struct pnv_ioda_pe
and was embedded in it. The pnv_ioda_pe was assigned to a PE on the bootup
stage. Since PEs are based on the hardware layout which is static in the
system, they will never get released. This means the iommu_table in the
pnv_ioda_pe will never get released either.
This no longer works for VF PE. VF PEs are created and released dynamically
when VFs are created and released. So we need to assign pnv_ioda_pe to VF
PEs respectively when VFs are enabled and clean up those resources for VF
PE when VFs are disabled. And iommu_table is one of the resources we need
to handle dynamically.
Current iommu_table is a static field in pnv_ioda_pe, which will face a
problem when freeing it. During the disabling of a VF,
pnv_pci_ioda2_release_dma_pe will call iommu_free_table to release the
iommu_table for this PE. A static iommu_table will fail in
iommu_free_table.
According to these requirement, this patch allocates iommu_table
dynamically.
Signed-off-by: Wei Yang <weiyang@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
After d905c5df9a ("PPC: POWERNV: move iommu_add_device earlier"), the
refcnt on the kobject backing the IOMMU group for a PCI device is
elevated by each call to pci_dma_dev_setup_pSeriesLP() (via
set_iommu_table_base_and_group). When we go to dlpar a multi-function
PCI device out:
iommu_reconfig_notifier ->
iommu_free_table ->
iommu_group_put
BUG_ON(tbl->it_group)
We trip this BUG_ON, because there are still references on the table, so
it is not freed. Fix this by moving the powernv bus notifier to common
code and calling it for both powernv and pseries.
Fixes: d905c5df9a ("PPC: POWERNV: move iommu_add_device earlier")
Signed-off-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Tested-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The IOMMU-API gained support for a new iommu_map_sg
function. This causes compile failures on powerpc because
the function name is already globally used there.
This patch renames adds a ppc_ prefix to these functions to
solve the compile problem.
Signed-off-by: Joerg Roedel <jroedel@suse.de>
This patch adds the support for to create a direct iommu "bypass"
window on IODA2 bridges (such as Power8) allowing to bypass iommu
page translation completely for 64-bit DMA capable devices, thus
significantly improving DMA performances.
Additionally, this adds a hook to the struct iommu_table so that
the IOMMU API / VFIO can disable the bypass when external ownership
is requested, since in that case, the device will be used by an
environment such as userspace or a KVM guest which must not be
allowed to bypass translations.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch updates the generic iommu backend code to use the
it_page_shift field to determine the iommu page size instead of
using hardcoded values.
Signed-off-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch adds a it_page_shift field to struct iommu_table and
initiliases it to 4K for all platforms.
Signed-off-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The powerpc iommu uses a hardcoded page size of 4K. This patch changes
the name of the IOMMU_PAGE_* macros to reflect the hardcoded values. A
future patch will use the existing names to support dynamic page
sizes.
Signed-off-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The current implementation of IOMMU on sPAPR does not use iommu_ops
and therefore does not call IOMMU API's bus_set_iommu() which
1) sets iommu_ops for a bus
2) registers a bus notifier
Instead, PCI devices are added to IOMMU groups from
subsys_initcall_sync(tce_iommu_init) which does basically the same
thing without using iommu_ops callbacks.
However Freescale PAMU driver (https://lkml.org/lkml/2013/7/1/158)
implements iommu_ops and when tce_iommu_init is called, every PCI device
is already added to some group so there is a conflict.
This patch does 2 things:
1. removes the loop in which PCI devices were added to groups and
adds explicit iommu_add_device() calls to add devices as soon as they get
the iommu_table pointer assigned to them.
2. moves a bus notifier to powernv code in order to avoid conflict with
the notifier from Freescale driver.
iommu_add_device() and iommu_del_device() are public now.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
pci_iommu_init() and pci_direct_iommu_init() are not referenced anywhere,
so remove them.
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This initializes IOMMU groups based on the IOMMU configuration
discovered during the PCI scan on POWERNV (POWER non virtualized)
platform. The IOMMU groups are to be used later by the VFIO driver,
which is used for PCI pass through.
It also implements an API for mapping/unmapping pages for
guest PCI drivers and providing DMA window properties.
This API is going to be used later by QEMU-VFIO to handle
h_put_tce hypercalls from the KVM guest.
The iommu_put_tce_user_mode() does only a single page mapping
as an API for adding many mappings at once is going to be
added later.
Although this driver has been tested only on the POWERNV
platform, it should work on any platform which supports
TCE tables. As h_put_tce hypercall is received by the host
kernel and processed by the QEMU (what involves calling
the host kernel again), performance is not the best -
circa 220MB/s on 10Gb ethernet network.
To enable VFIO on POWER, enable SPAPR_TCE_IOMMU config
option and configure VFIO as required.
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
At the moment all queues in a multiqueue adapter will serialise
against the IOMMU table lock. This is proving to be a big issue,
especially with 10Gbit ethernet.
This patch creates 4 pools and tries to spread the load across
them. If the table is under 1GB in size we revert back to the
original behaviour of 1 pool and 1 largealloc pool.
We create a hash to map CPUs to pools. Since we prefer interrupts to
be affinitised to primary CPUs, without some form of hashing we are
very likely to end up using the same pool. As an example, POWER7
has 4 way SMT and with 4 pools all primary threads will map to the
same pool.
The largealloc pool is reduced from 1/2 to 1/4 of the space to
partially offset the overhead of breaking the table up into pools.
Some performance numbers were obtained with a Chelsio T3 adapter on
two POWER7 boxes, running a 100 session TCP round robin test.
Performance improved 69% with this patch applied.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Sometimes this is used to hold a simple offset, and sometimes
it is used to hold a pointer. This patch changes it to a union containing
void * and dma_addr_t. get/set accessors are also provided, because it was
getting a bit ugly to get to the actual data.
Signed-off-by: Becky Bruce <beckyb@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
As <asm/iommu.h> doesn't contain any other hardware specific definitions
but only interfaces.
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Geert Uytterhoeven <Geert.Uytterhoeven@sonycom.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Both arch/powerpc/platforms/cell/iommu.c and arch/powerpc/platforms/ps3/mm.c
contain the same Cell IOMMU page table entry definitions. Extract them and move
them to <asm/iommu.h>, while adding a CBE_ prefix.
This also allows them to be used by drivers.
Signed-off-by: Geert Uytterhoeven <Geert.Uytterhoeven@sonycom.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
After the merge of the 32 and 64bit DMA code, dma_direct_ops lost
their map/unmap_single() functions but gained map/unmap_page(). This
caused a problem for Cell because Cell's dma_iommu_fixed_ops called
the dma_direct_ops if the fixed linear mapping was to be used or the
iommu ops if the dynamic window was to be used. So in order to fix
this problem we need to update the 64bit DMA code to use
map/unmap_page.
First, we update the generic IOMMU code so that iommu_map_single()
becomes iommu_map_page() and iommu_unmap_single() becomes
iommu_unmap_page(). Then we propagate these changes up through all
the callers of these two functions and in the process update all the
dma_mapping_ops so that they have map/unmap_page rahter than
map/unmap_single. We can do this because on 64bit there is no HIGHMEM
memory so map/unmap_page ends up performing exactly the same function
as map/unmap_single, just taking different arguments.
This has no affect on drivers because the dma_map_single_attrs() just
ends up calling the map_page() function of the appropriate
dma_mapping_ops and similarly the dma_unmap_single_attrs() calls
unmap_page().
This fixes an oops on Cell blades, which oops on boot without this
because they call dma_direct_ops.map_single, which is NULL.
Signed-off-by: Mark Nelson <markn@au1.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
from include/asm-powerpc. This is the result of a
mkdir arch/powerpc/include/asm
git mv include/asm-powerpc/* arch/powerpc/include/asm
Followed by a few documentation/comment fixups and a couple of places
where <asm-powepc/...> was being used explicitly. Of the latter only
one was outside the arch code and it is a driver only built for powerpc.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Linus Torvalds