We store 2 multilevel tables in iommu_table - one for the hardware and
one with the corresponding userspace addresses. Before allocating
the tables, the iommu_table_group_ops::get_table_size() hook returns
the combined size of the two and VFIO SPAPR TCE IOMMU driver adjusts
the locked_vm counter correctly. When the table is actually allocated,
the amount of allocated memory is stored in iommu_table::it_allocated_size
and used to decrement the locked_vm counter when we release the memory
used by the table; .get_table_size() and .create_table() calculate it
independently but the result is expected to be the same.
However the allocator does not add the userspace table size to
.it_allocated_size so when we destroy the table because of VFIO PCI
unplug (i.e. VFIO container is gone but the userspace keeps running),
we decrement locked_vm by just a half of size of memory we are
releasing.
To make things worse, since we enabled on-demand allocation of
indirect levels, it_allocated_size contains only the amount of memory
actually allocated at the table creation time which can just be a
fraction. It is not a problem with incrementing locked_vm (as
get_table_size() value is used) but it is with decrementing.
As the result, we leak locked_vm and may not be able to allocate more
IOMMU tables after few iterations of hotplug/unplug.
This sets it_allocated_size in the pnv_pci_ioda2_ops::create_table()
hook to what pnv_pci_ioda2_get_table_size() returns so from now on we
have a single place which calculates the maximum memory a table can
occupy. The original meaning of it_allocated_size is somewhat lost now
though.
We do not ditch it_allocated_size whatsoever here and we do not call
get_table_size() from vfio_iommu_spapr_tce.c when decrementing
locked_vm as we may have multiple IOMMU groups per container and even
though they all are supposed to have the same get_table_size()
implementation, there is a small chance for failure or confusion.
Fixes: 090bad39b2 ("powerpc/powernv: Add indirect levels to it_userspace")
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 iommu_table pointer stored in iommu_table_group may get stale
by accident, this adds referencing and removes a redundant comment
about this.
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 powernv platform maintains 2 TCE tables for VFIO - a hardware TCE
table and a table with userspace addresses; the latter is used for
marking pages dirty when corresponging TCEs are unmapped from
the hardware table.
a68bd1267b ("powerpc/powernv/ioda: Allocate indirect TCE levels
on demand") enabled on-demand allocation of the hardware table,
however it missed the other table so it has still been fully allocated
at the boot time. This fixes the issue by allocating a single level,
just like we do for the hardware table.
Fixes: a68bd1267b ("powerpc/powernv/ioda: Allocate indirect TCE levels on demand")
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>
mpe: This was fixed originally in commit d3d4ffaae4
("powerpc/powernv/ioda2: Reduce upper limit for DMA window size"), but
contrary to what the merge commit says was inadvertently lost by me in
commit ce57c6610c ("Merge branch 'topic/ppc-kvm' into next") which
brought in changes that moved the code to a new file. So reapply it to
the new file.
Original commit message follows:
We use PHB in mode1 which uses bit 59 to select a correct DMA window.
However there is mode2 which uses bits 59:55 and allows up to 32 DMA
windows per a PE.
Even though documentation does not clearly specify that, it seems that
the actual hardware does not support bits 59:55 even in mode1, in
other words we can create a window as big as 1<<58 but DMA simply
won't work.
This reduces the upper limit from 59 to 55 bits to let the userspace
know about the hardware limits.
Fixes: ce57c6610c ("Merge branch 'topic/ppc-kvm' into next")
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
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>
This moves actual pages allocation to a separate function which is going
to be reused later in on-demand TCE allocation.
While we are at it, remove unnecessary level size round up as the caller
does this already.
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 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>
Right now we have allocation code in pci-ioda.c and traversing code in
pci.c, let's keep them toghether. However both files are big enough
already so let's move this business to a new file.
While we at it, move the code which links IOMMU table groups to
IOMMU tables as it is not specific to any PNV PHB model.
These puts exported symbols from the new file together.
This fixes several warnings from checkpatch.pl like this:
"WARNING: Prefer 'unsigned int' to bare use of 'unsigned'".
As this is almost cut-n-paste, there should be no behavioral change.
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>
Alexey Kardashevskiy