This reverts:
commit bee44f294e
Author: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Date: Fri Sep 12 19:42:35 2008 +0900
x86: make GART to respect device's dma_mask about virtual mappings
I wrote the above commit to fix a GART alloc_coherent regression, that
can't handle a device having dma_masks > 24bit < 32bits, introduced by
the alloc_coherent rewrite:
http://lkml.org/lkml/2008/8/12/200
After the alloc_coherent rewrite, GART alloc_coherent tried to
allocate pages with GFP_DMA32. If GART got an address that a device
can't access to, GART mapped the address to a virtual I/O address. But
GART mapping mechanism didn't take account of dma mask, so GART could
use a virtual I/O address that the device can't access to again.
Alan pointed out:
" This is indeed a specific problem found with things like older
AACRAID where control blocks must be below 31bits and the GART
is above 0x80000000. "
The above commit modified GART mapping mechanism to take care of dma
mask. But Andi pointed out, "The GART is somewhere in the 4GB range so
you cannot use it to map anything < 4GB. Also GART is pretty small."
http://lkml.org/lkml/2008/9/12/43
That means it's possible that GART doesn't have virtual I/O address
space that a device can access to. The above commit (to modify GART
mapping mechanism to take care of dma mask) can't fix the regression
reliably so let's avoid making GART more complicated.
We need a solution that always works for dma_masks > 24bit <
32bits. That's how GART worked before the alloc_coherent rewrite.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Acked-by: Alan Cox <alan@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch against tip/x86/iommu virtually reverts
2842e5bf31. But just reverting the
commit breaks AMD IOMMU so this patch also includes some fixes.
The above commit adds new two options to x86 IOMMU generic kernel boot
options, fullflush and nofullflush. But such change that affects all
the IOMMUs needs more discussion (all IOMMU parties need the chance to
discuss it):
http://lkml.org/lkml/2008/9/19/106
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
x86 has set_bit_string() that does the exact same thing that
set_bit_area() in lib/iommu-helper.c does.
This patch exports set_bit_area() in lib/iommu-helper.c as
iommu_area_reserve(), converts GART, Calgary, and AMD IOMMU to use it.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The alloc_coherent implementation for AMD IOMMU currently uses
*dev->dma_mask per default. This patch changes it to prefer
dev->coherent_dma_mask if it is set.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The command buffer release function uses the CMD_BUF_SIZE macro for
get_order. Replace this with iommu->cmd_buf_size which is more reliable
about the actual size of the buffer.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current calculation of the IVHD entry size is hard to read. So move
this code to a seperate function to make it more clear what this
calculation does.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The ctrl variable is only u32 and readl also returns a 32 bit value. So
the cast to u64 is pointless. Remove it with this patch.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The amd_iommu_pd_alloc_bitmap is allocated with a calculated order and
freed with order 1. This is not a bug since the calculated order always
evaluates to 1, but its unclean code. So replace the 1 with the
calculation in the release path.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current calculation is very complicated. This patch replaces it with
a much simpler version.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Remove the memset and use __GFP_ZERO at allocation time instead.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
x86's common alloc_coherent (dma_alloc_coherent in dma-mapping.h) sets
up the gfp flag according to the device dma_mask but AMD IOMMU doesn't
need it for devices that the IOMMU can do virtual mappings for. This
patch avoids unnecessary low zone allocation.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Remove some magic numbers and split the pte_root using standard
functions.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In isolation mode the protection domains for the devices are
preallocated and preassigned. This is bad if a device should be passed
to a virtualization guest because the IOMMU code does not know if it is
in use by a driver. This patch changes the code to assign the device to
the preallocated domain only if there are dma mapping requests for it.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This function determines if the AMD IOMMU implementation is responsible
for a given device. So the DMA layer can get this information from the
driver.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
There is a bit in the device entry to suppress all IO page faults
generated by a device. This bit was set until now because there was no
event logging. Now that there is event logging this patch allows IO page
faults from devices to see them in the kernel log.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The code to log IOMMU events is in place now. So enable event logging
with this patch.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds code for polling and printing out events generated by
the AMD IOMMU.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The AMD IOMMU can generate interrupts for various reasons. This patch
adds the basic interrupt enabling infrastructure to the driver.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We need the pci_dev later anyways to enable MSI for the IOMMU hardware.
So remove the devid pointing to the BDF and replace it with the pci_dev
structure where the IOMMU is implemented.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds the pci_seg field to the amd_iommu structure and fills
it with the corresponding value from the ACPI table.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds the allocation of a event buffer for each AMD IOMMU in
the system. The hardware will log events like device page faults or
other errors to this buffer once this is enabled.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The API definition for dma_alloc_coherent states that the bus address
has to be aligned to the next power of 2 boundary greater than the
allocation size. This is violated by AMD IOMMU so far and this patch
fixes it.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds branch hints to the cecks if a completion_wait is
necessary. The completion_waits in the mapping paths are unlikly because
they will only happen on software implementations of AMD IOMMU which
don't exists today or with lazy IO/TLB flushing when the allocator wraps
around the address space. With lazy IO/TLB flushing the completion_wait
in the unmapping path is unlikely too.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The IO/TLB flushing on every unmaping operation is the most expensive
part in AMD IOMMU code and not strictly necessary. It is sufficient to
do the flush before any entries are reused. This is patch implements
lazy IO/TLB flushing which does exactly this.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The GART currently implements the iommu=[no]fullflush command line
parameters which influence its IO/TLB flushing strategy. This patch
makes these parameters generic so that they can be used by the AMD IOMMU
too.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch moves the invocation of the flushing functions to the
map/unmap helpers because its common code in all dma_ops relevant
mapping/unmapping code.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently AMD IOMMU code triggers a BUG_ON if NULL is passed as the
device. This is inconsistent with other IOMMU implementations.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
gart alloc_coherent need to do virtual mapppings only when an
allocated buffer is not DMA-capable for a device.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
x86's common alloc_coherent (dma_alloc_coherent in dma-mapping.h) sets
up the gfp flag according to the device dma_mask but Calgary doesn't
need it because of virtual mappings. This patch avoids unnecessary low
zone allocation.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Muli Ben-Yehuda <muli@il.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently, GART IOMMU ingores device's dma_mask when it does virtual
mappings. So it could give a device a virtual address that the device
can't access to.
This patch fixes the above problem.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
As well as discard fake accessed bit and dirty bit of EPT.
Signed-off-by: Sheng Yang <sheng.yang@intel.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Accesses to CR4 are intercepted even with Nested Paging enabled. But the code
does not check if the guest wants to do a global TLB flush. So this flush gets
lost. This patch adds the check and the flush to svm_set_cr4.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
This patch introduces a guest TLB flush on every NPF exit in KVM. This fixes
random segfaults and #UD exceptions in the guest seen under some workloads
(e.g. long running compile workloads or tbench). A kernbench run with and
without that fix showed that it has a slowdown lower than 0.5%
Cc: stable@kernel.org
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Avi Kivity <avi@qumranet.com>
On 32-bit, at least the generic nops are fairly reasonable, but the
default nops for 64-bit really look pretty sad, and the P6 nops really do
look better.
So I would suggest perhaps moving the static P6 nop selection into the
CONFIG_X86_64 thing.
The alternative is to just get rid of that static nop selection, and just
have two cases: 32-bit and 64-bit, and just pick obviously safe cases for
them.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Non real IOMMU implemenations (which doesn't do virtual mappings,
e.g. swiotlb, pci-nommu, etc) need to use proper gfp flags and
dma_mask to allocate pages in their own dma_alloc_coherent()
(allocated page need to be suitable for device's coherent_dma_mask).
This patch makes dma_alloc_coherent do this job so that IOMMUs don't
need to take care of it any more.
Real IOMMU implemenataions can simply ignore the gfp flags.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We need to use __GFP_DMA for NULL device argument (fallback_dev) with
pci-nommu. It's a hack for ISA (and some old code) so we need to use
GFP_DMA.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The check to see if dev->dma_mask is NULL in pci-nommu is more
appropriate for dma_alloc_coherent().
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
x86: cpu_init(): fix memory leak when using CPU hotplug
x86: pda_init(): fix memory leak when using CPU hotplug
x86, xen: Use native_pte_flags instead of native_pte_val for .pte_flags
x86: move mtrr cpu cap setting early in early_init_xxxx
x86: delay early cpu initialization until cpuid is done
x86: use X86_FEATURE_NOPL in alternatives
x86: add NOPL as a synthetic CPU feature bit
x86: boot: stub out unimplemented CPU feature words
Exception stacks are allocated each time a CPU is set online.
But the allocated space is never freed. Thus with one CPU hotplug
offline/online cycle there is a memory leak of 24K (6 pages) for
a CPU.
Fix is to allocate exception stacks only once -- when the CPU is
set online for the first time.
Signed-off-by: Andreas Herrmann <andreas.herrmann3@amd.com>
Cc: akpm@linux-foundation.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
pda->irqstackptr is allocated whenever a CPU is set online.
But it is never freed. This results in a memory leak of 16K
for each CPU offline/online cycle.
Fix is to allocate pda->irqstackptr only once.
Signed-off-by: Andreas Herrmann <andreas.herrmann3@amd.com>
Cc: akpm@linux-foundation.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Using native_pte_val triggers the BUG_ON() in the paravirt_ops
version of pte_flags().
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Acked-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Krzysztof Helt found MTRR is not detected on k6-2
root cause:
we moved mtrr_bp_init() early for mtrr trimming,
and in early_detect we only read the CPU capability from cpuid,
so some cpu doesn't have that bit in cpuid.
So we need to add early_init_xxxx to preset those bit before mtrr_bp_init
for those earlier cpus.
this patch is for v2.6.27
Reported-by: Krzysztof Helt <krzysztof.h1@wp.pl>
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
