[PATCH v9 00/13] Fix the on-flight DMA issue on system with amd iommu
From: Baoquan He
Date: Tue Aug 01 2017 - 07:37:42 EST
When kernel panicked and jump into the kdump kernel, DMA started by the
1st kernel is not stopped, this is called on-flight DMA. In the current
code it will disable iommu and build new translation table and attach
device to it. This will cause:
1. IO_PAGE_FAULT warning message can be seen.
2. transfer data to or from incorrect areas of memory.
Sometime it causes the dump failure or kernel hang.
The principle of the fix is to copy the old device table to let the old-flight
DMA continue looking up to get correct address translation and irq remap result,
meanwhile to defer the assignment of device to domain to device driver initializtion
stage. The old domain ids used in 1st kernel are reserved. And a new call-back
is_attach_deferred() is added to iommu-ops, will check whether we need defer the
domain attach/detach in iommu-core code. If defer is needed, just return directly
from amd iommu attach/detach function. The attachment will be done in device driver
initializaiton stage when calling get_domain().
Made changes according to Joerg's reviewing comments and suggestions:
- Check if all IOMMUs are pre-enabled, otherwise do not copy dev table
and just continue as normal kernel does.
- Add a new global old_dev_tbl_cpy to point to a newly allocated device
table. The content of old device table will be copied to the specific
device table for copying which old_dev_tbl_cpy points at. If copy failed
we can still use the amd_iommu_dev_table which is allocated in
early_amd_iommu_init(). This is for better rolling back if copy failed,
the amd_iommu_dev_table has got necessary initialization since iommu init.
- Always allocate device table with GFP_DMA32 flag to make sure that they
are under 4G. This tries to work around the issue mentioned in patch 10/13.
Meanwhile double check if the address of device table is above 4G since
it could be touched accidentally in corrupted 1st kernel and not trustworthy
Rebase patchset v7 on the latest v4.13-rc1.
- And re-enable printing IO_PAGE_FAULT message in kdump kernel.
- Only disable iommu if amd_iommu=off is specified in kdump kernel.
Two main changes are made according to Joerg's suggestion:
- Add is_attach_deferred call-back to iommu-ops. With this domain
can be deferred to device driver init cleanly.
- Allocate memory below 4G for dev table if translation pre-enabled.
AMD engineer pointed out that it's unsafe to update the device-table
while iommu is enabled. device-table pointer update is split up into
two 32bit writes in the IOMMU hardware. So updating it while the IOMMU
is enabled could have some nasty side effects.
According to Joerg's comments made several below main changes:
- Add sanity check when copy old dev tables.
- If a device is set up with guest translations (DTE.GV=1), then don't
copy that information but move the device over to an empty guest-cr3
table and handle the faults in the PPR log (which just answer them
bnx2 NIC can't reset itself during driver init. Post patch to reset
it during driver init. IO_PAGE_FAULT can't be seen anymore.
Below is link of v5 post.
Baoquan He (13):
iommu/amd: Detect pre enabled translation
iommu/amd: add several helper functions
Revert "iommu/amd: Suppress IO_PAGE_FAULTs in kdump kernel"
iommu/amd: Define bit fields for DTE particularly
iommu/amd: Add function copy_dev_tables()
iommu/amd: copy old trans table from old kernel
iommu/amd: Do sanity check for address translation and irq remap of
old dev table entry
iommu: Add is_attach_deferred call-back to iommu-ops
iommu/amd: Use is_attach_deferred call-back
iommu/amd: Allocate memory below 4G for dev table if translation
iommu/amd: Don't copy GCR3 table root pointer
iommu/amd: Clear out the GV flag when handle deferred domain attach
iommu/amd: Disable iommu only if amd_iommu=off is specified
drivers/iommu/amd_iommu.c | 81 ++++++++-------
drivers/iommu/amd_iommu_init.c | 222 ++++++++++++++++++++++++++++++++++++----
drivers/iommu/amd_iommu_proto.h | 2 +
drivers/iommu/amd_iommu_types.h | 55 +++++++++-
drivers/iommu/amd_iommu_v2.c | 18 +++-
drivers/iommu/iommu.c | 8 ++
include/linux/iommu.h | 1 +
7 files changed, 323 insertions(+), 64 deletions(-)