An issue was introduced with "cxl: Unmap MMIO regions when detaching a
context" (b123429e6a) where closing a
context normally could also unmap the problem state area of other
contexts currently using the AFU.
It was also discovered that after a context's MMIO space had been
unmapped it would read 0s when accessing it, whereas the expected
behaviour was for the access to fail altogether.
In order to address these issues, this patch does two things:
- Forced mmap unmapping is only done when we are forcefully detaching
all contexts, and not in the normal detach path. Since the normal
context close path is tied to the file release any mmaps must have
already been released so we don't need to worry in that case.
- The mmap path now uses a vm_operations_struct with a fault handler.
The fault handler ensures that the context is in started state,
otherwise it fails the access attempt with a SIGBUS.
Fixes: b123429e6a ("cxl: Unmap MMIO regions when detaching a context")
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When we deactivate the AFU directed mode we free the scheduled process
area, but did not clear the register in the hardware that has a pointer
to it.
This should be fine since we will have already cleared out every context
and we won't do anything that would cause the hardware to access it
until after we have allocated a new one, but just to be safe this patch
clears out the register when we free the page.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Upon inspection of the implementation specific registers, it was
discovered that the high bit of the implementation specific RXCTL
register was enabled, which enables the DEADB00F debug feature.
The debug feature causes MMIO reads to a disabled AFU to respond with
0xDEADB00F instead of all Fs. In general this should not be visible as
the kernel will only allow MMIO access to enabled AFUs, but there may be
some circumstances where an AFU may become disabled while it is use.
One such case would be an AFU designed to only be used in the dedicated
process mode and to disable itself after it has completed it's work
(however even in that case the effects of this debug flag would be
limited as the userspace application must have completed any required
MMIO accesses before the AFU disables itself with or without the flag).
This patch removes the debug flag and replaces the magic value
programmed into this register with a preprocessor define so it is
clearer what the rest of this initialisation does.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If a context is being detached and we get a translation fault for it
there is little point getting it's mm and handling the fault, so just
respond with an address error and return earlier.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In this particular error path we have already allocated the AFU
interrupts, but have not yet set the status to STARTED. The detach
context code will only attempt to release the interrupts if the context
is in state STARTED, so in this case the interrupts would remain
allocated.
This patch releases the AFU interrupts immediately if the attach call
fails to prevent them leaking.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If we need to force detach a context (e.g. due to EEH or simply force
unbinding the driver) we should prevent the userspace contexts from
being able to access the Problem State Area MMIO region further, which
they may have mapped with mmap().
This patch unmaps any mapped MMIO regions when detaching a userspace
context.
Cc: stable@vger.kernel.org
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In the event that something goes wrong in the hardware and it is unable
to complete a process element comment we would end up polling forever,
effectively making the associated process unkillable.
This patch adds a timeout to the process element command code path, so
that we will give up if the hardware does not respond in a reasonable
time.
Cc: stable@vger.kernel.org
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We had a known sleep while atomic bug if a CXL device was forcefully
unbound while it was in use. This could occur as a result of EEH, or
manually induced with something like this while the device was in use:
echo 0000:01:00.0 > /sys/bus/pci/drivers/cxl-pci/unbind
The issue was that in this code path we iterated over each context and
forcefully detached it with the contexts_lock spin lock held, however
the detach also needed to take the spu_mutex, and call schedule.
This patch changes the contexts_lock to a mutex so that we are not in
atomic context while doing the detach, thereby avoiding the sleep while
atomic.
Also delete the related TODO comment, which suggested an alternate
solution which turned out to not be workable.
Cc: stable@vger.kernel.org
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
upatepp can get called for a nohpte fault when we find from the linux
page table that the translation was hashed before. In that case
we are sure that there is no existing translation, hence we could
avoid doing tlbie.
We could possibly race with a parallel fault filling the TLB. But
that should be ok because updatepp is only ever relaxing permissions.
We also look at linux pte permission bits when filling hash pte
permission bits. We also hold the linux pte busy bits while
inserting/updating a hashpte entry, hence a paralle update of
linux pte is not possible. On the other hand mprotect involves
ptep_modify_prot_start which cause a hpte invalidate and not updatepp.
Performance number:
We use randbox_access_bench written by Anton.
Kernel with THP disabled and smaller hash page table size.
86.60% random_access_b [kernel.kallsyms] [k] .native_hpte_updatepp
2.10% random_access_b random_access_bench [.] doit
1.99% random_access_b [kernel.kallsyms] [k] .do_raw_spin_lock
1.85% random_access_b [kernel.kallsyms] [k] .native_hpte_insert
1.26% random_access_b [kernel.kallsyms] [k] .native_flush_hash_range
1.18% random_access_b [kernel.kallsyms] [k] .__delay
0.69% random_access_b [kernel.kallsyms] [k] .native_hpte_remove
0.37% random_access_b [kernel.kallsyms] [k] .clear_user_page
0.34% random_access_b [kernel.kallsyms] [k] .__hash_page_64K
0.32% random_access_b [kernel.kallsyms] [k] fast_exception_return
0.30% random_access_b [kernel.kallsyms] [k] .hash_page_mm
With Fix:
27.54% random_access_b random_access_bench [.] doit
22.90% random_access_b [kernel.kallsyms] [k] .native_hpte_insert
5.76% random_access_b [kernel.kallsyms] [k] .native_hpte_remove
5.20% random_access_b [kernel.kallsyms] [k] fast_exception_return
5.12% random_access_b [kernel.kallsyms] [k] .__hash_page_64K
4.80% random_access_b [kernel.kallsyms] [k] .hash_page_mm
3.31% random_access_b [kernel.kallsyms] [k] data_access_common
1.84% random_access_b [kernel.kallsyms] [k] .trace_hardirqs_on_caller
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently all interrupts generated by cxl are named "cxl". This is not very
informative as we can't distinguish between cards, AFUs, error interrupts, user
contexts and user interrupts numbers. Being able to distinguish them is useful
for setting affinity.
This patch gives each of these names in /proc/interrupts.
A two card CAPI system, with afu0.0 having 2 active contexts each with 4 user
IRQs each, will now look like this:
% grep cxl /proc/interrupts
444: 0 OPAL ICS 141312 Level cxl-card1-err
445: 0 OPAL ICS 141313 Level cxl-afu1.0-err
446: 0 OPAL ICS 141314 Level cxl-afu1.0
462: 0 OPAL ICS 2052 Level cxl-afu0.0-pe0-1
463: 75517 OPAL ICS 2053 Level cxl-afu0.0-pe0-2
468: 0 OPAL ICS 2054 Level cxl-afu0.0-pe0-3
469: 0 OPAL ICS 2055 Level cxl-afu0.0-pe0-4
470: 0 OPAL ICS 2056 Level cxl-afu0.0-pe1-1
471: 75506 OPAL ICS 2057 Level cxl-afu0.0-pe1-2
472: 0 OPAL ICS 2058 Level cxl-afu0.0-pe1-3
473: 0 OPAL ICS 2059 Level cxl-afu0.0-pe1-4
502: 1066 OPAL ICS 2050 Level cxl-afu0.0
514: 0 OPAL ICS 2048 Level cxl-card0-err
515: 0 OPAL ICS 2049 Level cxl-afu0.0-err
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If an AFU has a hardware bug that causes it to acknowledge a context
terminate or remove while that context has outstanding transactions, it
is possible for the kernel to receive an interrupt for that context
after we have removed it from the context list.
The kernel will not be able to demultiplex the interrupt (or worse - if
we have already reallocated the process handle we could mis-attribute it
to the new context), and printed a big scary warning.
It did not acknowledge the interrupt, which would effectively halt
further translation fault processing on the PSL.
This patch makes the warning clearer about the likely cause of the issue
(i.e. hardware bug) to make it obvious to future AFU designers of what
needs to be fixed. It also prints out the process handle which can then
be matched up with hardware and software traces for debugging.
It also acknowledges the interrupt to the PSL with either an address
error or acknowledge, so that the PSL can continue with other
translations.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In certain circumstances the PSL (Power Service Layer, which provides
translation services for CXL hardware) can send an interrupt for a
segment miss that the kernel has already handled. This can happen if
multiple translations for the same segment are queued in the PSL before
the kernel has restarted the first translation.
The CXL driver does not expect this situation and does not check if a
segment had already been handled. This could cause a duplicate segment
table entry which in turn caused a PSL error taking down the card.
This patch fixes the issue by checking for existing entries in the
segment table that match the segment we are trying to insert, so as to
avoid inserting duplicate entries.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This moves the segment table hash calculation from cxl_load_segment()
into find_free_sste() since that is the only place it is actually used.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch simplifies the process of finding a free segment table entry
by disabling the secondary hash. This reduces the number of possible
entries in the segment table for a given address from 16 to 8.
Due to the large segment sizes we use it is extremely unlikely that the
secondary hash would ever have been used in practice, so this should not
have any negative impacts and may even improve performance due to the
reduced number of comparisons that software & hardware need to perform.
This patch clears the SC bit in the hardware's state register
(CXL_PSL_SR_An) to disable the secondary hash in the hardware since we
can no longer fill out entries using it.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If afu_read() returned due to a signal or the AFU file descriptor being
opened non-blocking it would not call finish_wait() before returning,
which could lead to a crash later when something else wakes up the wait
queue.
This patch restructures the wait logic to ensure that the cleanup is
done correctly.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This is the core of the cxl driver.
It adds support for using cxl cards in the powernv environment only (ie POWER8
bare metal). It allows access to cxl accelerators by userspace using the
/dev/cxl/afuM.N char devices.
The kernel driver has no knowledge of the function implemented by the
accelerator. It provides services to userspace via the /dev/cxl/afuM.N
devices. When a program opens this device and runs the start work IOCTL, the
accelerator will have coherent access to that processes memory using the same
virtual addresses. That process may mmap the device to access any MMIO space
the accelerator provides. Also, reads on the device will allow interrupts to
be received. These services are further documented in a later patch in
Documentation/powerpc/cxl.txt.
Documentation of the cxl hardware architecture and userspace API is provided in
subsequent patches.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This adds the base cxl support that cannot be built as a module. Specifically
it adds the cxl callbacks that are called from the core powerpc mm code which
must always exist irrespective of if the cxl module is loaded or not. This is
similar to how cell works with CONFIG_SPU_BASE.
This adds a cxl_slbia() call (similar to spu_flush_all_slbs()) which checks if
the cxl module is loaded and in use, returning immediately if it is not. If it
is in use it calls into the cxl SLB invalidation code.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Ian Munsie