Replace the bare-bones h8300 direct dma mapping implementation with
the fully featured generic dma-direct one.
Signed-off-by: Christoph Hellwig <hch@lst.de>
cris currently has an incomplete direct mapping dma_map_ops implementation
if PCI support is enabled. Replace it with the fully feature generic
dma-direct implementation.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Jesper Nilsson <jesper.nilsson@axis.com>
So that they don't need to indirect through the operation vector.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Vladimir Murzin <vladimir.murzin@arm.com>
If an attempt to allocate memory succeeded, but isn't inside the
supported DMA mask, retry the allocation with GFP_DMA set as a
last resort.
Based on the x86 code, but an off by one error in what is now
dma_coherent_ok has been fixed vs the x86 code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
This allows to dip into zones for lower memory if they are available.
If one of the zones is not available the corresponding GFP_* flag
will evaluate to 0 so they won't change anything. We provide an
arch tunable for those architectures that do not use GFP_DMA for
the lowest 24-bits, given that there are a few.
Roughly based on the x86 code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
This means it uses whatever linear remapping scheme that the architecture
provides is used in the generic dma_direct ops.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Vladimir Murzin <vladimir.murzin@arm.com>
The trivial direct mapping implementation already does a virtual to
physical translation which isn't strictly a noop, and will soon learn
to do non-direct but linear physical to dma translations through the
device offset and a few small tricks. Rename it to a better fitting
name.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Vladimir Murzin <vladimir.murzin@arm.com>
For architectures that just use the generic dma_noop_ops we can provide
a generic version of dma-mapping.h.
Signed-off-by: Christoph Hellwig <hch@lst.de>
To implement the x86 forbid_dac and iommu_sac_force we want an arch hook
so that it can apply the global options across all dma_map_ops
implementations.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Lift the code from x86 so that we behave consistently. In the future we
should probably warn if any of these is set.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Jesper Nilsson <jesper.nilsson@axis.com>
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org> [m68k]
When creating a new radio on the fly, hwsim allows this
to be done with an arbitrary number of channels, but
cfg80211 only supports a limited number of simultaneous
channels, leading to a warning.
Fix this by validating the number - this requires moving
the define for the maximum out to a visible header file.
Reported-by: syzbot+8dd9051ff19940290931@syzkaller.appspotmail.com
Fixes: b59ec8dd43 ("mac80211_hwsim: fix number of channels in interface combinations")
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
These days all devices should have a DMA coherent mask, and most dma_ops
implementations rely on that fact. But just to be sure add an assert to
ring the warning bell if that is not the case.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Vladimir Murzin <vladimir.murzin@arm.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
When closing multiple wmediumd instances with many radios and try to
unload the mac80211_hwsim module, it may happen that the work items live
longer than the module. To wait especially for this deletion work items,
add a work queue, otherwise flush_scheduled_work would be necessary.
Signed-off-by: Benjamin Beichler <benjamin.beichler@uni-rostock.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
As ieee80211_bss_get_ie() derefences an RCU to return ssid_ie, both
the call to this function and any operation on this variable need
protection by the RCU read lock.
Fixes: 44905265bc ("nl80211: don't expose wdev->ssid for most interfaces")
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Paul reported that he got a report about undefined behaviour
that seems to me to originate in using uninitialized memory
when the channel structure here is used in the event code in
nl80211 later.
He never reported whether this fixed it, and I wasn't able
to trigger this so far, but we should do the right thing and
fully initialize the on-stack structure anyway.
Reported-by: Paul Menzel <pmenzel+linux-wireless@molgen.mpg.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Sparse warns that 'sprd_dma_prep_dma_memcpy' should be static so make it
static.
drivers/dma/sprd-dma.c:713:32: warning:
symbol'sprd_dma_prep_dma_memcpy' was not declared. Should it be static?
Reviewed-by: Baolin Wang <baolin.wang@linaro.org>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>
Booting Linux on a mx6q based board leads to the following warning:
(NULL device *): hwmon_device_register() is deprecated. Please convert the
driver to use hwmon_device_register_with_info().
, so do the conversion as suggested.
Also, this results in the core taking care of creating the 'name'
attribute, so drop the code doing that from the thermal driver.
The initial attempt to convert this driver to
hwmon_device_register_with_info() caused issues on the N900 platform
in commit 7611fb6806 ("thermal: thermal_hwmon: Convert to
hwmon_device_register_with_info()"):
bq27xxx-battery 2-0055: failed to register battery
bq27xxx-battery: probe of 2-0055 failed with error -22
...
rx51-battery: probe of n900-battery failed with error -22
, leading to a revert in commit 3feb479cea ("Revert "thermal:
thermal_hwmon: Convert to hwmon_device_register_with_info()"").
The probe errors happened due to the '-' character being present in
the name of the power supply devices: bq27200-0 and rx51-battery.
Since commit 74d3b64197 ("hwmon: Relax name attribute validation
for new APIs") hwmon will no longer treat these names as errors,
allowing the transition for hwmon_device_register_with_info() to
happen in a safely manner.
Cc: Pavel Machek <pavel@ucw.cz>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Fabio Estevam <fabio.estevam@nxp.com>
Acked-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Zhang Rui <rui.zhang@intel.com>
When reading the reply from the server, insert an explicit
cond_resched() to avoid starving higher priority tasks.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Since nfs4_create_referral_server was the only call site of
rpc_protocol, rpc_protocol can now be removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
After traversing a referral or recovering from a migration event,
ensure that the server port reported in /proc/mounts is updated
to the correct port setting for the new submount.
Reported-by: Helen Chao <helen.chao@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Helen Chao <helen.chao@oracle.com> noticed that when a user
traverses a referral on an NFS/RDMA mount, the resulting submount
always uses TCP.
This behavior does not match the vers= setting when traversing
a referral (vers=4.1 is preserved). It also does not match the
behavior of crossing from the pseudofs into a real filesystem
(proto=rdma is preserved in that case).
The Linux NFS client does not currently support the
fs_locations_info attribute. The situation is similar for all
NFSv4 servers I know of. Therefore until the community has broad
support for fs_locations_info, when following a referral:
- First try to connect with RPC-over-RDMA. This will fail quickly
if the client has no RDMA-capable interfaces.
- If connecting with RPC-over-RDMA fails, or the RPC-over-RDMA
transport is not available, use TCP.
Reported-by: Helen Chao <helen.chao@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The NFS/RDMA port assignment is specified in Section 9 of RFC 8267.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
atomic_t variables are currently used to implement reference
counters with the following properties:
- counter is initialized to 1 using atomic_set()
- a resource is freed upon counter reaching zero
- once counter reaches zero, its further
increments aren't allowed
- counter schema uses basic atomic operations
(set, inc, inc_not_zero, dec_and_test, etc.)
Such atomic variables should be converted to a newly provided
refcount_t type and API that prevents accidental counter overflows
and underflows. This is important since overflows and underflows
can lead to use-after-free situation and be exploitable.
The variable nlm_rqst.a_count is used as pure reference counter.
Convert it to refcount_t and fix up the operations.
**Important note for maintainers:
Some functions from refcount_t API defined in lib/refcount.c
have different memory ordering guarantees than their atomic
counterparts.
The full comparison can be seen in
https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon
in state to be merged to the documentation tree.
Normally the differences should not matter since refcount_t provides
enough guarantees to satisfy the refcounting use cases, but in
some rare cases it might matter.
Please double check that you don't have some undocumented
memory guarantees for this variable usage.
For the nlm_rqst.a_count it might make a difference
in following places:
- nlmclnt_release_call() and nlmsvc_release_call(): decrement
in refcount_dec_and_test() only
provides RELEASE ordering and control dependency on success
vs. fully ordered atomic counterpart
Suggested-by: Kees Cook <keescook@chromium.org>
Reviewed-by: David Windsor <dwindsor@gmail.com>
Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com>
Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
atomic_t variables are currently used to implement reference
counters with the following properties:
- counter is initialized to 1 using atomic_set()
- a resource is freed upon counter reaching zero
- once counter reaches zero, its further
increments aren't allowed
- counter schema uses basic atomic operations
(set, inc, inc_not_zero, dec_and_test, etc.)
Such atomic variables should be converted to a newly provided
refcount_t type and API that prevents accidental counter overflows
and underflows. This is important since overflows and underflows
can lead to use-after-free situation and be exploitable.
The variable nlm_lockowner.count is used as pure reference counter.
Convert it to refcount_t and fix up the operations.
**Important note for maintainers:
Some functions from refcount_t API defined in lib/refcount.c
have different memory ordering guarantees than their atomic
counterparts.
The full comparison can be seen in
https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon
in state to be merged to the documentation tree.
Normally the differences should not matter since refcount_t provides
enough guarantees to satisfy the refcounting use cases, but in
some rare cases it might matter.
Please double check that you don't have some undocumented
memory guarantees for this variable usage.
For the nlm_lockowner.count it might make a difference
in following places:
- nlm_put_lockowner(): decrement in refcount_dec_and_lock() only
provides RELEASE ordering, control dependency on success and
holds a spin lock on success vs. fully ordered atomic counterpart.
No changes in spin lock guarantees.
Suggested-by: Kees Cook <keescook@chromium.org>
Reviewed-by: David Windsor <dwindsor@gmail.com>
Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com>
Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
atomic_t variables are currently used to implement reference
counters with the following properties:
- counter is initialized to 1 using atomic_set()
- a resource is freed upon counter reaching zero
- once counter reaches zero, its further
increments aren't allowed
- counter schema uses basic atomic operations
(set, inc, inc_not_zero, dec_and_test, etc.)
Such atomic variables should be converted to a newly provided
refcount_t type and API that prevents accidental counter overflows
and underflows. This is important since overflows and underflows
can lead to use-after-free situation and be exploitable.
The variable nsm_handle.sm_count is used as pure reference counter.
Convert it to refcount_t and fix up the operations.
**Important note for maintainers:
Some functions from refcount_t API defined in lib/refcount.c
have different memory ordering guarantees than their atomic
counterparts.
The full comparison can be seen in
https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon
in state to be merged to the documentation tree.
Normally the differences should not matter since refcount_t provides
enough guarantees to satisfy the refcounting use cases, but in
some rare cases it might matter.
Please double check that you don't have some undocumented
memory guarantees for this variable usage.
For the nsm_handle.sm_count it might make a difference
in following places:
- nsm_release(): decrement in refcount_dec_and_lock() only
provides RELEASE ordering, control dependency on success
and holds a spin lock on success vs. fully ordered atomic
counterpart. No change for the spin lock guarantees.
Suggested-by: Kees Cook <keescook@chromium.org>
Reviewed-by: David Windsor <dwindsor@gmail.com>
Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com>
Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
atomic_t variables are currently used to implement reference
counters with the following properties:
- counter is initialized to 1 using atomic_set()
- a resource is freed upon counter reaching zero
- once counter reaches zero, its further
increments aren't allowed
- counter schema uses basic atomic operations
(set, inc, inc_not_zero, dec_and_test, etc.)
Such atomic variables should be converted to a newly provided
refcount_t type and API that prevents accidental counter overflows
and underflows. This is important since overflows and underflows
can lead to use-after-free situation and be exploitable.
The variable nlm_host.h_count is used as pure reference counter.
Convert it to refcount_t and fix up the operations.
**Important note for maintainers:
Some functions from refcount_t API defined in lib/refcount.c
have different memory ordering guarantees than their atomic
counterparts.
The full comparison can be seen in
https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon
in state to be merged to the documentation tree.
Normally the differences should not matter since refcount_t provides
enough guarantees to satisfy the refcounting use cases, but in
some rare cases it might matter.
Please double check that you don't have some undocumented
memory guarantees for this variable usage.
For the nlm_host.h_count it might make a difference
in following places:
- nlmsvc_release_host(): decrement in refcount_dec()
provides RELEASE ordering, while original atomic_dec()
was fully unordered. Since the change is for better, it
should not matter.
- nlmclnt_release_host(): decrement in refcount_dec_and_test() only
provides RELEASE ordering and control dependency on success
vs. fully ordered atomic counterpart. It doesn't seem to
matter in this case since object freeing happens under mutex
lock anyway.
Suggested-by: Kees Cook <keescook@chromium.org>
Reviewed-by: David Windsor <dwindsor@gmail.com>
Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com>
Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Currently when falling back to doing I/O through the MDS (via
pnfs_{read|write}_through_mds), the client frees the nfs_pgio_header
without releasing the reference taken on the dreq
via pnfs_generic_pg_{read|write}pages -> nfs_pgheader_init ->
nfs_direct_pgio_init. It then takes another reference on the dreq via
nfs_generic_pg_pgios -> nfs_pgheader_init -> nfs_direct_pgio_init and
as a result the requester will become stuck in inode_dio_wait. Once
that happens, other processes accessing the inode will become stuck as
well.
Ensure that pnfs_read_through_mds() and pnfs_write_through_mds() clean
up correctly by calling hdr->completion_ops->completion() instead of
calling hdr->release() directly.
This can be reproduced (sometimes) by performing "storage failover
takeover" commands on NetApp filer while doing direct I/O from a client.
This can also be reproduced using SystemTap to simulate a failure while
doing direct I/O from a client (from Dave Wysochanski
<dwysocha@redhat.com>):
stap -v -g -e 'probe module("nfs_layout_nfsv41_files").function("nfs4_fl_prepare_ds").return { $return=NULL; exit(); }'
Suggested-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Scott Mayhew <smayhew@redhat.com>
Fixes: 1ca018d28d ("pNFS: Fix a memory leak when attempted pnfs fails")
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
PNFS block/SCSI layouts should gracefully handle cases where block devices
are not available when a layout is retrieved, or the block devices are
removed while the client holds a layout.
While setting up a layout segment, keep a record of an unavailable or
un-parsable block device in cache with a flag so that subsequent layouts do
not spam the server with GETDEVINFO. We can reuse the current
NFS_DEVICEID_UNAVAILABLE handling with one variation: instead of reusing
the device, we will discard it and send a fresh GETDEVINFO after the
timeout, since the lookup and validation of the device occurs within the
GETDEVINFO response handling.
A lookup of a layout segment that references an unavailable device will
return a segment with the NFS_LSEG_UNAVAILABLE flag set. This will allow
the pgio layer to mark the layout with the appropriate fail bit, which
forces subsequent IO to the MDS, and prevents spamming the server with
LAYOUTGET, LAYOUTRETURN.
Finally, when IO to a block device fails, look up the block device(s)
referenced by the pgio header, and mark them as unavailable.
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
If there's an error doing I/O to block device, and the client resends the
I/O to the MDS, the MDS must recall the layout from the client before
processing the I/O. Let's preempt that exchange by returning the layout
before falling back to the MDS when there's an error.
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The blocklayout module contains the client support for both block and SCSI
layouts. Add a module alias for the SCSI layout type so that the module
will be loaded for SCSI layouts.
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
nfs_pgio_rpcsetup() is always called with an offset of 0, so we should be
able to drop the arguement altogether.
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
There are 2 comments in the NFSv4 code which suggest that
SIGLOST should possibly be sent to a process. In these
cases a lock has been lost.
The current practice is to set NFS_LOCK_LOST so that
read/write returns EIO when a lock is lost.
So change these comments to code when sets NFS_LOCK_LOST.
One case is when lock recovery after apparent server restart
fails with NFS4ERR_DENIED, NFS4ERR_RECLAIM_BAD, or
NFS4ERRO_RECLAIM_CONFLICT. The other case is when a lock
attempt as part of lease recovery fails with NFS4ERR_DENIED.
In an ideal world, these should not happen. However I have
a packet trace showing an NFSv4.1 session getting
NFS4ERR_BADSESSION after an extended network parition. The
NFSv4.1 client treats this like server reboot until/unless
it get NFS4ERR_NO_GRACE, in which case it switches over to
"nograce" recovery mode. In this network trace, the client
attempts to recover a lock and the server (incorrectly)
reports NFS4ERR_DENIED rather than NFS4ERR_NO_GRACE. This
leads to the ineffective comment and the client then
continues to write using the OPEN stateid.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
This code can never be used as the IS_AUTOMOUNT(inode)
case has already been handled.
So remove it to avoid confusion.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Support the query flags AT_STATX_FORCE_SYNC by forcing an attribute
revalidation, and AT_STATX_DONT_SYNC by returning cached attributes
only.
Use the mask to optimise away server revalidation for attributes
that are not being requested by the user.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The LOOKUPP operation was inserted into the nfs4_procedures array
rather than being appended, which put /proc/net/rpc/nfs out of
whack, and broke the nfsstat utility.
Fix by moving the LOOKUPP operation to the end of the array, and
by ensuring that it keeps the same length whether or not NFSV4.1
and NFSv4.2 are compiled in.
Fixes: 5b5faaf6df ("nfs4: add NFSv4 LOOKUPP handlers")
Cc: stable@vger.kernel.org # v4.13+
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Convert CLOSE so that it specifies the correct stateid and
inode for the error handling.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Convert CLOSE so that it specifies the correct stateid, state and
inode for the error handling.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The commit list can get very large, and so we need a cond_resched()
in nfs_commit_release_pages() in order to ensure we don't hog the CPU
for excessive periods of time.
Reported-by: Mike Galbraith <efault@gmx.de>
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
One of the limitations of pm_runtime_force_suspend/resume() is that
if a parent driver wants to use these functions, all of its child
drivers generally have to do that too because of the parent usage
counter manipulations necessary to get the correct state of the parent
during system-wide transitions to the working state (system resume).
However, that limitation turns out to be artificial, so remove it.
Namely, pm_runtime_force_suspend() only needs to update the children
counter of its parent (if there's is a parent) when the device can
stay in suspend after the subsequent system resume transition, as
that counter is correct already otherwise. Now, if the parent's
children counter is not updated, it is not necessary to increment
the parent's usage counter in that case any more, as long as the
children counters of devices are checked along with their usage
counters in order to decide whether or not the devices may be left
in suspend after the subsequent system resume transition.
Accordingly, modify pm_runtime_force_suspend() to only call
pm_runtime_set_suspended() for devices whose usage and children
counters are at the "no references" level (the runtime PM status
of the device needs to be updated to "suspended" anyway in case
this function is called once again for the same device during the
transition under way), drop the parent usage counter incrementation
from it and update pm_runtime_force_resume() to compensate for these
changes.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
There are problems with calling pm_runtime_force_suspend/resume()
to "stop" and "start" devices in genpd_finish_suspend() and
genpd_resume_noirq() (and in analogous hibernation-specific genpd
callbacks) after commit 122a22377a (PM / Domains: Stop/start
devices during system PM suspend/resume in genpd) as those routines
do much more than just "stopping" and "starting" devices (which was
the stated purpose of that commit) unnecessarily and may not play
well with system-wide PM driver callbacks.
First, consider the pm_runtime_force_suspend() in
genpd_finish_suspend(). If the current runtime PM status of the
device is "suspended", that function most likely does the right thing
by ignoring the device, because it should have been "stopped" already
and whatever needed to be done to deactivate it shoud have been done.
In turn, if the runtime PM status of the device is "active",
genpd_runtime_suspend() is called for it (indirectly) and (1) runs
the ->runtime_suspend callback provided by the device's driver
(assuming no bus type with ->runtime_suspend of its own), (2) "stops"
the device and (3) checks if the domain can be powered down, and then
(4) the device's runtime PM status is changed to "suspended". Out of
the four actions above (1) is not necessary and it may be outright
harmful, (3) is pointless and (4) is questionable. The only
operation that needs to be carried out here is (2).
The reason why (1) is not necessary is because the system-wide
PM callbacks provided by the device driver for the transition in
question have been run and they should have taken care of the
driver's part of device suspend already. Moreover, it may be
harmful, because the ->runtime_suspend callback may want to
access the device which is partially suspended at that point
and may not be responsive. Also, system-wide PM callbacks may
have been run already (in the previous phases of the system
transition under way) for the device's parent or for its supplier
devices (if any) and the device may not be accessible because of
that.
There also is no reason to do (3), because genpd_finish_suspend()
will repeat it anyway, and (4) potentially causes confusion to ensue
during the subsequent system transition to the working state.
Consider pm_runtime_force_resume() in genpd_resume_noirq() now.
It runs genpd_runtime_resume() for all devices with runtime PM
status set to "suspended", which includes all of the devices
whose runtime PM status was changed by pm_runtime_force_suspend()
before and may include some devices already suspended when the
pm_runtime_force_suspend() was running, which may be confusing. The
genpd_runtime_resume() first tries to power up the domain, which
(again) is pointless, because genpd_resume_noirq() has done that
already. Then, it "starts" the device and runs the ->runtime_resume
callback (from the driver, say) for it. If all is well, the device
is left with the runtime PM status set to "active".
Unfortunately, running the driver's ->runtime_resume callback
before its system-wide PM callbacks and possibly before some
system-wide PM callbacks of the parent device's driver (let
alone supplier drivers) is asking for trouble, especially if
the device had been suspended before pm_runtime_force_suspend()
ran previously or if the callbacks in question expect to be run
back-to-back with their suspend-side counterparts. It also should
not be necessary, because the system-wide PM driver callbacks that
will be invoked for the device subsequently should take care of
resuming it just fine.
[Running the driver's ->runtime_resume callback in the "noirq"
phase of the transition to the working state may be problematic
even for devices whose drivers do use pm_runtime_force_resume()
in (or as) their system-wide PM callbacks if they have suppliers
other than their parents, because it may cause the supplier to
be resumed after the consumer in some cases.]
Because of the above, modify genpd as follows:
1. Change genpd_finish_suspend() to only "stop" devices with
runtime PM status set to "active" (without invoking runtime PM
callbacks for them, changing their runtime PM status and so on).
That doesn't change the handling of devices whose drivers use
pm_runtime_force_suspend/resume() in (or as) their system-wide
PM callbacks and addresses the issues described above for the
other devices.
2. Change genpd_resume_noirq() to only "start" devices with
runtime PM status set to "active" (without invoking runtime PM
callbacks for them, changing their runtime PM status and so on).
Again, that doesn't change the handling of devices whose drivers
use pm_runtime_force_suspend/resume() in (or as) their system-wide
PM callbacks and addresses the described issues for the other
devices. Devices with runtime PM status set to "suspended"
are not started with the assumption that they will be resumed
later, either by pm_runtime_force_resume() or via runtime PM.
3. Change genpd_restore_noirq() to follow genpd_resume_noirq().
That causes devices already suspended before hibernation to be
left alone (which also is the case without the change) and
avoids running the ->runtime_resume driver callback too early
for the other devices.
4. Change genpd_freeze_noirq() and genpd_thaw_noirq() in accordance
with the above modifications.
Fixes: 122a22377a (PM / Domains: Stop/start devices during system PM suspend/resume in genpd)
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Ulf Hansson <ulf.hansson@linaro.org>
The PAUSE instruction is currently used in the retpoline and RSB filling
macros as a speculation trap. The use of PAUSE was originally suggested
because it showed a very, very small difference in the amount of
cycles/time used to execute the retpoline as compared to LFENCE. On AMD,
the PAUSE instruction is not a serializing instruction, so the pause/jmp
loop will use excess power as it is speculated over waiting for return
to mispredict to the correct target.
The RSB filling macro is applicable to AMD, and, if software is unable to
verify that LFENCE is serializing on AMD (possible when running under a
hypervisor), the generic retpoline support will be used and, so, is also
applicable to AMD. Keep the current usage of PAUSE for Intel, but add an
LFENCE instruction to the speculation trap for AMD.
The same sequence has been adopted by GCC for the GCC generated retpolines.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Borislav Petkov <bp@alien8.de>
Acked-by: David Woodhouse <dwmw@amazon.co.uk>
Acked-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Jiri Kosina <jikos@kernel.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Greg Kroah-Hartman <gregkh@linux-foundation.org>
Cc: Kees Cook <keescook@google.com>
Link: https://lkml.kernel.org/r/20180113232730.31060.36287.stgit@tlendack-t1.amdoffice.net
On context switch from a shallow call stack to a deeper one, as the CPU
does 'ret' up the deeper side it may encounter RSB entries (predictions for
where the 'ret' goes to) which were populated in userspace.
This is problematic if neither SMEP nor KPTI (the latter of which marks
userspace pages as NX for the kernel) are active, as malicious code in
userspace may then be executed speculatively.
Overwrite the CPU's return prediction stack with calls which are predicted
to return to an infinite loop, to "capture" speculation if this
happens. This is required both for retpoline, and also in conjunction with
IBRS for !SMEP && !KPTI.
On Skylake+ the problem is slightly different, and an *underflow* of the
RSB may cause errant branch predictions to occur. So there it's not so much
overwrite, as *filling* the RSB to attempt to prevent it getting
empty. This is only a partial solution for Skylake+ since there are many
other conditions which may result in the RSB becoming empty. The full
solution on Skylake+ is to use IBRS, which will prevent the problem even
when the RSB becomes empty. With IBRS, the RSB-stuffing will not be
required on context switch.
[ tglx: Added missing vendor check and slighty massaged comments and
changelog ]
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: gnomes@lxorguk.ukuu.org.uk
Cc: Rik van Riel <riel@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: thomas.lendacky@amd.com
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Jiri Kosina <jikos@kernel.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kees Cook <keescook@google.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Greg Kroah-Hartman <gregkh@linux-foundation.org>
Cc: Paul Turner <pjt@google.com>
Link: https://lkml.kernel.org/r/1515779365-9032-1-git-send-email-dwmw@amazon.co.uk
