[ Upstream commit 2997d94b5dd0e8b10076f5e0b6f18410c73e28bd ]
When writing a task id to the "tasks" file in an rdtgroup,
rdtgroup_tasks_write() treats the pid as a number in the current pid
namespace. But when reading the "tasks" file, rdtgroup_tasks_show() shows
the list of global pids from the init namespace, which is confusing and
incorrect.
To be more robust, let the "tasks" file only show pids in the current pid
namespace.
Fixes: e02737d5b8 ("x86/intel_rdt: Add tasks files")
Signed-off-by: Shawn Wang <shawnwang@linux.alibaba.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Acked-by: Reinette Chatre <reinette.chatre@intel.com>
Acked-by: Fenghua Yu <fenghua.yu@intel.com>
Tested-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/all/20230116071246.97717-1-shawnwang@linux.alibaba.com/
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 7fef099702527c3b2c5234a2ea6a24411485a13a upstream.
The implementation of 'current' on x86 is very intentionally special: it
is a very common thing to look up, and it uses 'this_cpu_read_stable()'
to get the current thread pointer efficiently from per-cpu storage.
And the keyword in there is 'stable': the current thread pointer never
changes as far as a single thread is concerned. Even if when a thread
is preempted, or moved to another CPU, or even across an explicit call
'schedule()' that thread will still have the same value for 'current'.
It is, after all, the kernel base pointer to thread-local storage.
That's why it's stable to begin with, but it's also why it's important
enough that we have that special 'this_cpu_read_stable()' access for it.
So this is all done very intentionally to allow the compiler to treat
'current' as a value that never visibly changes, so that the compiler
can do CSE and combine multiple different 'current' accesses into one.
However, there is obviously one very special situation when the
currently running thread does actually change: inside the scheduler
itself.
So the scheduler code paths are special, and do not have a 'current'
thread at all. Instead there are _two_ threads: the previous and the
next thread - typically called 'prev' and 'next' (or prev_p/next_p)
internally.
So this is all actually quite straightforward and simple, and not all
that complicated.
Except for when you then have special code that is run in scheduler
context, that code then has to be aware that 'current' isn't really a
valid thing. Did you mean 'prev'? Did you mean 'next'?
In fact, even if then look at the code, and you use 'current' after the
new value has been assigned to the percpu variable, we have explicitly
told the compiler that 'current' is magical and always stable. So the
compiler is quite free to use an older (or newer) value of 'current',
and the actual assignment to the percpu storage is not relevant even if
it might look that way.
Which is exactly what happened in the resctl code, that blithely used
'current' in '__resctrl_sched_in()' when it really wanted the new
process state (as implied by the name: we're scheduling 'into' that new
resctl state). And clang would end up just using the old thread pointer
value at least in some configurations.
This could have happened with gcc too, and purely depends on random
compiler details. Clang just seems to have been more aggressive about
moving the read of the per-cpu current_task pointer around.
The fix is trivial: just make the resctl code adhere to the scheduler
rules of using the prev/next thread pointer explicitly, instead of using
'current' in a situation where it just wasn't valid.
That same code is then also used outside of the scheduler context (when
a thread resctl state is explicitly changed), and then we will just pass
in 'current' as that pointer, of course. There is no ambiguity in that
case.
The fix may be trivial, but noticing and figuring out what went wrong
was not. The credit for that goes to Stephane Eranian.
Reported-by: Stephane Eranian <eranian@google.com>
Link: https://lore.kernel.org/lkml/20230303231133.1486085-1-eranian@google.com/
Link: https://lore.kernel.org/lkml/alpine.LFD.2.01.0908011214330.3304@localhost.localdomain/
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Tested-by: Stephane Eranian <eranian@google.com>
Tested-by: Babu Moger <babu.moger@amd.com>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit fe1f0714385fbcf76b0cbceb02b7277d842014fc ]
When the user moves a running task to a new rdtgroup using the task's
file interface or by deleting its rdtgroup, the resulting change in
CLOSID/RMID must be immediately propagated to the PQR_ASSOC MSR on the
task(s) CPUs.
x86 allows reordering loads with prior stores, so if the task starts
running between a task_curr() check that the CPU hoisted before the
stores in the CLOSID/RMID update then it can start running with the old
CLOSID/RMID until it is switched again because __rdtgroup_move_task()
failed to determine that it needs to be interrupted to obtain the new
CLOSID/RMID.
Refer to the diagram below:
CPU 0 CPU 1
----- -----
__rdtgroup_move_task():
curr <- t1->cpu->rq->curr
__schedule():
rq->curr <- t1
resctrl_sched_in():
t1->{closid,rmid} -> {1,1}
t1->{closid,rmid} <- {2,2}
if (curr == t1) // false
IPI(t1->cpu)
A similar race impacts rdt_move_group_tasks(), which updates tasks in a
deleted rdtgroup.
In both cases, use smp_mb() to order the task_struct::{closid,rmid}
stores before the loads in task_curr(). In particular, in the
rdt_move_group_tasks() case, simply execute an smp_mb() on every
iteration with a matching task.
It is possible to use a single smp_mb() in rdt_move_group_tasks(), but
this would require two passes and a means of remembering which
task_structs were updated in the first loop. However, benchmarking
results below showed too little performance impact in the simple
approach to justify implementing the two-pass approach.
Times below were collected using `perf stat` to measure the time to
remove a group containing a 1600-task, parallel workload.
CPU: Intel(R) Xeon(R) Platinum P-8136 CPU @ 2.00GHz (112 threads)
# mkdir /sys/fs/resctrl/test
# echo $$ > /sys/fs/resctrl/test/tasks
# perf bench sched messaging -g 40 -l 100000
task-clock time ranges collected using:
# perf stat rmdir /sys/fs/resctrl/test
Baseline: 1.54 - 1.60 ms
smp_mb() every matching task: 1.57 - 1.67 ms
[ bp: Massage commit message. ]
Fixes: ae28d1aae48a ("x86/resctrl: Use an IPI instead of task_work_add() to update PQR_ASSOC MSR")
Fixes: 0efc89be94 ("x86/intel_rdt: Update task closid immediately on CPU in rmdir and unmount")
Signed-off-by: Peter Newman <peternewman@google.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Reviewed-by: Babu Moger <babu.moger@amd.com>
Cc: <stable@kernel.org>
Link: https://lore.kernel.org/r/20221220161123.432120-1-peternewman@google.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e0ad6dc8969f790f14bddcfd7ea284b7e5f88a16 ]
James reported in [1] that there could be two tasks running on the same CPU
with task_struct->on_cpu set. Using task_struct->on_cpu as a test if a task
is running on a CPU may thus match the old task for a CPU while the
scheduler is running and IPI it unnecessarily.
task_curr() is the correct helper to use. While doing so move the #ifdef
check of the CONFIG_SMP symbol to be a C conditional used to determine
if this helper should be used to ensure the code is always checked for
correctness by the compiler.
[1] https://lore.kernel.org/lkml/a782d2f3-d2f6-795f-f4b1-9462205fd581@arm.com
Reported-by: James Morse <james.morse@arm.com>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/e9e68ce1441a73401e08b641cc3b9a3cf13fe6d4.1608243147.git.reinette.chatre@intel.com
Stable-dep-of: fe1f0714385f ("x86/resctrl: Fix task CLOSID/RMID update race")
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 64e87d4bd3201bf8a4685083ee4daf5c0d001452 upstream.
domain_add_cpu() is called whenever a CPU is brought online. The
earlier call to domain_setup_ctrlval() allocates the control value
arrays.
If domain_setup_mon_state() fails, the control value arrays are not
freed.
Add the missing kfree() calls.
Fixes: 1bd2a63b4f ("x86/intel_rdt/mba_sc: Add initialization support")
Fixes: edf6fa1c4a ("x86/intel_rdt/cqm: Add RMID (Resource monitoring ID) management")
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Acked-by: Reinette Chatre <reinette.chatre@intel.com>
Cc: <stable@vger.kernel.org>
Link: https://lkml.kernel.org/r/20210917165958.28313-1-james.morse@arm.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 527f721478bce3f49b513a733bacd19d6f34b08c upstream.
The recent commit
064855a69003 ("x86/resctrl: Fix default monitoring groups reporting")
caused a RHEL build failure with an uninitialized variable warning
treated as an error because it removed the default case snippet.
The RHEL Makefile uses '-Werror=maybe-uninitialized' to force possibly
uninitialized variable warnings to be treated as errors. This is also
reported by smatch via the 0day robot.
The error from the RHEL build is:
arch/x86/kernel/cpu/resctrl/monitor.c: In function ‘__mon_event_count’:
arch/x86/kernel/cpu/resctrl/monitor.c:261:12: error: ‘m’ may be used
uninitialized in this function [-Werror=maybe-uninitialized]
m->chunks += chunks;
^~
The upstream Makefile does not build using '-Werror=maybe-uninitialized'.
So, the problem is not seen there. Fix the problem by putting back the
default case snippet.
[ bp: note that there's nothing wrong with the code and other compilers
do not trigger this warning - this is being done just so the RHEL compiler
is happy. ]
Fixes: 064855a69003 ("x86/resctrl: Fix default monitoring groups reporting")
Reported-by: Terry Bowman <Terry.Bowman@amd.com>
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Babu Moger <babu.moger@amd.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/162949631908.23903.17090272726012848523.stgit@bmoger-ubuntu
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 064855a69003c24bd6b473b367d364e418c57625 upstream.
Creating a new sub monitoring group in the root /sys/fs/resctrl leads to
getting the "Unavailable" value for mbm_total_bytes and mbm_local_bytes
on the entire filesystem.
Steps to reproduce:
1. mount -t resctrl resctrl /sys/fs/resctrl/
2. cd /sys/fs/resctrl/
3. cat mon_data/mon_L3_00/mbm_total_bytes
23189832
4. Create sub monitor group:
mkdir mon_groups/test1
5. cat mon_data/mon_L3_00/mbm_total_bytes
Unavailable
When a new monitoring group is created, a new RMID is assigned to the
new group. But the RMID is not active yet. When the events are read on
the new RMID, it is expected to report the status as "Unavailable".
When the user reads the events on the default monitoring group with
multiple subgroups, the events on all subgroups are consolidated
together. Currently, if any of the RMID reads report as "Unavailable",
then everything will be reported as "Unavailable".
Fix the issue by discarding the "Unavailable" reads and reporting all
the successful RMID reads. This is not a problem on Intel systems as
Intel reports 0 on Inactive RMIDs.
Fixes: d89b737901 ("x86/intel_rdt/cqm: Add mon_data")
Reported-by: Paweł Szulik <pawel.szulik@intel.com>
Signed-off-by: Babu Moger <Babu.Moger@amd.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Acked-by: Reinette Chatre <reinette.chatre@intel.com>
Cc: stable@vger.kernel.org
Link: https://bugzilla.kernel.org/show_bug.cgi?id=213311
Link: https://lkml.kernel.org/r/162793309296.9224.15871659871696482080.stgit@bmoger-ubuntu
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit ae28d1aae48a1258bd09a6f707ebb4231d79a761 upstream.
Currently, when moving a task to a resource group the PQR_ASSOC MSR is
updated with the new closid and rmid in an added task callback. If the
task is running, the work is run as soon as possible. If the task is not
running, the work is executed later in the kernel exit path when the
kernel returns to the task again.
Updating the PQR_ASSOC MSR as soon as possible on the CPU a moved task
is running is the right thing to do. Queueing work for a task that is
not running is unnecessary (the PQR_ASSOC MSR is already updated when
the task is scheduled in) and causing system resource waste with the way
in which it is implemented: Work to update the PQR_ASSOC register is
queued every time the user writes a task id to the "tasks" file, even if
the task already belongs to the resource group.
This could result in multiple pending work items associated with a
single task even if they are all identical and even though only a single
update with most recent values is needed. Specifically, even if a task
is moved between different resource groups while it is sleeping then it
is only the last move that is relevant but yet a work item is queued
during each move.
This unnecessary queueing of work items could result in significant
system resource waste, especially on tasks sleeping for a long time.
For example, as demonstrated by Shakeel Butt in [1] writing the same
task id to the "tasks" file can quickly consume significant memory. The
same problem (wasted system resources) occurs when moving a task between
different resource groups.
As pointed out by Valentin Schneider in [2] there is an additional issue
with the way in which the queueing of work is done in that the task_struct
update is currently done after the work is queued, resulting in a race with
the register update possibly done before the data needed by the update is
available.
To solve these issues, update the PQR_ASSOC MSR in a synchronous way
right after the new closid and rmid are ready during the task movement,
only if the task is running. If a moved task is not running nothing
is done since the PQR_ASSOC MSR will be updated next time the task is
scheduled. This is the same way used to update the register when tasks
are moved as part of resource group removal.
[1] https://lore.kernel.org/lkml/CALvZod7E9zzHwenzf7objzGKsdBmVwTgEJ0nPgs0LUFU3SN5Pw@mail.gmail.com/
[2] https://lore.kernel.org/lkml/20201123022433.17905-1-valentin.schneider@arm.com
[ bp: Massage commit message and drop the two update_task_closid_rmid()
variants. ]
Fixes: e02737d5b8 ("x86/intel_rdt: Add tasks files")
Reported-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Fenghua Yu <fenghua.yu@intel.com>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: James Morse <james.morse@arm.com>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/17aa2fb38fc12ce7bb710106b3e7c7b45acb9e94.1608243147.git.reinette.chatre@intel.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The MBA software controller (mba_sc) is a feedback loop which
periodically reads MBM counters and tries to restrict the bandwidth
below a user-specified value. It tags along the MBM counter overflow
handler to do the updates with 1s interval in mbm_update() and
update_mba_bw().
The purpose of mbm_update() is to periodically read the MBM counters to
make sure that the hardware counter doesn't wrap around more than once
between user samplings. mbm_update() calls __mon_event_count() for local
bandwidth updating when mba_sc is not enabled, but calls mbm_bw_count()
instead when mba_sc is enabled. __mon_event_count() will not be called
for local bandwidth updating in MBM counter overflow handler, but it is
still called when reading MBM local bandwidth counter file
'mbm_local_bytes', the call path is as below:
rdtgroup_mondata_show()
mon_event_read()
mon_event_count()
__mon_event_count()
In __mon_event_count(), m->chunks is updated by delta chunks which is
calculated from previous MSR value (m->prev_msr) and current MSR value.
When mba_sc is enabled, m->chunks is also updated in mbm_update() by
mistake by the delta chunks which is calculated from m->prev_bw_msr
instead of m->prev_msr. But m->chunks is not used in update_mba_bw() in
the mba_sc feedback loop.
When reading MBM local bandwidth counter file, m->chunks was changed
unexpectedly by mbm_bw_count(). As a result, the incorrect local
bandwidth counter which calculated from incorrect m->chunks is shown to
the user.
Fix this by removing incorrect m->chunks updating in mbm_bw_count() in
MBM counter overflow handler, and always calling __mon_event_count() in
mbm_update() to make sure that the hardware local bandwidth counter
doesn't wrap around.
Test steps:
# Run workload with aggressive memory bandwidth (e.g., 10 GB/s)
git clone https://github.com/intel/intel-cmt-cat && cd intel-cmt-cat
&& make
./tools/membw/membw -c 0 -b 10000 --read
# Enable MBA software controller
mount -t resctrl resctrl -o mba_MBps /sys/fs/resctrl
# Create control group c1
mkdir /sys/fs/resctrl/c1
# Set MB throttle to 6 GB/s
echo "MB:0=6000;1=6000" > /sys/fs/resctrl/c1/schemata
# Write PID of the workload to tasks file
echo `pidof membw` > /sys/fs/resctrl/c1/tasks
# Read local bytes counters twice with 1s interval, the calculated
# local bandwidth is not as expected (approaching to 6 GB/s):
local_1=`cat /sys/fs/resctrl/c1/mon_data/mon_L3_00/mbm_local_bytes`
sleep 1
local_2=`cat /sys/fs/resctrl/c1/mon_data/mon_L3_00/mbm_local_bytes`
echo "local b/w (bytes/s):" `expr $local_2 - $local_1`
Before fix:
local b/w (bytes/s): 11076796416
After fix:
local b/w (bytes/s): 5465014272
Fixes: ba0f26d852 (x86/intel_rdt/mba_sc: Prepare for feedback loop)
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Cc: <stable@vger.kernel.org>
Link: https://lkml.kernel.org/r/1607063279-19437-1-git-send-email-xiaochen.shen@intel.com
When the AMD QoS feature CDP (code and data prioritization) is enabled
or disabled, the CDP bit in MSR 0000_0C81 is written on one of the CPUs
in an L3 domain (core complex). That is not correct - the CDP bit needs
to be updated on all the logical CPUs in the domain.
This was not spelled out clearly in the spec earlier. The specification
has been updated and the updated document, "AMD64 Technology Platform
Quality of Service Extensions Publication # 56375 Revision: 1.02 Issue
Date: October 2020" is available now. Refer the section: Code and Data
Prioritization.
Fix the issue by adding a new flag arch_has_per_cpu_cfg in rdt_cache
data structure.
The documentation can be obtained at:
https://developer.amd.com/wp-content/resources/56375.pdf
Link: https://bugzilla.kernel.org/show_bug.cgi?id=206537
[ bp: Massage commit message. ]
Fixes: 4d05bf71f1 ("x86/resctrl: Introduce AMD QOS feature")
Signed-off-by: Babu Moger <babu.moger@amd.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/160675180380.15628.3309402017215002347.stgit@bmoger-ubuntu
On resource group creation via a mkdir an extra kernfs_node reference is
obtained by kernfs_get() to ensure that the rdtgroup structure remains
accessible for the rdtgroup_kn_unlock() calls where it is removed on
deletion. Currently the extra kernfs_node reference count is only
dropped by kernfs_put() in rdtgroup_kn_unlock() while the rdtgroup
structure is removed in a few other locations that lack the matching
reference drop.
In call paths of rmdir and umount, when a control group is removed,
kernfs_remove() is called to remove the whole kernfs nodes tree of the
control group (including the kernfs nodes trees of all child monitoring
groups), and then rdtgroup structure is freed by kfree(). The rdtgroup
structures of all child monitoring groups under the control group are
freed by kfree() in free_all_child_rdtgrp().
Before calling kfree() to free the rdtgroup structures, the kernfs node
of the control group itself as well as the kernfs nodes of all child
monitoring groups still take the extra references which will never be
dropped to 0 and the kernfs nodes will never be freed. It leads to
reference count leak and kernfs_node_cache memory leak.
For example, reference count leak is observed in these two cases:
(1) mount -t resctrl resctrl /sys/fs/resctrl
mkdir /sys/fs/resctrl/c1
mkdir /sys/fs/resctrl/c1/mon_groups/m1
umount /sys/fs/resctrl
(2) mkdir /sys/fs/resctrl/c1
mkdir /sys/fs/resctrl/c1/mon_groups/m1
rmdir /sys/fs/resctrl/c1
The same reference count leak issue also exists in the error exit paths
of mkdir in mkdir_rdt_prepare() and rdtgroup_mkdir_ctrl_mon().
Fix this issue by following changes to make sure the extra kernfs_node
reference on rdtgroup is dropped before freeing the rdtgroup structure.
(1) Introduce rdtgroup removal helper rdtgroup_remove() to wrap up
kernfs_put() and kfree().
(2) Call rdtgroup_remove() in rdtgroup removal path where the rdtgroup
structure is about to be freed by kfree().
(3) Call rdtgroup_remove() or kernfs_put() as appropriate in the error
exit paths of mkdir where an extra reference is taken by kernfs_get().
Fixes: f3cbeacaa0 ("x86/intel_rdt/cqm: Add rmdir support")
Fixes: e02737d5b8 ("x86/intel_rdt: Add tasks files")
Fixes: 60cf5e101f ("x86/intel_rdt: Add mkdir to resctrl file system")
Reported-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/1604085088-31707-1-git-send-email-xiaochen.shen@intel.com
Willem reported growing of kernfs_node_cache entries in slabtop when
repeatedly creating and removing resctrl subdirectories as well as when
repeatedly mounting and unmounting the resctrl filesystem.
On resource group (control as well as monitoring) creation via a mkdir
an extra kernfs_node reference is obtained to ensure that the rdtgroup
structure remains accessible for the rdtgroup_kn_unlock() calls where it
is removed on deletion. The kernfs_node reference count is dropped by
kernfs_put() in rdtgroup_kn_unlock().
With the above explaining the need for one kernfs_get()/kernfs_put()
pair in resctrl there are more places where a kernfs_node reference is
obtained without a corresponding release. The excessive amount of
reference count on kernfs nodes will never be dropped to 0 and the
kernfs nodes will never be freed in the call paths of rmdir and umount.
It leads to reference count leak and kernfs_node_cache memory leak.
Remove the superfluous kernfs_get() calls and expand the existing
comments surrounding the remaining kernfs_get()/kernfs_put() pair that
remains in use.
Superfluous kernfs_get() calls are removed from two areas:
(1) In call paths of mount and mkdir, when kernfs nodes for "info",
"mon_groups" and "mon_data" directories and sub-directories are
created, the reference count of newly created kernfs node is set to 1.
But after kernfs_create_dir() returns, superfluous kernfs_get() are
called to take an additional reference.
(2) kernfs_get() calls in rmdir call paths.
Fixes: 17eafd0762 ("x86/intel_rdt: Split resource group removal in two")
Fixes: 4af4a88e0c ("x86/intel_rdt/cqm: Add mount,umount support")
Fixes: f3cbeacaa0 ("x86/intel_rdt/cqm: Add rmdir support")
Fixes: d89b737901 ("x86/intel_rdt/cqm: Add mon_data")
Fixes: c7d9aac613 ("x86/intel_rdt/cqm: Add mkdir support for RDT monitoring")
Fixes: 5dc1d5c6ba ("x86/intel_rdt: Simplify info and base file lists")
Fixes: 60cf5e101f ("x86/intel_rdt: Add mkdir to resctrl file system")
Fixes: 4e978d06de ("x86/intel_rdt: Add "info" files to resctrl file system")
Reported-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Tested-by: Willem de Bruijn <willemb@google.com>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/1604085053-31639-1-git-send-email-xiaochen.shen@intel.com
A previous commit changed the notification mode from true/false to an
int, allowing notify-no, notify-yes, or signal-notify. This was
backwards compatible in the sense that any existing true/false user
would translate to either 0 (on notification sent) or 1, the latter
which mapped to TWA_RESUME. TWA_SIGNAL was assigned a value of 2.
Clean this up properly, and define a proper enum for the notification
mode. Now we have:
- TWA_NONE. This is 0, same as before the original change, meaning no
notification requested.
- TWA_RESUME. This is 1, same as before the original change, meaning
that we use TIF_NOTIFY_RESUME.
- TWA_SIGNAL. This uses TIF_SIGPENDING/JOBCTL_TASK_WORK for the
notification.
Clean up all the callers, switching their 0/1/false/true to using the
appropriate TWA_* mode for notifications.
Fixes: e91b481623 ("task_work: teach task_work_add() to do signal_wake_up()")
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull x86 cache resource control updates from Borislav Petkov:
- Misc cleanups to the resctrl code in preparation for the ARM side
(James Morse)
- Add support for controlling per-thread memory bandwidth throttling
delay values on hw which supports it (Fenghua Yu)
* tag 'x86_cache_for_v5.10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/resctrl: Enable user to view thread or core throttling mode
x86/resctrl: Enumerate per-thread MBA controls
cacheinfo: Move resctrl's get_cache_id() to the cacheinfo header file
x86/resctrl: Add struct rdt_cache::arch_has_{sparse, empty}_bitmaps
x86/resctrl: Merge AMD/Intel parse_bw() calls
x86/resctrl: Add struct rdt_membw::arch_needs_linear to explain AMD/Intel MBA difference
x86/resctrl: Use is_closid_match() in more places
x86/resctrl: Include pid.h
x86/resctrl: Use container_of() in delayed_work handlers
x86/resctrl: Fix stale comment
x86/resctrl: Remove struct rdt_membw::max_delay
x86/resctrl: Remove unused struct mbm_state::chunks_bw
Early Intel hardware implementations of Memory Bandwidth Allocation (MBA)
could only control bandwidth at the processor core level. This meant that
when two processes with different bandwidth allocations ran simultaneously
on the same core the hardware had to resolve this difference. It did so by
applying the higher throttling value (lower bandwidth) to both processes.
Newer implementations can apply different throttling values to each
thread on a core.
Introduce a new resctrl file, "thread_throttle_mode", on Intel systems
that shows to the user how throttling values are allocated, per-core or
per-thread.
On systems that support per-core throttling, the file will display "max".
On newer systems that support per-thread throttling, the file will display
"per-thread".
AMD confirmed in [1] that AMD bandwidth allocation is already at thread
level but that the AMD implementation does not use a memory delay
throttle mode. So to avoid confusion the thread throttling mode would be
UNDEFINED on AMD systems and the "thread_throttle_mode" file will not be
visible.
Originally-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Fenghua Yu <fenghua.yu@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/1598296281-127595-3-git-send-email-fenghua.yu@intel.com
Link: [1] https://lore.kernel.org/lkml/18d277fd-6523-319c-d560-66b63ff606b8@amd.com
resctrl/core.c defines get_cache_id() for use in its cpu-hotplug
callbacks. This gets the id attribute of the cache at the corresponding
level of a CPU.
Later rework means this private function needs to be shared. Move
it to the header file.
The name conflicts with a different definition in intel_cacheinfo.c,
name it get_cpu_cacheinfo_id() to show its relation with
get_cpu_cacheinfo().
Now this is visible on other architectures, check the id attribute
has actually been set.
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Babu Moger <babu.moger@amd.com>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/20200708163929.2783-11-james.morse@arm.com
Intel CPUs expect the cache bitmap provided by user-space to have on a
single span of 1s, whereas AMD can support bitmaps like 0xf00f. Arm's
MPAM support also allows sparse bitmaps.
Similarly, Intel CPUs check at least one bit set, whereas AMD CPUs are
quite happy with an empty bitmap. Arm's MPAM allows an empty bitmap.
To move resctrl out to /fs/, platform differences like this need to be
explained.
Add two resource properties arch_has_{empty,sparse}_bitmaps. Test these
around the relevant parts of cbm_validate().
Merging the validate calls causes AMD to gain the min_cbm_bits test
needed for Haswell, but as it always sets this value to 1, it will never
match.
[ bp: Massage commit message. ]
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Babu Moger <babu.moger@amd.com>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/20200708163929.2783-10-james.morse@arm.com
Now after arch_needs_linear has been added, the parse_bw() calls are
almost the same between AMD and Intel.
The difference is '!is_mba_sc()', which is not checked on AMD. This
will always be true on AMD CPUs as mba_sc cannot be enabled as
is_mba_linear() is false.
Removing this duplication means user-space visible behaviour and
error messages are not validated or generated in different places.
Reviewed-by : Babu Moger <babu.moger@amd.com>
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/20200708163929.2783-9-james.morse@arm.com
The configuration values user-space provides to the resctrl filesystem
are ABI. To make this work on another architecture, all the ABI bits
should be moved out of /arch/x86 and under /fs.
To do this, the differences between AMD and Intel CPUs needs to be
explained to resctrl via resource properties, instead of function
pointers that let the arch code accept subtly different values on
different platforms/architectures.
For MBA, Intel CPUs reject configuration attempts for non-linear
resources, whereas AMD ignore this field as its MBA resource is never
linear. To merge the parse/validate functions, this difference needs to
be explained.
Add struct rdt_membw::arch_needs_linear to indicate the arch code needs
the linear property to be true to configure this resource. AMD can set
this and delay_linear to false. Intel can set arch_needs_linear to
true to keep the existing "No support for non-linear MB domains" error
message for affected platforms.
[ bp: convert "we" etc to passive voice. ]
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Reviewed-by: Babu Moger <babu.moger@amd.com>
Link: https://lkml.kernel.org/r/20200708163929.2783-8-james.morse@arm.com
rdtgroup_tasks_assigned() and show_rdt_tasks() loop over threads testing
for a CTRL/MON group match by closid/rmid with the provided rdtgrp.
Further down the file are helpers to do this, move these further up and
make use of them here.
These helpers additionally check for alloc/mon capable. This is harmless
as rdtgroup_mkdir() tests these capable flags before allowing the config
directories to be created.
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/20200708163929.2783-7-james.morse@arm.com
mbm_handle_overflow() and cqm_handle_limbo() are both provided with
the domain's work_struct when called, but use get_domain_from_cpu()
to find the domain, along with the appropriate error handling.
container_of() saves some list walking and bitmap testing, use that
instead.
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lkml.kernel.org/r/20200708163929.2783-5-james.morse@arm.com
The callers don't expect *d_cdp to be set to an error pointer, they only
check for NULL. This leads to a static checker warning:
arch/x86/kernel/cpu/resctrl/rdtgroup.c:2648 __init_one_rdt_domain()
warn: 'd_cdp' could be an error pointer
This would not trigger a bug in this specific case because
__init_one_rdt_domain() calls it with a valid domain that would not have
a negative id and thus not trigger the return of the ERR_PTR(). If this
was a negative domain id then the call to rdt_find_domain() in
domain_add_cpu() would have returned the ERR_PTR() much earlier and the
creation of the domain with an invalid id would have been prevented.
Even though a bug is not triggered currently the right and safe thing to
do is to set the pointer to NULL because that is what can be checked for
when the caller is handling the CDP and non-CDP cases.
Fixes: 52eb74339a ("x86/resctrl: Fix rdt_find_domain() return value and checks")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Acked-by: Reinette Chatre <reinette.chatre@intel.com>
Acked-by: Fenghua Yu <fenghua.yu@intel.com>
Link: https://lkml.kernel.org/r/20200602193611.GA190851@mwanda
Memory bandwidth is calculated reading the monitoring counter
at two intervals and calculating the delta. It is the software’s
responsibility to read the count often enough to avoid having
the count roll over _twice_ between reads.
The current code hardcodes the bandwidth monitoring counter's width
to 24 bits for AMD. This is due to default base counter width which
is 24. Currently, AMD does not implement the CPUID 0xF.[ECX=1]:EAX
to adjust the counter width. But, the AMD hardware supports much
wider bandwidth counter with the default width of 44 bits.
Kernel reads these monitoring counters every 1 second and adjusts the
counter value for overflow. With 24 bits and scale value of 64 for AMD,
it can only measure up to 1GB/s without overflowing. For the rates
above 1GB/s this will fail to measure the bandwidth.
Fix the issue setting the default width to 44 bits by adjusting the
offset.
AMD future products will implement CPUID 0xF.[ECX=1]:EAX.
[ bp: Let the line stick out and drop {}-brackets around a single
statement. ]
Fixes: 4d05bf71f1 ("x86/resctrl: Introduce AMD QOS feature")
Signed-off-by: Babu Moger <babu.moger@amd.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/159129975546.62538.5656031125604254041.stgit@naples-babu.amd.com
The original Memory Bandwidth Monitoring (MBM) architectural
definition defines counters of up to 62 bits in the
IA32_QM_CTR MSR while the first-generation MBM implementation
uses statically defined 24 bit counters.
The MBM CPUID enumeration properties have been expanded to include
the MBM counter width, encoded as an offset from 24 bits.
While eight bits are available for the counter width offset IA32_QM_CTR
MSR only supports 62 bit counters. Add a sanity check, with warning
printed when encountered, to ensure counters cannot exceed the 62 bit
limit.
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/69d52abd5b14794d3a0f05ba7c755ed1f4c0d5ed.1588715690.git.reinette.chatre@intel.com
The original Memory Bandwidth Monitoring (MBM) architectural
definition defines counters of up to 62 bits in the
IA32_QM_CTR MSR while the first-generation MBM implementation
uses statically defined 24 bit counters.
Expand the MBM CPUID enumeration properties to include the MBM
counter width. The previously undefined EAX output register contains,
in bits [7:0], the MBM counter width encoded as an offset from
24 bits. Enumerating this property is only specified for Intel
CPUs.
Suggested-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/afa3af2f753f6bc301fb743bc8944e749cb24afa.1588715690.git.reinette.chatre@intel.com
The original Memory Bandwidth Monitoring (MBM) architectural
definition defines counters of up to 62 bits in the IA32_QM_CTR MSR,
and the first-generation MBM implementation uses 24 bit counters.
Software is required to poll at 1 second or faster to ensure that
data is retrieved before a counter rollover occurs more than once
under worst conditions.
As system bandwidths scale the software requirement is maintained with
the introduction of a per-resource enumerable MBM counter width.
In preparation for supporting hardware with an enumerable MBM counter
width the current globally static MBM counter width is moved to a
per-resource MBM counter width. Currently initialized to 24 always
to result in no functional change.
In essence there is one function, mbm_overflow_count() that needs to
know the counter width to handle rollovers. The static value
used within mbm_overflow_count() will be replaced with a value
discovered from the hardware. Support for learning the MBM counter
width from hardware is added in the change that follows.
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/e36743b9800f16ce600f86b89127391f61261f23.1588715690.git.reinette.chatre@intel.com
Cache and memory bandwidth monitoring are features that are part of
x86 CPU resource control that is supported by the resctrl subsystem.
The monitoring properties are obtained via CPUID from every CPU
and only used within the resctrl subsystem where the properties are
only read from boot_cpu_data.
Obtain the monitoring properties once, placed in boot_cpu_data, via the
->c_bsp_init() helpers of the vendors that support X86_FEATURE_CQM_LLC.
Suggested-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/6d74a6ac3e69f4b7a8b4115835f9455faf0f468d.1588715690.git.reinette.chatre@intel.com
The function determining a platform's support and properties of cache
occupancy and memory bandwidth monitoring (properties of
X86_FEATURE_CQM_LLC) can be found among the common CPU code. After
the feature's properties is populated in the per-CPU data the resctrl
subsystem is the only consumer (via boot_cpu_data).
Move the function that obtains the CPU information used by resctrl to
the resctrl subsystem and rename it from init_cqm() to
resctrl_cpu_detect(). The function continues to be called from the
common CPU code. This move is done in preparation of the addition of some
vendor specific code.
No functional change.
Suggested-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/38433b99f9d16c8f4ee796f8cc42b871531fa203.1588715690.git.reinette.chatre@intel.com
Resctrl assumes that all CPUs are online when the filesystem is mounted,
and that CPUs remember their CDP-enabled state over CPU hotplug.
This goes wrong when resctrl's CDP-enabled state changes while all the
CPUs in a domain are offline.
When a domain comes online, enable (or disable!) CDP to match resctrl's
current setting.
Fixes: 5ff193fbde ("x86/intel_rdt: Add basic resctrl filesystem support")
Suggested-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Cc: <stable@vger.kernel.org>
Link: https://lkml.kernel.org/r/20200221162105.154163-1-james.morse@arm.com
The default resource group ("rdtgroup_default") is associated with the
root of the resctrl filesystem and should never be removed. New resource
groups can be created as subdirectories of the resctrl filesystem and
they can be removed from user space.
There exists a safeguard in the directory removal code
(rdtgroup_rmdir()) that ensures that only subdirectories can be removed
by testing that the directory to be removed has to be a child of the
root directory.
A possible deadlock was recently fixed with
334b0f4e9b ("x86/resctrl: Fix a deadlock due to inaccurate reference").
This fix involved associating the private data of the "mon_groups"
and "mon_data" directories to the resource group to which they belong
instead of NULL as before. A consequence of this change was that
the original safeguard code preventing removal of "mon_groups" and
"mon_data" found in the root directory failed resulting in attempts to
remove the default resource group that ends in a BUG:
kernel BUG at mm/slub.c:3969!
invalid opcode: 0000 [#1] SMP PTI
Call Trace:
rdtgroup_rmdir+0x16b/0x2c0
kernfs_iop_rmdir+0x5c/0x90
vfs_rmdir+0x7a/0x160
do_rmdir+0x17d/0x1e0
do_syscall_64+0x55/0x1d0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
Fix this by improving the directory removal safeguard to ensure that
subdirectories of the resctrl root directory can only be removed if they
are a child of the resctrl filesystem's root _and_ not associated with
the default resource group.
Fixes: 334b0f4e9b ("x86/resctrl: Fix a deadlock due to inaccurate reference")
Reported-by: Sai Praneeth Prakhya <sai.praneeth.prakhya@intel.com>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Tested-by: Sai Praneeth Prakhya <sai.praneeth.prakhya@intel.com>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/884cbe1773496b5dbec1b6bd11bb50cffa83603d.1584461853.git.reinette.chatre@intel.com
Pull vfs file system parameter updates from Al Viro:
"Saner fs_parser.c guts and data structures. The system-wide registry
of syntax types (string/enum/int32/oct32/.../etc.) is gone and so is
the horror switch() in fs_parse() that would have to grow another case
every time something got added to that system-wide registry.
New syntax types can be added by filesystems easily now, and their
namespace is that of functions - not of system-wide enum members. IOW,
they can be shared or kept private and if some turn out to be widely
useful, we can make them common library helpers, etc., without having
to do anything whatsoever to fs_parse() itself.
And we already get that kind of requests - the thing that finally
pushed me into doing that was "oh, and let's add one for timeouts -
things like 15s or 2h". If some filesystem really wants that, let them
do it. Without somebody having to play gatekeeper for the variants
blessed by direct support in fs_parse(), TYVM.
Quite a bit of boilerplate is gone. And IMO the data structures make a
lot more sense now. -200LoC, while we are at it"
* 'merge.nfs-fs_parse.1' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (25 commits)
tmpfs: switch to use of invalfc()
cgroup1: switch to use of errorfc() et.al.
procfs: switch to use of invalfc()
hugetlbfs: switch to use of invalfc()
cramfs: switch to use of errofc() et.al.
gfs2: switch to use of errorfc() et.al.
fuse: switch to use errorfc() et.al.
ceph: use errorfc() and friends instead of spelling the prefix out
prefix-handling analogues of errorf() and friends
turn fs_param_is_... into functions
fs_parse: handle optional arguments sanely
fs_parse: fold fs_parameter_desc/fs_parameter_spec
fs_parser: remove fs_parameter_description name field
add prefix to fs_context->log
ceph_parse_param(), ceph_parse_mon_ips(): switch to passing fc_log
new primitive: __fs_parse()
switch rbd and libceph to p_log-based primitives
struct p_log, variants of warnf() et.al. taking that one instead
teach logfc() to handle prefices, give it saner calling conventions
get rid of cg_invalf()
...
Pull x86 fixes from Ingo Molnar:
"Misc fixes:
- three fixes and a cleanup for the resctrl code
- a HyperV fix
- a fix to /proc/kcore contents in live debugging sessions
- a fix for the x86 decoder opcode map"
* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/decoder: Add TEST opcode to Group3-2
x86/resctrl: Clean up unused function parameter in mkdir path
x86/resctrl: Fix a deadlock due to inaccurate reference
x86/resctrl: Fix use-after-free due to inaccurate refcount of rdtgroup
x86/resctrl: Fix use-after-free when deleting resource groups
x86/hyper-v: Add "polling" bit to hv_synic_sint
x86/crash: Define arch_crash_save_vmcoreinfo() if CONFIG_CRASH_CORE=y
Pull x86 resource control updates from Ingo Molnar:
"The main change in this tree is the extension of the resctrl procfs
ABI with a new file that helps tooling to navigate from tasks back to
resctrl groups: /proc/{pid}/cpu_resctrl_groups.
Also fix static key usage for certain feature combinations and
simplify the task exit resctrl case"
* 'x86-cache-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/resctrl: Add task resctrl information display
x86/resctrl: Check monitoring static key in the MBM overflow handler
x86/resctrl: Do not reconfigure exiting tasks
There is a race condition which results in a deadlock when rmdir and
mkdir execute concurrently:
$ ls /sys/fs/resctrl/c1/mon_groups/m1/
cpus cpus_list mon_data tasks
Thread 1: rmdir /sys/fs/resctrl/c1
Thread 2: mkdir /sys/fs/resctrl/c1/mon_groups/m1
3 locks held by mkdir/48649:
#0: (sb_writers#17){.+.+}, at: [<ffffffffb4ca2aa0>] mnt_want_write+0x20/0x50
#1: (&type->i_mutex_dir_key#8/1){+.+.}, at: [<ffffffffb4c8c13b>] filename_create+0x7b/0x170
#2: (rdtgroup_mutex){+.+.}, at: [<ffffffffb4a4389d>] rdtgroup_kn_lock_live+0x3d/0x70
4 locks held by rmdir/48652:
#0: (sb_writers#17){.+.+}, at: [<ffffffffb4ca2aa0>] mnt_want_write+0x20/0x50
#1: (&type->i_mutex_dir_key#8/1){+.+.}, at: [<ffffffffb4c8c3cf>] do_rmdir+0x13f/0x1e0
#2: (&type->i_mutex_dir_key#8){++++}, at: [<ffffffffb4c86d5d>] vfs_rmdir+0x4d/0x120
#3: (rdtgroup_mutex){+.+.}, at: [<ffffffffb4a4389d>] rdtgroup_kn_lock_live+0x3d/0x70
Thread 1 is deleting control group "c1". Holding rdtgroup_mutex,
kernfs_remove() removes all kernfs nodes under directory "c1"
recursively, then waits for sub kernfs node "mon_groups" to drop active
reference.
Thread 2 is trying to create a subdirectory "m1" in the "mon_groups"
directory. The wrapper kernfs_iop_mkdir() takes an active reference to
the "mon_groups" directory but the code drops the active reference to
the parent directory "c1" instead.
As a result, Thread 1 is blocked on waiting for active reference to drop
and never release rdtgroup_mutex, while Thread 2 is also blocked on
trying to get rdtgroup_mutex.
Thread 1 (rdtgroup_rmdir) Thread 2 (rdtgroup_mkdir)
(rmdir /sys/fs/resctrl/c1) (mkdir /sys/fs/resctrl/c1/mon_groups/m1)
------------------------- -------------------------
kernfs_iop_mkdir
/*
* kn: "m1", parent_kn: "mon_groups",
* prgrp_kn: parent_kn->parent: "c1",
*
* "mon_groups", parent_kn->active++: 1
*/
kernfs_get_active(parent_kn)
kernfs_iop_rmdir
/* "c1", kn->active++ */
kernfs_get_active(kn)
rdtgroup_kn_lock_live
atomic_inc(&rdtgrp->waitcount)
/* "c1", kn->active-- */
kernfs_break_active_protection(kn)
mutex_lock
rdtgroup_rmdir_ctrl
free_all_child_rdtgrp
sentry->flags = RDT_DELETED
rdtgroup_ctrl_remove
rdtgrp->flags = RDT_DELETED
kernfs_get(kn)
kernfs_remove(rdtgrp->kn)
__kernfs_remove
/* "mon_groups", sub_kn */
atomic_add(KN_DEACTIVATED_BIAS, &sub_kn->active)
kernfs_drain(sub_kn)
/*
* sub_kn->active == KN_DEACTIVATED_BIAS + 1,
* waiting on sub_kn->active to drop, but it
* never drops in Thread 2 which is blocked
* on getting rdtgroup_mutex.
*/
Thread 1 hangs here ---->
wait_event(sub_kn->active == KN_DEACTIVATED_BIAS)
...
rdtgroup_mkdir
rdtgroup_mkdir_mon(parent_kn, prgrp_kn)
mkdir_rdt_prepare(parent_kn, prgrp_kn)
rdtgroup_kn_lock_live(prgrp_kn)
atomic_inc(&rdtgrp->waitcount)
/*
* "c1", prgrp_kn->active--
*
* The active reference on "c1" is
* dropped, but not matching the
* actual active reference taken
* on "mon_groups", thus causing
* Thread 1 to wait forever while
* holding rdtgroup_mutex.
*/
kernfs_break_active_protection(
prgrp_kn)
/*
* Trying to get rdtgroup_mutex
* which is held by Thread 1.
*/
Thread 2 hangs here ----> mutex_lock
...
The problem is that the creation of a subdirectory in the "mon_groups"
directory incorrectly releases the active protection of its parent
directory instead of itself before it starts waiting for rdtgroup_mutex.
This is triggered by the rdtgroup_mkdir() flow calling
rdtgroup_kn_lock_live()/rdtgroup_kn_unlock() with kernfs node of the
parent control group ("c1") as argument. It should be called with kernfs
node "mon_groups" instead. What is currently missing is that the
kn->priv of "mon_groups" is NULL instead of pointing to the rdtgrp.
Fix it by pointing kn->priv to rdtgrp when "mon_groups" is created. Then
it could be passed to rdtgroup_kn_lock_live()/rdtgroup_kn_unlock()
instead. And then it operates on the same rdtgroup structure but handles
the active reference of kernfs node "mon_groups" to prevent deadlock.
The same changes are also made to the "mon_data" directories.
This results in some unused function parameters that will be cleaned up
in follow-up patch as the focus here is on the fix only in support of
backporting efforts.
Fixes: c7d9aac613 ("x86/intel_rdt/cqm: Add mkdir support for RDT monitoring")
Suggested-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/1578500886-21771-4-git-send-email-xiaochen.shen@intel.com
There is a race condition in the following scenario which results in an
use-after-free issue when reading a monitoring file and deleting the
parent ctrl_mon group concurrently:
Thread 1 calls atomic_inc() to take refcount of rdtgrp and then calls
kernfs_break_active_protection() to drop the active reference of kernfs
node in rdtgroup_kn_lock_live().
In Thread 2, kernfs_remove() is a blocking routine. It waits on all sub
kernfs nodes to drop the active reference when removing all subtree
kernfs nodes recursively. Thread 2 could block on kernfs_remove() until
Thread 1 calls kernfs_break_active_protection(). Only after
kernfs_remove() completes the refcount of rdtgrp could be trusted.
Before Thread 1 calls atomic_inc() and kernfs_break_active_protection(),
Thread 2 could call kfree() when the refcount of rdtgrp (sentry) is 0
instead of 1 due to the race.
In Thread 1, in rdtgroup_kn_unlock(), referring to earlier rdtgrp memory
(rdtgrp->waitcount) which was already freed in Thread 2 results in
use-after-free issue.
Thread 1 (rdtgroup_mondata_show) Thread 2 (rdtgroup_rmdir)
-------------------------------- -------------------------
rdtgroup_kn_lock_live
/*
* kn active protection until
* kernfs_break_active_protection(kn)
*/
rdtgrp = kernfs_to_rdtgroup(kn)
rdtgroup_kn_lock_live
atomic_inc(&rdtgrp->waitcount)
mutex_lock
rdtgroup_rmdir_ctrl
free_all_child_rdtgrp
/*
* sentry->waitcount should be 1
* but is 0 now due to the race.
*/
kfree(sentry)*[1]
/*
* Only after kernfs_remove()
* completes, the refcount of
* rdtgrp could be trusted.
*/
atomic_inc(&rdtgrp->waitcount)
/* kn->active-- */
kernfs_break_active_protection(kn)
rdtgroup_ctrl_remove
rdtgrp->flags = RDT_DELETED
/*
* Blocking routine, wait for
* all sub kernfs nodes to drop
* active reference in
* kernfs_break_active_protection.
*/
kernfs_remove(rdtgrp->kn)
rdtgroup_kn_unlock
mutex_unlock
atomic_dec_and_test(
&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
kernfs_unbreak_active_protection(kn)
kfree(rdtgrp)
mutex_lock
mon_event_read
rdtgroup_kn_unlock
mutex_unlock
/*
* Use-after-free: refer to earlier rdtgrp
* memory which was freed in [1].
*/
atomic_dec_and_test(&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
/* kn->active++ */
kernfs_unbreak_active_protection(kn)
kfree(rdtgrp)
Fix it by moving free_all_child_rdtgrp() to after kernfs_remove() in
rdtgroup_rmdir_ctrl() to ensure it has the accurate refcount of rdtgrp.
Fixes: f3cbeacaa0 ("x86/intel_rdt/cqm: Add rmdir support")
Suggested-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/1578500886-21771-3-git-send-email-xiaochen.shen@intel.com
A resource group (rdtgrp) contains a reference count (rdtgrp->waitcount)
that indicates how many waiters expect this rdtgrp to exist. Waiters
could be waiting on rdtgroup_mutex or some work sitting on a task's
workqueue for when the task returns from kernel mode or exits.
The deletion of a rdtgrp is intended to have two phases:
(1) while holding rdtgroup_mutex the necessary cleanup is done and
rdtgrp->flags is set to RDT_DELETED,
(2) after releasing the rdtgroup_mutex, the rdtgrp structure is freed
only if there are no waiters and its flag is set to RDT_DELETED. Upon
gaining access to rdtgroup_mutex or rdtgrp, a waiter is required to check
for the RDT_DELETED flag.
When unmounting the resctrl file system or deleting ctrl_mon groups,
all of the subdirectories are removed and the data structure of rdtgrp
is forcibly freed without checking rdtgrp->waitcount. If at this point
there was a waiter on rdtgrp then a use-after-free issue occurs when the
waiter starts running and accesses the rdtgrp structure it was waiting
on.
See kfree() calls in [1], [2] and [3] in these two call paths in
following scenarios:
(1) rdt_kill_sb() -> rmdir_all_sub() -> free_all_child_rdtgrp()
(2) rdtgroup_rmdir() -> rdtgroup_rmdir_ctrl() -> free_all_child_rdtgrp()
There are several scenarios that result in use-after-free issue in
following:
Scenario 1:
-----------
In Thread 1, rdtgroup_tasks_write() adds a task_work callback
move_myself(). If move_myself() is scheduled to execute after Thread 2
rdt_kill_sb() is finished, referring to earlier rdtgrp memory
(rdtgrp->waitcount) which was already freed in Thread 2 results in
use-after-free issue.
Thread 1 (rdtgroup_tasks_write) Thread 2 (rdt_kill_sb)
------------------------------- ----------------------
rdtgroup_kn_lock_live
atomic_inc(&rdtgrp->waitcount)
mutex_lock
rdtgroup_move_task
__rdtgroup_move_task
/*
* Take an extra refcount, so rdtgrp cannot be freed
* before the call back move_myself has been invoked
*/
atomic_inc(&rdtgrp->waitcount)
/* Callback move_myself will be scheduled for later */
task_work_add(move_myself)
rdtgroup_kn_unlock
mutex_unlock
atomic_dec_and_test(&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
mutex_lock
rmdir_all_sub
/*
* sentry and rdtgrp are freed
* without checking refcount
*/
free_all_child_rdtgrp
kfree(sentry)*[1]
kfree(rdtgrp)*[2]
mutex_unlock
/*
* Callback is scheduled to execute
* after rdt_kill_sb is finished
*/
move_myself
/*
* Use-after-free: refer to earlier rdtgrp
* memory which was freed in [1] or [2].
*/
atomic_dec_and_test(&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
kfree(rdtgrp)
Scenario 2:
-----------
In Thread 1, rdtgroup_tasks_write() adds a task_work callback
move_myself(). If move_myself() is scheduled to execute after Thread 2
rdtgroup_rmdir() is finished, referring to earlier rdtgrp memory
(rdtgrp->waitcount) which was already freed in Thread 2 results in
use-after-free issue.
Thread 1 (rdtgroup_tasks_write) Thread 2 (rdtgroup_rmdir)
------------------------------- -------------------------
rdtgroup_kn_lock_live
atomic_inc(&rdtgrp->waitcount)
mutex_lock
rdtgroup_move_task
__rdtgroup_move_task
/*
* Take an extra refcount, so rdtgrp cannot be freed
* before the call back move_myself has been invoked
*/
atomic_inc(&rdtgrp->waitcount)
/* Callback move_myself will be scheduled for later */
task_work_add(move_myself)
rdtgroup_kn_unlock
mutex_unlock
atomic_dec_and_test(&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
rdtgroup_kn_lock_live
atomic_inc(&rdtgrp->waitcount)
mutex_lock
rdtgroup_rmdir_ctrl
free_all_child_rdtgrp
/*
* sentry is freed without
* checking refcount
*/
kfree(sentry)*[3]
rdtgroup_ctrl_remove
rdtgrp->flags = RDT_DELETED
rdtgroup_kn_unlock
mutex_unlock
atomic_dec_and_test(
&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
kfree(rdtgrp)
/*
* Callback is scheduled to execute
* after rdt_kill_sb is finished
*/
move_myself
/*
* Use-after-free: refer to earlier rdtgrp
* memory which was freed in [3].
*/
atomic_dec_and_test(&rdtgrp->waitcount)
&& (flags & RDT_DELETED)
kfree(rdtgrp)
If CONFIG_DEBUG_SLAB=y, Slab corruption on kmalloc-2k can be observed
like following. Note that "0x6b" is POISON_FREE after kfree(). The
corrupted bits "0x6a", "0x64" at offset 0x424 correspond to
waitcount member of struct rdtgroup which was freed:
Slab corruption (Not tainted): kmalloc-2k start=ffff9504c5b0d000, len=2048
420: 6b 6b 6b 6b 6a 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkjkkkkkkkkkkk
Single bit error detected. Probably bad RAM.
Run memtest86+ or a similar memory test tool.
Next obj: start=ffff9504c5b0d800, len=2048
000: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk
010: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk
Slab corruption (Not tainted): kmalloc-2k start=ffff9504c58ab800, len=2048
420: 6b 6b 6b 6b 64 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkdkkkkkkkkkkk
Prev obj: start=ffff9504c58ab000, len=2048
000: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk
010: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk
Fix this by taking reference count (waitcount) of rdtgrp into account in
the two call paths that currently do not do so. Instead of always
freeing the resource group it will only be freed if there are no waiters
on it. If there are waiters, the resource group will have its flags set
to RDT_DELETED.
It will be left to the waiter to free the resource group when it starts
running and finding that it was the last waiter and the resource group
has been removed (rdtgrp->flags & RDT_DELETED) since. (1) rdt_kill_sb()
-> rmdir_all_sub() -> free_all_child_rdtgrp() (2) rdtgroup_rmdir() ->
rdtgroup_rmdir_ctrl() -> free_all_child_rdtgrp()
Fixes: f3cbeacaa0 ("x86/intel_rdt/cqm: Add rmdir support")
Fixes: 60cf5e101f ("x86/intel_rdt: Add mkdir to resctrl file system")
Suggested-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/1578500886-21771-2-git-send-email-xiaochen.shen@intel.com
Monitoring tools that want to find out which resctrl control and monitor
groups a task belongs to must currently read the "tasks" file in every
group until they locate the process ID.
Add an additional file /proc/{pid}/cpu_resctrl_groups to provide this
information:
1) res:
mon:
resctrl is not available.
2) res:/
mon:
Task is part of the root resctrl control group, and it is not associated
to any monitor group.
3) res:/
mon:mon0
Task is part of the root resctrl control group and monitor group mon0.
4) res:group0
mon:
Task is part of resctrl control group group0, and it is not associated
to any monitor group.
5) res:group0
mon:mon1
Task is part of resctrl control group group0 and monitor group mon1.
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Tested-by: Jinshi Chen <jinshi.chen@intel.com>
Link: https://lkml.kernel.org/r/20200115092851.14761-1-yu.c.chen@intel.com
Currently, there are three static keys in the resctrl file system:
rdt_mon_enable_key and rdt_alloc_enable_key indicate if the monitoring
feature and the allocation feature are enabled, respectively. The
rdt_enable_key is enabled when either the monitoring feature or the
allocation feature is enabled.
If no monitoring feature is present (either hardware doesn't support a
monitoring feature or the feature is disabled by the kernel command line
option "rdt="), rdt_enable_key is still enabled but rdt_mon_enable_key
is disabled.
MBM is a monitoring feature. The MBM overflow handler intends to
check if the monitoring feature is not enabled for fast return.
So check the rdt_mon_enable_key in it instead of the rdt_enable_key as
former is the more accurate check.
[ bp: Massage commit message. ]
Fixes: e33026831b ("x86/intel_rdt/mbm: Handle counter overflow")
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lkml.kernel.org/r/1576094705-13660-1-git-send-email-xiaochen.shen@intel.com
