Re: [PATCH v8 -tip 25/26] Documentation: Add core scheduling documentation

From: Joel Fernandes
Date: Thu Nov 12 2020 - 11:11:16 EST


On Mon, Oct 19, 2020 at 08:36:25PM -0700, Randy Dunlap wrote:
> Hi Joel,
>
> On 10/19/20 6:43 PM, Joel Fernandes (Google) wrote:
> > Document the usecases, design and interfaces for core scheduling.
> >
> > Co-developed-by: Vineeth Pillai <viremana@xxxxxxxxxxxxxxxxxxx>
> > Tested-by: Julien Desfossez <jdesfossez@xxxxxxxxxxxxxxxx>
> > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>

All fixed as below, updated and thanks!

---8<-----------------------

From: "Joel Fernandes (Google)" <joel@xxxxxxxxxxxxxxxxx>
Subject: [PATCH] Documentation: Add core scheduling documentation

Document the usecases, design and interfaces for core scheduling.

Co-developed-by: Vineeth Pillai <viremana@xxxxxxxxxxxxxxxxxxx>
Tested-by: Julien Desfossez <jdesfossez@xxxxxxxxxxxxxxxx>
Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
---
.../admin-guide/hw-vuln/core-scheduling.rst | 313 ++++++++++++++++++
Documentation/admin-guide/hw-vuln/index.rst | 1 +
2 files changed, 314 insertions(+)
create mode 100644 Documentation/admin-guide/hw-vuln/core-scheduling.rst

diff --git a/Documentation/admin-guide/hw-vuln/core-scheduling.rst b/Documentation/admin-guide/hw-vuln/core-scheduling.rst
new file mode 100644
index 000000000000..c7399809c74d
--- /dev/null
+++ b/Documentation/admin-guide/hw-vuln/core-scheduling.rst
@@ -0,0 +1,313 @@
+Core Scheduling
+***************
+Core scheduling support allows userspace to define groups of tasks that can
+share a core. These groups can be specified either for security usecases (one
+group of tasks don't trust another), or for performance usecases (some
+workloads may benefit from running on the same core as they don't need the same
+hardware resources of the shared core).
+
+Security usecase
+----------------
+A cross-HT attack involves the attacker and victim running on different
+Hyper Threads of the same core. MDS and L1TF are examples of such attacks.
+Without core scheduling, the only full mitigation of cross-HT attacks is to
+disable Hyper Threading (HT). Core scheduling allows HT to be turned on safely
+by ensuring that trusted tasks can share a core. This increase in core sharing
+can improvement performance, however it is not guaranteed that performance will
+always improve, though that is seen to be the case with a number of real world
+workloads. In theory, core scheduling aims to perform at least as good as when
+Hyper Threading is disabled. In practice, this is mostly the case though not
+always: as synchronizing scheduling decisions across 2 or more CPUs in a core
+involves additional overhead - especially when the system is lightly loaded
+(``total_threads <= N/2``, where N is the total number of CPUs).
+
+Usage
+-----
+Core scheduling support is enabled via the ``CONFIG_SCHED_CORE`` config option.
+Using this feature, userspace defines groups of tasks that trust each other.
+The core scheduler uses this information to make sure that tasks that do not
+trust each other will never run simultaneously on a core, while doing its best
+to satisfy the system's scheduling requirements.
+
+There are 2 ways to use core-scheduling:
+
+CGroup
+######
+Core scheduling adds additional files to the CPU controller CGroup:
+
+* ``cpu.tag``
+Writing ``1`` into this file results in all tasks in the group getting tagged.
+This results in all the CGroup's tasks allowed to run concurrently on a core's
+hyperthreads (also called siblings).
+
+The file being a value of ``0`` means the tag state of the CGroup is inherited
+from its parent hierarchy. If any ancestor of the CGroup is tagged, then the
+group is tagged.
+
+.. note:: Once a CGroup is tagged via cpu.tag, it is not possible to set this
+ for any descendant of the tagged group. For finer grained control, the
+ ``cpu.tag_color`` file described next may be used.
+
+.. note:: When a CGroup is not tagged, all the tasks within the group can share
+ a core with kernel threads and untagged system threads. For this reason,
+ if a group has ``cpu.tag`` of 0, it is considered to be trusted.
+
+* ``cpu.tag_color``
+For finer grained control over core sharing, a color can also be set in
+addition to the tag. This allows to further control core sharing between child
+CGroups within an already tagged CGroup. The color and the tag are both used to
+generate a `cookie` which is used by the scheduler to identify the group.
+
+Up to 256 different colors can be set (0-255) by writing into this file.
+
+A sample real-world usage of this file follows:
+
+Google uses DAC controls to make ``cpu.tag`` writable only by root and the
+``cpu.tag_color`` can be changed by anyone.
+
+The hierarchy looks like this:
+::
+ Root group
+ / \
+ A B (These are created by the root daemon - borglet).
+ / \ \
+ C D E (These are created by AppEngine within the container).
+
+A and B are containers for 2 different jobs or apps that are created by a root
+daemon called borglet. borglet then tags each of these group with the ``cpu.tag``
+file. The job itself can create additional child CGroups which are colored by
+the container's AppEngine with the ``cpu.tag_color`` file.
+
+The reason why Google uses this 2-level tagging system is that AppEngine wants to
+allow a subset of child CGroups within a tagged parent CGroup to be co-scheduled on a
+core while not being co-scheduled with other child CGroups. Think of these
+child CGroups as belonging to the same customer or project. Because these
+child CGroups are created by AppEngine, they are not tracked by borglet (the
+root daemon), therefore borglet won't have a chance to set a color for them.
+That's where cpu.tag_color file comes in. A color could be set by AppEngine,
+and once set, the normal tasks within the subcgroup would not be able to
+overwrite it. This is enforced by promoting the permission of the
+``cpu.tag_color`` file in cgroupfs.
+
+The color is an 8-bit value allowing for up to 256 unique colors.
+
+.. note:: Once a CGroup is colored, none of its descendants can be re-colored. Also
+ coloring of a CGroup is possible only if either the group or one of its
+ ancestors was tagged via the ``cpu.tag`` file.
+
+prctl interface
+###############
+A ``prtcl(2)`` command ``PR_SCHED_CORE_SHARE`` is available to a process to request
+sharing a core with another process. For example, consider 2 processes ``P1``
+and ``P2`` with PIDs 100 and 200. If process ``P1`` calls
+``prctl(PR_SCHED_CORE_SHARE, 200)``, the kernel makes ``P1`` share a core with ``P2``.
+The kernel performs ptrace access mode checks before granting the request.
+
+.. note:: This operation is not commutative. P1 calling
+ ``prctl(PR_SCHED_CORE_SHARE, pidof(P2)`` is not the same as P2 calling the
+ same for P1. The former case is P1 joining P2's group of processes
+ (which P2 would have joined with ``prctl(2)`` prior to P1's ``prctl(2)``).
+
+.. note:: The core-sharing granted with prctl(2) will be subject to
+ core-sharing restrictions specified by the CGroup interface. For example
+ if P1 and P2 are a part of 2 different tagged CGroups, then they will
+ not share a core even if a prctl(2) call is made. This is analogous
+ to how affinities are set using the cpuset interface.
+
+It is important to note that, on a ``CLONE_THREAD`` ``clone(2)`` syscall, the child
+will be assigned the same tag as its parent and thus be allowed to share a core
+with them. This design choice is because, for the security usecase, a
+``CLONE_THREAD`` child can access its parent's address space anyway, so there's
+no point in not allowing them to share a core. If a different behavior is
+desired, the child thread can call ``prctl(2)`` as needed. This behavior is
+specific to the ``prctl(2)`` interface. For the CGroup interface, the child of a
+fork always shares a core with its parent. On the other hand, if a parent
+was previously tagged via ``prctl(2)`` and does a regular ``fork(2)`` syscall, the
+child will receive a unique tag.
+
+Design/Implementation
+---------------------
+Each task that is tagged is assigned a cookie internally in the kernel. As
+mentioned in `Usage`_, tasks with the same cookie value are assumed to trust
+each other and share a core.
+
+The basic idea is that, every schedule event tries to select tasks for all the
+siblings of a core such that all the selected tasks running on a core are
+trusted (same cookie) at any point in time. Kernel threads are assumed trusted.
+The idle task is considered special, as it trusts everything and everything
+trusts it.
+
+During a schedule() event on any sibling of a core, the highest priority task on
+the sibling's core is picked and assigned to the sibling calling schedule(), if
+the sibling has the task enqueued. For rest of the siblings in the core,
+highest priority task with the same cookie is selected if there is one runnable
+in their individual run queues. If a task with same cookie is not available,
+the idle task is selected. Idle task is globally trusted.
+
+Once a task has been selected for all the siblings in the core, an IPI is sent to
+siblings for whom a new task was selected. Siblings on receiving the IPI will
+switch to the new task immediately. If an idle task is selected for a sibling,
+then the sibling is considered to be in a `forced idle` state. I.e., it may
+have tasks on its on runqueue to run, however it will still have to run idle.
+More on this in the next section.
+
+Forced-idling of tasks
+----------------------
+The scheduler tries its best to find tasks that trust each other such that all
+tasks selected to be scheduled are of the highest priority in a core. However,
+it is possible that some runqueues had tasks that were incompatible with the
+highest priority ones in the core. Favoring security over fairness, one or more
+siblings could be forced to select a lower priority task if the highest
+priority task is not trusted with respect to the core wide highest priority
+task. If a sibling does not have a trusted task to run, it will be forced idle
+by the scheduler (idle thread is scheduled to run).
+
+When the highest priority task is selected to run, a reschedule-IPI is sent to
+the sibling to force it into idle. This results in 4 cases which need to be
+considered depending on whether a VM or a regular usermode process was running
+on either HT::
+
+ HT1 (attack) HT2 (victim)
+ A idle -> user space user space -> idle
+ B idle -> user space guest -> idle
+ C idle -> guest user space -> idle
+ D idle -> guest guest -> idle
+
+Note that for better performance, we do not wait for the destination CPU
+(victim) to enter idle mode. This is because the sending of the IPI would bring
+the destination CPU immediately into kernel mode from user space, or VMEXIT
+in the case of guests. At best, this would only leak some scheduler metadata
+which may not be worth protecting. It is also possible that the IPI is received
+too late on some architectures, but this has not been observed in the case of
+x86.
+
+Kernel protection from untrusted tasks
+--------------------------------------
+The scheduler on its own cannot protect the kernel executing concurrently with
+an untrusted task in a core. This is because the scheduler is unaware of
+interrupts/syscalls at scheduling time. To mitigate this, an IPI is sent to
+siblings on kernel entry (syscall and IRQ). This IPI forces the sibling to enter
+kernel mode and wait before returning to user until all siblings of the
+core have left kernel mode. This process is also known as stunning. For good
+performance, an IPI is sent only to a sibling only if it is running a tagged
+task. If a sibling is running a kernel thread or is idle, no IPI is sent.
+
+The kernel protection feature can be turned off on the kernel command line by
+passing ``sched_core_protect_kernel=0``.
+
+Other alternative ideas discussed for kernel protection are listed below just
+for completeness. They all have limitations:
+
+1. Changing interrupt affinities to a trusted core which does not execute untrusted tasks
+#########################################################################################
+By changing the interrupt affinities to a designated safe-CPU which runs
+only trusted tasks, IRQ data can be protected. One issue is this involves
+giving up a full CPU core of the system to run safe tasks. Another is that,
+per-cpu interrupts such as the local timer interrupt cannot have their
+affinity changed. Also, sensitive timer callbacks such as the random entropy timer
+can run in softirq on return from these interrupts and expose sensitive
+data. In the future, that could be mitigated by forcing softirqs into threaded
+mode by utilizing a mechanism similar to ``CONFIG_PREEMPT_RT``.
+
+Yet another issue with this is, for multiqueue devices with managed
+interrupts, the IRQ affinities cannot be changed; however it could be
+possible to force a reduced number of queues which would in turn allow to
+shield one or two CPUs from such interrupts and queue handling for the price
+of indirection.
+
+2. Running IRQs as threaded-IRQs
+################################
+This would result in forcing IRQs into the scheduler which would then provide
+the process-context mitigation. However, not all interrupts can be threaded.
+Also this does nothing about syscall entries.
+
+3. Kernel Address Space Isolation
+#################################
+System calls could run in a much restricted address space which is
+guaranteed not to leak any sensitive data. There are practical limitation in
+implementing this - the main concern being how to decide on an address space
+that is guaranteed to not have any sensitive data.
+
+4. Limited cookie-based protection
+##################################
+On a system call, change the cookie to the system trusted cookie and initiate a
+schedule event. This would be better than pausing all the siblings during the
+entire duration for the system call, but still would be a huge hit to the
+performance.
+
+Trust model
+-----------
+Core scheduling maintains trust relationships amongst groups of tasks by
+assigning the tag of them with the same cookie value.
+When a system with core scheduling boots, all tasks are considered to trust
+each other. This is because the core scheduler does not have information about
+trust relationships until userspace uses the above mentioned interfaces, to
+communicate them. In other words, all tasks have a default cookie value of 0.
+and are considered system-wide trusted. The stunning of siblings running
+cookie-0 tasks is also avoided.
+
+Once userspace uses the above mentioned interfaces to group sets of tasks, tasks
+within such groups are considered to trust each other, but do not trust those
+outside. Tasks outside the group also don't trust tasks within.
+
+Limitations
+-----------
+Core scheduling tries to guarantee that only trusted tasks run concurrently on a
+core. But there could be small window of time during which untrusted tasks run
+concurrently or kernel could be running concurrently with a task not trusted by
+kernel.
+
+1. IPI processing delays
+########################
+Core scheduling selects only trusted tasks to run together. IPI is used to notify
+the siblings to switch to the new task. But there could be hardware delays in
+receiving of the IPI on some arch (on x86, this has not been observed). This may
+cause an attacker task to start running on a CPU before its siblings receive the
+IPI. Even though cache is flushed on entry to user mode, victim tasks on siblings
+may populate data in the cache and micro architectural buffers after the attacker
+starts to run and this is a possibility for data leak.
+
+Open cross-HT issues that core scheduling does not solve
+--------------------------------------------------------
+1. For MDS
+##########
+Core scheduling cannot protect against MDS attacks between an HT running in
+user mode and another running in kernel mode. Even though both HTs run tasks
+which trust each other, kernel memory is still considered untrusted. Such
+attacks are possible for any combination of sibling CPU modes (host or guest mode).
+
+2. For L1TF
+###########
+Core scheduling cannot protect against an L1TF guest attacker exploiting a
+guest or host victim. This is because the guest attacker can craft invalid
+PTEs which are not inverted due to a vulnerable guest kernel. The only
+solution is to disable EPT (Extended Page Tables).
+
+For both MDS and L1TF, if the guest vCPU is configured to not trust each
+other (by tagging separately), then the guest to guest attacks would go away.
+Or it could be a system admin policy which considers guest to guest attacks as
+a guest problem.
+
+Another approach to resolve these would be to make every untrusted task on the
+system to not trust every other untrusted task. While this could reduce
+parallelism of the untrusted tasks, it would still solve the above issues while
+allowing system processes (trusted tasks) to share a core.
+
+Use cases
+---------
+The main use case for Core scheduling is mitigating the cross-HT vulnerabilities
+with SMT enabled. There are other use cases where this feature could be used:
+
+- Isolating tasks that needs a whole core: Examples include realtime tasks, tasks
+ that uses SIMD instructions etc.
+- Gang scheduling: Requirements for a group of tasks that needs to be scheduled
+ together could also be realized using core scheduling. One example is vCPUs of
+ a VM.
+
+Future work
+-----------
+Skipping per-HT mitigations if task is trusted
+##############################################
+If core scheduling is enabled, by default all tasks trust each other as
+mentioned above. In such scenario, it may be desirable to skip the same-HT
+mitigations on return to the trusted user-mode to improve performance.
diff --git a/Documentation/admin-guide/hw-vuln/index.rst b/Documentation/admin-guide/hw-vuln/index.rst
index 21710f8609fe..361ccbbd9e54 100644
--- a/Documentation/admin-guide/hw-vuln/index.rst
+++ b/Documentation/admin-guide/hw-vuln/index.rst
@@ -16,3 +16,4 @@ are configurable at compile, boot or run time.
multihit.rst
special-register-buffer-data-sampling.rst
l1d_flush.rst
+ core-scheduling.rst
--
2.29.2.222.g5d2a92d10f8-goog