diff mbox series

[RFC,13/13] Documentation: Document PowerPC kernel DEXCR interface

Message ID 20221128024458.46121-14-bgray@linux.ibm.com (mailing list archive)
State Handled Elsewhere
Headers show
Series Add DEXCR support | expand

Commit Message

Benjamin Gray Nov. 28, 2022, 2:44 a.m. UTC
Describe the DEXCR and document how to interact with it via the
prctl and sysctl interfaces.

Signed-off-by: Benjamin Gray <bgray@linux.ibm.com>
---
 Documentation/powerpc/dexcr.rst | 183 ++++++++++++++++++++++++++++++++
 Documentation/powerpc/index.rst |   1 +
 2 files changed, 184 insertions(+)
 create mode 100644 Documentation/powerpc/dexcr.rst

Comments

Nicholas Piggin March 7, 2023, 5:40 a.m. UTC | #1
On Mon Nov 28, 2022 at 12:44 PM AEST, Benjamin Gray wrote:
> Describe the DEXCR and document how to interact with it via the
> prctl and sysctl interfaces.

Oh you've got the docs here, sorry. Thanks for that. I don't know enough
yet to give much useful feedback on the API. I think at least all the
mechanism stuff up to the prctl API looks pretty straightfoward so would
like to get that merged if we can.

Might need a bit more time and discussion on the API. Interestingly
because the hashchk aspect is architectural, we may not be able to
necessarily sanely enable that, because if it was disabled to start
out with, our callchain up to the prctl call I think would have no
return hashes set so we'd immediately fail on our first return.

Thanks,
Nick
Benjamin Gray March 7, 2023, 5:52 a.m. UTC | #2
On Tue, 2023-03-07 at 15:40 +1000, Nicholas Piggin wrote:
> Might need a bit more time and discussion on the API. Interestingly
> because the hashchk aspect is architectural, we may not be able to
> necessarily sanely enable that, because if it was disabled to start
> out with, our callchain up to the prctl call I think would have no
> return hashes set so we'd immediately fail on our first return.

I assumed it could eventually be supported in whatever startup wrapper
programs are built with, so either as one of the first things before
any calls, or the compiler could skip putting hash instructions in the
wrapper altogether. The ELF file itself might even be able to request
bits be enabled, so the kernel would start the process correctly.

As it is inherited, it's also possible for a wrapper program to set a
specific DEXCR before a child runs. E.g.,

	fork();
	prctl(...NPHIE...);
	exec();

I hadn't thought of the prctl call itself causing an unbalanced hashchk
when it returns, but that should be solvable with an inline syscall.
diff mbox series

Patch

diff --git a/Documentation/powerpc/dexcr.rst b/Documentation/powerpc/dexcr.rst
new file mode 100644
index 000000000000..3c995f4b9fe0
--- /dev/null
+++ b/Documentation/powerpc/dexcr.rst
@@ -0,0 +1,183 @@ 
+==========================================
+DEXCR (Dynamic Execution Control Register)
+==========================================
+
+Overview
+========
+
+The DEXCR is a privileged special purpose register (SPR) introduced in
+PowerPC ISA 3.1B (Power10) that allows per-cpu control over several dynamic
+execution behaviours. These behaviours include speculation (e.g., indirect
+branch target prediction) and enabling return-oriented programming (ROP)
+protection instructions.
+
+The execution control is exposed in hardware as up to 32 bits ('aspects') in
+the DEXCR. Each aspect controls a certain behaviour, and can be set or cleared
+to enable/disable the aspect. There are several variants of the DEXCR for
+different purposes:
+
+DEXCR
+    A priviliged SPR that can control aspects for userspace and kernel space
+HDEXCR
+    A hypervisor-privileged SPR that can control aspects for the hypervisor and
+    enforce aspects for the kernel and userspace.
+UDEXCR
+    An optional ultravisor-privileged SPR that can control aspects for the ultravisor.
+
+Userspace can examine the current DEXCR state using a dedicated SPR that
+provides a non-privileged read-only view of the userspace DEXCR aspects.
+There is also an SPR that provides a read-only view of the hypervisor enforced
+aspects, which ORed with the userspace DEXCR view gives the effective DEXCR
+state for a process.
+
+
+User API
+========
+
+prctl()
+-------
+
+A process can control its own userspace DEXCR value using the
+``PR_PPC_GET_DEXCR`` and ``PR_PPC_SET_DEXCR`` pair of
+:manpage:`prctl(2)` commands. These calls have the form::
+
+    prctl(PR_PPC_GET_DEXCR, unsigned long aspect, 0, 0, 0);
+    prctl(PR_PPC_SET_DEXCR, unsigned long aspect, unsigned long flags, 0, 0);
+
+Where ``aspect`` (``arg1``) is a constant and ``flags`` (``arg2``) is a bifield.
+The possible aspect and flag values are as follows. Note there is no relation
+between aspect value and ``prctl()`` constant value.
+
+.. flat-table::
+   :header-rows: 1
+   :widths: 2 7 1
+
+   * - ``prctl()`` constant
+     - Aspect name
+     - Aspect bit
+
+   * - ``PR_PPC_DEXCR_SBHE``
+     - Speculative Branch Hint Enable (SBHE)
+     - 0
+
+   * - ``PR_PPC_DEXCR_IBRTPD``
+     - Indirect Branch Recurrent Target Prediction Disable (IBRTPD)
+     - 3
+
+   * - ``PR_PPC_DEXCR_SRAPD``
+     - Subroutine Return Address Prediction Disable (SRAPD)
+     - 4
+
+   * - ``PR_PPC_DEXCR_NPHIE``
+     - Non-Privileged Hash Instruction Enable (NPHIE)
+     - 5
+
+.. flat-table::
+   :header-rows: 1
+   :widths: 2 8
+
+   * - ``prctl()`` flag
+     - Meaning
+
+   * - ``PR_PPC_DEXCR_PRCTL``
+     - This aspect can be configured with ``prctl(PR_PPC_SET_DEXCR, ...)``
+
+   * - ``PR_PPC_DEXCR_SET_ASPECT``
+     - This aspect is set
+
+   * - ``PR_PPC_DEXCR_FORCE_SET_ASPECT``
+     - This aspect is set and cannot be undone. A subsequent
+       ``prctl(..., PR_PPC_DEXCR_CLEAR_ASPECT)`` will fail.
+
+   * - ``PR_PPC_DEXCR_CLEAR_ASPECT``
+     - This aspect is clear
+
+Note that
+
+* The ``*_SET_ASPECT`` / ``*_CLEAR_ASPECT`` refers to setting/clearing the bit in the DEXCR.
+  For example::
+
+      prctl(PR_PPC_SET_DEXCR, PR_PPC_DEXCR_IBRTPD, PR_PPC_DEXCR_SET_ASPECT, 0, 0);
+
+  will set the IBRTPD aspect bit in the DEXCR, causing indirect branch prediction
+  to be disabled.
+
+* The status returned by ``PR_PPC_GET_DEXCR`` does not include any alternative
+  config overrides. To see the true DEXCR state software should read the appropriate
+  SPRs directly.
+
+* A forced aspect will still report ``PR_PPC_DEXCR_PRCTL`` if it would
+  otherwise be editable.
+
+* The aspect state when starting a process is copied from the parent's
+  state on :manpage:`fork(2)` and :manpage:`execve(2)`. Aspects may also be set
+  or cleared by the kernel on process creation.
+
+Use ``PR_PPC_SET_DEXCR`` with one of ``PR_PPC_DEXCR_SET_ASPECT``,
+``PR_PPC_DEXCR_FORCE_SET_ASPECT``, or ``PR_PPC_DEXCR_CLEAR_ASPECT`` to edit a
+ given aspect.
+
+Common error codes for both getting and setting the DEXCR are as follows:
+
+.. flat-table::
+   :header-rows: 1
+   :widths: 2 8
+
+   * - Error
+     - Meaning
+
+   * - ``EINVAL``
+     - The DEXCR is not supported by the kernel.
+
+   * - ``ENODEV``
+     - The aspect is not recognised by the kernel or not supported by the hardware.
+
+``PR_PPC_SET_DEXCR`` may also report the following error codes:
+
+.. flat-table::
+   :header-rows: 1
+   :widths: 2 8
+
+   * - Error
+     - Meaning
+
+   * - ``ERANGE``
+     - ``arg2`` is incorrect. E.g., it does not select an action (set/clear),
+       or the flags are not recognised by the kernel.
+
+   * - ``ENXIO``
+     - The aspect is not editable via ``prctl()``.
+
+   * - ``EPERM``
+     - The process does not have sufficient privilege to modify this aspect,
+       or the aspect has been force set and cannot be modified.
+
+
+sysctl
+------
+
+Some aspects can be modified globally via :manpage:`sysctl(8)` entries. Such global
+modifications are applied after any process modifications. Any ``prctl()`` call to
+an overridden aspect this aspect may still report it as editable. The prctl setting
+will take effect again if the global override is restored to its default state.
+
+A global SBHE config is exposed in ``/proc/sys/kernel/speculative_branch_hint_enable``.
+Any process can read the current config value from it. Privileged processes can
+write to it to change the config. The new config is applied to all current and future
+processes (though note the kernel cannot override any hypervisor enforced aspects).
+
+.. flat-table::
+   :header-rows: 1
+   :widths: 2 8
+
+   * - Value
+     - Meaning
+
+   * - ``-1``
+     - Do not change from default or ``prctl()`` config.
+
+   * - ``0``
+     - Force clear aspect.
+
+   * - ``1``
+     - Force set aspect.
diff --git a/Documentation/powerpc/index.rst b/Documentation/powerpc/index.rst
index 85e80e30160b..d33b554ca7ba 100644
--- a/Documentation/powerpc/index.rst
+++ b/Documentation/powerpc/index.rst
@@ -15,6 +15,7 @@  powerpc
     cxl
     cxlflash
     dawr-power9
+    dexcr
     dscr
     eeh-pci-error-recovery
     elf_hwcaps