mirror of
git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
synced 2025-09-04 20:19:47 +08:00
7011d72588
The "First Fault Register" (FFR) is an SVE register that mimics a
predicate register, but clears bits when a load or store fails to handle
an element of a vector. The supposed usage scenario is to initialise
this register (using SETFFR), then *read* it later on to learn about
elements that failed to load or store. Explicit writes to this register
using the WRFFR instruction are only supposed to *restore* values
previously read from the register (for context-switching only).
As the manual describes, this register holds only certain values, it:
"... contains a monotonic predicate value, in which starting from bit 0
there are zero or more 1 bits, followed only by 0 bits in any remaining
bit positions."
Any other value is UNPREDICTABLE and is not supposed to be "restored"
into the register.
The SVE test currently tries to write a signature pattern into the
register, which is *not* a canonical FFR value. Apparently the existing
setups treat UNPREDICTABLE as "read-as-written", but a new
implementation actually only stores canonical values. As a consequence,
the sve-test fails immediately when comparing the FFR value:
-----------
# ./sve-test
Vector length: 128 bits
PID: 207
Mismatch: PID=207, iteration=0, reg=48
Expected [cf00]
Got [0f00]
Aborted
-----------
Fix this by only populating the FFR with proper canonical values.
Effectively the requirement described above limits us to 17 unique
values over 16 bits worth of FFR, so we condense our signature down to 4
bits (2 bits from the PID, 2 bits from the generation) and generate the
canonical pattern from it. Any bits describing elements above the
minimum 128 bit are set to 0.
This aligns the FFR usage to the architecture and fixes the test on
microarchitectures implementing FFR in a more restricted way.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Reviwed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20210319120128.29452-1-andre.przywara@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
|
||
|---|---|---|
| .. | ||
| .gitignore | ||
| asm-offsets.h | ||
| assembler.h | ||
| fpsimd-stress | ||
| fpsimd-test.S | ||
| Makefile | ||
| README | ||
| sve-probe-vls.c | ||
| sve-ptrace-asm.S | ||
| sve-ptrace.c | ||
| sve-stress | ||
| sve-test.S | ||
| vlset.c | ||
This directory contains a mix of tests integrated with kselftest and standalone stress tests. kselftest tests =============== sve-probe-vls - Checks the SVE vector length enumeration interface sve-ptrace - Checks the SVE ptrace interface Running the non-kselftest tests =============================== sve-stress performs an SVE context switch stress test, as described below. (The fpsimd-stress test works the same way; just substitute "fpsimd" for "sve" in the following commands.) The test runs until killed by the user. If no context switch error was detected, you will see output such as the following: $ ./sve-stress (wait for some time) ^C Vector length: 512 bits PID: 1573 Terminated by signal 15, no error, iterations=9467, signals=1014 Vector length: 512 bits PID: 1575 Terminated by signal 15, no error, iterations=9448, signals=1028 Vector length: 512 bits PID: 1577 Terminated by signal 15, no error, iterations=9436, signals=1039 Vector length: 512 bits PID: 1579 Terminated by signal 15, no error, iterations=9421, signals=1039 Vector length: 512 bits PID: 1581 Terminated by signal 15, no error, iterations=9403, signals=1039 Vector length: 512 bits PID: 1583 Terminated by signal 15, no error, iterations=9385, signals=1036 Vector length: 512 bits PID: 1585 Terminated by signal 15, no error, iterations=9376, signals=1039 Vector length: 512 bits PID: 1587 Terminated by signal 15, no error, iterations=9361, signals=1039 Vector length: 512 bits PID: 1589 Terminated by signal 15, no error, iterations=9350, signals=1039 If an error was detected, details of the mismatch will be printed instead of "no error". Ideally, the test should be allowed to run for many minutes or hours to maximise test coverage. KVM stress testing ================== To try to reproduce the bugs that we have been observing, sve-stress should be run in parallel in two KVM guests, while simultaneously running on the host. 1) Start 2 guests, using the following command for each: $ lkvm run --console=virtio -pconsole=hvc0 --sve Image (Depending on the hardware GIC implementation, you may also need --irqchip=gicv3. New kvmtool defaults to that if appropriate, but I can't remember whether my branch is new enough for that. Try without the option first.) Kvmtool occupies the terminal until you kill it (Ctrl+A x), or until the guest terminates. It is therefore recommended to run each instance in separate terminal (use screen or ssh etc.) This allows multiple guests to be run in parallel while running other commands on the host. Within the guest, the host filesystem is accessible, mounted on /host. 2) Run the sve-stress on *each* guest with the Vector-Length set to 32: guest$ ./vlset --inherit 32 ./sve-stress 3) Run the sve-stress on the host with the maximum Vector-Length: host$ ./vlset --inherit --max ./sve-stress Again, the test should be allowed to run for many minutes or hours to maximise test coverage. If no error is detected, you will see output from each sve-stress instance similar to that illustrated above; otherwise details of the observed mismatches will be printed.