scripts: Fix accel handling
cover scripts: Fix accel handling
When running on a system without KVM but with a /dev/kvm file or when /dev/kvm has the wrong permissions we will think that we have the kvm accelerator because we only check if /dev/kvm exists. To fix that we instead start a qemu with the kvm accel and check the exit value we can check if kvm is available.
Also we only compare the accel specified in unittests.conf with the env ACCEL. That won't help if we don't have kvm but a test has KVM as a requirement in unittests.conf.
My bash knowledge is rather limited, so maybe there's a better solution?
Janosch Frank (2):
scripts: Check kvm availability by asking qemu
scripts: Set ACCEL in run_tests.sh if empty
arm/run | 4 +-- powerpc/run | 4 +-- run_tests.sh | 6 +++++ s390x/run | 10 ++++--- scripts/accel.bash | 63 +++++++++++++++++++++++++++++++++++++++++++ scripts/arch-run.bash | 63 ++----------------------------------------- scripts/runtime.bash | 2 +- x86/run | 4 +-- 8 files changed, 85 insertions(+), 71 deletions(-) create mode 100644 scripts/accel.bash
Signature · 1 line
2.27.0
RFC 1/2 scripts: Check kvm availability by asking qemu
The existence of the /dev/kvm character device doesn't imply that the kvm module is part of the kernel or that it's always magically loaded when needed.
--- arm/run | 4 ++-- powerpc/run | 4 ++-- s390x/run | 4 ++-- scripts/arch-run.bash | 7 +++++-- x86/run | 4 ++-- 5 files changed, 13 insertions(+), 10 deletions(-)
arm/run +2 -2
@@ -10,10 +10,10 @@ if [ -z "$STANDALONE" ]; then fi processor="$PROCESSOR" -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $? -qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $? if ! $qemu -machine '?' 2>&1 | grep 'ARM Virtual Machine' > /dev/null; then
powerpc/run +2 -2
@@ -9,10 +9,10 @@ if [ -z "$STANDALONE" ]; then source scripts/arch-run.bash fi -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $? -qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $? if ! $qemu -machine '?' 2>&1 | grep 'pseries' > /dev/null; then
s390x/run +2 -2
@@ -9,10 +9,10 @@ if [ -z "$STANDALONE" ]; then source scripts/arch-run.bash fi -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $? -qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $? M='-machine s390-ccw-virtio'
scripts/arch-run.bash +5 -2
@@ -342,8 +342,11 @@ trap_exit_push () kvm_available () { - [ -c /dev/kvm ] || - return 1 + if $($qemu -accel kvm 2> /dev/null); then + return 0; + else + return 1; + fi [ "$HOST" = "$ARCH_NAME" ] || ( [ "$HOST" = aarch64 ] && [ "$ARCH" = arm ] ) ||
x86/run +2 -2
@@ -9,10 +9,10 @@ if [ -z "$STANDALONE" ]; then source scripts/arch-run.bash fi -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $? -qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $? if ! ${qemu} -device '?' 2>&1 | grep -F -e \"testdev\" -e \"pc-testdev\" > /dev/null;
Signature · 1 line
2.27.0
Re: [kvm-unit-tests RFC 1/2] scripts: Check kvm availability by asking qemu
Quoted 6 lines · Janosch Frank
scripts/arch-run.bash +0
-0@@ -342,8 +342,11 @@ trap_exit_push ()
scripts/arch-run.bash trap_exit_push () · line 350
Quoted 10 lines
scripts/arch-run.bash +5
-2kvm_available () { - [ -c /dev/kvm ] || - return 1 + if $($qemu -accel kvm 2> /dev/null); then + return 0; + else + return 1; + fi
Hi Janosch,
Are we sure that even old QEMU supports this type of probing? If not, then we should probably keep the /dev/kvm test as a fallback.
Thanks,
drew
Re: [kvm-unit-tests RFC 1/2] scripts: Check kvm availability by asking qemu
In reply to Janosch FrankQuoted 18 lines · Janosch Frank
The existence of the
/dev/kvmcharacter device doesn't imply that the kvm module is part of the kernel or that it's always magically loaded when needed.Signed-off-by Janosch Frank <[email protected]>--- arm/run | 4 ++-- powerpc/run | 4 ++-- s390x/run | 4 ++-- scripts/arch-run.bash | 7 +++++-- x86/run | 4 ++-- 5 files changed, 13 insertions(+), 10 deletions(-)arm/run +0
-0
arm/run if [ -z "$STANDALONE" ]; then · line 17
Quoted 10 lines
arm/run +2
-2@@ -10,10 +10,10 @@ if [ -z "$STANDALONE" ]; then fi processor="$PROCESSOR" -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $? -qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $?
How about renaming search_qemu_binary() to set_qemu_accelerator(), which would also ensure QEMU is set (if it doesn't error out on failure) and then call that from get_qemu_accelerator()? That way we don't need to worry about this order of calls nor this lowercase 'qemu' variable being set. Also, we can rename get_qemu_accelerator() to set_qemu_accelerator() and ensure it sets ACCEL.
Thanks,
drew
Re: [kvm-unit-tests RFC 1/2] scripts: Check kvm availability by asking qemu
Quoted 26 lines · Andrew Jones
Quoted 25 lines · Janosch Frank
The existence of the
/dev/kvmcharacter device doesn't imply that the kvm module is part of the kernel or that it's always magically loaded when needed.Signed-off-by Janosch Frank <[email protected]>--- arm/run | 4 ++-- powerpc/run | 4 ++-- s390x/run | 4 ++-- scripts/arch-run.bash | 7 +++++-- x86/run | 4 ++-- 5 files changed, 13 insertions(+), 10 deletions(-)arm/run +1
-1@@ -10,10 +10,10 @@ if [ -z "$STANDALONE" ]; then fi processor="$PROCESSOR" -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $?
Quoted 10 lines
arm/run if [ -z "$STANDALONE" ]; then · line 17
Quoted 3 lines
arm/run +1
-1-qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $?How about renaming
search_qemu_binary()toset_qemu_accelerator(), which would also ensure QEMU is set (if it doesn't error out on failure) and then call that fromget_qemu_accelerator()? That way we don't need to worry about this order of calls nor this lowercase 'qemu' variable being set. Also, we can renameget_qemu_accelerator()toset_qemu_accelerator()and ensure it sets ACCEL.
Sure, I was already considering that.
The fact that qemu and ACCEL are basically global and accel is not the same as ACCEL makes all of this harder to understand than it should be.
Quoted 4 lines
Thanks,
drew
Re: [kvm-unit-tests RFC 1/2] scripts: Check kvm availability by asking qemu
In reply to Andrew JonesQuoted 37 lines · Andrew Jones
Quoted 18 lines · Janosch Frank
The existence of the
/dev/kvmcharacter device doesn't imply that the kvm module is part of the kernel or that it's always magically loaded when needed.Signed-off-by Janosch Frank <[email protected]>--- arm/run | 4 ++-- powerpc/run | 4 ++-- s390x/run | 4 ++-- scripts/arch-run.bash | 7 +++++-- x86/run | 4 ++-- 5 files changed, 13 insertions(+), 10 deletions(-)arm/run +0
-0arm/run if [ -z "$STANDALONE" ]; then · line 17
Quoted 10 lines
arm/run +2
-2@@ -10,10 +10,10 @@ if [ -z "$STANDALONE" ]; then fi processor="$PROCESSOR" -ACCEL=$(get_qemu_accelerator) || +qemu=$(search_qemu_binary) || exit $? -qemu=$(search_qemu_binary) || +ACCEL=$(get_qemu_accelerator) || exit $?How about renaming
search_qemu_binary()toset_qemu_accelerator(), which would also ensure QEMU is set (if it doesn't error out on failure) and then call that fromget_qemu_accelerator()? That way we don't need to worry about this order of calls nor this lowercase 'qemu' variable being set. Also, we can renameget_qemu_accelerator()toset_qemu_accelerator()and ensure it sets ACCEL.
Quoted 3 lines
Thanks,
drew
I've already broken off the accel.bash change and fixed the arch/run problem but this here will have two wait until after my easter vacation so don't wait for patches the next ~2 weeks.
RFC 2/2 scripts: Set ACCEL in run_tests.sh if empty
In reply to Janosch FrankThe current checks compare the env ACCEL to the unittests.conf provided accel. That's all fine as long as the user always specifies the ACCEL env variable.
If that's not the case and KVM is not available or if a test specifies tcg and we start a qemu with the kvm acceleration as it's the default we'll run into problems.
So let's fetch the accelerator before calling the arch/run script and check it against the test's specified accel. Yes, we now do that twice, once in the run_tests.sh and one in arch/run, but I don't think there's a good way around it since you can execute arch/run without run_tests.sh.
--- run_tests.sh | 6 ++++ s390x/run | 6 +++- scripts/accel.bash | 63 +++++++++++++++++++++++++++++++++++++++++ scripts/arch-run.bash | 66 ++----------------------------------------- scripts/runtime.bash | 2 +- 5 files changed, 77 insertions(+), 66 deletions(-) create mode 100644 scripts/accel.bash
run_tests.sh +6 -0
@@ -10,6 +10,7 @@ if [ ! -f config.mak ]; then fi source config.mak source scripts/common.bash +source scripts/accel.bash function usage() {
@@ -164,6 +165,11 @@ if [[ $tap_output == "yes" ]]; then echo "TAP version 13" fi +qemu=$(search_qemu_binary) +if [ -z "$ACCEL" ]; then + ACCEL=$(get_qemu_accelerator) +fi + trap "wait; exit 130" SIGINT (
s390x/run +5 -1
@@ -12,8 +12,12 @@ fi qemu=$(search_qemu_binary) || exit $? -ACCEL=$(get_qemu_accelerator) || +if [ -z "$DEF_ACCEL "]; then + ACCEL=$(get_qemu_accelerator) || exit $? +else + ACCEL=$DEF_ACCEL +fi M='-machine s390-ccw-virtio' M+=",accel=$ACCEL"
scripts/accel.bash new file +63 -0
63 lines not on this page · raw message
scripts/arch-run.bash +2 -64
81 lines not on this page · raw message
scripts/runtime.bash +1 -1
@@ -30,7 +30,7 @@ premature_failure() get_cmdline() { local kernel=$1 - echo "TESTNAME=$testname TIMEOUT=$timeout ACCEL=$accel $RUNTIME_arch_run $kernel -smp $smp $opts" + echo "TESTNAME=$testname TIMEOUT=$timeout DEF_ACCEL=$accel $RUNTIME_arch_run $kernel -smp $smp $opts" } skip_nodefault()
Signature · 1 line
2.27.0
Re: [kvm-unit-tests RFC 2/2] scripts: Set ACCEL in run_tests.sh if empty
Quoted 228 lines · Janosch Frank
The current checks compare the env ACCEL to the unittests.conf provided accel. That's all fine as long as the user always specifies the ACCEL env variable.
If that's not the case and KVM is not available or if a test specifies tcg and we start a qemu with the kvm acceleration as it's the default we'll run into problems.
So let's fetch the accelerator before calling the
arch/runscript and check it against the test's specified accel. Yes, we now do that twice, once in therun_tests.shand one inarch/run, but I don't think there's a good way around it since you can executearch/runwithoutrun_tests.sh.Signed-off-by Janosch Frank <[email protected]>--- run_tests.sh | 6 ++++ s390x/run | 6 +++- scripts/accel.bash | 63 +++++++++++++++++++++++++++++++++++++++++ scripts/arch-run.bash | 66 ++----------------------------------------- scripts/runtime.bash | 2 +- 5 files changed, 77 insertions(+), 66 deletions(-) create mode 100644 scripts/accel.bashrun_tests.sh +6
-0@@ -10,6 +10,7 @@ if [ ! -f config.mak ]; then fi source config.mak source scripts/common.bash +source scripts/accel.bash function usage() {@@ -164,6 +165,11 @@ if [[ $tap_output == "yes" ]]; then echo "TAP version 13" fi +qemu=$(search_qemu_binary) +if [ -z "$ACCEL" ]; then + ACCEL=$(get_qemu_accelerator) +fi + trap "wait; exit 130" SIGINT (s390x/run +5
-1@@ -12,8 +12,12 @@ fi qemu=$(search_qemu_binary) || exit $? -ACCEL=$(get_qemu_accelerator) || +if [ -z "$DEF_ACCEL "]; then + ACCEL=$(get_qemu_accelerator) || exit $? +else + ACCEL=$DEF_ACCEL +fi M='-machine s390-ccw-virtio' M+=",accel=$ACCEL"scripts/accel.bash new file +63
-0@@ -0,0 +1,63 @@ +search_qemu_binary () +{ + local save_path=$PATH + local qemucmd qemu + + export PATH=$PATH:/usr/libexec + for qemucmd in ${QEMU:-qemu-system-$ARCH_NAME qemu-kvm}; do + if $qemucmd --help 2>/dev/null | grep -q 'QEMU'; then + qemu="$qemucmd" + break + fi + done + + if [ -z "$qemu" ]; then + echo "A QEMU binary was not found." >&2 + echo "You can set a custom location by using the QEMU=<path> environment variable." >&2 + return 2 + fi + command -v $qemu + export PATH=$save_path +} + +kvm_available () +{ + if $($qemu -accel kvm 2> /dev/null); then + return 0; + else + return 1; + fi + + [ "$HOST" = "$ARCH_NAME" ] || + ( [ "$HOST" = aarch64 ] && [ "$ARCH" = arm ] ) || + ( [ "$HOST" = x86_64 ] && [ "$ARCH" = i386 ] ) +} + +hvf_available () +{ + [ "$(sysctl -n kern.hv_support 2>/dev/null)" = "1" ] || return 1 + [ "$HOST" = "$ARCH_NAME" ] || + ( [ "$HOST" = x86_64 ] && [ "$ARCH" = i386 ] ) +} + +get_qemu_accelerator () +{ + if [ "$ACCEL" = "kvm" ] && ! kvm_available; then + echo "KVM is needed, but not available on this host" >&2 + return 2 + fi + if [ "$ACCEL" = "hvf" ] && ! hvf_available; then + echo "HVF is needed, but not available on this host" >&2 + return 2 + fi + + if [ "$ACCEL" ]; then + echo $ACCEL + elif kvm_available; then + echo kvm + elif hvf_available; then + echo hvf + else + echo tcg + fi +}scripts/arch-run.bash +2
-64@@ -24,6 +24,8 @@ # 1..127 - FAILURE (could be QEMU, a run script, or the unittest) # >= 128 - Signal (signum = status - 128) ############################################################################## +source scripts/accel.bash + run_qemu () { local stdout errors ret sig@@ -171,28 +173,6 @@ migration_cmd () fi } -search_qemu_binary () -{ - local save_path=$PATH - local qemucmd qemu - - export PATH=$PATH:/usr/libexec - for qemucmd in ${QEMU:-qemu-system-$ARCH_NAME qemu-kvm}; do - if $qemucmd --help 2>/dev/null | grep -q 'QEMU'; then - qemu="$qemucmd" - break - fi - done - - if [ -z "$qemu" ]; then - echo "A QEMU binary was not found." >&2 - echo "You can set a custom location by using the QEMU=<path> environment variable." >&2 - return 2 - fi - command -v $qemu - export PATH=$save_path -} - initrd_create () { if [ "$ENVIRON_DEFAULT" = "yes" ]; then@@ -339,45 +319,3 @@ trap_exit_push () local old_exit=$(trap -p EXIT | sed "s/^[^']*'//;s/'[^']*$//") trap -- "$1; $old_exit" EXIT } - -kvm_available () -{ - if $($qemu -accel kvm 2> /dev/null); then - return 0; - else - return 1; - fi - - [ "$HOST" = "$ARCH_NAME" ] || - ( [ "$HOST" = aarch64 ] && [ "$ARCH" = arm ] ) || - ( [ "$HOST" = x86_64 ] && [ "$ARCH" = i386 ] ) -} - -hvf_available () -{ - [ "$(sysctl -n kern.hv_support 2>/dev/null)" = "1" ] || return 1 - [ "$HOST" = "$ARCH_NAME" ] || - ( [ "$HOST" = x86_64 ] && [ "$ARCH" = i386 ] ) -} - -get_qemu_accelerator () -{ - if [ "$ACCEL" = "kvm" ] && ! kvm_available; then - echo "KVM is needed, but not available on this host" >&2 - return 2 - fi - if [ "$ACCEL" = "hvf" ] && ! hvf_available; then - echo "HVF is needed, but not available on this host" >&2 - return 2 - fi - - if [ "$ACCEL" ]; then - echo $ACCEL - elif kvm_available; then - echo kvm - elif hvf_available; then - echo hvf - else - echo tcg - fi -}scripts/runtime.bash +0
-0
scripts/runtime.bash premature_failure() · line 33
Quoted 5 lines
scripts/runtime.bash +1
-1@@ -30,7 +30,7 @@ premature_failure() get_cmdline() { local kernel=$1 - echo "TESTNAME=$testname TIMEOUT=$timeout ACCEL=$accel $RUNTIME_arch_run $kernel -smp $smp $opts" + echo "TESTNAME=$testname TIMEOUT=$timeout DEF_ACCEL=$accel $RUNTIME_arch_run $kernel -smp $smp $opts"
I think this name change can break any $TEST_DIR/run files that aren't updated to check DEF_ACCEL. This patch only updates the runner for s390.
Also, it'd be better to do the code movement (scripts/accel.bash) of this patch separately with a patch that does not have any functional change.
Thanks,
drew