Message ID | 20211116172926.587062-1-mcgrof@kernel.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | blktests: replace module removal with patient module removal | expand |
On 11/16/21 09:29, Luis Chamberlain wrote: > A long time ago, in a galaxy far, far away... > > I ran into some odd scsi_debug false positives with fstests. This > prompted me to look into them given these false positives prevents > me from moving forward with establishing a test baseline with high > number of cycles. That is, this stupid issue was prevening creating > high confidence in testing. > > I reported it [0] and exchanged some ideas with Doug. However, in > the end, despite efforts to help things with scsi_debug there were > still issues lingering which seemed to defy our expectations upstream. > One of the last hanging fruit issues is and always has been that > userspace expectations for proper module removal has been broken, > so in the end I have demonstrated this is a generic issue [1]. > > Long ago a WAIT option for module removal was added... that was then > removed as it was deemed not needed as folks couldn't figure out when > these races happened. The races are actually pretty easy to trigger, it > was just never properly documented. A simpe blkdev_open() will easily > bump a module refcnt, and these days many thing scan do that sort of > thing. > > The proper solution is to implement then a patient module removal > on kmod and patches have been sent for that and those patches are > under review. In the meantime we need a work around to open code a > similar solution for users of old versions of kmod. I sent an open > coded solution for fstests about since August 19th and has been used > there for a few months now. Now that that stuff is merged and tested > in fstests with more exposure, its time to match parity on blktests. > > I've tested blktests with this for things which I can run virtually > for a while now. More wider testig is welcomed. > > [0] https://bugzilla.kernel.org/show_bug.cgi?id=212337 > [1] https://bugzilla.kernel.org/show_bug.cgi?id=214015 > Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> > Thanks for having this work done and explaining the importance of it. Give me couple of days, I'll provide you a feedback after I finish my testing of your patch.
On Thu, Nov 18, 2021 at 01:38:49AM +0000, Chaitanya Kulkarni wrote: > On 11/16/21 09:29, Luis Chamberlain wrote: > > A long time ago, in a galaxy far, far away... > > > > I ran into some odd scsi_debug false positives with fstests. This > > prompted me to look into them given these false positives prevents > > me from moving forward with establishing a test baseline with high > > number of cycles. That is, this stupid issue was prevening creating > > high confidence in testing. > > > > I reported it [0] and exchanged some ideas with Doug. However, in > > the end, despite efforts to help things with scsi_debug there were > > still issues lingering which seemed to defy our expectations upstream. > > One of the last hanging fruit issues is and always has been that > > userspace expectations for proper module removal has been broken, > > so in the end I have demonstrated this is a generic issue [1]. > > > > Long ago a WAIT option for module removal was added... that was then > > removed as it was deemed not needed as folks couldn't figure out when > > these races happened. The races are actually pretty easy to trigger, it > > was just never properly documented. A simpe blkdev_open() will easily > > bump a module refcnt, and these days many thing scan do that sort of > > thing. > > > > The proper solution is to implement then a patient module removal > > on kmod and patches have been sent for that and those patches are > > under review. In the meantime we need a work around to open code a > > similar solution for users of old versions of kmod. I sent an open > > coded solution for fstests about since August 19th and has been used > > there for a few months now. Now that that stuff is merged and tested > > in fstests with more exposure, its time to match parity on blktests. > > > > I've tested blktests with this for things which I can run virtually > > for a while now. More wider testig is welcomed. > > > > [0] https://bugzilla.kernel.org/show_bug.cgi?id=212337 > > [1] https://bugzilla.kernel.org/show_bug.cgi?id=214015 > > Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> > > > > Thanks for having this work done and explaining the importance of it. > > Give me couple of days, I'll provide you a feedback after I finish my > testing of your patch. How did testing go? Luis
On Wed, Dec 08, 2021 at 04:53:04AM -0800, Luis Chamberlain wrote: > On Thu, Nov 18, 2021 at 01:38:49AM +0000, Chaitanya Kulkarni wrote: > > On 11/16/21 09:29, Luis Chamberlain wrote: > > > A long time ago, in a galaxy far, far away... > > > > > > I ran into some odd scsi_debug false positives with fstests. This > > > prompted me to look into them given these false positives prevents > > > me from moving forward with establishing a test baseline with high > > > number of cycles. That is, this stupid issue was prevening creating > > > high confidence in testing. > > > > > > I reported it [0] and exchanged some ideas with Doug. However, in > > > the end, despite efforts to help things with scsi_debug there were > > > still issues lingering which seemed to defy our expectations upstream. > > > One of the last hanging fruit issues is and always has been that > > > userspace expectations for proper module removal has been broken, > > > so in the end I have demonstrated this is a generic issue [1]. > > > > > > Long ago a WAIT option for module removal was added... that was then > > > removed as it was deemed not needed as folks couldn't figure out when > > > these races happened. The races are actually pretty easy to trigger, it > > > was just never properly documented. A simpe blkdev_open() will easily > > > bump a module refcnt, and these days many thing scan do that sort of > > > thing. > > > > > > The proper solution is to implement then a patient module removal > > > on kmod and patches have been sent for that and those patches are > > > under review. In the meantime we need a work around to open code a > > > similar solution for users of old versions of kmod. I sent an open > > > coded solution for fstests about since August 19th and has been used > > > there for a few months now. Now that that stuff is merged and tested > > > in fstests with more exposure, its time to match parity on blktests. > > > > > > I've tested blktests with this for things which I can run virtually > > > for a while now. More wider testig is welcomed. > > > > > > [0] https://bugzilla.kernel.org/show_bug.cgi?id=212337 > > > [1] https://bugzilla.kernel.org/show_bug.cgi?id=214015 > > > Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> > > > > > > > Thanks for having this work done and explaining the importance of it. > > > > Give me couple of days, I'll provide you a feedback after I finish my > > testing of your patch. > > How did testing go? *Friendly poke* Luis
On 11/16/21 9:29 AM, Luis Chamberlain wrote: > A long time ago, in a galaxy far, far away... > > I ran into some odd scsi_debug false positives with fstests. This > prompted me to look into them given these false positives prevents > me from moving forward with establishing a test baseline with high > number of cycles. That is, this stupid issue was prevening creating > high confidence in testing. > > I reported it [0] and exchanged some ideas with Doug. However, in > the end, despite efforts to help things with scsi_debug there were > still issues lingering which seemed to defy our expectations upstream. > One of the last hanging fruit issues is and always has been that > userspace expectations for proper module removal has been broken, > so in the end I have demonstrated this is a generic issue [1]. > > Long ago a WAIT option for module removal was added... that was then > removed as it was deemed not needed as folks couldn't figure out when > these races happened. The races are actually pretty easy to trigger, it > was just never properly documented. A simpe blkdev_open() will easily > bump a module refcnt, and these days many thing scan do that sort of > thing. > > The proper solution is to implement then a patient module removal > on kmod and patches have been sent for that and those patches are > under review. In the meantime we need a work around to open code a > similar solution for users of old versions of kmod. I sent an open > coded solution for fstests about since August 19th and has been used > there for a few months now. Now that that stuff is merged and tested > in fstests with more exposure, its time to match parity on blktests. > > I've tested blktests with this for things which I can run virtually > for a while now. More wider testig is welcomed. > > [0] https://bugzilla.kernel.org/show_bug.cgi?id=212337 > [1] https://bugzilla.kernel.org/show_bug.cgi?id=214015 > Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> > --- This looks good to me, I'd wait Bart (CCd here) to review the srp side. Looks good. Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
On Wed, Jan 26, 2022 at 05:39:13AM +0000, Chaitanya Kulkarni wrote: > On 11/16/21 9:29 AM, Luis Chamberlain wrote: > > A long time ago, in a galaxy far, far away... > > > > I ran into some odd scsi_debug false positives with fstests. This > > prompted me to look into them given these false positives prevents > > me from moving forward with establishing a test baseline with high > > number of cycles. That is, this stupid issue was prevening creating > > high confidence in testing. > > > > I reported it [0] and exchanged some ideas with Doug. However, in > > the end, despite efforts to help things with scsi_debug there were > > still issues lingering which seemed to defy our expectations upstream. > > One of the last hanging fruit issues is and always has been that > > userspace expectations for proper module removal has been broken, > > so in the end I have demonstrated this is a generic issue [1]. > > > > Long ago a WAIT option for module removal was added... that was then > > removed as it was deemed not needed as folks couldn't figure out when > > these races happened. The races are actually pretty easy to trigger, it > > was just never properly documented. A simpe blkdev_open() will easily > > bump a module refcnt, and these days many thing scan do that sort of > > thing. > > > > The proper solution is to implement then a patient module removal > > on kmod and patches have been sent for that and those patches are > > under review. In the meantime we need a work around to open code a > > similar solution for users of old versions of kmod. I sent an open > > coded solution for fstests about since August 19th and has been used > > there for a few months now. Now that that stuff is merged and tested > > in fstests with more exposure, its time to match parity on blktests. > > > > I've tested blktests with this for things which I can run virtually > > for a while now. More wider testig is welcomed. > > > > [0] https://bugzilla.kernel.org/show_bug.cgi?id=212337 > > [1] https://bugzilla.kernel.org/show_bug.cgi?id=214015 > > Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> > > --- > > > This looks good to me, I'd wait Bart (CCd here) to review the > srp side. > > Looks good. > > Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com> Bart, *poke* Luis
On Tue, Nov 16, 2021 at 09:29:26AM -0800, Luis Chamberlain wrote: > diff --git a/common/rc b/common/rc > index 85e6314..2b551bf 100644 > --- a/common/rc > +++ b/common/rc > @@ -321,15 +321,143 @@ _uptime_s() { > awk '{ print int($1) }' /proc/uptime > } > > -# Arguments: module to unload ($1) and retry count ($2). > -unload_module() { > - local i m=$1 rc=${2:-1} > - > - [ ! -e "/sys/module/$m" ] && return 0 > - for ((i=rc;i>0;i--)); do > - modprobe -r "$m" > - [ ! -e "/sys/module/$m" ] && return 0 > - sleep .1 > +# Set MODPROBE_PATIENT_RM_TIMEOUT_SECONDS to "forever" if you want the patient > +# modprobe removal to run forever trying to remove a module. > +MODPROBE_REMOVE_PATIENT="" > +modprobe --help | grep -q -1 "remove-patiently" Could you add the same guard I suggested for fstests for ancient versions of modprobe? e.g.,: modprobe --help >& /dev/null && modprobe --help | grep -q -1 "remove-patiently" Thanks! - Ted
On 2/4/22 18:33, Luis Chamberlain wrote:
> Bart, *poke*
My 2y old son is really good at taking up most of my spare time. Anyway,
I will try to find some time to review this patch.
Bart.
On 11/16/21 09:29, Luis Chamberlain wrote: > for a while now. More wider testig is welcomed. ^^^^^^ testing? > @@ -427,14 +428,8 @@ stop_soft_rdma() { > echo "$i ..." > rdma link del "${i}" || echo "Failed to remove ${i}" > done > - if ! unload_module rdma_rxe 10; then > - echo "Unloading rdma_rxe failed" > - return 1 > - fi > - if ! unload_module siw 10; then > - echo "Unloading siw failed" > - return 1 > - fi > + _patient_rmmod rdma_rxe || return 1 > + _patient_rmmod 1siw || return 1 The above clearly has not been tested. There is an easy to spot typo in the above change. Please test your changes before publishing these. > + if [[ ! -z "$MODPROBE_PATIENT_RM_TIMEOUT_SECONDS" ]]; then Please use -n instead of ! -z. > + if [[ -f /sys/module/$module/refcnt ]]; then Has this patch been analyzed with 'make check'? I expect checkpatch to complain about missing double quotes for the above statement. > +# Patiently tries to wait to remove a module by ensuring first ^^^^^^^^^^^^^^^^^^^^^^^ Tries to wait patiently? > + if [[ ! -z $MODPROBE_REMOVE_PATIENT ]]; then Please use -n instead of ! -z and surround the variable expansion with double quotes. > + # If we have extra time left. Use the time left to now try to > + # persistently remove the module. We do this because although through What is persistent module removal? Is that perhaps removing a module such that it cannot be loaded again? > for m in ib_srpt iscsi_target_mod target_core_pscsi target_core_iblock \ > target_core_file target_core_stgt target_core_user \ > target_core_mod > - do > - unload_module $m 10 || return $? > - done > + _patient_rmmod $m || return $1 Please proofread your changes and test your changes before posting these. I think that both "do" and "done" should have been preserved above. Thanks, Bart.
On Mon, Feb 07, 2022 at 05:10:24PM -0800, Bart Van Assche wrote: > On 11/16/21 09:29, Luis Chamberlain wrote: > > for a while now. More wider testig is welcomed. > ^^^^^^ > testing? > > > @@ -427,14 +428,8 @@ stop_soft_rdma() { > > echo "$i ..." > > rdma link del "${i}" || echo "Failed to remove ${i}" > > done > > - if ! unload_module rdma_rxe 10; then > > - echo "Unloading rdma_rxe failed" > > - return 1 > > - fi > > - if ! unload_module siw 10; then > > - echo "Unloading siw failed" > > - return 1 > > - fi > > + _patient_rmmod rdma_rxe || return 1 > > + _patient_rmmod 1siw || return 1 > > The above clearly has not been tested. There is an easy to spot typo in the > above change. Please test your changes before publishing these. Thanks, fixed. Please I stated: I've tested blktests with this for things which I can run virtually for a while now. More wider testig is welcomed. > > + if [[ ! -z "$MODPROBE_PATIENT_RM_TIMEOUT_SECONDS" ]]; then > > Please use -n instead of ! -z. Sure... > > + if [[ -f /sys/module/$module/refcnt ]]; then > > Has this patch been analyzed with 'make check'? I expect checkpatch to > complain about missing double quotes for the above statement. No, I was not aware this was a thing. I fixed all the stupid complaints however I should note at least tests/srp/rc has existing complaints before my changes. I won't address those. I can't say I agree with the changes to remove checks for $?, but whatever. Consinstancy is better than not having one. Perhaps we should embrace this on fstests as well. Who knows how many crufty bugs lie there. > > +# Patiently tries to wait to remove a module by ensuring first > ^^^^^^^^^^^^^^^^^^^^^^^ > Tries to wait patiently? Sure.. > > + if [[ ! -z $MODPROBE_REMOVE_PATIENT ]]; then > > Please use -n instead of ! -z and surround the variable expansion with > double quotes. Sure.. > > + # If we have extra time left. Use the time left to now try to > > + # persistently remove the module. We do this because although through > > What is persistent module removal? Is that perhaps removing a module such > that it cannot be loaded again? No, I'll clarify to this: # Tries to wait patiently to remove a module by ensuring first # the refcnt is 0 and then trying to remove the module over and over # again within the time allowed > > for m in ib_srpt iscsi_target_mod target_core_pscsi target_core_iblock \ > > target_core_file target_core_stgt target_core_user \ > > target_core_mod > > - do > > - unload_module $m 10 || return $? > > - done > > + _patient_rmmod $m || return $1 > > Please proofread your changes and test your changes before posting these. I > think that both "do" and "done" should have been preserved above. And the return value shoudl be $? not $1. I couldn't test srp so wanted someone who would to help review, but sure, I should have caught this. Thanks for the review, I'll post a v2. Luis
On 2/8/22 13:21, Luis Chamberlain wrote: > On Mon, Feb 07, 2022 at 05:10:24PM -0800, Bart Van Assche wrote: >> On 11/16/21 09:29, Luis Chamberlain wrote: >>> + if [[ -f /sys/module/$module/refcnt ]]; then >> >> Has this patch been analyzed with 'make check'? I expect checkpatch to >> complain about missing double quotes for the above statement. > > No, I was not aware this was a thing. I fixed all the stupid complaints > however I should note at least tests/srp/rc has existing complaints > before my changes. I won't address those. Please consider opening a pull request for the https://github.com/osandov/blktests repository. That will cause the blktests presubmit tests to be run. See also .github/workflows/ci.yml in the blktests source code. > I can't say I agree with the changes to remove checks for $?, but > whatever. Consinstancy is better than not having one. Perhaps we should > embrace this on fstests as well. +1 > No, I'll clarify to this: > > # Tries to wait patiently to remove a module by ensuring first > # the refcnt is 0 and then trying to remove the module over and over > # again within the time allowed Looks better to me :-) >>> for m in ib_srpt iscsi_target_mod target_core_pscsi target_core_iblock \ >>> target_core_file target_core_stgt target_core_user \ >>> target_core_mod >>> - do >>> - unload_module $m 10 || return $? >>> - done >>> + _patient_rmmod $m || return $1 >> >> Please proofread your changes and test your changes before posting these. I >> think that both "do" and "done" should have been preserved above. > > And the return value shoudl be $? not $1. I couldn't test srp so wanted > someone who would to help review, but sure, I should have caught this. Hmm ... the SRP tests can be run inside a virtual machine so it is not clear to me what prevents to run these tests? Thanks, Bart.
On Tue, Feb 08, 2022 at 02:26:38PM -0800, Bart Van Assche wrote: > On 2/8/22 13:21, Luis Chamberlain wrote: > > And the return value shoudl be $? not $1. I couldn't test srp so wanted > > someone who would to help review, but sure, I should have caught this. > > Hmm ... the SRP tests can be run inside a virtual machine so it is not clear > to me what prevents to run these tests? OK I've managed to get time to run SRP tests on a virtual environement but I'm a bit surprised as to the failures found. I need this baseline to ensure the change does not regress anything. I'll report the failures next.. Luis
diff --git a/common/multipath-over-rdma b/common/multipath-over-rdma index cef05ec..c2af372 100644 --- a/common/multipath-over-rdma +++ b/common/multipath-over-rdma @@ -4,6 +4,7 @@ # # Functions and global variables used by both the srp and nvmeof-mp tests. +. common/rc . common/shellcheck . common/null_blk @@ -427,14 +428,8 @@ stop_soft_rdma() { echo "$i ..." rdma link del "${i}" || echo "Failed to remove ${i}" done - if ! unload_module rdma_rxe 10; then - echo "Unloading rdma_rxe failed" - return 1 - fi - if ! unload_module siw 10; then - echo "Unloading siw failed" - return 1 - fi + _patient_rmmod rdma_rxe || return 1 + _patient_rmmod 1siw || return 1 } >>"$FULL" } diff --git a/common/null_blk b/common/null_blk index 6611db0..6c81657 100644 --- a/common/null_blk +++ b/common/null_blk @@ -4,6 +4,8 @@ # # null_blk helper functions. +. common/rc + _have_null_blk() { _have_modules null_blk } @@ -21,9 +23,8 @@ _init_null_blk() { local zoned="" if (( RUN_FOR_ZONED )); then zoned="zoned=1"; fi - if ! modprobe -r null_blk || ! modprobe null_blk "$@" "${zoned}" ; then - return 1 - fi + _patient_rmmod null_blk || return 1 + modprobe null_blk "$@" "${zoned}" || 1 udevadm settle return 0 @@ -46,5 +47,5 @@ _configure_null_blk() { _exit_null_blk() { _remove_null_blk_devices udevadm settle - modprobe -r null_blk + _patient_rmmod null_blk } diff --git a/common/rc b/common/rc index 85e6314..2b551bf 100644 --- a/common/rc +++ b/common/rc @@ -321,15 +321,143 @@ _uptime_s() { awk '{ print int($1) }' /proc/uptime } -# Arguments: module to unload ($1) and retry count ($2). -unload_module() { - local i m=$1 rc=${2:-1} - - [ ! -e "/sys/module/$m" ] && return 0 - for ((i=rc;i>0;i--)); do - modprobe -r "$m" - [ ! -e "/sys/module/$m" ] && return 0 - sleep .1 +# Set MODPROBE_PATIENT_RM_TIMEOUT_SECONDS to "forever" if you want the patient +# modprobe removal to run forever trying to remove a module. +MODPROBE_REMOVE_PATIENT="" +modprobe --help | grep -q -1 "remove-patiently" +if [[ $? -ne 0 ]]; then + if [[ -z "$MODPROBE_PATIENT_RM_TIMEOUT_SECONDS" ]]; then + # We will open code our own implementation of patient module + # remover in blktests. Use a 50 second default. + export MODPROBE_PATIENT_RM_TIMEOUT_SECONDS="50" + fi +else + MODPROBE_RM_PATIENT_TIMEOUT_ARGS="" + if [[ ! -z "$MODPROBE_PATIENT_RM_TIMEOUT_SECONDS" ]]; then + if [[ "$MODPROBE_PATIENT_RM_TIMEOUT_SECONDS" != "forever" ]]; then + MODPROBE_PATIENT_RM_TIMEOUT_MS="$((MODPROBE_PATIENT_RM_TIMEOUT_SECONDS * 1000))" + MODPROBE_RM_PATIENT_TIMEOUT_ARGS="-t $MODPROBE_PATIENT_RM_TIMEOUT_MS" + fi + else + # We export MODPROBE_PATIENT_RM_TIMEOUT_SECONDS here for parity + # with environments without support for modprobe -p, but we + # only really need it exported right now for environments which + # don't have support for modprobe -p to implement our own + # patient module removal support within fstests. + export MODPROBE_PATIENT_RM_TIMEOUT_SECONDS="50" + MODPROBE_PATIENT_RM_TIMEOUT_MS="$((MODPROBE_PATIENT_RM_TIMEOUT_SECONDS * 1000))" + MODPROBE_RM_PATIENT_TIMEOUT_ARGS="-t $MODPROBE_PATIENT_RM_TIMEOUT_MS" + fi + MODPROBE_REMOVE_PATIENT="modprobe -p $MODPROBE_RM_TIMEOUT_ARGS" +fi +export MODPROBE_REMOVE_PATIENT + +# checks the refcount and returns 0 if we can safely remove the module. rmmod +# does this check for us, but we can use this to also iterate checking for this +# refcount before we even try to remove the module. This is useful when using +# debug test modules which take a while to quiesce. +_patient_rmmod_check_refcnt() +{ + local module=$1 + local refcnt=0 + + if [[ -f /sys/module/$module/refcnt ]]; then + refcnt=$(cat /sys/module/$module/refcnt 2>/dev/null) + if [[ $? -ne 0 || $refcnt -eq 0 ]]; then + return 0 + fi + return 1 + fi + return 0 +} + +# Patiently tries to wait to remove a module by ensuring first +# the refcnt is 0 and then trying to persistently remove the module within +# the time allowed. The timeout is configurable per test, just set +# MODPROBE_PATIENT_RM_TIMEOUT_SECONDS prior to including this file. +# If you want this to try forever just set MODPROBE_PATIENT_RM_TIMEOUT_SECONDS +# to the special value of "forever". This applies to both cases where kmod +# supports the patient module remover (modrobe -p) and where it does not. +# +# If your version of kmod supports modprobe -p, we instead use that +# instead. Otherwise we have to implement a patient module remover +# ourselves. +_patient_rmmod() +{ + local module=$1 + local max_tries_max=$MODPROBE_PATIENT_RM_TIMEOUT_SECONDS + local max_tries=0 + local mod_ret=0 + local refcnt_is_zero=0 + + if [[ ! -z $MODPROBE_REMOVE_PATIENT ]]; then + $MODPROBE_REMOVE_PATIENT $module + mod_ret=$? + if [[ $mod_ret -ne 0 ]]; then + echo "kmod patient module removal for $module timed out waiting for refcnt to become 0 using timeout of $max_tries_max returned $mod_ret" + fi + return $mod_ret + fi + + max_tries=$max_tries_max + + # We must use a string check as otherwise if max_tries is set to + # "forever" and we don't use a string check we can end up skipping + # entering this loop. + while [[ "$max_tries" != "0" ]]; do + _patient_rmmod_check_refcnt $module + if [[ $? -eq 0 ]]; then + refcnt_is_zero=1 + break + fi + sleep 1 + if [[ "$max_tries" == "forever" ]]; then + continue + fi + let max_tries=$max_tries-1 done - return 1 + + if [[ $refcnt_is_zero -ne 1 ]]; then + echo "custom patient module removal for $module timed out waiting for refcnt to become 0 using timeout of $max_tries_max" + return -1 + fi + + # If we ran out of time but our refcnt check confirms we had + # a refcnt of 0, just try to remove the module once. + if [[ "$max_tries" == "0" ]]; then + modprobe -r $module + return $? + fi + + # If we have extra time left. Use the time left to now try to + # persistently remove the module. We do this because although through + # the above we found refcnt to be 0, removal can still fail since + # userspace can always race to bump the refcnt. An example is any + # blkdev_open() calls against a block device. These issues have been + # tracked and documented in the following bug reports, which justifies + # our need to do this in userspace: + # https://bugzilla.kernel.org/show_bug.cgi?id=212337 + # https://bugzilla.kernel.org/show_bug.cgi?id=214015 + while [[ $max_tries != 0 ]]; do + if [[ -d /sys/module/$module ]]; then + modprobe -r $module 2> /dev/null + mod_ret=$? + if [[ $mod_ret == 0 ]]; then + break; + fi + sleep 1 + if [[ "$max_tries" == "forever" ]]; then + continue + fi + let max_tries=$max_tries-1 + else + break + fi + done + + if [[ $mod_ret -ne 0 ]]; then + echo "custom patient module removal for $module timed out trying to remove $module using timeout of $max_tries_max last try returned $mod_ret" + fi + + return $mod_ret } diff --git a/common/scsi_debug b/common/scsi_debug index b48cdc9..9987286 100644 --- a/common/scsi_debug +++ b/common/scsi_debug @@ -4,14 +4,15 @@ # # scsi_debug helper functions. +. common/rc + _have_scsi_debug() { _have_modules scsi_debug } _init_scsi_debug() { - if ! modprobe -r scsi_debug || ! modprobe scsi_debug "$@"; then - return 1 - fi + _patient_rmmod scsi_debug || return 1 + modprobe scsi_debug "$@" || return 1 udevadm settle @@ -51,5 +52,5 @@ _exit_scsi_debug() { unset SCSI_DEBUG_TARGETS unset SCSI_DEBUG_DEVICES udevadm settle - modprobe -r scsi_debug + _patient_rmmod scsi_debug } diff --git a/tests/nvme/rc b/tests/nvme/rc index 1c27cde..2fb3c0d 100644 --- a/tests/nvme/rc +++ b/tests/nvme/rc @@ -125,11 +125,11 @@ _cleanup_nvmet() { shopt -u nullglob trap SIGINT - modprobe -r nvme-"${nvme_trtype}" 2>/dev/null + _patient_rmmod nvme-"${nvme_trtype}" 2>/dev/null if [[ "${nvme_trtype}" != "loop" ]]; then - modprobe -r nvmet-"${nvme_trtype}" 2>/dev/null + _patient_rmmod nvmet-"${nvme_trtype}" 2>/dev/null fi - modprobe -r nvmet 2>/dev/null + _patient_rmmod nvmet 2>/dev/null if [[ "${nvme_trtype}" == "rdma" ]]; then stop_soft_rdma fi diff --git a/tests/nvmeof-mp/rc b/tests/nvmeof-mp/rc index dcb2e3c..54d8c95 100755 --- a/tests/nvmeof-mp/rc +++ b/tests/nvmeof-mp/rc @@ -164,12 +164,13 @@ start_nvme_client() { } stop_nvme_client() { - unload_module nvme-rdma || return $? - unload_module nvme-fabrics || return $? + _patient_rmmod nvme-rdma || return 1 + _patient_rmmod nvme-fabrics || return 1 # Ignore nvme and nvme-core unload errors - this test may be run on a # system equipped with one or more NVMe SSDs. - unload_module nvme >&/dev/null - unload_module nvme-core >&/dev/null + export MODPROBE_PATIENT_RM_TIMEOUT_SECONDS=5 + _patient_rmmod nvme + _patient_rmmod nvme-core return 0 } @@ -269,9 +270,9 @@ stop_nvme_target() { rmdir "$d" done ) - unload_module nvmet_rdma && - unload_module nvmet && - _exit_null_blk + _patient_rmmod nvmet_rdma || return 1 + _patient_rmmod nvmet || return 1 + _exit_null_blk } start_target() { diff --git a/tests/srp/rc b/tests/srp/rc index 586f007..c3bba24 100755 --- a/tests/srp/rc +++ b/tests/srp/rc @@ -326,16 +326,9 @@ stop_srp_ini() { local i log_out - for ((i=40;i>=0;i--)); do - remove_mpath_devs || return $? - unload_module ib_srp >/dev/null 2>&1 && break - sleep 1 - done - if [ -e /sys/module/ib_srp ]; then - echo "Error: unloading kernel module ib_srp failed" - return 1 - fi - unload_module scsi_transport_srp || return $? + remove_mpath_devs || return $? + _patient_rmmod ib_srp || return 1 + _patient_rmmod scsi_transport_srp || return $1 } # Associate the LIO device with name $1/$2 with file $3 and SCSI serial $4. @@ -554,9 +547,7 @@ stop_lio_srpt() { for m in ib_srpt iscsi_target_mod target_core_pscsi target_core_iblock \ target_core_file target_core_stgt target_core_user \ target_core_mod - do - unload_module $m 10 || return $? - done + _patient_rmmod $m || return $1 } # Load and configure the SRP target driver
A long time ago, in a galaxy far, far away... I ran into some odd scsi_debug false positives with fstests. This prompted me to look into them given these false positives prevents me from moving forward with establishing a test baseline with high number of cycles. That is, this stupid issue was prevening creating high confidence in testing. I reported it [0] and exchanged some ideas with Doug. However, in the end, despite efforts to help things with scsi_debug there were still issues lingering which seemed to defy our expectations upstream. One of the last hanging fruit issues is and always has been that userspace expectations for proper module removal has been broken, so in the end I have demonstrated this is a generic issue [1]. Long ago a WAIT option for module removal was added... that was then removed as it was deemed not needed as folks couldn't figure out when these races happened. The races are actually pretty easy to trigger, it was just never properly documented. A simpe blkdev_open() will easily bump a module refcnt, and these days many thing scan do that sort of thing. The proper solution is to implement then a patient module removal on kmod and patches have been sent for that and those patches are under review. In the meantime we need a work around to open code a similar solution for users of old versions of kmod. I sent an open coded solution for fstests about since August 19th and has been used there for a few months now. Now that that stuff is merged and tested in fstests with more exposure, its time to match parity on blktests. I've tested blktests with this for things which I can run virtually for a while now. More wider testig is welcomed. [0] https://bugzilla.kernel.org/show_bug.cgi?id=212337 [1] https://bugzilla.kernel.org/show_bug.cgi?id=214015 Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> --- common/multipath-over-rdma | 11 +-- common/null_blk | 9 ++- common/rc | 148 ++++++++++++++++++++++++++++++++++--- common/scsi_debug | 9 ++- tests/nvme/rc | 6 +- tests/nvmeof-mp/rc | 15 ++-- tests/srp/rc | 17 +---- 7 files changed, 166 insertions(+), 49 deletions(-)