nvme: Refactor and expose per-controller timeout configuration
Subsystems (3)
v4 cover nvme: Refactor and expose per-controller timeout configuration
This patchset tries to address some limitations in how the NVMe driver handles command timeouts.
Currently, the driver relies heavily on global module parameters (NVME_IO_TIMEOUT and NVME_ADMIN_TIMEOUT), making it difficult for users to tune timeouts for specific controllers that may have very different characteristics. Also, in some cases, manual changes to sysfs timeout values are ignored by the driver logic.
For example this patchset removes the unconditional timeout assignment in nvme_init_request. This allows the block layer to correctly apply the request queue's timeout settings, ensuring that user-initiated changes via sysfs are actually respected for all requests.
It introduces new sysfs attributes (admin_timeout and io_timeout) to the NVMe controller. This allows users to configure distinct timeout requirements for different controllers rather than relying on global module parameters.
Some examples:
Changes to the controller's io_timeout gets propagated to all the associated namespaces' queues:
# find /sys -name 'io_timeout' /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n1/queue/io_timeout /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n2/queue/io_timeout /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n3/queue/io_timeout /sys/devices/virtual/nvme-fabrics/ctl/nvme0/io_timeout
# echo 27000 > /sys/devices/virtual/nvme-fabrics/ctl/nvme0/io_timeout # cat /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n1/queue/io_timeout 27000 # cat /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n2/queue/io_timeout 27000 # cat /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n3/queue/io_timeout 27000
When adding a namespace target-side, the io_timeout is inherited from the controller's preferred timeout:
* target side *
# nvmetcli /> cd subsystems/test-nqn/namespaces/4 /subsystems/t.../namespaces/4> enable The Namespace has been enabled.
************
* Host-side *
nvme nvme0: rescanning namespaces.
# find /sys -name 'io_timeout' /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n1/queue/io_timeout /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n2/queue/io_timeout /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n3/queue/io_timeout /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n4/queue/io_timeout <-- new namespace /sys/devices/virtual/nvme-fabrics/ctl/nvme0/io_timeout
# cat /sys/devices/virtual/nvme-fabrics/ctl/nvme0/nvme0c0n4/queue/io_timeout 27000 ***********
io_timeout and admin_timeout module parameters are used as default values for new controllers:
# nvme connect -t tcp -a 10.37.153.138 -s 8000 -n test-nqn2 connecting to device: nvme1
# cat /sys/devices/virtual/nvme-fabrics/ctl/nvme1/nvme1c1n1/queue/io_timeout 30000 # cat /sys/devices/virtual/nvme-fabrics/ctl/nvme1/admin_timeout 60000
V4: - Refactor the nvmet-loop patch as suggested by Daniel Wagner
- Add a patch to fix a potential race condition between connected uevent and STARTED_ONCE controller's flag.
V3: - Rebase on top of nvme 7.1 branch
- add an admin_timeout variable to nvme_ctrl structure
- Wait until the controller has reached the LIVE state
for the first time before allowing the user to modify
the timeouts, this prevents the dereferencing of the admin_q, fabrics_q and connect_q queues before their initialization.
- move blk_put_queue(fabrics_q) to nvme_free_ctrl() to align it to admin_q teardown
- modify nvmet-loop to avoid deleting and re-creating the admin_q queue when the controller enters the resetting state.
- add a warning if nvme_alloc_admin_tag_set() is called twice against the same controller
V2: - Drop the RFC tag
- apply the timeout settings to fabrics_q and connect_q too
- Code style fixes
- remove unnecessary check for null admin_q in __nvme_delete_io_queues()
- Use DEVICE_ATTR() macro
Heyne, Maximilian (1):
nvme: Let the blocklayer set timeouts for requests
Maurizio Lombardi (8):
nvme: add sysfs attribute to change admin timeout per nvme controller
nvme: fix race condition between connected uevent and STARTED_ONCE
flag
nvme: pci: use admin queue timeout over NVME_ADMIN_TIMEOUT
nvme: add sysfs attribute to change IO timeout per controller
nvme: use per controller timeout waits over depending on global
default
nvme-core: align fabrics_q teardown with admin_q in nvme_free_ctrl
nvmet-loop: do not alloc admin tag set during reset
nvme-core: warn on allocating admin tag set with existing queue
drivers/nvme/host/apple.c | 2 +- drivers/nvme/host/core.c | 23 +++++----- drivers/nvme/host/nvme.h | 4 +- drivers/nvme/host/pci.c | 4 +- drivers/nvme/host/rdma.c | 2 +- drivers/nvme/host/sysfs.c | 89 ++++++++++++++++++++++++++++++++++++++ drivers/nvme/host/tcp.c | 2 +- drivers/nvme/target/loop.c | 31 ++++++------- 8 files changed, 123 insertions(+), 34 deletions(-)
Signature · 1 line
2.54.0
PATCH v4 1/9 nvme: Let the blocklayer set timeouts for requests
From: "Heyne, Maximilian" <[email protected]>
When initializing an nvme request which is about to be send to the block layer, we do not need to initialize its timeout. If it's left uninitialized at 0 the block layer will use the request queue's timeout in blk_add_timer (via nvme_start_request which is called from nvme_*_queue_rq). These timeouts are setup to either NVME_IO_TIMEOUT or NVME_ADMIN_TIMEOUT when the request queues were created.
Because the io_timeout of the IO queues can be modified via sysfs, the following situation can occur:
1) NVME_IO_TIMEOUT = 30 (default module parameter)
2) nvme1n1 is probed. IO queues default timeout is 30 s
3) manually change the IO timeout to 90 s
echo 90000 > /sys/class/nvme/nvme1/nvme1n1/queue/io_timeout
4) Any call of __submit_sync_cmd on nvme1n1 to an IO queue will issue commands with the 30 s timeout instead of the wanted 90 s which might be more suitable for this device.
Commit 470e900c8036 ("nvme: refactor nvme_alloc_request") silently changed the behavior for ioctl's already because it unconditionally overrides the request's timeout that was set in nvme_init_request. If it was unset by the user of the ioctl if will be overridden with 0 meaning the block layer will pick the request queue's IO timeout.
Following up on that, this patch further improves the consistency of IO timeout usage. However, there are still uses of NVME_IO_TIMEOUT which could be inconsistent with what is set in the device's request_queue by the user.
--- drivers/nvme/host/core.c | 2 -- 1 file changed, 2 deletions(-)
drivers/nvme/host/core.c +0 -2
@@ -729,10 +729,8 @@ void nvme_init_request(struct request *req, struct nvme_command *cmd)struct nvme_ns *ns = req->q->disk->private_data; logging_enabled = ns->head->passthru_err_log_enabled; - req->timeout =NVME_IO_TIMEOUT ; }else {/* no queuedata implies admin queue */ logging_enabled = nr->ctrl->passthru_err_log_enabled; - req->timeout =NVME_ADMIN_TIMEOUT ; }if (!logging_enabled)
Signature · 1 line
2.54.0
Re: [PATCH V4 1/9] nvme: Let the blocklayer set timeouts for requests
Quoted 31 lines · Maurizio Lombardi
From: "Heyne, Maximilian" <[email protected]>
When initializing an nvme request which is about to be send to the block layer, we do not need to initialize its timeout. If it's left uninitialized at 0 the block layer will use the request queue's timeout in
blk_add_timer(vianvme_start_requestwhich is called from nvme_*_queue_rq). These timeouts are setup to eitherNVME_IO_TIMEOUTorNVME_ADMIN_TIMEOUTwhen the request queues were created.Because the
io_timeoutof the IO queues can be modified via sysfs, the following situation can occur:1)
NVME_IO_TIMEOUT= 30 (default module parameter)
2) nvme1n1 is probed. IO queues default timeout is 30 s
3) manually change the IO timeout to 90 secho 90000 > /sys/class/nvme/nvme1/nvme1n1/queue/io_timeout4) Any call of
__submit_sync_cmdon nvme1n1 to an IO queue will issue commands with the 30 s timeout instead of the wanted 90 s which might be more suitable for this device.Commit
470e900c8036("nvme: refactornvme_alloc_request") silently changed the behavior for ioctl's already because it unconditionally overrides the request's timeout that was set innvme_init_request. If it was unset by the user of the ioctl if will be overridden with 0 meaning the block layer will pick the request queue's IO timeout.Following up on that, this patch further improves the consistency of IO timeout usage. However, there are still uses of
NVME_IO_TIMEOUTwhich could be inconsistent with what is set in the device'srequest_queueby the user.
Quoted 9 lines
Reviewed-by Mohamed Khalfella <[email protected]>Reviewed-by Christoph Hellwig <[email protected]>Reviewed-by Daniel Wagner <[email protected]>Signed-off-by Maximilian Heyne <[email protected]>Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/core.c | 2 -- 1 file changed, 2 deletions(-)
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH v4 2/9 nvme: add sysfs attribute to change admin timeout per nvme controller
In reply to Maurizio LombardiCurrently, there is no method to adjust the timeout values on a per controller basis with nvme admin queues. Add an admin_timeout attribute to nvme so that different nvme controllers which may have different timeout requirements can have custom admin timeouts set.
--- drivers/nvme/host/core.c | 1 + drivers/nvme/host/nvme.h | 1 + drivers/nvme/host/sysfs.c | 42 +++++++++++++++++++++++++++++++++++++++ 3 files changed, 44 insertions(+)
drivers/nvme/host/core.c +1 -0
@@ -5140,6 +5140,7 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,memset (&ctrl->ka_cmd,0 ,sizeof (ctrl->ka_cmd)); ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive; ctrl->ka_last_check_time = jiffies; + ctrl->admin_timeout =NVME_ADMIN_TIMEOUT ;BUILD_BUG_ON (NVME_DSM_MAX_RANGES *sizeof (struct nvme_dsm_range ) >PAGE_SIZE );
drivers/nvme/host/nvme.h +1 -0
@@ -370,6 +370,7 @@ struct nvme_ctrl {u16 mtfa;u32 ctrl_config;u32 queue_count; +u32 admin_timeout;u64 cap;u32 max_hw_sectors;
drivers/nvme/host/sysfs.c +42 -0
54 lines not on this page · raw message
Signature · 1 line
2.54.0
Re: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
Quoted 7 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme admin queues. Add an
admin_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom admin timeouts set.Signed-off-by Maurizio Lombardi <[email protected]>
Re: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
In reply to Maurizio LombardiChristoph Hellwig
maintainerRe: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
In reply to Maurizio LombardiQuoted 5 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme admin queues. Add an
admin_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom admin timeouts set.
Please use up all 73 characters for commit messages.
drivers/nvme/host/sysfs.c static ssize_t quirks_show(struct device *dev, struct device_attribute *attr, · line 627
Quoted 3 lines
drivers/nvme/host/sysfs.c +2
-0+static ssize_t nvme_admin_timeout_show (struct device *dev, +struct device_attribute *attr,char *buf)
The indent for prototype continuations should use two, not three tabs.
drivers/nvme/host/sysfs.c static ssize_t quirks_show(struct device *dev, struct device_attribute *attr, · line 646
Quoted 3 lines
drivers/nvme/host/sysfs.c +3
-0+* Wait until the controller reaches the LIVE state +* to be sure that admin_q and fabrics_q are +* properly initialized.
Please use up all 80 characters for comments.
drivers/nvme/host/sysfs.c static ssize_t quirks_show(struct device *dev, struct device_attribute *attr, · line 659
Quoted 13 lines
drivers/nvme/host/sysfs.c +13
-0+*/ +if (!test_bit (NVME_CTRL_STARTED_ONCE , &ctrl->flags)) +return -EBUSY ; + + err =kstrtou32 (buf,10 , &timeout); +if (err || !timeout) +return -EINVAL ; + + ctrl->admin_timeout =msecs_to_jiffies (timeout); + +blk_queue_rq_timeout (ctrl->admin_q, ctrl->admin_timeout); +if (ctrl->fabrics_q) +blk_queue_rq_timeout (ctrl->fabrics_q, ctrl->admin_timeout);
Do we really want to apply the the admin timeout for the fabrics queue? If so can you document here why we do that?
Re: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
Quoted 15 lines · Christoph Hellwig
Quoted 5 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme admin queues. Add an
admin_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom admin timeouts set.Please use up all 73 characters for commit messages.
drivers/nvme/host/sysfs.c static ssize_t quirks_show(struct device *dev, struct device_attribute *attr, · line 627
Quoted 3 lines
drivers/nvme/host/sysfs.c +2
-0+static ssize_t nvme_admin_timeout_show (struct device *dev, +struct device_attribute *attr,char *buf)The indent for prototype continuations should use two, not three tabs.
Quoted 5 lines
drivers/nvme/host/sysfs.c static ssize_t quirks_show(struct device *dev, struct device_attribute *attr, · line 646
Quoted 3 lines
drivers/nvme/host/sysfs.c +3
-0+* Wait until the controller reaches the LIVE state +* to be sure that admin_q and fabrics_q are +* properly initialized. Please use up all 80 characters for comments.
Ok.
Quoted 17 lines
Quoted 4 lines
drivers/nvme/host/sysfs.c +3
-0+*/ +if (!test_bit (NVME_CTRL_STARTED_ONCE , &ctrl->flags)) +return -EBUSY ;drivers/nvme/host/sysfs.c static ssize_t quirks_show(struct device *dev, struct device_attribute *attr, · line 659
Quoted 10 lines
drivers/nvme/host/sysfs.c +10
-0+ + err =kstrtou32 (buf,10 , &timeout); +if (err || !timeout) +return -EINVAL ; + + ctrl->admin_timeout =msecs_to_jiffies (timeout); + +blk_queue_rq_timeout (ctrl->admin_q, ctrl->admin_timeout); +if (ctrl->fabrics_q) +blk_queue_rq_timeout (ctrl->fabrics_q, ctrl->admin_timeout);Do we really want to apply the the admin timeout for the fabrics queue? If so can you document here why we do that?
Both admin_q and fabrics_q are initialized to share the same NVME_ADMIN_TIMEOUT value, therefore keeping them in sync maintains the consistency.
If we didn't apply the timeout to fabrics_q, it would end up operating under a different timeout than the standard admin_q, is there any reason why we would want that?
Maurizio
Christoph Hellwig
maintainerRe: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
Quoted 7 lines · Maurizio Lombardi
Both
admin_qandfabrics_qare initialized to share the sameNVME_ADMIN_TIMEOUTvalue, therefore keeping them in sync maintains the consistency.If we didn't apply the timeout to
fabrics_q, it would end up operating under a different timeout than the standardadmin_q, is there any reason why we would want that?
Not sure. I just want to make sure everything is properly documented.
Re: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
In reply to Maurizio LombardiQuoted 18 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme admin queues. Add an
admin_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom admin timeouts set.Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/core.c | 1 + drivers/nvme/host/nvme.h | 1 + drivers/nvme/host/sysfs.c | 42 +++++++++++++++++++++++++++++++++++++++ 3 files changed, 44 insertions(+)drivers/nvme/host/core.c +0
-0
drivers/nvme/host/core.c int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, · line 5143
Quoted 4 lines
drivers/nvme/host/core.c +1
-0@@ -5140,6 +5140,7 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,memset (&ctrl->ka_cmd,0 ,sizeof (ctrl->ka_cmd)); ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive; ctrl->ka_last_check_time = jiffies; + ctrl->admin_timeout =NVME_ADMIN_TIMEOUT ;
Why do you remove the default timeout?
Shouldn't admin requests run with the admin timeout per default?
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
Re: [PATCH V4 2/9] nvme: add sysfs attribute to change admin timeout per nvme controller
Quoted 14 lines · Hannes Reinecke
Quoted 13 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme admin queues. Add an
admin_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom admin timeouts set.Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/core.c | 1 + drivers/nvme/host/nvme.h | 1 + drivers/nvme/host/sysfs.c | 42 +++++++++++++++++++++++++++++++++++++++ 3 files changed, 44 insertions(+)
Quoted 11 lines
drivers/nvme/host/core.c int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, · line 5143
Quoted 9 lines
drivers/nvme/host/core.c +1
-0@@ -5140,6 +5140,7 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,memset (&ctrl->ka_cmd,0 ,sizeof (ctrl->ka_cmd)); ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive; ctrl->ka_last_check_time = jiffies; + ctrl->admin_timeout =NVME_ADMIN_TIMEOUT ;Why do you remove the default timeout?
nvme_init_ctrl() initializes a new controller, so it must inherit the default global NVME_ADMIN_TIMEOUT settings
Quoted 1 line
Shouldn't admin requests run with the admin timeout per default?
They do, the default admin timeout for new controllers remains NVME_ADMIN_TIMEOUT and it's applied to both admin_q and fabrics_q queues when they are initialized.
Maurizio
PATCH v4 3/9 nvme: fix race condition between connected uevent and STARTED_ONCE flag
In reply to Maurizio LombardiWhen a controller connects, nvme_start_ctrl() emits the "NVME_EVENT=connected" uevent and sets the NVME_CTRL_STARTED_ONCE flag. Currently, the uevent is emitted before the flag is set.
This creates a race condition for userspace tools (like udev rules) that might rely on the "connected" event to configure sysfs attributes. Specifically, if a udev rule attempts to set the newly introduced admin_timeout attribute immediately after receiving the uevent, the sysfs store function might evaluate the NVME_CTRL_STARTED_ONCE bit before it is actually set, resulting in spurious -EBUSY error.
Swap the order of operations in nvme_start_ctrl() so that the NVME_CTRL_STARTED_ONCE flag is set before the uevent is sent. This guarantees that the admin_timeout can already be changed when userspace is notified.
--- drivers/nvme/host/core.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/nvme/host/core.c +1 -1
@@ -5039,8 +5039,8 @@ void nvme_start_ctrl(struct nvme_ctrl *ctrl)nvme_mpath_update (ctrl); } -nvme_change_uevent (ctrl,"NVME_EVENT=connected" );set_bit (NVME_CTRL_STARTED_ONCE , &ctrl->flags); +nvme_change_uevent (ctrl,"NVME_EVENT=connected" ); }EXPORT_SYMBOL_GPL (nvme_start_ctrl);
Signature · 1 line
2.54.0
Re: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
Quoted 11 lines · Maurizio Lombardi
When a controller connects,
nvme_start_ctrl()emits the "NVME_EVENT=connected" uevent and sets theNVME_CTRL_STARTED_ONCEflag. Currently, the uevent is emitted before the flag is set.This creates a race condition for userspace tools (like udev rules) that might rely on the "connected" event to configure sysfs attributes. Specifically, if a udev rule attempts to set the newly introduced
admin_timeoutattribute immediately after receiving the uevent, the sysfs store function might evaluate theNVME_CTRL_STARTED_ONCEbit before it is actually set, resulting in spurious-EBUSYerror.
Quoted 6 lines
Swap the order of operations in
nvme_start_ctrl()so that theNVME_CTRL_STARTED_ONCEflag is set before the uevent is sent. This guarantees that theadmin_timeoutcan already be changed when userspace is notified.Signed-off-by Maurizio Lombardi <[email protected]>
Re: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
In reply to Maurizio LombardiThis should even go as a separate patch:
Christoph Hellwig
maintainerRe: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
Quoted 1 line · Sagi Grimberg
This should even go as a separate patch:
You mean a stadalone fix? Because it already is a separate patch.
Sagi Grimberg
maintainerRe: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
Quoted 4 lines · Sagi Grimberg
Quoted 1 line
This should even go as a separate patch:
You mean a stadalone fix? Because it already is a separate patch.
Yes, this is a race condition fix that can go separately.
Keith Busch
maintainerRe: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
Quoted 6 lines · Sagi Grimberg
Quoted 4 lines · Sagi Grimberg
Quoted 1 line
This should even go as a separate patch:
You mean a stadalone fix? Because it already is a separate patch.
Yes, this is a race condition fix that can go separately.
Sure thing, I've applied patch 3/9 to nvme-7.1 as a separate bug fix.
Re: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
Quoted 9 lines · Keith Busch
Quoted 6 lines · Sagi Grimberg
Quoted 4 lines · Sagi Grimberg
Quoted 1 line
This should even go as a separate patch:
You mean a stadalone fix? Because it already is a separate patch.
Yes, this is a race condition fix that can go separately.
Sure thing, I've applied patch 3/9 to nvme-7.1 as a separate bug fix.
Oh, that was way faster than I expected.
The issue is that, in explaining the rationale behind PATCH 3, I mentioned in its commit message the admin_timeout sysfs attribute introduced by this patchset.
You may want to remove that reference if you plan to merge patch 3 indipendently of the rest of this patchset.
In the vanilla kernel, the only sysfs attribute that could be affected by this race condition is nvme_sysfs_delete().
Maurizio
Keith Busch
maintainerRe: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
Quoted 17 lines · Maurizio Lombardi
Quoted 9 lines · Keith Busch
Quoted 6 lines · Sagi Grimberg
Quoted 4 lines · Sagi Grimberg
Quoted 1 line
This should even go as a separate patch:
You mean a stadalone fix? Because it already is a separate patch.
Yes, this is a race condition fix that can go separately.
Sure thing, I've applied patch 3/9 to nvme-7.1 as a separate bug fix.
Oh, that was way faster than I expected.
The issue is that, in explaining the rationale behind PATCH 3, I mentioned in its commit message the
admin_timeoutsysfs attribute introduced by this patchset.
Quoted 5 lines
You may want to remove that reference if you plan to merge patch 3 indipendently of the rest of this patchset.
In the vanilla kernel, the only sysfs attribute that could be affected by this race condition is
nvme_sysfs_delete().
Ah, thanks. I got to this one quicker because I had some free time for patch maintenance. I'll fix up the message.
Re: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
In reply to Maurizio LombardiLooks good:
Re: [PATCH V4 3/9] nvme: fix race condition between connected uevent and STARTED_ONCE flag
In reply to Maurizio LombardiQuoted 26 lines · Maurizio Lombardi
When a controller connects,
nvme_start_ctrl()emits the "NVME_EVENT=connected" uevent and sets theNVME_CTRL_STARTED_ONCEflag. Currently, the uevent is emitted before the flag is set.This creates a race condition for userspace tools (like udev rules) that might rely on the "connected" event to configure sysfs attributes. Specifically, if a udev rule attempts to set the newly introduced
admin_timeoutattribute immediately after receiving the uevent, the sysfs store function might evaluate theNVME_CTRL_STARTED_ONCEbit before it is actually set, resulting in spurious-EBUSYerror.Swap the order of operations in
nvme_start_ctrl()so that theNVME_CTRL_STARTED_ONCEflag is set before the uevent is sent. This guarantees that theadmin_timeoutcan already be changed when userspace is notified.Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/core.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/nvme/host/core.c +0
-0
drivers/nvme/host/core.c void nvme_start_ctrl(struct nvme_ctrl *ctrl) · line 5046
Quoted 9 lines
drivers/nvme/host/core.c +1
-1@@ -5039,8 +5039,8 @@ void nvme_start_ctrl(struct nvme_ctrl *ctrl)nvme_mpath_update (ctrl); } -nvme_change_uevent (ctrl,"NVME_EVENT=connected" );set_bit (NVME_CTRL_STARTED_ONCE , &ctrl->flags); +nvme_change_uevent (ctrl,"NVME_EVENT=connected" ); }EXPORT_SYMBOL_GPL (nvme_start_ctrl);
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH v4 4/9 nvme: pci: use admin queue timeout over NVME_ADMIN_TIMEOUT
In reply to Maurizio LombardiWhile tearing down its queues, nvme-pci uses NVME_ADMIN_TIMEOUT as its timeout target. Instead, use the configured admin queue's timeout value to match the device's existing timeout setting.
--- drivers/nvme/host/pci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/nvme/host/pci.c +1 -1
@@ -3094,7 +3094,7 @@ static bool __nvme_delete_io_queues(struct nvme_dev *dev, u8 opcode)unsigned long timeout; retry: - timeout =NVME_ADMIN_TIMEOUT ; + timeout = dev->ctrl.admin_timeout;while (nr_queues >0 ) {if (nvme_delete_queue (&dev->queues[nr_queues], opcode))break ;
Signature · 1 line
2.54.0
Re: [PATCH V4 4/9] nvme: pci: use admin queue timeout over NVME_ADMIN_TIMEOUT
Christoph Hellwig
maintainerRe: [PATCH V4 4/9] nvme: pci: use admin queue timeout over NVME_ADMIN_TIMEOUT
In reply to Maurizio LombardiQuoted 3 lines · Maurizio Lombardi
While tearing down its queues, nvme-pci uses
NVME_ADMIN_TIMEOUTas its timeout target. Instead, use the configured admin queue's timeout value to match the device's existing timeout setting.
Shouldn't this go into the patch introducing the configurable timeout?
Otherwise looks good.
Re: [PATCH V4 4/9] nvme: pci: use admin queue timeout over NVME_ADMIN_TIMEOUT
In reply to Maurizio LombardiQuoted 15 lines · Maurizio Lombardi
While tearing down its queues, nvme-pci uses
NVME_ADMIN_TIMEOUTas its timeout target. Instead, use the configured admin queue's timeout value to match the device's existing timeout setting.Reviewed-by Daniel Wagner <[email protected]>Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/pci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/nvme/host/pci.c +0
-0
drivers/nvme/host/pci.c static bool __nvme_delete_io_queues(struct nvme_dev *dev, u8 opcode) · line 3100
Quoted 8 lines
drivers/nvme/host/pci.c +1
-1@@ -3094,7 +3094,7 @@ static bool __nvme_delete_io_queues(struct nvme_dev *dev, u8 opcode)unsigned long timeout; retry: - timeout =NVME_ADMIN_TIMEOUT ; + timeout = dev->ctrl.admin_timeout;while (nr_queues >0 ) {if (nvme_delete_queue (&dev->queues[nr_queues], opcode))break ;
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH v4 5/9 nvme: add sysfs attribute to change IO timeout per controller
In reply to Maurizio LombardiCurrently, there is no method to adjust the timeout values on a per controller basis with nvme I/O queues.
Add an io_timeout attribute to nvme so that different nvme controllers which may have different timeout requirements can have custom I/O timeouts set.
--- drivers/nvme/host/core.c | 2 ++ drivers/nvme/host/nvme.h | 1 + drivers/nvme/host/sysfs.c | 47 +++++++++++++++++++++++++++++++++++++++ 3 files changed, 50 insertions(+)
drivers/nvme/host/core.c +2 -0
@@ -4203,6 +4203,7 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, struct nvme_ns_info *info)mutex_unlock (&ctrl->namespaces_lock); goto out_unlink_ns; } +blk_queue_rq_timeout (ns->queue, ctrl->io_timeout);nvme_ns_add_to_ctrl_list (ns);mutex_unlock (&ctrl->namespaces_lock);synchronize_srcu (&ctrl->srcu);
@@ -5141,6 +5142,7 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive; ctrl->ka_last_check_time = jiffies; ctrl->admin_timeout =NVME_ADMIN_TIMEOUT ; + ctrl->io_timeout =NVME_IO_TIMEOUT ;BUILD_BUG_ON (NVME_DSM_MAX_RANGES *sizeof (struct nvme_dsm_range ) >PAGE_SIZE );
drivers/nvme/host/nvme.h +1 -0
@@ -371,6 +371,7 @@ struct nvme_ctrl {u32 ctrl_config;u32 queue_count;u32 admin_timeout; +u32 io_timeout;u64 cap;u32 max_hw_sectors;
drivers/nvme/host/sysfs.c +47 -0
59 lines not on this page · raw message
Signature · 1 line
2.54.0
Re: [PATCH V4 5/9] nvme: add sysfs attribute to change IO timeout per controller
Quoted 7 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme I/O queues.
Add anio_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom I/O timeouts set.Signed-off-by Maurizio Lombardi <[email protected]>
I was wondering if namespaces would ever want different timeouts but given that at this point it was always the default one and this change scopes to controllers is properly enough.
Re: [PATCH V4 5/9] nvme: add sysfs attribute to change IO timeout per controller
Quoted 7 lines · Daniel Wagner
Quoted 6 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme I/O queues.
Add anio_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom I/O timeouts set.
Quoted 5 lines
Quoted 1 line
Signed-off-by Maurizio Lombardi <[email protected]>I was wondering if namespaces would ever want different timeouts but given that at this point it was always the default one and this change scopes to controllers is properly enough.
In theory, that might be possible, since each namespace could be backed by very different storage backends, ranging from a ramdisk to a file over an NFS export. However, in such cases, the user could simply ignore io_timeout and set his preferred timeout value on the queue associated with the namespace.
Maurizio
Sagi Grimberg
maintainerRe: [PATCH V4 5/9] nvme: add sysfs attribute to change IO timeout per controller
In reply to Maurizio LombardiQuoted 5 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme I/O queues.
Add anio_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom I/O timeouts set.
Why is this needed? Why not simply change the timeout on the namespaces themselves?
Re: [PATCH V4 5/9] nvme: add sysfs attribute to change IO timeout per controller
Quoted 11 lines · Sagi Grimberg
Quoted 5 lines · Maurizio Lombardi
Currently, there is no method to adjust the timeout values on a per controller basis with nvme I/O queues.
Add anio_timeoutattribute to nvme so that different nvme controllers which may have different timeout requirements can have custom I/O timeouts set.Why is this needed? Why not simply change the timeout on the namespaces themselves?
One reason is convenience, having an io_timeout default value at the controller level means that a newly discovered or hot-plugged namespace will automatically inherit the controller-specific timeout.
Additionally, for nvmeof, the controller allocates a dedicated I/O queue (ctrl->connect_q) that doesn't have a block device representation. The io_timeout attribute will allow userspace to change the timeout for this internal queue.
Maurizio
PATCH v4 6/9 nvme: use per controller timeout waits over depending on global default
In reply to Maurizio LombardiInstead of passing NVME_IO_TIMEOUT as a parameter with every call to nvme_wait_freeze_timeout, use the controller's preferred timeout.
--- drivers/nvme/host/apple.c | 2 +- drivers/nvme/host/core.c | 3 ++- drivers/nvme/host/nvme.h | 2 +- drivers/nvme/host/pci.c | 2 +- drivers/nvme/host/rdma.c | 2 +- drivers/nvme/host/tcp.c | 2 +- 6 files changed, 7 insertions(+), 6 deletions(-)
drivers/nvme/host/apple.c +1 -1
@@ -858,7 +858,7 @@ static void apple_nvme_disable(struct apple_nvme *anv, bool shutdown)* doing a safe shutdown. */ if (!dead && shutdown && freeze) -nvme_wait_freeze_timeout (&anv->ctrl,NVME_IO_TIMEOUT ); +nvme_wait_freeze_timeout (&anv->ctrl);nvme_quiesce_io_queues (&anv->ctrl);
drivers/nvme/host/core.c +2 -1
@@ -5249,8 +5249,9 @@ void nvme_unfreeze(struct nvme_ctrl *ctrl) }EXPORT_SYMBOL_GPL (nvme_unfreeze); -int nvme_wait_freeze_timeout (struct nvme_ctrl *ctrl,long timeout) +int nvme_wait_freeze_timeout (struct nvme_ctrl *ctrl) { +unsigned long timeout = ctrl->io_timeout;struct nvme_ns *ns;int srcu_idx;
drivers/nvme/host/nvme.h +1 -1
@@ -902,7 +902,7 @@ void nvme_sync_queues(struct nvme_ctrl *ctrl);void nvme_sync_io_queues (struct nvme_ctrl *ctrl);void nvme_unfreeze (struct nvme_ctrl *ctrl);void nvme_wait_freeze (struct nvme_ctrl *ctrl); -int nvme_wait_freeze_timeout (struct nvme_ctrl *ctrl,long timeout); +int nvme_wait_freeze_timeout (struct nvme_ctrl *ctrl);void nvme_start_freeze (struct nvme_ctrl *ctrl);static inline enum req_op nvme_req_op (struct nvme_command *cmd)
drivers/nvme/host/pci.c +1 -1
@@ -3276,7 +3276,7 @@ static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)* if doing a safe shutdown. */ if (!dead && shutdown) -nvme_wait_freeze_timeout (&dev->ctrl,NVME_IO_TIMEOUT ); +nvme_wait_freeze_timeout (&dev->ctrl); }nvme_quiesce_io_queues (&dev->ctrl);
drivers/nvme/host/rdma.c +1 -1
@@ -888,7 +888,7 @@ static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)if (!new) {nvme_start_freeze (&ctrl->ctrl);nvme_unquiesce_io_queues (&ctrl->ctrl); -if (!nvme_wait_freeze_timeout (&ctrl->ctrl,NVME_IO_TIMEOUT )) { +if (!nvme_wait_freeze_timeout (&ctrl->ctrl)) {/* * If we timed out waiting for freeze we are likely to * be stuck. Fail the controller initialization just
drivers/nvme/host/tcp.c +1 -1
@@ -2208,7 +2208,7 @@ static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)if (!new) {nvme_start_freeze (ctrl);nvme_unquiesce_io_queues (ctrl); -if (!nvme_wait_freeze_timeout (ctrl,NVME_IO_TIMEOUT )) { +if (!nvme_wait_freeze_timeout (ctrl)) {/* * If we timed out waiting for freeze we are likely to * be stuck. Fail the controller initialization just
Signature · 1 line
2.54.0
Re: [PATCH V4 6/9] nvme: use per controller timeout waits over depending on global default
Quoted 2 lines · Maurizio Lombardi
Instead of passing
NVME_IO_TIMEOUTas a parameter with every call tonvme_wait_freeze_timeout, use the controller's preferred timeout.
This should probably go very early in the series before even adding configurable timeouts as it just drops a pointless paramter.
The patch itself looks good:
Re: [PATCH V4 6/9] nvme: use per controller timeout waits over depending on global default
Quoted 6 lines · Christoph Hellwig
Quoted 2 lines · Maurizio Lombardi
Instead of passing
NVME_IO_TIMEOUTas a parameter with every call tonvme_wait_freeze_timeout, use the controller's preferred timeout.This should probably go very early in the series before even adding configurable timeouts as it just drops a pointless paramter.
It drops the parameter but uses the timeout value stored in ctrl->io_timeout, so it actually depends on the patch introducing the per-controller io_timeout.
Maurizio
Christoph Hellwig
maintainerRe: [PATCH V4 6/9] nvme: use per controller timeout waits over depending on global default
Quoted 11 lines · Maurizio Lombardi
Quoted 6 lines · Christoph Hellwig
Quoted 2 lines · Maurizio Lombardi
Instead of passing
NVME_IO_TIMEOUTas a parameter with every call tonvme_wait_freeze_timeout, use the controller's preferred timeout.This should probably go very early in the series before even adding configurable timeouts as it just drops a pointless paramter.
It drops the parameter but uses the timeout value stored in
ctrl->io_timeout, so it actually depends on the patch introducing the per-controllerio_timeout.
Well, switch to hardcoded NVME_IO_TIMEOUT early, and then replace it with the variable timeout like the other instances of NVME_IO_TIMEOUT.
Quoted 2 lines
Maurizio
---end quoted text---
Re: [PATCH V4 6/9] nvme: use per controller timeout waits over depending on global default
In reply to Maurizio LombardiQuoted 6 lines · Maurizio Lombardi
Instead of passing
NVME_IO_TIMEOUTas a parameter with every call tonvme_wait_freeze_timeout, use the controller's preferred timeout.Reviewed-by Daniel Wagner <[email protected]>Reviewed-by Mohamed Khalfella <[email protected]>Signed-off-by Maurizio Lombardi <[email protected]>
Quoted 9 lines
--- drivers/nvme/host/apple.c | 2 +- drivers/nvme/host/core.c | 3 ++- drivers/nvme/host/nvme.h | 2 +- drivers/nvme/host/pci.c | 2 +- drivers/nvme/host/rdma.c | 2 +- drivers/nvme/host/tcp.c | 2 +- 6 files changed, 7 insertions(+), 6 deletions(-)
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH v4 7/9 nvme-core: align fabrics_q teardown with admin_q in nvme_free_ctrl
In reply to Maurizio LombardiCurrently, the final reference for the fabrics admin queue (fabrics_q) is dropped inside nvme_remove_admin_tag_set(). However, the primary admin queue (admin_q) defers dropping its final reference until nvme_free_ctrl().
Move the blk_put_queue() call for fabrics_q from nvme_remove_admin_tag_set() to nvme_free_ctrl(). This aligns the lifecycle management of both admin queues, ensuring they are freed symmetrically when the controller is finally torn down.
--- drivers/nvme/host/core.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/nvme/host/core.c +3 -3
@@ -4932,10 +4932,8 @@ void nvme_remove_admin_tag_set(struct nvme_ctrl *ctrl)*/ nvme_stop_keep_alive (ctrl);blk_mq_destroy_queue (ctrl->admin_q); -if (ctrl->ops->flags &NVME_F_FABRICS ) { +if (ctrl->ops->flags &NVME_F_FABRICS )blk_mq_destroy_queue (ctrl->fabrics_q); -blk_put_queue (ctrl->fabrics_q); - }blk_mq_free_tag_set (ctrl->admin_tagset); }EXPORT_SYMBOL_GPL (nvme_remove_admin_tag_set);
@@ -5077,6 +5075,8 @@ static void nvme_free_ctrl(struct device *dev)if (ctrl->admin_q)blk_put_queue (ctrl->admin_q); +if (ctrl->fabrics_q) +blk_put_queue (ctrl->fabrics_q);if (!subsys || ctrl->instance != subsys->instance)ida_free (&nvme_instance_ida, ctrl->instance);nvme_free_cels (ctrl);
Signature · 1 line
2.54.0
Re: [PATCH V4 7/9] nvme-core: align fabrics_q teardown with admin_q in nvme_free_ctrl
Re: [PATCH V4 7/9] nvme-core: align fabrics_q teardown with admin_q in nvme_free_ctrl
In reply to Maurizio LombardiLooks good:
Re: [PATCH V4 7/9] nvme-core: align fabrics_q teardown with admin_q in nvme_free_ctrl
In reply to Maurizio LombardiQuoted 33 lines · Maurizio Lombardi
Currently, the final reference for the fabrics admin queue (
fabrics_q) is dropped insidenvme_remove_admin_tag_set(). However, the primary admin queue (admin_q) defers dropping its final reference untilnvme_free_ctrl().Move the
blk_put_queue()call forfabrics_qfromnvme_remove_admin_tag_set()tonvme_free_ctrl(). This aligns the lifecycle management of both admin queues, ensuring they are freed symmetrically when the controller is finally torn down.Reviewed-by Daniel Wagner <[email protected]>Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/core.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/nvme/host/core.c +1
-3@@ -4932,10 +4932,8 @@ void nvme_remove_admin_tag_set(struct nvme_ctrl *ctrl)*/ nvme_stop_keep_alive (ctrl);blk_mq_destroy_queue (ctrl->admin_q); -if (ctrl->ops->flags &NVME_F_FABRICS ) { +if (ctrl->ops->flags &NVME_F_FABRICS )blk_mq_destroy_queue (ctrl->fabrics_q); -blk_put_queue (ctrl->fabrics_q); - }blk_mq_free_tag_set (ctrl->admin_tagset); }EXPORT_SYMBOL_GPL (nvme_remove_admin_tag_set);
drivers/nvme/host/core.c static void nvme_free_ctrl(struct device *dev) · line 5082
Quoted 8 lines
drivers/nvme/host/core.c +2
-0@@ -5077,6 +5075,8 @@ static void nvme_free_ctrl(struct device *dev)if (ctrl->admin_q)blk_put_queue (ctrl->admin_q); +if (ctrl->fabrics_q) +blk_put_queue (ctrl->fabrics_q);if (!subsys || ctrl->instance != subsys->instance)ida_free (&nvme_instance_ida, ctrl->instance);nvme_free_cels (ctrl);
One wonders why we check for 'flags' in the first hunk, but for the existence of 'fabrics_q' in the second hunk. But anyway.
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
Re: [PATCH V4 7/9] nvme-core: align fabrics_q teardown with admin_q in nvme_free_ctrl
Quoted 35 lines · Hannes Reinecke
Quoted 34 lines · Maurizio Lombardi
Currently, the final reference for the fabrics admin queue (
fabrics_q) is dropped insidenvme_remove_admin_tag_set(). However, the primary admin queue (admin_q) defers dropping its final reference untilnvme_free_ctrl().Move the
blk_put_queue()call forfabrics_qfromnvme_remove_admin_tag_set()tonvme_free_ctrl(). This aligns the lifecycle management of both admin queues, ensuring they are freed symmetrically when the controller is finally torn down.Reviewed-by Daniel Wagner <[email protected]>Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/host/core.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/nvme/host/core.c +1
-3@@ -4932,10 +4932,8 @@ void nvme_remove_admin_tag_set(struct nvme_ctrl *ctrl)*/ nvme_stop_keep_alive (ctrl);blk_mq_destroy_queue (ctrl->admin_q); -if (ctrl->ops->flags &NVME_F_FABRICS ) { +if (ctrl->ops->flags &NVME_F_FABRICS )blk_mq_destroy_queue (ctrl->fabrics_q); -blk_put_queue (ctrl->fabrics_q); - }blk_mq_free_tag_set (ctrl->admin_tagset); }EXPORT_SYMBOL_GPL (nvme_remove_admin_tag_set);@@ -5077,6 +5075,8 @@ static void nvme_free_ctrl(struct device *dev)
Quoted 11 lines
drivers/nvme/host/core.c static void nvme_free_ctrl(struct device *dev) · line 5082
Quoted 7 lines
drivers/nvme/host/core.c +2
-0if (ctrl->admin_q)blk_put_queue (ctrl->admin_q); +if (ctrl->fabrics_q) +blk_put_queue (ctrl->fabrics_q);if (!subsys || ctrl->instance != subsys->instance)ida_free (&nvme_instance_ida, ctrl->instance);nvme_free_cels (ctrl);One wonders why we check for 'flags' in the first hunk, but for the existence of '
fabrics_q' in the second hunk.
But anyway.
That is true,
I have to send a V5 anyway to address other comments, so I will change it to use if(ctrl->fabrics_q).
Maurizio
PATCH v4 8/9 nvmet-loop: do not alloc admin tag set during reset
In reply to Maurizio LombardiCurrently, resetting a loopback controller unconditionally invokes nvme_alloc_admin_tag_set() inside nvme_loop_configure_admin_queue(). Doing so drops the old queue and allocates a new one. Consequently, this reverts the admin queue's timeout (q->rq_timeout) back to the module default (NVME_ADMIN_TIMEOUT), completely wiping out any custom timeout values the user may have configured via sysfs and potentially racing against the sysfs nvme_admin_timeout_store() function that may dereference the admin_q pointer during the RESETTING state.
Decouple the admin tag set lifecycle from the admin queue configuration and destruction paths, which are executed during resets; Specifically:
* Move nvme_alloc_admin_tag_set() into nvme_loop_create_ctrl() so it is only allocated once during the initial controller creation.
* Defer the destruction of the admin tag set to nvme_loop_delete_ctrl_host() and the terminal error-handling paths of nvme_loop_reset_ctrl_work() and nvme_loop_create_ctrl().
--- drivers/nvme/target/loop.c | 31 ++++++++++++++++--------------- 1 file changed, 16 insertions(+), 15 deletions(-)
drivers/nvme/target/loop.c +16 -15
@@ -274,7 +274,6 @@ static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)nvmet_sq_destroy (&ctrl->queues[0 ].nvme_sq);nvmet_cq_put (&ctrl->queues[0 ].nvme_cq); -nvme_remove_admin_tag_set (&ctrl->ctrl); }static void nvme_loop_free_ctrl (struct nvme_ctrl *nctrl)
@@ -375,25 +374,18 @@ static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl) } ctrl->ctrl.queue_count =1 ; - error =nvme_alloc_admin_tag_set (&ctrl->ctrl, &ctrl->admin_tag_set, - &nvme_loop_admin_mq_ops, -sizeof (struct nvme_loop_iod ) + -NVME_INLINE_SG_CNT *sizeof (struct scatterlist )); -if (error) - goto out_free_sq; -/* reset stopped state for the fresh admin queue */ clear_bit (NVME_CTRL_ADMIN_Q_STOPPED , &ctrl->ctrl.flags); error =nvmf_connect_admin_queue (&ctrl->ctrl);if (error) - goto out_cleanup_tagset; + goto out_free_sq;set_bit (NVME_LOOP_Q_LIVE , &ctrl->queues[0 ].flags); error =nvme_enable_ctrl (&ctrl->ctrl);if (error) - goto out_cleanup_tagset; + goto out_free_sq; ctrl->ctrl.max_hw_sectors = (NVME_LOOP_MAX_SEGMENTS -1 ) <<PAGE_SECTORS_SHIFT ;
@@ -402,14 +394,12 @@ static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl) error =nvme_init_ctrl_finish (&ctrl->ctrl, false);if (error) - goto out_cleanup_tagset; + goto out_free_sq;return 0 ; -out_cleanup_tagset: -clear_bit (NVME_LOOP_Q_LIVE , &ctrl->queues[0 ].flags); -nvme_remove_admin_tag_set (&ctrl->ctrl); out_free_sq: +clear_bit (NVME_LOOP_Q_LIVE , &ctrl->queues[0 ].flags);nvmet_sq_destroy (&ctrl->queues[0 ].nvme_sq);nvmet_cq_put (&ctrl->queues[0 ].nvme_cq);return error;
@@ -432,6 +422,7 @@ static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)static void nvme_loop_delete_ctrl_host (struct nvme_ctrl *ctrl) {nvme_loop_shutdown_ctrl (to_loop_ctrl (ctrl)); +nvme_remove_admin_tag_set (ctrl); }static void nvme_loop_delete_ctrl (struct nvmet_ctrl *nctrl)
@@ -494,6 +485,7 @@ static void nvme_loop_reset_ctrl_work(struct work_struct *work)nvme_cancel_admin_tagset (&ctrl->ctrl);nvme_loop_destroy_admin_queue (ctrl); out_disable: +nvme_remove_admin_tag_set (&ctrl->ctrl);dev_warn (ctrl->ctrl.device,"Removing after reset failure\n" );nvme_uninit_ctrl (&ctrl->ctrl); }
@@ -594,10 +586,17 @@ static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,if (!ctrl->queues) goto out_uninit_ctrl; - ret =nvme_loop_configure_admin_queue (ctrl); + ret =nvme_alloc_admin_tag_set (&ctrl->ctrl, &ctrl->admin_tag_set, + &nvme_loop_admin_mq_ops, +sizeof (struct nvme_loop_iod ) + +NVME_INLINE_SG_CNT *sizeof (struct scatterlist ));if (ret) goto out_free_queues; + ret =nvme_loop_configure_admin_queue (ctrl); +if (ret) + goto out_remove_admin_tagset; +if (opts->queue_size > ctrl->ctrl.maxcmd) {/* warn if maxcmd is lower than queue_size */ dev_warn(ctrl->ctrl.device,
@@ -633,6 +632,8 @@ static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,nvme_quiesce_admin_queue (&ctrl->ctrl);nvme_cancel_admin_tagset (&ctrl->ctrl);nvme_loop_destroy_admin_queue (ctrl); +out_remove_admin_tagset: +nvme_remove_admin_tag_set (&ctrl->ctrl); out_free_queues:kfree (ctrl->queues); out_uninit_ctrl:
Signature · 1 line
2.54.0
Re: [PATCH V4 8/9] nvmet-loop: do not alloc admin tag set during reset
Quoted 16 lines · Maurizio Lombardi
Currently, resetting a loopback controller unconditionally invokes
nvme_alloc_admin_tag_set()insidenvme_loop_configure_admin_queue(). Doing so drops the old queue and allocates a new one. Consequently, this reverts the admin queue's timeout (q->rq_timeout) back to the module default (NVME_ADMIN_TIMEOUT), completely wiping out any custom timeout values the user may have configured via sysfs and potentially racing against the sysfsnvme_admin_timeout_store()function that may dereference theadmin_qpointer during the RESETTING state.Decouple the admin tag set lifecycle from the admin queue configuration and destruction paths, which are executed during resets; Specifically:
* Move
nvme_alloc_admin_tag_set()intonvme_loop_create_ctrl()so it is only allocated once during the initial controller creation.
Quoted 6 lines
* Defer the destruction of the admin tag set to
nvme_loop_delete_ctrl_host()and the terminal error-handling paths ofnvme_loop_reset_ctrl_work()andnvme_loop_create_ctrl().Signed-off-by Maurizio Lombardi <[email protected]>
Re: [PATCH V4 8/9] nvmet-loop: do not alloc admin tag set during reset
In reply to Maurizio LombardiRe: [PATCH V4 8/9] nvmet-loop: do not alloc admin tag set during reset
In reply to Maurizio LombardiLooks good:
(I still wish we could consolidate some of the alloc/reset flow a little more instead of having to do this in every driver)
Re: [PATCH V4 8/9] nvmet-loop: do not alloc admin tag set during reset
In reply to Maurizio LombardiQuoted 21 lines · Maurizio Lombardi
Currently, resetting a loopback controller unconditionally invokes
nvme_alloc_admin_tag_set()insidenvme_loop_configure_admin_queue(). Doing so drops the old queue and allocates a new one. Consequently, this reverts the admin queue's timeout (q->rq_timeout) back to the module default (NVME_ADMIN_TIMEOUT), completely wiping out any custom timeout values the user may have configured via sysfs and potentially racing against the sysfsnvme_admin_timeout_store()function that may dereference theadmin_qpointer during the RESETTING state.Decouple the admin tag set lifecycle from the admin queue configuration and destruction paths, which are executed during resets; Specifically:
* Move
nvme_alloc_admin_tag_set()intonvme_loop_create_ctrl()so it is only allocated once during the initial controller creation.* Defer the destruction of the admin tag set to
nvme_loop_delete_ctrl_host()and the terminal error-handling paths ofnvme_loop_reset_ctrl_work()andnvme_loop_create_ctrl().
Quoted 5 lines
Signed-off-by Maurizio Lombardi <[email protected]>--- drivers/nvme/target/loop.c | 31 ++++++++++++++++--------------- 1 file changed, 16 insertions(+), 15 deletions(-)
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH v4 9/9 nvme-core: warn on allocating admin tag set with existing queue
In reply to Maurizio LombardiCurrently, nvme_alloc_admin_tag_set() silently drops and releases the existing admin_q if it called on a controller that already had one (e.g., during a controller reset).
However, transport drivers should not be reallocating the admin tag set and queue during a reset. Dropping the old queue and allocating a new one destroys user-configured timeouts and may race against nvme_admin_timeout_store()
Since all transport drivers are now expected to preserve the admin queue across resets, calling nvme_alloc_admin_tag_set() when ctrl->admin_q is already populated is a bug.
Remove the silent cleanup and replace it with a WARN_ON_ONCE() to explicitly catch any transport drivers that violate this lifecycle rule
--- drivers/nvme/host/core.c | 7 +------ 1 file changed, 1 insertion(+), 6 deletions(-)
drivers/nvme/host/core.c +1 -6
@@ -4889,12 +4889,7 @@ int nvme_alloc_admin_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,if (ret)return ret; -/* -* If a previous admin queue exists (e.g., from before a reset), -* put it now before allocating a new one to avoid orphaning it. -*/ -if (ctrl->admin_q) -blk_put_queue (ctrl->admin_q); +WARN_ON_ONCE (ctrl->admin_q); ctrl->admin_q =blk_mq_alloc_queue (set,NULL ,NULL );if (IS_ERR (ctrl->admin_q)) {
Signature · 1 line
2.54.0
Re: [PATCH V4 9/9] nvme-core: warn on allocating admin tag set with existing queue