New: Scsi host number of Adaptec RAID controller changes upon a PCIe hotplug and re-insert
A message that is not in the archive (sent off-list or not yet received): [email protected]/
From bugzilla. Reported against PCI, but I think the SCSI host number is determined by SCSI, not by PCI, so I don't see a PCI issue here.
---------- Forwarded message ---------
From: <[email protected]>
Date: Thu, Jun 2, 2022 at 1:53 AM
Subject: [Bug 216059] New: Scsi host number of Adaptec RAID controller changes upon a PCIe hotplug and re-insert
To: <[email protected]>
https://bugzilla.kernel.org/show_bug.cgi?id=216059
Bug ID: 216059
Summary: Scsi host number of Adaptec RAID controller changes
upon a PCIe hotplug and re-insert
Product: Drivers
Version: 2.5
Kernel Version: 4.18.11
Hardware: All
OS: Linux
Tree: Mainline
Status: NEW
Severity: normal
Priority: P1
Component: PCI
Assignee: [email protected]
Reporter: [email protected]
Regression: NoCreated attachment 301088
--> https://bugzilla.kernel.org/attachment.cgi?id=301088&action=edit
The attachments contain the log files which capture before and after cases for a hotplug host number change
Summary:
This issue talks of the smartpqi driver for Adaptec controller, PCIe hotplug and the corresponding SCSI host number
The Linux message log shows the host number (e.g. [14:2:0:0] storage - /dev/sg27) unexpectedly changing when PCIe hot remove is rapidly followed by PCIe hot add. The problem appears when the two PCIe events occur in quick succession (i.e. less than 2 minutes). Because of the timing factor, the issue can appear to be intermittent. The problem has been root caused as a kernel issue.
Investigation:
Kernel (4.18.11-hotplug-patch) debug prints were added in the “scsi_add_host( )” and “scsi_remove_host ( )” routines. Per the debug prints in the log, the scsi host number is released after the PCIe hot add event, which forces the kernel use a different host number.
(debug prints)
Line 48: [ 1811.461055] smartpqi 0000:b3:00.0: Debuggg . . . pqi_unregister_scsi function, before scsi_remove_host, shost->host_num=14 //smartpqi requests host num 14 to be removed
Line 83: [ 2012.125750] (null): Debuggg . . shost->host_no before dev_set_name = host15
Line 84: [ 2012.126709] smartpqi 0000:b3:00.0: Debuggg . . . before scsi_add_host, shost->host_num=15 //upon hot add, kernel allocates host number 15, it should be 14
Line 132: [ 2014.181784] scsi host14: Debuggg . . in scsi_host_dev_release function shost_host_no to be removed = 14 //kernel finally frees host number 14, but it’s too late
Conclusion:
The kernel is not releasing the host number immediately when the smartpqi driver calls the scsi_remove_host() routine. If the PCIe cable is added back within 2 minutes, the kernel can unexpectedly return a different host number. This can lead to applications accessing the wrong device.
This is a Linux kernel issue and we will be raising a bugzilla on the linux kernel.
Questions:
Will this be a problem for Amazon? (Wouldn’t they take several minutes to do this, they have to be very careful when hot plugging?)
Do we need to consider other customers that might use PCIe hot plug in the future?
The problem is observed in kernel V4.18.11, but would V5.04/V5.10 make a difference (should we test it ourselves)?
Consequence:
Application accesses wrong device. Rebooting system may still result in wrong host number.
--
You may reply to this email to add a comment.
You are receiving this mail because:
You are watching the assignee of the bug.
Quoted 3 lines · Bjorn Helgaas
From bugzilla. Reported against PCI, but I think the SCSI host number is determined by SCSI, not by PCI, so I don't see a PCI issue here.
Agree this is SCSI. However, can we be clear about what the expectation is? Host Number looks like it should be expected to change on hot plug/hot unplug, so what is the actual problem?
I get that the driver not releasing the host is causing this, but even if it did do instant release, when you hot plug two SCSI devices, you stand a good chance of getting a different host number anyway.
James