On Wed, 2023-04-19 at 10:58 -0700, Bart Van Assche wrote: > On 4/19/23 07:02, James Bottomley wrote: > > device_shutdown() goes in reverse devices_kset->list order, so it > > looks like it would do the PCI device then the SCSI device then the > > ULD then block, so we can use the queues in SCSI for emergency > > actions (like flush or stop) before block goes down. Although this > > isn't guaranteed; there are things, like device_link_add, which > > reorder this kset, so we'd need to make sure the above assumption > > is correct. > > Hi James, > > My understanding is that both the block device associated with > /dev/sd<x> and the struct scsi_disk associated with the same SCSI > device have the sdev_gendev member of struct scsi_device as parent. > In other words, without creating device links, there are no > guarantees about the order in which the .shutdown() methods of struct > block_device.bd_device and struct scsi_disk.disk_dev are called. > Adding device links seems like an unnecessary complexity to me. Hence > my preference to make sd_shutdown() responsible for quiescing future > SCSI command activity. So if we can't reliably get it right, let's not do it at all. The previous argument you made was about problems with I/O to shutdown PCI devices. That's not SCSI specific, so either we fix it for everything or decide it's not really a problem for anything. James
