On 09/05/2022 14:20, Krzysztof Kozlowski wrote:
On 09/05/2022 13:28, John Garry wrote:
For some reason I cannot fetch your git due to "error: RPC failed ..."
which I think is a timeout. I seem to have this problem recently
whenever a linux.git clone has branches based on linux-next.git . Maybe
a git config issue for me...
Just to be sure - the link was not a git remote, but direct link for a
web browser. The repo is:
https://github.com/krzk/linux.git
branch: n/qcom-ufs-opp-cleanups-v2
OK, I'll double check. Regardless I'll sort my own IT issues :)
However this does not solve the problem. The SHT has "module" which gets
incremented/decremented. Exactly like in case of other drivers
(driver->owner).
Ah, I missed that this could be a problem. So we have this member to
stop the SCSI host driver being removed when we have disks mounted, etc.
But isn't scsi_host_template.module just a pointer to the local driver
module data (and that data gets incremented/decremented)? I am looking
at the THIS_MODULE definition in export.h:
Yes, it is just a pointer, just like 'struct device_driver.owner' is.
extern stuct module __this_module;
#define THIS_MODULE(&__this_module)
However I do see scsi_device_dev_release(), which does:
sdp->host->hostt->module = NULL
I am not sure how necessary that really is. I would need to check further.
Did you see if there other places which set hostt->module dynamically?
I think all SHT set it statically.
Incidentally I did notice that the ata ahci driver does not set sht->module.
Then it is being dynamically
incremented when needed and in scsi_device_dev_release() is being
nullified (as you mentioned above).
I guess this SHT->module is actually duplicating what most of drivers
(e.g. PCI, platform, USB) are doing. If I understand correctly, the
Scsi_Host is instantiated by the probe of a driver (pci_driver,
virtio_driver etc), therefore the SHT->module could be simply stored in
Scsi_Host.
If you check scsi_device_dev_release(), we try to do a 'get' - if it
fails, then we nullify hostt->module. I think that is important as then
we call execute_in_process_context(), whose worker does the 'put'.
However, the 'put' will get upset if the refcnt was 0, which it would be
if the earlier 'get' fails - hence the nullify is to avoid that
possibility. So whatever you do needs to handle that. Details are in
f2b85040
Thanks,
john
