Hi all, I am wondered how does SCSI EH guarantee there is no UAF of scsi_cmnd
if host reset failed. If host reset failed and eh_cmd_q of shost is not empty,
these command in eh_cmd_q would be added to done_q in scsi_eh_offline_sdevs()
and finished by scsi_eh_flush_done_q(). So these scsi_cmnd and it's related
request would be freed.
While since host reset failed, we can not guarantee the LLDDs has cleared all
references to these commands in eh_cmd_q. Is there any possibility that the
LLDDs reference to these commands? If this happened, then a using after free
issue would occur.
Waiting for your response.
On 4/19/22 16:28, Wenchao Hao wrote:
> Hi all, I am wondered how does SCSI EH guarantee there is no UAF of scsi_cmnd
> if host reset failed. If host reset failed and eh_cmd_q of shost is not empty,
> these command in eh_cmd_q would be added to done_q in scsi_eh_offline_sdevs()
> and finished by scsi_eh_flush_done_q(). So these scsi_cmnd and it's related
> request would be freed.
>
Yes.
> While since host reset failed, we can not guarantee the LLDDs has cleared all
> references to these commands in eh_cmd_q. Is there any possibility that the
> LLDDs reference to these commands? If this happened, then a using after free
> issue would occur.
>
If host reset has failed there are _no_ assumptions we can make about
commands, and not even about the PCI device itself.
So in effect, once host_reset failed the system is hosed.
We _might_ be able to resurrect the system by doing PCI EEH, but not
many systems nor drivers implement that.
Cheers,
Hannes
On 2022/4/19 22:56, Hannes Reinecke wrote:
> On 4/19/22 16:28, Wenchao Hao wrote:
>> Hi all, I am wondered how does SCSI EH guarantee there is no UAF of scsi_cmnd
>> if host reset failed. If host reset failed and eh_cmd_q of shost is not empty,
>> these command in eh_cmd_q would be added to done_q in scsi_eh_offline_sdevs()
>> and finished by scsi_eh_flush_done_q(). So these scsi_cmnd and it's related
>> request would be freed.
>>
> Yes.
>
>> While since host reset failed, we can not guarantee the LLDDs has cleared all
>> references to these commands in eh_cmd_q. Is there any possibility that the
>> LLDDs reference to these commands? If this happened, then a using after free
>> issue would occur.
>>
> If host reset has failed there are _no_ assumptions we can make about commands, and not even about the PCI device itself.
> So in effect, once host_reset failed the system is hosed.
>
> We _might_ be able to resurrect the system by doing PCI EEH, but not many systems nor drivers implement that.
>
> Cheers,
>
> Hannes
>
> .
Thanks a lot for your reply. I am writing single LUN reset EH flow which discussed in previous mail
and testing it, if the test result is good I would post it.
By the way, you said you would make EH rework, are you doing it? Could you tell us your plan
if it's not secret.