Re: MFC fixes for uninitialized kernel stack variables in sys/cam or do direct fix for pvscsi driver

From: Zhenlei Huang <zlei_at_FreeBSD.org>
Date: Mon, 20 Jan 2025 12:48:47 UTC

> On Jan 15, 2025, at 12:37 PM, Warner Losh <imp@bsdimp.com> wrote:
> 
> Great. This looks good.

MFSing is done, along with the following commits those clear the stack allocated CCBs.

076686fe0703 cam: make sure to clear CCBs allocated on the stack
ec5325dbca62 cam: make sure to clear even more CCBs allocated on the stack
0f206cc91279 cam: add missing zeroing of a stack-allocated CCB.
616a676a0535 cam: clear stack-allocated CCB in the target layer

Everything looks good so far.

Best regards,
Zhenlei

> 
> Warner
> 
> On Tue, Jan 14, 2025, 9:36 PM Zhenlei Huang <zlei@freebsd.org <mailto:zlei@freebsd.org>> wrote:
> 
> 
>> On Jan 13, 2025, at 5:06 PM, Zhenlei Huang <zlei@FreeBSD.org <mailto:zlei@FreeBSD.org>> wrote:
>> 
>> 
>> 
>>> On Jan 13, 2025, at 4:22 PM, Zhenlei Huang <zlei@FreeBSD.org <mailto:zlei@FreeBSD.org>> wrote:
>>> 
>>> 
>>> 
>>>> On Dec 28, 2024, at 6:03 AM, Warner Losh <imp@bsdimp.com <mailto:imp@bsdimp.com>> wrote:
>>>> 
>>>> 
>>>> 
>>>> On Mon, Dec 2, 2024 at 2:41 AM Zhenlei Huang <zlei@freebsd.org <mailto:zlei@freebsd.org>> wrote:
>>>> Hi Warner,
>>>> 
>>>> Recently I upgraded some ESXi vms from 13.3 to 13.4 and noticed weird report for sas speed.
>>>> The boot console has the following,
>>>> 
>>>> ```
>>>> da0 at pvscsi0 bus 0 scbus2 target 0 lun 0
>>>> da0: <VMware Virtual disk 2.0> Fixed Direct Access SPC-4 SCSI device
>>>> da0: 4294967.295MB/s transfers
>>>> ```
>>>> But camcontrol report the correct value,
>>>> ```
>>>> # camcontrol inquiry da0 -R
>>>> pass1: 750.000MB/s transfers, Command Queueing Enabled
>>>> ```
>>>> 
>>>> The `4294967.295MB` is actually 0xffff_ffff or -1 but I do not see any logic set those values.
>>>> 
>>>> Finally I managed to get the stack trace,
>>>> ```
>>>> _scsi_announce_periph
>>>> scsi_announce_periph_sbuf
>>>> xpt_announce_periph_sbuf
>>>> dadone_proberc
>>>> xpt_done_process
>>>> xpt_done_td
>>>> fork_exit
>>>> fork_trampoline
>>>> ```
>>>> 
>>>> and noticed that the last param `cts` of `_scsi_announce_periph(struct cam_periph *periph, u_int *speed, u_int *freq, struct ccb_trans_settings *cts)`
>>>> is from kernel stack and is not properly initialized, latter I found some commits related to this,
>>>> 
>>>> 076686fe0703 cam: make sure to clear CCBs allocated on the stack
>>>> ec5325dbca62 cam: make sure to clear even more CCBs allocated on the stack
>>>> 0f206cc91279 cam: add missing zeroing of a stack-allocated CCB.
>>>> 616a676a0535 cam: clear stack-allocated CCB in the target layer
>>>> 
>>>> I applied them to stable/13, rebuild and reboot, now the speed of da0 is reported correctly. I also tried to patch the pvscsi driver with few lines and
>>>> it also works as intended.
>>>> 
>>>> ```
>>>> --- a/sys/dev/vmware/pvscsi/pvscsi.c
>>>> +++ b/sys/dev/vmware/pvscsi/pvscsi.c
>>>> @@ -1444,6 +1444,10 @@ pvscsi_action(struct cam_sim *sim, union ccb *ccb)
>>>>                 cts->proto_specific.scsi.flags = CTS_SCSI_FLAGS_TAG_ENB;
>>>>                 cts->proto_specific.scsi.valid = CTS_SCSI_VALID_TQ;
>>>> 
>>>> +               /* Prefer connection speed over sas port speed */
>>>> +               /* cts->xport_specific.sas.bitrate = 0; */
>>>> +               cts->xport_specific.sas.valid = 0;
>>>> +
>>>>                 ccb_h->status = CAM_REQ_CMP;
>>>>                 xpt_done(ccb);
>>>> ```
>>>> 
>>>> Things come clear and I know why this weird speed happens, now it is time to decide how to fix it.
>>>> 
>>>> Fixing the consumer of cam, aka pvscsi driver, is quite simple and promising. I did a quick search it appears other consumers set `cts->xport_specific.sas.valid` correctly. It does not convince me as I'm quite new to cam subsystem.
>>>> 
>>>> Yes. sas.valid is set when the sas.bitrate and other data has been set correctly. Setting it to 0 like this ensures it's ignored.  So if you know the speed, set sas.bitrate to that speed and sas.valid to 1.
>>> 
>>> That is clear.
>>> 
>>>> 
>>>> I'm not sure I answered the question right, but if not, please clarify or point out what I missed and I'll try again.
>>>>  
>>>> Which one do you prefer, MFC commits to stable/13, or do direct fix for pvscsi driver to stable/13 ?
>>>> 
>>>> [[ Sorry for the delay, I missed this all month ]]
>>>> 
>>>> I generally prefer a MFC, unless the code is no longer in -current. 
>>> 
>>> The code live in -current and all supported stable branches.
>>> 
>>>> Even if there's two different fixes for this logical bug, fixing it in current, then MFCing that (with the current hash) is fine, even if the stable/13 changes are completely different. 
>>> 
>>> The bug does not exist in current and stable/14 but only in stable/13, due to Edward's work ( commits those zero stack-allocated CCBs ) were not MFCed into stable/13 branch.
>>> 
>>>> For stable/13 I guess it matters a bit less than stable/14 since I'll be merging to it less, but if it's a commit from -current that doesn't need to be made to -stable because of the new commit on stable, I tend to include the MFC hash text.
>>> 
>>> Do you mean the `cherry picked from commit` commit message ?
>>> 
>>>> 
>>>> Warner
>>> 
>>> 
>>> I'm preparing and testing the MFCing. Bless me to not make things messed up :)
>> 
>> And the individual fix for pvscsi is posted to https://reviews.freebsd.org/D48438 <https://reviews.freebsd.org/D48438> .
> 
> Landed in -current. Will be MFCed to stable/13 after a few days.
> 
>> 
>>> 
>>> Best regards,
>>> Zhenlei
> 
> 
>