Skip to contentSkip to conversation

Smatch reports for drivers/scsi/scsi.c

DISCUSSION Andrew Goodbody 2 participants 2 messages Jul 2, 2025 – Jul 4, 2025

lore
Share & tools
lore
#0 lore raw

Hi,

Smatch reports the following for drivers/scsi/scsi.c

drivers/scsi/scsi.c:165 scsi_setup_erase_ext() warn: right shifting more than type allows 16 vs 24 drivers/scsi/scsi.c:166 scsi_setup_erase_ext() warn: right shifting more than type allows 16 vs 16

I just wanted some ideas on the best way to resolve this. There are differing patterns even within this file about how the number of blocks parameter should be handled in calls to static functions. In all cases the value starts as a lbaint_t which can be a long or a 64 bit type.

1) in scsi_read there are calls to scsi_setup_read_ext whose third parameter is an unsigned short. The lbaint_t variable is assigned to an unsigned short (smallblks), which will simply truncate the value, before smallblks is passed as the 3rd parameter to scsi_setup_read_ext. In scsi_setup_read_ext the third parameter (blocks) is split into 2 bytes. A third byte is set to 0 and I am guessing that if appropriate this could have taken a higher order byte if that had made sense.

2) scsi_write is similar to scsi_read

3) scsi_erase calls scsi_setup_erase_ext with lbaint_t as the third parameter. scsi_setup_erase_ext takes unsigned short so there is an implicit truncation here. Here is where Smatch complains as this unsigned short is shifted 16 or 24 bits to the right, both of which exceed the size of this type.

So possible fixes are to make all three functions consistent and take an lbaint_t as the third parameter. Or to make scsi_setup_erase_ext like the other two and just hard code those bytes with excessive shifts to be 0. Can probably remove the use of the smallblks variable at the same time.

Suggestions?

Andrew

#1 lore raw
Quoted 30 lines · Andrew Goodbody

Hi,

Smatch reports the following for drivers/scsi/scsi.c

drivers/scsi/scsi.c:165 scsi_setup_erase_ext() warn: right shifting more than type allows 16 vs 24
drivers/scsi/scsi.c:166 scsi_setup_erase_ext() warn: right shifting more than type allows 16 vs 16

I just wanted some ideas on the best way to resolve this. There are differing patterns even within this file about how the number of blocks parameter should be handled in calls to static functions. In all cases the value starts as a lbaint_t which can be a long or a 64 bit type.

1) in scsi_read there are calls to scsi_setup_read_ext whose third parameter is an unsigned short. The lbaint_t variable is assigned to an unsigned short (smallblks), which will simply truncate the value, before smallblks is passed as the 3rd parameter to scsi_setup_read_ext.
In scsi_setup_read_ext the third parameter (blocks) is split into 2 bytes. A third byte is set to 0 and I am guessing that if appropriate this could have taken a higher order byte if that had made sense.

2) scsi_write is similar to scsi_read

3) scsi_erase calls scsi_setup_erase_ext with lbaint_t as the third parameter. scsi_setup_erase_ext takes unsigned short so there is an implicit truncation here. Here is where Smatch complains as this unsigned short is shifted 16 or 24 bits to the right, both of which exceed the size of this type.

Quoted 4 lines

So possible fixes are to make all three functions consistent and take an lbaint_t as the third parameter. Or to make scsi_setup_erase_ext like the other two and just hard code those bytes with excessive shifts to be 0. Can probably remove the use of the smallblks variable at the same time.

I suspect the best functional answer would be to use lbaint_t all around and cleanup/make sure we're good when SYS_64BIT_LBA is enabled.

Signature · 1 line
Tom