On Tue, Jun 17, 2025 at 03:25:21PM +0800, Ming Lei wrote: > > @@ -473,7 +473,9 @@ struct Scsi_Host *scsi_host_alloc(const struct scsi_host_template *sht, int priv > > else > > shost->max_sectors = SCSI_DEFAULT_MAX_SECTORS; > > > > - if (sht->max_segment_size) > > + if (sht->virt_boundary_mask) > > + shost->virt_boundary_mask = sht->virt_boundary_mask; > > + else if (sht->max_segment_size) > > shost->max_segment_size = sht->max_segment_size; > > else > > shost->max_segment_size = BLK_MAX_SEGMENT_SIZE; > > This way works, but I prefer to set it explicitly in driver, instead of > making block layer more fragile to deal with def ->max_segment_size > if ->virt_boundary_mask is defined The block layer already enforces this as it is a requirement. It is just the SCSI wrapper that broke it. Without this proper fix iser is still broken, and srp might or might not. > - for low level driver, if ->virt_boundary_mask is defined, ->max_segment_size > should be UINT_MAX obviously since it implies single `virt segment`. > Setting UINT_MAX in driver has document benefit too. No, it doesn't. It means you need to cargo cult copy and paste code instead of solving the problem in the proper place. > - for logical block device(md, dm, ...), both ->virt_boundary_mask and > ->max_segment_size may be set, and it is fine since logical block device > driver needn't to deal with sg Stacked devices should not inherit the hardware limits at all because we split below them, but that's a separate story. Either way this is not relevant here as we don't have stacking drivers that use the SCSI layer.
