Hi Martin, Thanks for looking into this. The problem I'm trying to solve is that, I want to have lower layer driver to setup max_sectors bigger than BLK_DEF_MAX_SECTORS. In Hyper-v, we use 2MB max transfer I/O size, in future version the max transfer I/O size will increase to 8MB. The implementation of sd.c limits the maximum value of max_sectors to BLK_DEF_MAX_SECTORS. Because sd_revalidate_disk is called late in the SCSI disk initialization process, there is no way for a lower layer driver to set this value to its "bigger" optimal size.
The reason why I think it may not be necessary for sd.c to setup max_sectors, it's because this value may have already been setup twice before reaching the code in sd.c: 1. When this disk device is first scanned, or re-scanned (in scsi_scan.c), where it eventually calls __scsi_init_queue(), and use the max_sectors in the scsi_host_template. 2. in slave_configure of scsi_host_template, when the lower layer driver implements this function in its template and it can change this value there. Long > -----Original Message----- > From: Martin K. Petersen [mailto:martin.peter...@oracle.com] > Sent: Monday, June 6, 2016 8:42 PM > To: Long Li <lon...@microsoft.com> > Cc: Tom Yan <tom.t...@gmail.com>; James E.J. Bottomley > <j...@linux.vnet.ibm.com>; Martin K. Petersen > <martin.peter...@oracle.com>; linux-s...@vger.kernel.org; linux- > ker...@vger.kernel.org > Subject: Re: [PATCH] sd: remove redundant check for > BLK_DEF_MAX_SECTORS > > >>>>> "Long" == Long Li <lon...@microsoft.com> writes: > > Long, > > Long> The reason is that, max_sectors already has value at this point, > Long> the default value is SCSI_DEFAULT_MAX_SECTORS > Long> (include/scsi/scsi_host.h). The lower layer host driver can change > Long> this value in its template. > > The LLD sets max_hw_sectors which indicates the capabilities of the > controller DMA hardware. Whereas the max_sectors limit is set by sd to > either follow advise by the device or--if not provided--use the block layer > default. max_sectors governs the size of READ/WRITE requests and do not > reflect the capabilities of the DMA hardware. > > Long> I think the drivers care about this value have already set it. So > Long> it's better not to change it again. If they want max_sectors to be > Long> set by sd, they can use BLOCK LIMITS VPD to tell it to do so. > > Most drivers don't have the luxury of being able to generate VPDs for their > attached target devices :) > > -- > Martin K. Petersen Oracle Linux Engineering
