About BLKSECTGET in sg
Subsystems (4)
Hi,
Is BLKSECTGET in sg intentionally kept "broken" (i.e. it gives out max_sectors in bytes) or is it just a miss? I'm just wondering if I should send a patch to fix it (and implement BLKSSZGET).
Also, shouldn't max_sectors_bytes() in both drivers/scsi/sg.c and block/scsi_ioctl.c use queue_logical_block_size() instead?
Regards,
Tom
Douglas Gilbert
maintainerQuoted 8 lines · Tom Yan
Hi,
Is BLKSECTGET in sg intentionally kept "broken" (i.e. it gives out
max_sectorsin bytes) or is it just a miss? I'm just wondering if I should send a patch to fix it (and implement BLKSSZGET).Also, shouldn't
max_sectors_bytes()in bothdrivers/scsi/sg.candblock/scsi_ioctl.cusequeue_logical_block_size()instead?
Tom,
I have no idea! One of the great things about maintaining a driver in Linux is that virtually anyone can send patches and get them accepted without and Ack from the maintainer. And when bugs are found in those patches, the culprits can't be found or have no inclination to fix them.
'git blame' will show you that I had nothing to do with the BLK* ioctl()s added to the sg driver. That said, if they are broken, they should be fixed. I would be interested in getting some test code as I'm not familiar with using those ioctl()s.
Would you like to send some patches?
Doug Gilbert
Hi Doug,
Quoted 17 lines · Douglas Gilbert
Quoted 8 lines · Tom Yan
Hi,
Is BLKSECTGET in sg intentionally kept "broken" (i.e. it gives out
max_sectorsin bytes) or is it just a miss? I'm just wondering if I should send a patch to fix it (and implement BLKSSZGET).Also, shouldn't
max_sectors_bytes()in bothdrivers/scsi/sg.candblock/scsi_ioctl.cusequeue_logical_block_size()instead?Tom,
I have no idea! One of the great things about maintaining a driver in Linux is that virtually anyone can send patches and get them accepted without and Ack from the maintainer. And when bugs are found in those patches, the culprits can't be found or have no inclination to fix them.
Quoted 3 lines
'git blame' will show you that I had nothing to do with the BLK* ioctl()s added to the sg driver. That said, if they are broken, they should be fixed. I would be interested in getting some test code as
You can test them easily with blockdev (--getmaxsect and --getss) in util-linux.
max_sectors_bytes() are used in sg_add_sfp() and the SG_{GET,SET}_RESERVED_SIZE ioctls. I am not familiar with those but I don't see why they shouldn't use the queue_logical_sector_size() (when they use queue_max_sectors()).
For the BLKSECTGET fix, I think the sg version needs to be bumped (so that userspace programs like qemu can work as properly as possible with different versions). Can you point me to some guidance on how exactly it should be bumped (e.g. which component(s), the "step size"...)? Or would you do that after the patches are accepted?
Quoted 3 lines
I'm not familiar with using those ioctl()s.
Would you like to send some patches?
They are on the way.
Quoted 4 lines
Doug Gilbert
Regards,
Tom
PATCH 1/4 scsi: sg: fix BLKSECTGET ioctl
It should give out the maximum number of sectors per request instead of maximum number of bytes.
--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +4 -2
@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));
@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH 2/4 scsi: sg: implement BLKSSZGET
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
--- drivers/scsi/sg.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/scsi/sg.c +2 -0
@@ -1118,6 +1118,8 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, max_sectors = min_t(unsigned int, USHRT_MAX,queue_max_sectors (sdp->device->request_queue));return put_user (max_sectors, ip); +case BLKSSZGET : +return put_user (queue_logical_block_size (sdp->device->request_queue), ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
Re: [PATCH 2/4] scsi: sg: implement BLKSSZGET
Quoted 6 lines · Tom Yan
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
Signed-off-by Tom Yan <[email protected]>
Quoted 18 lines
--- drivers/scsi/sg.c | 2 ++ 1 file changed, 2 insertions(+)drivers/scsi/sg.c +2
-0@@ -1118,6 +1118,8 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, max_sectors = min_t(unsigned int, USHRT_MAX,queue_max_sectors (sdp->device->request_queue));return put_user (max_sectors, ip); +case BLKSSZGET : +return put_user (queue_logical_block_size (sdp->device->request_queue), ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
PATCH 3/4 scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom Yan--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +3 -2
@@ -848,10 +848,11 @@ static int srp_done(Sg_fd *sfp, Sg_request *srp)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; } static void
Signature · 1 line
2.28.0
Re: [PATCH 3/4] scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
Quoted 1 line · Tom Yan
Signed-off-by Tom Yan <[email protected]>
Quoted 23 lines
--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)drivers/scsi/sg.c +3
-2@@ -848,10 +848,11 @@ static int srp_done(Sg_fd *sfp, Sg_request *srp)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; } static void
PATCH 4/4 block/scsi_ioctl.c: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom Yan--- block/scsi_ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
block/scsi_ioctl.c +3 -2
@@ -73,10 +73,11 @@ static int sg_set_timeout(struct request_queue *q, int __user *p)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; }static int sg_get_reserved_size (struct request_queue *q,int __user *p)
Signature · 1 line
2.28.0
Re: [PATCH 4/4] block/scsi_ioctl.c: use queue_logical_sector_size() in max_sectors_bytes()
Quoted 1 line · Tom Yan
Signed-off-by Tom Yan <[email protected]>
Quoted 23 lines
--- block/scsi_ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)block/scsi_ioctl.c +3
-2@@ -73,10 +73,11 @@ static int sg_set_timeout(struct request_queue *q, int __user *p)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; }static int sg_get_reserved_size (struct request_queue *q,int __user *p)
Re: [PATCH 1/4] scsi: sg: fix BLKSECTGET ioctl
In reply to Tom YanQuoted 4 lines · Tom Yan
It should give out the maximum number of sectors per request instead of maximum number of bytes.
Signed-off-by Tom Yan <[email protected]>
Quoted 29 lines
--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/scsi/sg.c +4
-2@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Re: [PATCH 1/4] scsi: sg: fix BLKSECTGET ioctl
In reply to Tom YanQuoted 22 lines · Tom Yan
It should give out the maximum number of sectors per request instead of maximum number of bytes.
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/scsi/sg.c +1
-0@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;
drivers/scsi/sg.c sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, · line 1123
Quoted 10 lines
drivers/scsi/sg.c +3
-2return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Is this perhaps a backwards-incompatible change to the kernel ABI, the kind of change Linus totally disagrees with?
Additionally, please Cc linux-api for patches that modify the kernel ABI. From https://www.kernel.org/doc/man-pages/linux-api-ml.html: "The kernel source file Documentation/SubmitChecklist notes that all Linux kernel patches that change userspace interfaces should be CCed to [email protected], so that the various parties who are interested in API changes are informed. For further information, see
https://www.kernel.org/doc/man-pages/linux-api-ml.html"
Thanks,
Bart.
Douglas Gilbert
maintainerRe: [PATCH 1/4] scsi: sg: fix BLKSECTGET ioctl
Quoted 37 lines · Bart Van Assche
Quoted 22 lines · Tom Yan
It should give out the maximum number of sectors per request instead of maximum number of bytes.
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/scsi/sg.c +1
-0@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;drivers/scsi/sg.c sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, · line 1123
Quoted 10 lines
drivers/scsi/sg.c +3
-2return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,Is this perhaps a backwards-incompatible change to the kernel ABI, the kind of change Linus totally disagrees with?
Quoted 7 lines
Additionally, please Cc linux-api for patches that modify the kernel ABI.
Quoted 1 line
From https://www.kernel.org/doc/man-pages/linux-api-ml.html: "The kernel
source file
Documentation/SubmitChecklistnotes that all Linux kernel patches that change userspace interfaces should be CCed to [email protected], so that the various parties who are interested in API changes are informed. For further information, see
https://www.kernel.org/doc/man-pages/linux-api-ml.html"
Hmm,
The BLK* ioctl()s in the sg driver were an undocumented addition by others.
Plus it is not clear to me why a char device such as the sg driver should
be supporting BLK* ioctl(2)s. For example, how should an enclosure react to those ioctls or a WLUN?
If they are going to be supported for storage devices and /dev/sdb and /dev/sg2 are the same device then if:
blockdev --getmaxsect /dev/sg1
gives a different result to:
blockdev --getmaxsect /dev/sdb
then I would consider that a bug. So fixing it is making the kernel ABI more consistent :-)
Doug Gilbert
Re: [PATCH 1/4] scsi: sg: fix BLKSECTGET ioctl
It was my concern as well, that for this sort of "backwards-incompatible reason", it has been kept broken intentionally.
I am not sure if queue_max_sectors() or BLKSECTGET has ever been implemented in the block layer to give out the limit in bytes, but it certainly isn't the case anymore.
I am not in position to say whether it's "right" or "wrong" to implement BLKSECTGET/BLKSSZGET in the sg driver, but they is definitely useful in some cases (as long as the queue_* functions work for the given underlying device). Is it not okay if they don't ultimately work on *some* devices, even when they aren't named SG_*?
Perhaps we can consider having them removed as well (and implement them as e.g. SG_GET_MAX_SECTORS and SG_GET_LBS; but IMHO that only makes a point if they can be made to work on more than block devices).
Quoted 55 lines · Douglas Gilbert
Quoted 37 lines · Bart Van Assche
Quoted 22 lines · Tom Yan
It should give out the maximum number of sectors per request instead of maximum number of bytes.
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/scsi/sg.c +1
-0@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;drivers/scsi/sg.c sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, · line 1123
Quoted 10 lines
drivers/scsi/sg.c +3
-2return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,Is this perhaps a backwards-incompatible change to the kernel ABI, the kind of change Linus totally disagrees with?
Quoted 7 lines
Additionally, please Cc linux-api for patches that modify the kernel ABI.
Quoted 1 line
From https://www.kernel.org/doc/man-pages/linux-api-ml.html: "The kernel
source file
Documentation/SubmitChecklistnotes that all Linux kernel patches that change userspace interfaces should be CCed to [email protected], so that the various parties who are interested in API changes are informed. For further information, see
https://www.kernel.org/doc/man-pages/linux-api-ml.html"Hmm,
The BLK* ioctl()s in the sg driver were an undocumented addition by others. Plus it is not clear to me why a char device such as the sg driver should be supporting BLK* ioctl(2)s. For example, how should an enclosure react to those ioctls or a WLUN?If they are going to be supported for storage devices and
/dev/sdband/dev/sg2are the same device then if:
Quoted 7 lines
blockdev --getmaxsect /dev/sg1gives a different result to:
blockdev --getmaxsect /dev/sdbthen I would consider that a bug. So fixing it is making the kernel ABI more consistent :-)
That's exactly my point. They should work identically as the ones implemented in the block layer, because of their names.
With that said, sg_version needs to be bumped once the fix gets in, so that there's a way to differentiate the "different implementations" in userspace.
Quoted 5 lines
Doug Gilbert
Regards,
Tom
PATCH RESEND 1/4 scsi: sg: fix BLKSECTGET ioctl
It should give out the maximum number of sectors per request instead of maximum number of bytes.
--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +4 -2
@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));
@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH RESEND 2/4 scsi: sg: implement BLKSSZGET
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
--- drivers/scsi/sg.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/scsi/sg.c +2 -0
@@ -1118,6 +1118,8 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, max_sectors = min_t(unsigned int, USHRT_MAX,queue_max_sectors (sdp->device->request_queue));return put_user (max_sectors, ip); +case BLKSSZGET : +return put_user (queue_logical_block_size (sdp->device->request_queue), ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH RESEND 3/4 scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom Yan--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +3 -2
@@ -848,10 +848,11 @@ static int srp_done(Sg_fd *sfp, Sg_request *srp)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; } static void
Signature · 1 line
2.28.0
PATCH RESEND 4/4 block/scsi_ioctl.c: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom Yan--- block/scsi_ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
block/scsi_ioctl.c +3 -2
@@ -73,10 +73,11 @@ static int sg_set_timeout(struct request_queue *q, int __user *p)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; }static int sg_get_reserved_size (struct request_queue *q,int __user *p)
Signature · 1 line
2.28.0
PATCH RESEND 1/4 scsi: sg: fix BLKSECTGET ioctl
In reply to Tom YanIt should give out the maximum number of sectors per request instead of maximum number of bytes.
--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +4 -2
@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));
@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH RESEND 2/4 scsi: sg: implement BLKSSZGET
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
--- drivers/scsi/sg.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/scsi/sg.c +2 -0
@@ -1118,6 +1118,8 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, max_sectors = min_t(unsigned int, USHRT_MAX,queue_max_sectors (sdp->device->request_queue));return put_user (max_sectors, ip); +case BLKSSZGET : +return put_user (queue_logical_block_size (sdp->device->request_queue), ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 4 lines · Tom Yan
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
I don't think this has any business going in. /dev/sg is a generic interface that is not specific to block based command sets. Just issue your command set specific command to query the logical block size if you care about it.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Feel free to omit this. But then you will probably want to ditch BLKSECTGET as well, and then any usage of queue_max_sectors(), and maybe more/all queue_*().
I'm not really interested in discussing/arguing whether general/ideally-speaking it's appropriate/necessary to keep BLKSECTGET / add BLKSSZGET. The only reason I added this is that, when BLKSECTGET was introduced to sg long time ago, it was wrongly implemented to gives out the limit in bytes, so now when I'm fixing it, I'm merely making sure that whatever has been relying on the ioctl (e.g. qemu) will only need to do one more ioctl (instead of e.g. doing SCSI in its non-SCSI-specific part), if they want/need the limit in bytes. If they can be implemented more "generic"-ly, feel free to improve/extend them to make them "SG_*-qualified".
Even if you can do SCSI from the userspace, or even should, I don't see any reason that we shouldn't provide an ioctl to do queue_logical_block_size() *while we provide one to do queue_max_sectors()*.
Quoted 11 lines · Christoph Hellwig
Quoted 4 lines · Tom Yan
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
I don't think this has any business going in.
/dev/sgis a generic interface that is not specific to block based command sets. Just issue your command set specific command to query the logical block size if you care about it.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 15 lines · Tom Yan
Feel free to omit this. But then you will probably want to ditch BLKSECTGET as well, and then any usage of
queue_max_sectors(), and maybe more/all queue_*().I'm not really interested in discussing/arguing whether general/ideally-speaking it's appropriate/necessary to keep BLKSECTGET / add BLKSSZGET. The only reason I added this is that, when BLKSECTGET was introduced to sg long time ago, it was wrongly implemented to gives out the limit in bytes, so now when I'm fixing it, I'm merely making sure that whatever has been relying on the ioctl (e.g. qemu) will only need to do one more ioctl (instead of e.g. doing SCSI in its non-SCSI-specific part), if they want/need the limit in bytes. If they can be implemented more "generic"-ly, feel free to improve/extend them to make them "SG_*-qualified".
Quoted 4 lines
Even if you can do SCSI from the userspace, or even should, I don't see any reason that we shouldn't provide an ioctl to do
queue_logical_block_size()*while we provide one to do queue_max_sectors()*.
Well, the different definition in bytes for sg actually makes sense to me, as a bytes based limit is what fundamentally makes sense for the passthrough interface. Only that it reuses the same cmd value is a bit confusing. So instead of changing anything and potentially breaking applications I'd suggest to just better document the semantics.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
If we rename it to e.g. SG_GET_MAX_XFER_BYTES, it will still break applications unless we also keep the wrong/ugly/confusing name (and you lose the advantage/generality that the two ioctls can be used on both sg and "pure" block devices; which seems to be the case of some SG_* ioctls as well).
I don't see what it has to do with passthrough. Either way, it's just a matter of whether you want to decouple it and make things more flexible. The only real disadvantage is, you will have to do two ioctls instead of one, but no more than that, and for good reasons.
I don't really care enough though. I mean, I'm okay with SG_GET_MAX_XFER_BYTES *and* NO "improper" BLKSECTGET. If that will get the patch series in, I am willing to send a new version. If not, I'm just gonna drop this.
Quoted 27 lines · Christoph Hellwig
Quoted 15 lines · Tom Yan
Feel free to omit this. But then you will probably want to ditch BLKSECTGET as well, and then any usage of
queue_max_sectors(), and maybe more/all queue_*().I'm not really interested in discussing/arguing whether general/ideally-speaking it's appropriate/necessary to keep BLKSECTGET / add BLKSSZGET. The only reason I added this is that, when BLKSECTGET was introduced to sg long time ago, it was wrongly implemented to gives out the limit in bytes, so now when I'm fixing it, I'm merely making sure that whatever has been relying on the ioctl (e.g. qemu) will only need to do one more ioctl (instead of e.g. doing SCSI in its non-SCSI-specific part), if they want/need the limit in bytes. If they can be implemented more "generic"-ly, feel free to improve/extend them to make them "SG_*-qualified".
Quoted 4 lines
Even if you can do SCSI from the userspace, or even should, I don't see any reason that we shouldn't provide an ioctl to do
queue_logical_block_size()*while we provide one to do queue_max_sectors()*.Well, the different definition in bytes for sg actually makes sense to me, as a bytes based limit is what fundamentally makes sense for the passthrough interface. Only that it reuses the same cmd value is a bit confusing. So instead of changing anything and potentially breaking applications I'd suggest to just better document the semantics.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 5 lines · Tom Yan
If we rename it to e.g.
SG_GET_MAX_XFER_BYTES, it will still break applications unless we also keep the wrong/ugly/confusing name (and you lose the advantage/generality that the two ioctls can be used on both sg and "pure" block devices; which seems to be the case of some SG_* ioctls as well).
How is that an advantage? Applications that works with block devices don't really work with a magic passthrough character device.
Quoted 4 lines
I don't really care enough though. I mean, I'm okay with
SG_GET_MAX_XFER_BYTES*and* NO "improper" BLKSECTGET. If that will get the patch series in, I am willing to send a new version. If not, I'm just gonna drop this.
You must assume that there are applications already that depend in the "weird" BLKSECTGET on sg, given that it has been around so long. Any change to the semantics will break them.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 10 lines · Christoph Hellwig
Quoted 5 lines · Tom Yan
If we rename it to e.g.
SG_GET_MAX_XFER_BYTES, it will still break applications unless we also keep the wrong/ugly/confusing name (and you lose the advantage/generality that the two ioctls can be used on both sg and "pure" block devices; which seems to be the case of some SG_* ioctls as well).How is that an advantage? Applications that works with block devices don't really work with a magic passthrough character device.
You must assume that there are applications already assuming that work. (And it will, at least in some cases, if this series get merged.)
And you have not been giving me a solid point anyway, as I said, it's just queue_*() at the end of the day; regardless of whether those would work in all sg cases, we have been using them in the sg driver anyway.
And it's not like we have to guarantee that (the) ioctls can work in every case anyway, right? (Especially when they aren't named SG_*).
I mean, what's even your point? How do you propose we fix this? Should we just leave it broken/wrong/rotten as-is (including the case that you don't consider it being so)? If that's what you are trying to say then please just say it, I'll just walk away quietly. I'm really kind of done with this sort of looping that leads to nowhere.
Quoted 9 lines
Quoted 5 lines
I don't really care enough though. I mean, I'm okay with
SG_GET_MAX_XFER_BYTES*and* NO "improper" BLKSECTGET. If that will get the patch series in, I am willing to send a new version. If not, I'm just gonna drop this.You must assume that there are applications already that depend in the "weird" BLKSECTGET on sg, given that it has been around so long. Any change to the semantics will break them.
While you forgot to assume that there are applications assuming BLKSECTGET is as "weird" in the block layer at the same time. (That's exactly what has happened in qemu.) You don't get the slightest advantage in "trying not to break things" by doing the wrong thing. It only goes on making more things broken.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 6 lines · Tom Yan
Quoted 2 lines
How is that an advantage? Applications that works with block devices don't really work with a magic passthrough character device.
You must assume that there are applications already assuming that work. (And it will, at least in some cases, if this series get merged.)
Why "must" I assume that?
Quoted 7 lines
And you have not been giving me a solid point anyway, as I said, it's just queue_*() at the end of the day; regardless of whether those would work in all sg cases, we have been using them in the sg driver anyway.
And it's not like we have to guarantee that (the) ioctls can work in every case anyway, right? (Especially when they aren't named SG_*).
No. While it is unfortunte we have all kinds of cases of ioctls working differnetly on different devices.
Quoted 2 lines
I mean, what's even your point? How do you propose we fix this?
I propose to not "fix" anything, because nothing is broken except for maybe a lack of documentation.
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 18 lines · Christoph Hellwig
Quoted 6 lines · Tom Yan
Quoted 2 lines
How is that an advantage? Applications that works with block devices don't really work with a magic passthrough character device.
You must assume that there are applications already assuming that work. (And it will, at least in some cases, if this series get merged.)
Why "must" I assume that?
Quoted 7 lines
And you have not been giving me a solid point anyway, as I said, it's just queue_*() at the end of the day; regardless of whether those would work in all sg cases, we have been using them in the sg driver anyway.
And it's not like we have to guarantee that (the) ioctls can work in every case anyway, right? (Especially when they aren't named SG_*).
Quoted 8 lines
No. While it is unfortunte we have all kinds of cases of ioctls working differnetly on different devices.
Quoted 2 lines
I mean, what's even your point? How do you propose we fix this?
I propose to not "fix" anything, because nothing is broken except for maybe a lack of documentation.
Alan Stern are you reading this thread? Why do I ask, you may ask? Because 'git blame' fingers you:
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
commit 44ec95425c1d9dce6e4638c29e4362cfb44814e7
Author: Alan Stern <[email protected]> Date: Tue Feb 20 11:01:57 2007 -0500
[SCSI] sg: cap reserved_size values at max_sectors
This patch (as857) modifies the SG_GET_RESERVED_SIZE and SG_SET_RESERVED_SIZE ioctls in the sg driver, capping the values at the device's request_queue's max_sectors value. This will permit cdrecord to obtain a legal value for the maximum transfer length, fixing Bugzilla #7026.
The patch also caps the initial reserved_size value. There's no reason to have a reserved buffer larger than max_sectors, since it would be impossible to use the extra space.
The corresponding ioctls in the block layer are modified similarly, and the initial value for the reserved_size is set as large as possible. This will effectively make it default to max_sectors. Note that the actual value is meaningless anyway, since block devices don't have a reserved buffer.
Finally, the BLKSECTGET ioctl is added to sg, so that there will be a uniform way for users to determine the actual max_sectors value for any raw SCSI transport.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Oops, I ack-ed this patch from 2007:-) Anyway it would seem BLKSECTGET ioctl was meant to be a "uniform way to determine the actual max_sectors value for any raw SCSI transport." Given that the initial implementation of BLKSECTGET now seems to be at odds with other implementations, what should we do?
It is possible that it was correct on 2007 and the BLKSECTGET ioctl has changed elsewhere but failed to fix the sg driver's implementation.
If I get a vote then it would be for Tom Yan's approach: reduce entropy or it will overwhelm us :-)
So Christoph, it IS documented, both in the above commit message and:
https://doug-gilbert.github.io/sg_v40.html
in Table 8. So please stop with your "maybe a lack of documentation" line.
Doug Gilbert
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
Quoted 59 lines · Douglas Gilbert
Quoted 18 lines · Christoph Hellwig
Quoted 6 lines · Tom Yan
Quoted 2 lines
How is that an advantage? Applications that works with block devices don't really work with a magic passthrough character device.
You must assume that there are applications already assuming that work. (And it will, at least in some cases, if this series get merged.)
Why "must" I assume that?
Quoted 7 lines
And you have not been giving me a solid point anyway, as I said, it's just queue_*() at the end of the day; regardless of whether those would work in all sg cases, we have been using them in the sg driver anyway.
And it's not like we have to guarantee that (the) ioctls can work in every case anyway, right? (Especially when they aren't named SG_*).
Quoted 8 lines
No. While it is unfortunte we have all kinds of cases of ioctls working differnetly on different devices.
Quoted 2 lines
I mean, what's even your point? How do you propose we fix this?
I propose to not "fix" anything, because nothing is broken except for maybe a lack of documentation.
Alan Stern are you reading this thread? Why do I ask, you may ask? Because 'git blame' fingers you:
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
commit
44ec95425c1d9dce6e4638c29e4362cfb44814e7Author: Alan Stern <[email protected]> Date: Tue Feb 20 11:01:57 2007 -0500[SCSI] sg: cap
reserved_sizevalues atmax_sectorsThis patch (as857) modifies the
SG_GET_RESERVED_SIZEandSG_SET_RESERVED_SIZEioctls in the sg driver, capping the values at the device's request_queue'smax_sectorsvalue. This will permit cdrecord to obtain a legal value for the maximum transfer length, fixing Bugzilla #7026.The patch also caps the initial
reserved_sizevalue. There's no reason to have a reserved buffer larger thanmax_sectors, since it would be impossible to use the extra space.The corresponding ioctls in the block layer are modified similarly, and the initial value for the
reserved_sizeis set as large as possible. This will effectively make it default tomax_sectors. Note that the actual value is meaningless anyway, since block devices don't have a reserved buffer.Finally, the BLKSECTGET ioctl is added to sg, so that there will be a uniform way for users to determine the actual
max_sectorsvalue for any raw SCSI transport.
Quoted 8 lines
Signed-off-by Alan Stern <[email protected]>Acked-by Jens Axboe <[email protected]>Acked-by Douglas Gilbert <[email protected]>Signed-off-by James Bottomley <[email protected]>^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Oops, I ack-ed this patch from 2007:-)
The Bugzilla entry it talks about is from 2006!
Quoted 3 lines
Anyway it would seem BLKSECTGET ioctl
was meant to be a "uniform way to determine the actualmax_sectorsvalue for any raw SCSI transport."
Right. The question at hand was: Given an open file descriptor for an SG device, how can a program determine the largest amount it can send in a single transfer? This ioctl seemed to be the best answer.
See comment #26 (https://bugzilla.kernel.org/show_bug.cgi?id=7026#c26)
and following for the viewpoint of the notoriously prickly author of cdrecord.
Quoted 5 lines
Given that the initial implementation of BLKSECTGET
now seems to be at odds with other implementations, what should we do?It is possible that it was correct on 2007 and the BLKSECTGET ioctl has changed elsewhere but failed to fix the sg driver's implementation.
Could be. Also, I'm not sure how careful people were back then to distinguish between logical and physical sector sizes.
Quoted 8 lines
If I get a vote then it would be for Tom Yan's approach: reduce entropy or it will overwhelm us :-)
So Christoph, it IS documented, both in the above commit message and:
https://doug-gilbert.github.io/sg_v40.htmlin Table 8. So please stop with your "maybe a lack of documentation" line.
My vote is not to change an interface which a program like cdrecord may currently rely on.
I can understand Christoph's point about documentation. It would be good to have something in the actual kernel source, rather than in the history or somebody's github files.
Alan Stern
Re: [PATCH RESEND 2/4] scsi: sg: implement BLKSSZGET
In reply to Christoph HellwigQuoted 10 lines · Christoph Hellwig
Quoted 6 lines · Tom Yan
Quoted 2 lines
How is that an advantage? Applications that works with block devices don't really work with a magic passthrough character device.
You must assume that there are applications already assuming that work. (And it will, at least in some cases, if this series get merged.)
Why "must" I assume that?
Because of the name. The whole discussion was about the name. (Well, also because 'I must assume that some applications has already been "relying" on the wrong name.')
Quoted 9 lines
Quoted 8 lines
And you have not been giving me a solid point anyway, as I said, it's just queue_*() at the end of the day; regardless of whether those would work in all sg cases, we have been using them in the sg driver anyway.
And it's not like we have to guarantee that (the) ioctls can work in every case anyway, right? (Especially when they aren't named SG_*).
Quoted 8 lines
No. While it is unfortunte we have all kinds of cases of ioctls working differnetly on different devices.
Quoted 2 lines
I mean, what's even your point? How do you propose we fix this?
I propose to not "fix" anything, because nothing is broken except for maybe a lack of documentation.
It won't really help anyway. Documenting something wrong won't make it correct anyway. It's just wrong, semantically wrong, logically wrong; expecting people to "rtfm and realize the difference" only makes it even more wrong. End of story.
This was never about whether it is "broken" or whether we should provide means to fetch the limit in the number of bytes or sectors. It was always just about the name.
Just one last question. So should we not even replace the bit shift with queue_logicial_block_size() in max_sectors_bytes()? Given that we "must assume that it has been that way for long enough so applications must have been dealing with the flaw on their own already and fixing it will only breaks them"?
PATCH RESEND 3/4 scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom Yan--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +3 -2
@@ -848,10 +848,11 @@ static int srp_done(Sg_fd *sfp, Sg_request *srp)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; } static void
Signature · 1 line
2.28.0
Re: [PATCH RESEND 3/4] scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
Quoted 4 lines · Tom Yan
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
I know I don't take patches without any changelog text at all, but maybe the scsi maintainers are more leniant.
You should write changelogs that explain _why_ you are doing what you are doing, you just say what you did, with no reasoning at all.
Same for another patch in this series.
thanks,
greg k-h
Re: [PATCH RESEND 3/4] scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
The resend is done to add [email protected] to the CC list. I also removed some CC because the email addresses no longer exist (and accidentally made a typo to one that still does, hence the second resend). I don't know if that counts as a change.
I didn't think 3/4 and 4/4 requires further explanation, as I thought they are self-explanatory enough (logical sector size has never been, or at least is no longer, necessarily 512). I can add that to the commit message.
P.S. Even I myself isn't exactly sure what/which clamping should be used in max_sectors_bytes(). The reason I picked USHRT_MAX is that BLKSECTGET in sg should work identically to its equivalence in the block layer, regardless of whether that is technically/programmatically necessary. So I decided to use the same clamping in max_sector_bytes(). But it seems fine/correct to clamp (max_sectors * logical_sector_size) to INT_MAX instead
(https://github.com/torvalds/linux/blob/v5.9-rc3/block/blk-mq.c#L3211).
(On the other hand, in that case it could end up being inconsistent to what BLKSECTGET + BLKSSZGET reports). Perhaps I should add my uncertainty / the alternative to the commit message.
Regards,
Tom
Quoted 18 lines · Greg KH
Quoted 4 lines · Tom Yan
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)I know I don't take patches without any changelog text at all, but maybe the scsi maintainers are more leniant.
You should write changelogs that explain _why_ you are doing what you are doing, you just say what you did, with no reasoning at all.
Same for another patch in this series.
thanks,
greg k-h
PATCH v2 1/4 scsi: sg: fix BLKSECTGET ioctl
It should give out the maximum number of sectors per request instead of maximum number of bytes, as that is what its equivalence in the block layer does (that is also the reason for the USHRT_MAX clamp; they should always have identical behaviour when possbile).
---
v2: add more details in commit messages
drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +4 -2
@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));
@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH v2 2/4 scsi: sg: implement BLKSSZGET
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
--- drivers/scsi/sg.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/scsi/sg.c +2 -0
@@ -1118,6 +1118,8 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, max_sectors = min_t(unsigned int, USHRT_MAX,queue_max_sectors (sdp->device->request_queue));return put_user (max_sectors, ip); +case BLKSSZGET : +return put_user (queue_logical_block_size (sdp->device->request_queue), ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH v2 3/4 scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom YanLogical sector size was never / is no longer necessarily 512.
Programatically speaking it may not be necessary for us to clamp max_sectors to USHRT_MAX here, but since such clamping is used in BLKSECTGET, it's probably a good idea to have it here too, so that what the function returns is consistent to what the ioctl reports.
Alternatively we can clamp (max_sectors * logical_block_size) to INT_MAX instead, or maybe even not clamping it at all.
P.S. sg_reserved_size is initially set to INT_MAX by blk_mq_init_allocated_queue().
--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +3 -2
@@ -848,10 +848,11 @@ static int srp_done(Sg_fd *sfp, Sg_request *srp)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; } static void
Signature · 1 line
2.28.0
PATCH v2 4/4 block/scsi_ioctl.c: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom YanProgramatically speaking it may not be necessary for us to clamp max_sectors to USHRT_MAX here, but since such clamping is used in BLKSECTGET, it's probably a good idea to have it here too, so that what the function returns is consistent to what the ioctl reports.
Alternatively we can clamp (max_sectors * logical_block_size) to INT_MAX instead, or maybe even not clamping it at all.
P.S. sg_reserved_size is initially set to INT_MAX by blk_mq_init_allocated_queue().
--- block/scsi_ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
block/scsi_ioctl.c +3 -2
@@ -73,10 +73,11 @@ static int sg_set_timeout(struct request_queue *q, int __user *p)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; }static int sg_get_reserved_size (struct request_queue *q,int __user *p)
Signature · 1 line
2.28.0
PATCH v3 1/4 scsi: sg: fix BLKSECTGET ioctl
In reply to Tom YanIt should give out the maximum number of sectors per request instead of maximum number of bytes, as that is what its equivalence in the block layer does (that is also the reason for the USHRT_MAX clamp; they should always have identical behaviour when possbile).
---
v2: add more details in commit messages
v3: add a missed line back to the commit message of the last patch
drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +4 -2
@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));
@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH v3 2/4 scsi: sg: implement BLKSSZGET
Provide an ioctl to get the logical sector size so that the maximum bytes per request can be derived.
Follow-up of the BLKSECTGET fix.
--- drivers/scsi/sg.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/scsi/sg.c +2 -0
@@ -1118,6 +1118,8 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, max_sectors = min_t(unsigned int, USHRT_MAX,queue_max_sectors (sdp->device->request_queue));return put_user (max_sectors, ip); +case BLKSSZGET : +return put_user (queue_logical_block_size (sdp->device->request_queue), ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,
Signature · 1 line
2.28.0
PATCH v3 3/4 scsi: sg: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom YanLogical sector size was never / is no longer necessarily 512.
Programatically speaking it may not be necessary for us to clamp max_sectors to USHRT_MAX here, but since such clamping is used in BLKSECTGET, it's probably a good idea to have it here too, so that what the function returns is consistent to what the ioctl reports.
Alternatively we can clamp (max_sectors * logical_block_size) to INT_MAX instead, or maybe even not clamping it at all.
P.S. sg_reserved_size is initially set to INT_MAX by blk_mq_init_allocated_queue().
--- drivers/scsi/sg.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/scsi/sg.c +3 -2
@@ -848,10 +848,11 @@ static int srp_done(Sg_fd *sfp, Sg_request *srp)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; } static void
Signature · 1 line
2.28.0
PATCH v3 4/4 block/scsi_ioctl.c: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom YanLogical sector size was never / is no longer necessarily 512.
Programatically speaking it may not be necessary for us to clamp max_sectors to USHRT_MAX here, but since such clamping is used in BLKSECTGET, it's probably a good idea to have it here too, so that what the function returns is consistent to what the ioctl reports.
Alternatively we can clamp (max_sectors * logical_block_size) to INT_MAX instead, or maybe even not clamping it at all.
P.S. sg_reserved_size is initially set to INT_MAX by blk_mq_init_allocated_queue().
--- block/scsi_ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
block/scsi_ioctl.c +3 -2
@@ -73,10 +73,11 @@ static int sg_set_timeout(struct request_queue *q, int __user *p)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; }static int sg_get_reserved_size (struct request_queue *q,int __user *p)
Signature · 1 line
2.28.0
PATCH RESEND 4/4 block/scsi_ioctl.c: use queue_logical_sector_size() in max_sectors_bytes()
In reply to Tom Yan--- block/scsi_ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
block/scsi_ioctl.c +3 -2
@@ -73,10 +73,11 @@ static int sg_set_timeout(struct request_queue *q, int __user *p)static int max_sectors_bytes (struct request_queue *q) {unsigned int max_sectors =queue_max_sectors (q); +unsigned int logical_block_size =queue_logical_block_size (q); - max_sectors = min_t(unsigned int, max_sectors, INT_MAX >> 9); + max_sectors = min_t(unsigned int, max_sectors, USHRT_MAX); -return max_sectors <<9 ; +return max_sectors * logical_block_size; }static int sg_get_reserved_size (struct request_queue *q,int __user *p)
Signature · 1 line
2.28.0
Douglas Gilbert
maintainerRe: [PATCH 1/4] scsi: sg: fix BLKSECTGET ioctl
In reply to Tom YanQuoted 3 lines · Tom Yan
It was my concern as well, that for this sort of "backwards-incompatible reason", it has been kept broken intentionally.
Bumping the sg driver version number is simple:
drivers/scsi/sg.c +2 -2
@@ -11,8 +11,8 @@* Copyright (C) 1998 - 2014 Douglas Gilbert */ -static int sg_version_num =30536 ;/* 2 digits for each component */ -#define SG_VERSION_STR "3.5.36" +static int sg_version_num =30537 ;/* 2 digits for each component */ +#define SG_VERSION_STR "3.5.37"/* * D. P. Gilbert ([email protected]), notes:
And bumping the version number is appropriate for an interface tweak/correction.
I'm rewriting the sg driver currently (12 months and counting) and am up to version 11 of the _first_ half. So far I'm using a sg_version_num of
Quoted 1 line
40000 for the rewritten code. Please keep away from version numbers
40000 and above.
The rewritten driver is documented here:
https://doug-gilbert.github.io/sg_v40.html
and its ioctls are listed in Table 8, including the BLK* ones. Perhaps you could suggest some corrections. Obviously BLKSSZGET needs to be added when your patches are accepted.
Doug Gilbert
Quoted 95 lines
I am not sure if
queue_max_sectors()or BLKSECTGET has ever been implemented in the block layer to give out the limit in bytes, but it certainly isn't the case anymore.I am not in position to say whether it's "right" or "wrong" to implement BLKSECTGET/BLKSSZGET in the sg driver, but they is definitely useful in some cases (as long as the queue_* functions work for the given underlying device). Is it not okay if they don't ultimately work on *some* devices, even when they aren't named SG_*?
Perhaps we can consider having them removed as well (and implement them as e.g.
SG_GET_MAX_SECTORSandSG_GET_LBS; but IMHO that only makes a point if they can be made to work on more than block devices).Quoted 55 lines · Douglas Gilbert
Quoted 37 lines · Bart Van Assche
Quoted 22 lines · Tom Yan
It should give out the maximum number of sectors per request instead of maximum number of bytes.
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/scsi/sg.c +1
-0@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;drivers/scsi/sg.c sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, · line 1123
Quoted 10 lines
drivers/scsi/sg.c +3
-2return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,Is this perhaps a backwards-incompatible change to the kernel ABI, the kind of change Linus totally disagrees with?
Quoted 7 lines
Additionally, please Cc linux-api for patches that modify the kernel ABI.
Quoted 1 line
From https://www.kernel.org/doc/man-pages/linux-api-ml.html: "The kernel
source file
Documentation/SubmitChecklistnotes that all Linux kernel patches that change userspace interfaces should be CCed to [email protected], so that the various parties who are interested in API changes are informed. For further information, see
https://www.kernel.org/doc/man-pages/linux-api-ml.html"Hmm,
The BLK* ioctl()s in the sg driver were an undocumented addition by others. Plus it is not clear to me why a char device such as the sg driver should be supporting BLK* ioctl(2)s. For example, how should an enclosure react to those ioctls or a WLUN?If they are going to be supported for storage devices and
/dev/sdband/dev/sg2are the same device then if:Quoted 7 lines
blockdev --getmaxsect /dev/sg1gives a different result to:
blockdev --getmaxsect /dev/sdbthen I would consider that a bug. So fixing it is making the kernel ABI more consistent :-)
That's exactly my point. They should work identically as the ones implemented in the block layer, because of their names.
With that said,
sg_versionneeds to be bumped once the fix gets in, so that there's a way to differentiate the "different implementations" in userspace.Quoted 5 lines
Doug Gilbert
Regards,
Tom
Re: [PATCH 1/4] scsi: sg: fix BLKSECTGET ioctl
Quoted 25 lines · Douglas Gilbert
Quoted 3 lines · Tom Yan
It was my concern as well, that for this sort of "backwards-incompatible reason", it has been kept broken intentionally.
Bumping the sg driver version number is simple:
drivers/scsi/sg.c +2
-2@@ -11,8 +11,8 @@* Copyright (C) 1998 - 2014 Douglas Gilbert */ -static int sg_version_num =30536 ;/* 2 digits for each component */ -#define SG_VERSION_STR "3.5.36" +static int sg_version_num =30537 ;/* 2 digits for each component */ +#define SG_VERSION_STR "3.5.37"/* * D. P. Gilbert ([email protected]), notes:
Quoted 2 lines
And bumping the version number is appropriate for an interface tweak/correction.
I wasn't (and still isn't) sure how much I should bump. Maybe 36000 / 3.6.0 will do? (Seems to deserve the bigger bump)
Quoted 4 lines
I'm rewriting the sg driver currently (12 months and counting) and am up to version 11 of the _first_ half. So far I'm using a
sg_version_numofQuoted 1 line
40000 for the rewritten code. Please keep away from version numbers
Quoted 8 lines
40000 and above.
The rewritten driver is documented here:
https://doug-gilbert.github.io/sg_v40.htmland its ioctls are listed in Table 8, including the BLK* ones. Perhaps you could suggest some corrections. Obviously BLKSSZGET needs to be added when your patches are accepted.
"this ioctl value replicates what a block layer device file (e.g. /dev/sda) will do with the same value. It calls the queue_max_sectors() helper on the owning device's command queue. The resulting number represents the maximum number of logical sectors of a single request that the block layer will accept."? and for BLKSSZGET "the resulting number represents the logical sector size"?
Quoted 99 lines
Doug Gilbert
Quoted 95 lines
I am not sure if
queue_max_sectors()or BLKSECTGET has ever been implemented in the block layer to give out the limit in bytes, but it certainly isn't the case anymore.I am not in position to say whether it's "right" or "wrong" to implement BLKSECTGET/BLKSSZGET in the sg driver, but they is definitely useful in some cases (as long as the queue_* functions work for the given underlying device). Is it not okay if they don't ultimately work on *some* devices, even when they aren't named SG_*?
Perhaps we can consider having them removed as well (and implement them as e.g.
SG_GET_MAX_SECTORSandSG_GET_LBS; but IMHO that only makes a point if they can be made to work on more than block devices).Quoted 55 lines · Douglas Gilbert
Quoted 37 lines · Bart Van Assche
Quoted 22 lines · Tom Yan
It should give out the maximum number of sectors per request instead of maximum number of bytes.
Signed-off-by Tom Yan <[email protected]>--- drivers/scsi/sg.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/scsi/sg.c +1
-0@@ -922,6 +922,7 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,int result, val, read_only;Sg_request *srp;unsigned long iflags; +unsigned int max_sectors;SCSI_LOG_TIMEOUT (3 ,sg_printk (KERN_INFO , sdp,"sg_ioctl: cmd=0x%x\n" , (int ) cmd_in));@@ -1114,8 +1115,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, sdp->sgdebug = (char ) val;drivers/scsi/sg.c sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp, · line 1123
Quoted 10 lines
drivers/scsi/sg.c +3
-2return 0 ;case BLKSECTGET : -return put_user (max_sectors_bytes (sdp->device->request_queue), - ip); + max_sectors = min_t(unsigned int, USHRT_MAX, +queue_max_sectors (sdp->device->request_queue)); +return put_user (max_sectors, ip);case BLKTRACESETUP : return blk_trace_setup(sdp->device->request_queue, sdp->disk->disk_name,Is this perhaps a backwards-incompatible change to the kernel ABI, the kind of change Linus totally disagrees with?
Quoted 7 lines
Additionally, please Cc linux-api for patches that modify the kernel ABI.
Quoted 1 line
From https://www.kernel.org/doc/man-pages/linux-api-ml.html: "The kernel
source file
Documentation/SubmitChecklistnotes that all Linux kernel patches that change userspace interfaces should be CCed to [email protected], so that the various parties who are interested in API changes are informed. For further information, see
https://www.kernel.org/doc/man-pages/linux-api-ml.html"Hmm,
The BLK* ioctl()s in the sg driver were an undocumented addition by others. Plus it is not clear to me why a char device such as the sg driver should be supporting BLK* ioctl(2)s. For example, how should an enclosure react to those ioctls or a WLUN?If they are going to be supported for storage devices and
/dev/sdband/dev/sg2are the same device then if:Quoted 7 lines
blockdev --getmaxsect /dev/sg1gives a different result to:
blockdev --getmaxsect /dev/sdbthen I would consider that a bug. So fixing it is making the kernel ABI more consistent :-)
That's exactly my point. They should work identically as the ones implemented in the block layer, because of their names.
With that said,
sg_versionneeds to be bumped once the fix gets in, so that there's a way to differentiate the "different implementations" in userspace.Quoted 5 lines
Doug Gilbert
Regards,
Tom