scsi_lib: make sure scsi_request.sense valid
Subsystems (2)
PATCH 1/1 scsi_lib: make sure scsi_request.sense valid
The block layer assumes scsi_request:sense is always a valid pointer. This is set up once in scsi_mq_init_request() and the containing scsi_cmnd object is used often, being re-initialized by scsi_init_command(). That works unless some code re-purposes part of the scsi_cmnd object for something else. And that is what bidi handling does in scsi_mq_prep_fn(). The result is an oops at some later time when the partly overwritten object is re-used. The overwrite is from d285203cf647d but 'git blame' does not show removed code, so that commit may not be the culprit.
---
This was found while injecting errors (thus generating sense data) into a sequence of bidi commands. At some later time the block layer blew up with a scsi_request::sense NULL dereference in sg_rq_end_io(). Without testing I'm confident the bsg driver, the osd ULD and exofs are exposed to this bug.
drivers/scsi/scsi_lib.c | 1 + 1 file changed, 1 insertion(+)
drivers/scsi/scsi_lib.c +1 -0
@@ -1175,6 +1175,7 @@ void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd) cmd->device = dev; cmd->sense_buffer = buf; + cmd->req.sense = buf; cmd->prot_sdb = prot; cmd->flags = flags;INIT_DELAYED_WORK (&cmd->abort_work, scmd_eh_abort_handler);
Signature · 1 line
2.17.1
Quoted 28 lines · Douglas Gilbert
The block layer assumes scsi_request:sense is always a valid pointer. This is set up once in
scsi_mq_init_request()and the containingscsi_cmndobject is used often, being re-initialized byscsi_init_command(). That works unless some code re-purposes part of thescsi_cmndobject for something else. And that is what bidi handling does inscsi_mq_prep_fn(). The result is an oops at some later time when the partly overwritten object is re-used. The overwrite is fromd285203cf647dbut 'git blame' does not show removed code, so that commit may not be the culprit.Signed-off-by Douglas Gilbert <[email protected]>>---This was found while injecting errors (thus generating sense data) into a sequence of bidi commands. At some later time the block layer blew up with a
scsi_request::senseNULL dereference insg_rq_end_io(). Without testing I'm confident the bsg driver, the osd ULD and exofs are exposed to this bug.
drivers/scsi/scsi_lib.c|1 +-1 file changed, 1 insertion(+-)drivers/scsi/scsi_lib.c +1
-1--- a/drivers/scsi/scsi_lib.c +-+-+- b/drivers/scsi/scsi_lib.c@@-1175,6 +-1175,7 @@void
scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
Quoted 1 line
cmd->device =dev;> cmd->sense_buffer =buf;>+- cmd->req.sense =buf;> cmd->prot_sdb =prot;> cmd->flags =flags;> INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
Hi Doug,
The description of this patch does not look correct to me. scsi_init_command() does not overwrite the sense pointer. From the body of that function:
/*zero out the cmd, except for the embedded scsi_request */ memset((char *cmd +- sizeof(cmd->req), 0, sizeof(*cmd) - sizeof(cmd->req) +- dev->host->hostt->cmd_size);
It is not clear to me which code overwrites the sense pointer. I think that needs to be figured out before discussion of this patch can continue.
Thanks,
Bart.
Quoted 42 lines · Bart Van Assche
Quoted 28 lines · Douglas Gilbert
The block layer assumes scsi_request:sense is always a valid pointer. This is set up once in
scsi_mq_init_request()and the containingscsi_cmndobject is used often, being re-initialized byscsi_init_command(). That works unless some code re-purposes part of thescsi_cmndobject for something else. And that is what bidi handling does inscsi_mq_prep_fn(). The result is an oops at some later time when the partly overwritten object is re-used. The overwrite is fromd285203cf647dbut 'git blame' does not show removed code, so that commit may not be the culprit.Signed-off-by Douglas Gilbert <[email protected]>---
This was found while injecting errors (thus generating sense data) into a sequence of bidi commands. At some later time the block layer blew up with a
scsi_request::senseNULL dereference insg_rq_end_io(). Without testing I'm confident the bsg driver, the osd ULD and exofs are exposed to this bug.drivers/scsi/scsi_lib.c | 1 + 1 file changed, 1 insertion(+)drivers/scsi/scsi_lib.c +0
-0drivers/scsi/scsi_lib.c void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd) · line 1181
Quoted 7 lines
drivers/scsi/scsi_lib.c +1
-0@@ -1175,6 +1175,7 @@ void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd) cmd->device = dev; cmd->sense_buffer = buf; + cmd->req.sense = buf; cmd->prot_sdb = prot; cmd->flags = flags;INIT_DELAYED_WORK (&cmd->abort_work, scmd_eh_abort_handler);Hi Doug,
The description of this patch does not look correct to me.
scsi_init_command()does not overwrite the sense pointer. From the body of that function:
Quoted 6 lines
/* zero out the cmd, except for the embedded scsi_request */ memset((char *)cmd + sizeof(cmd->req), 0, sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);It is not clear to me which code overwrites the sense pointer. I think that needs to be figured out before discussion of this patch can continue.
Bart,
Please re-read the explanation. The two sentences in the middle should tell you that you are looking at the wrong memset().
And I'm waiting for the person who wrote the questionable code to comment.
If you don't believe me, try sending a device reset to a scsi_debug device. Then use sg_raw *** to send a XDWRITREAD(10) bidi command to the same scsi_debug device, you should get a Unit Attention concerning that device reset. If you do, keep sending that bidi command. Make sure you have no important files open because that machine will lock solid. Basically what happens when a rq_end_io() callback de-references a NULL pointer. It looks like it has been there since 2014 and took me 2 days to find. Sorry that I can't explain it in one simple sentence.
Douglas
*** Use 'man sg_raw' and see the EXAMPLES section (second to last example)
You can use a bsg device as shown.
Quoted 49 lines · Douglas Gilbert
Quoted 32 lines · Bart Van Assche
Quoted 28 lines · Douglas Gilbert
The block layer assumes scsi_request:sense is always a valid pointer. This is set up once in
scsi_mq_init_request()and the containingscsi_cmndobject is used often, being re-initialized byscsi_init_command(). That works unless some code re-purposes part of thescsi_cmndobject for something else. And that is what bidi handling does inscsi_mq_prep_fn(). The result is an oops at some later time when the partly overwritten object is re-used. The overwrite is fromd285203cf647dbut 'git blame' does not show removed code, so that commit may not be the culprit.Signed-off-by Douglas Gilbert <[email protected]>>>>---This was found while injecting errors (thus generating sense data) into a sequence of bidi commands. At some later time the block layer blew up with a
scsi_request::senseNULL dereference insg_rq_end_io(). Without testing I'm confident the bsg driver, the osd ULD and exofs are exposed to this bug.drivers/scsi/scsi_lib.c |1 +- 1 file changed, 1 insertion(+-)drivers/scsi/scsi_lib.c +1
-1--- a/drivers/scsi/scsi_lib.c +-+-+- b/drivers/scsi/scsi_lib.c@@-1175,6 +-1175,7 @@void
scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)Quoted 1 line
cmd->device=dev;>>>cmd->sense_buffer=buf;>>>+-cmd->req.sense=buf;>>>cmd->prot_sdb=prot;>>>cmd->flags=flags;>>>INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);>>Hi Doug,
Quoted 8 lines
The description of this patch does not look correct to me.
scsi_init_command()does not overwrite the sense pointer. From the body of that function:/*zero out the cmd, except for the embedded scsi_request */ memset((char *cmd +- sizeof(cmd->req), 0, sizeof(*cmd) - sizeof(cmd->req) +- dev->host->hostt->cmd_size);>>It is not clear to me which code overwrites the sense pointer. I think that needs to be figured out before discussion of this patch can continue.
Bart,
Please re-read the explanation. The two sentences in the middle should tell you that you are looking at the wrongmemset().And I'm waiting for the person who wrote the questionable code to comment.
Quoted 8 lines
If you don't believe me, try sending a device reset to a
scsi_debugdevice. Then usesg_raw***to send a XDWRITREAD(10) bidi command to the samescsi_debugdevice, you should get a Unit Attention concerning that device reset. If you do, keep sending that bidi command. Make sure you have no important files open because that machine will lock solid. Basically what happens when arq_end_io()callback de-references a NULL pointer. It looks like it has been there since 2014 and took me 2 days to find. Sorry that I can't explain it in one simple sentence.
Hi Doug,
Is this perhaps the memset you are referring to?
memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
Bart.