Re: How to detect if a libata ioctl SG_IO had an error?

[Douglas Gilbert <dgilbert@interlog.com>]

On 12-06-07 02:46 AM, Norman Diamond wrote:
> My application wants to detect if an error occured, but in struct sg_io_hdr sg_io, sg_io.driver_status != DRIVER_OK does not say if an error occured, and (sg_io.info&  SG_INFO_OK_MASK) != SG_INFO_OK does not say if an error occured.
>
> In http://tldp.org/HOWTO/SCSI-Generic-HOWTO/x322.html:
> "A copy of these defines can be found in sg_err.h (see the utilities section):
> SG_ERR_DRIVER_OK [0x00] Typically no suggestion"
>
> Where's sg_err.h?  scsi.h defines DRIVER_OK but no one defines SG_ERR_DRIVER_OK.  Anyway, the real problem is that other values besides DRIVER_OK also mean driver ok.
>
> In http://tldp.org/HOWTO/SCSI-Generic-HOWTO/x364.html:
> "SG_INFO_CHECK [0x1] something abnormal happened."
>
> Sense data is abnormal but not an error, so I don't know if an error occured.
>
> In scsi.h:
> 443 #define DRIVER_OK       0x00    /* Driver status                           */
> 444
> 445 /*
> 446  *  These indicate the error that occurred, and what is available.
> 447  */
> 448
> 449 #define DRIVER_BUSY         0x01
> 450 #define DRIVER_SOFT         0x02
> 451 #define DRIVER_MEDIA        0x03
> 452 #define DRIVER_ERROR        0x04
> 453
> 454 #define DRIVER_INVALID      0x05
> 455 #define DRIVER_TIMEOUT      0x06
> 456 #define DRIVER_HARD         0x07
> 457 #define DRIVER_SENSE        0x08
>
> Is DRIVER_SENSE a bitwise value that can be or'ed with other values?
>
> In libata-scsi.c
> 523         if (driver_byte(cmd_result) == DRIVER_SENSE) {/* sense data available */
> 524                 u8 *desc = sensebuf + 8;
> 525                 cmd_result&= ~(0xFF<<24); /* DRIVER_SENSE is not an error */
>
> Line 523 does not test driver_byte(cmd_result)&  DRIVER_SENSE.  It looks like DRIVER_SENSE is a value all by itself, not bitwise.
>
> In sc.c:
> 463         if ((CHECK_CONDITION&  hp->masked_status) ||
> 464             (DRIVER_SENSE&  hp->driver_status))
>
> Line 464 does not test DRIVER_SENSE == hp->driver_status.  It looks like DRIVER_SENSE is bitwise, not a value all by itself.
>
> I sense inconsistency.
>
> In sc.c:
> 546         if (hp->masked_status || hp->host_status || hp->driver_status)
> 547                 hp->info |= SG_INFO_CHECK;
>
> So even though DRIVER_SENSE is not an error, the SG_INFO_CHECK bit gets set.  (The info field is undoubtedly bitwise.)
>
> How can I really test if an error occured?
>
> A bit more here.
>
> In scsi.h:
> 217 /*
> 218  *  SCSI Architecture Model (SAM) Status codes. Taken from SAM-3 draft
> 219  *  T10/1561-D Revision 4 Draft dated 7th November 2002.
> 220  */
> 221 #define SAM_STAT_GOOD            0x00
> 222 #define SAM_STAT_CHECK_CONDITION 0x02
> 223 #define SAM_STAT_CONDITION_MET   0x04
> 224 #define SAM_STAT_BUSY            0x08
> 225 #define SAM_STAT_INTERMEDIATE    0x10
> 226 #define SAM_STAT_INTERMEDIATE_CONDITION_MET 0x14
> 227 #define SAM_STAT_RESERVATION_CONFLICT 0x18
> 228 #define SAM_STAT_COMMAND_TERMINATED 0x22        /* obsolete in SAM-3 */
> 229 #define SAM_STAT_TASK_SET_FULL   0x28
> 230 #define SAM_STAT_ACA_ACTIVE      0x30
> 231 #define SAM_STAT_TASK_ABORTED    0x40
>
> Is SAM_STAT_CHECK_CONDITION a bitwise value?  It looks like the answer is mostly no.  Some of them could be, but not consistently.  Did the T10 committee say if these are bitwise?
>
> In scsi.h:
> 256 /*
> 257  *  Status codes. These are deprecated as they are shifted 1 bit right
> 258  *  from those found in the SCSI standards. This causes confusion for
> 259  *  applications that are ported to several OSes. Prefer SAM Status codes
> 260  *  above.
> 261  */
> 262
> 263 #define GOOD                 0x00
> 264 #define CHECK_CONDITION      0x01
>
> In sg.c:
> 533                 if ((CHECK_CONDITION&  hp->masked_status) ||
> 534                     (DRIVER_SENSE&  hp->driver_status)) {
>
> Why doesn't sg use SAM Status codes?
>
> Meanwhile, is CHECK_CONDITION a bitwise value?  It looks like the answer is mostly no, for the same reason as SAM_STAT_CHECK_CONDITION.  These bitwise tests look like bugs.

Please fetch the sg3_utils package and look at some user
space examples. If you are interested in libata then
following src/sg_sat_identify.c might be instructive. For
Linux, the pass-through code is in lib/sg_pt_linux.c .
Look at the get_scsi_pt_result_category() functions which
should answer your questions of how to check for sense
data. BTW the presence of sense data may indicate an error
or it may be just informative.

sg_pt_linux.c handles either of two SG_IO interfaces:
   - sg v3: as used by the sg driver and block devices
     (e.g. /dev/sdb)
   - sg v4: as used by the bsg driver which has device
     names like /dev/bsg/6:0:0:2

Doug Gilbert