Re: [PATCH] sym53c8xx: fix bad memset argument insym_set_cam_result_error

[James Bottomley]

On Sun, 2008-01-27 at 01:18 +0100, Mariusz Kozlowski wrote:
> Hello,
> 
> > On a big powerpc box I got the following oops with 2.6.24-git2:
> > 
> > sym0: <1010-66> rev 0x1 at pci 0000:d0:01.0 irq 215
> > sym0: No NVRAM, ID 7, Fast-80, LVD, parity checking
> > sym0: SCSI BUS has been reset.
> > scsi0 : sym-2.2.3
> >  target0:0:8: FAST-40 WIDE SCSI 80.0 MB/s ST (25 ns, offset 31)
> > scsi 0:0:8:0: Direct-Access     IBM      ST318305LC       C509 PQ: 0
> > ANSI: 3
> >  target0:0:8: tagged command queuing enabled, command queue depth 16.
> >  target0:0:8: Beginning Domain Validation
> >  target0:0:8: asynchronous
> >  target0:0:8: wide asynchronous
> >  target0:0:8: FAST-80 WIDE SCSI 160.0 MB/s DT (12.5 ns, offset 31)
> >  target0:0:8: FAST-80 WIDE SCSI 160.0 MB/s DT (12.5 ns, offset 31)
> > Unable to handle kernel paging request for data at address 0x00000000
> > Faulting instruction address: 0xc000000000038460
> > cpu 0x25: Vector: 300 (Data Access) at [c00000000f567840]
> >     pc: c000000000038460: .memcpy+0x60/0x280
> >     lr: d000000000050280: .sym_set_cam_result_error+0xfc/0x1e0 [sym53c8xx]
> >     sp: c00000000f567ac0
> >    msr: 8000000000009032
> >    dar: 0
> >  dsisr: 42000000
> >   current = 0xc000006d1e0af0a0
> >   paca    = 0xc0000000004afc00
> >     pid   = 0, comm = swapper
> > enter ? for help
> > [link register   ] d000000000050280
> > .sym_set_cam_result_error+0xfc/0x1e0 [sym53c8xx]
> > [c00000000f567ac0] c00000000f567b80 (unreliable)
> > [c00000000f567b80] d0000000000552b8 .sym_complete_error+0x12c/0x1bc [sym53c8xx]
> > [c00000000f567c20] d0000000000561a4 .sym_int_sir+0xaa4/0x1718 [sym53c8xx]
> > [c00000000f567d00] d000000000057e8c .sym_interrupt+0x4e4/0x6ec [sym53c8xx]
> > [c00000000f567dc0] d00000000004fdf4 .sym53c8xx_intr+0x6c/0xdc [sym53c8xx]
> > [c00000000f567e50] c0000000000a83e0 .handle_IRQ_event+0x7c/0xec
> > [c00000000f567ef0] c0000000000aa344 .handle_fasteoi_irq+0x130/0x1f0
> > [c00000000f567f90] c00000000002a538 .call_handle_irq+0x1c/0x2c
> > [c000004d5e0b3a90] c00000000000c320 .do_IRQ+0x108/0x1d0
> > [c000004d5e0b3b20] c000000000004790 hardware_interrupt_entry+0x18/0x1c
> > 
> > The memset() in sym_set_cam_result_error() would appear to be trashing
> > the scsi_cmnd struct instead of clearing sense_buffer.
> > 
> > Signed-off-by: Nathan Lynch <ntl@xxxxxxxxx>
> > ---
> >  drivers/scsi/sym53c8xx_2/sym_glue.c |    2 +-
> >  1 files changed, 1 insertions(+), 1 deletions(-)
> > 
> > diff --git a/drivers/scsi/sym53c8xx_2/sym_glue.c b/drivers/scsi/sym53c8xx_2/sym_glue.c
> > index 21e926d..e939f38 100644
> > --- a/drivers/scsi/sym53c8xx_2/sym_glue.c
> > +++ b/drivers/scsi/sym53c8xx_2/sym_glue.c
> > @@ -207,7 +207,7 @@ void sym_set_cam_result_error(struct sym_hcb *np, struct sym_ccb *cp, int resid)
> >  			/*
> >  			 *  Bounce back the sense data to user.
> >  			 */
> > -			memset(&cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
> > +			memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
> >  			memcpy(cmd->sense_buffer, cp->sns_bbuf,
> >  			       min(SCSI_SENSE_BUFFERSIZE, SYM_SNS_BBUF_LEN));
> >  #if 0
> 
> I was looking into it and found something interesting. What you see was just exposed by:
> 
> commit b80ca4f7ee36c26d300c5a8f429e73372d153379
> Author: FUJITA Tomonori <tomof@xxxxxxx>
> Date:   Sun Jan 13 15:46:13 2008 +0900
> 
>     [SCSI] replace sizeof sense_buffer with  
>     
>     This replaces sizeof sense_buffer with SCSI_SENSE_BUFFERSIZE in
>     several LLDs. It's a preparation for the future changes to remove
>     sense_buffer array in scsi_cmnd structure.
>     
>     Signed-off-by: FUJITA Tomonori <fujita.tomonori@xxxxxxxxxxxxx>
>     Signed-off-by: James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx>
> 
> But this is not a simple replace sizeof sense_buffer with SCSI_SENSE_BUFFERSIZE
> It has a side effect. Take a look a this portion of the commit:
> 
> --- a/drivers/scsi/sym53c8xx_2/sym_glue.c
> +++ b/drivers/scsi/sym53c8xx_2/sym_glue.c
> @@ -207,10 +207,9 @@ void sym_set_cam_result_error(struct sym_hcb *np, struct sym_ccb *cp, int resid)
>                         /*
>                          *  Bounce back the sense data to user.
>                          */
> -                       memset(&cmd->sense_buffer, 0, sizeof(cmd->sense_buffer));
> +                       memset(&cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
>                         memcpy(cmd->sense_buffer, cp->sns_bbuf,
> -                             min(sizeof(cmd->sense_buffer),
> -                                 (size_t)SYM_SNS_BBUF_LEN));
> +                              min(SCSI_SENSE_BUFFERSIZE, SYM_SNS_BBUF_LEN));
> 
> Before the patch:
> 1) memset sets cmd->sense_buffer pointer value to zero (only 4 bytes which is wrong)
> 2) memcpy copies min(sizeof(cmd->sense_buffer), (size_t)SYM_SNS_BBUF_LEN)) which
>    is equal to min(4, 32) so after this only first four bytes were copied from 32 bytes
>    long cp->sns_bbuf into the sense_buffer

This logic is wrong.  There was a commit
(de25deb18016f66dcdede165d07654559bb332bc) that changed
scsi_cmnd.sense_buffer from a char [96] to a char *, which is actually
what caused this to break.  sizeof(char [96]) is 96 and for some reason
in C &x->y == &x->y[0] if char y[something].

> After the patch:
> 1) memset zeroes 96 bytes of cmd structure starting at offset where sense_buffer in
>    cmd is and results in an oops which you fixed
> 2) memcpy copies min(SCSI_SENSE_BUFFERSIZE, SYM_SNS_BBUF_LEN)) is equal to
>    min(96, 32). cp->sns_bbuf is 32 bytes so I guess that's ok but that's different from
>    the code before b80ca4f7 which was wrong (copied only 4 bytes) but 'worked'.
> 
> Someone more experienced with that code should take a look at the logic of it.

James


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/