scsi: lpfc: fix double free bug in lpfc_bsg_write_ebuf_set

When error occurs, it frees dmabuf in both lpfc_bsg_write_ebuf_set
and lpfc_bsg_issue_mbox.

Fix it by removing free code in lpfc_bsg_write_ebuf_set.

Reported-by: Zheng Wang <hackerzheng666@gmail.com>
Reported-by: Zhuorao Yang <alex000young@gmail.com>

Fixes: 7ad20aa9d39a ("[SCSI] lpfc 8.3.24: Extend BSG infrastructure and add link diagnostics")

Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
---
 drivers/scsi/lpfc/lpfc_bsg.c | 17 +++--------------
 1 file changed, 3 insertions(+), 14 deletions(-)

Friendly ping :)

On 10/27/2022 10:07 PM, Zheng Wang wrote:
> When error occurs, it frees dmabuf in both lpfc_bsg_write_ebuf_set
> and lpfc_bsg_issue_mbox.
> 
> Fix it by removing free code in lpfc_bsg_write_ebuf_set.
> 
> Reported-by: Zheng Wang <hackerzheng666@gmail.com>
> Reported-by: Zhuorao Yang <alex000young@gmail.com>
> 
> Fixes: 7ad20aa9d39a ("[SCSI] lpfc 8.3.24: Extend BSG infrastructure and add link diagnostics")
> 
> Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
> ---
>   drivers/scsi/lpfc/lpfc_bsg.c | 17 +++--------------
>   1 file changed, 3 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/scsi/lpfc/lpfc_bsg.c b/drivers/scsi/lpfc/lpfc_bsg.c
> index ac0c7ccf2eae..7362d9c1a50b 100644
> --- a/drivers/scsi/lpfc/lpfc_bsg.c
> +++ b/drivers/scsi/lpfc/lpfc_bsg.c
> @@ -4439,15 +4439,13 @@ lpfc_bsg_write_ebuf_set(struct lpfc_hba *phba, struct bsg_job *job,
>   
>   		dd_data = kmalloc(sizeof(struct bsg_job_data), GFP_KERNEL);
>   		if (!dd_data) {
> -			rc = -ENOMEM;
> -			goto job_error;
> +			return -ENOMEM;
>   		}
>   
>   		/* mailbox command structure for base driver */
>   		pmboxq = mempool_alloc(phba->mbox_mem_pool, GFP_KERNEL);
>   		if (!pmboxq) {
> -			rc = -ENOMEM;
> -			goto job_error;
> +			return -ENOMEM;
>   		}
>   		memset(pmboxq, 0, sizeof(LPFC_MBOXQ_t));
>   		pbuf = (uint8_t *)phba->mbox_ext_buf_ctx.mbx_dmabuf->virt;

Minimally, just looking at this one snippet, by returning after the 
mempool_alloc() failure, we are leaking the dd_data memory just allocated.

> @@ -4480,8 +4478,7 @@ lpfc_bsg_write_ebuf_set(struct lpfc_hba *phba, struct bsg_job *job,
>   		lpfc_printf_log(phba, KERN_ERR, LOG_LIBDFC,
>   				"2970 Failed to issue SLI_CONFIG ext-buffer "
>   				"mailbox command, rc:x%x\n", rc);
> -		rc = -EPIPE;
> -		goto job_error;
> +		return -EPIPE;

and this leaks both the dd_data and pmboxq memory.

>   	}
>   
>   	/* wait for additional external buffers */
> @@ -4489,14 +4486,6 @@ lpfc_bsg_write_ebuf_set(struct lpfc_hba *phba, struct bsg_job *job,
>   	bsg_job_done(job, bsg_reply->result,
>   		       bsg_reply->reply_payload_rcv_len);
>   	return SLI_CONFIG_HANDLED;
> -
> -job_error:
> -	if (pmboxq)
> -		mempool_free(pmboxq, phba->mbox_mem_pool);
> -	lpfc_bsg_dma_page_free(phba, dmabuf);
> -	kfree(dd_data);
> -
> -	return rc;
>   }
>   
>   /**

all of these errors should cause:
   lpfc_bsg_write_ebuf_set() to return -Exxx
   causing lpfc_bsg_handle_sli_cfg_ebuf() to return -Exxx
   causing lpfc_bsg_handle_sli_cfg_ext() to return -Exxx
   which causes lpfc_bsg_issue_mbox() to jump to job_done

I understand the argument is that issue_mbox deletes them, but....

job_done:
   checks/frees pmboxq is allocated after the jump so it will be NULL
   frees dmabuf - which was allocated prior to the jump; is freed
      in freedlpfc_bsg_handle_sli_cfg_ebuf() but only in a block
      that returns SLI_CONFIG_HANDLED, which is not the block that
      invokes lpfc_bsg_write_ebuf_set. So it's valid to delete.
      Note: there's a special case for SLI_CONFIG_HANDLED which skips
      over these deletes so it's ok.
   frees dd_data - which is allocated after the jump so it too will
      be NULL

So - the code is fine.  The SLI_CONFIG_HANDLED is a little weird, but 
the logic is fine. If the patch were added it would leak memory.

I take it this was identified by some tool ?

-- james

James Smart <jsmart2021@gmail.com> 于2022年11月3日周四 00:37写道：

> Minimally, just looking at this one snippet, by returning after the
> mempool_alloc() failure, we are leaking the dd_data memory just allocated.
>

Yes, this is a bad patch.

> > @@ -4480,8 +4478,7 @@ lpfc_bsg_write_ebuf_set(struct lpfc_hba *phba, struct bsg_job *job,
> >               lpfc_printf_log(phba, KERN_ERR, LOG_LIBDFC,
> >                               "2970 Failed to issue SLI_CONFIG ext-buffer "
> >                               "mailbox command, rc:x%x\n", rc);
> > -             rc = -EPIPE;
> > -             goto job_error;
> > +             return -EPIPE;
>
> and this leaks both the dd_data and pmboxq memory.

So Here it is.

>
> all of these errors should cause:
>    lpfc_bsg_write_ebuf_set() to return -Exxx
>    causing lpfc_bsg_handle_sli_cfg_ebuf() to return -Exxx
>    causing lpfc_bsg_handle_sli_cfg_ext() to return -Exxx
>    which causes lpfc_bsg_issue_mbox() to jump to job_done
>

Hi James! Really apprecaite for your reply. I was not sure if it it a
really issue. Sorry for the bad patch.

> I understand the argument is that issue_mbox deletes them, but....
>
> job_done:
>    checks/frees pmboxq is allocated after the jump so it will be NULL
>    frees dmabuf - which was allocated prior to the jump; is freed
>       in freedlpfc_bsg_handle_sli_cfg_ebuf() but only in a block
>       that returns SLI_CONFIG_HANDLED, which is not the block that
>       invokes lpfc_bsg_write_ebuf_set. So it's valid to delete.
>       Note: there's a special case for SLI_CONFIG_HANDLED which skips
>       over these deletes so it's ok.
>    frees dd_data - which is allocated after the jump so it too will
>       be NULL

I understood your point. Here is a call chain :
 lpfc_bsg_issue_mbox->lpfc_bsg_handle_sli_cfg_ext->lpfc_bsg_handle_sli_cfg_ebuf->lpfc_bsg_write_ebuf_set->lpfc_bsg_dma_page_free->kfree(dmabuf)
It leads to another kfree in lpfc_bsg_mbox_cmd :
job_done:
    /* common exit for error or job completed inline */
    if (pmboxq)
        mempool_free(pmboxq, phba->mbox_mem_pool);
    [7] lpfc_bsg_dma_page_free(phba, dmabuf);
    kfree(dd_data);

So the key point is whether phba->mbox_ext_buf_ctx.mboxType can be
mbox_wr. If not, just as you illustrated, all is fine.
It will get into SLI_CONFIG_HANDLED path and handle dmabuf as expected.

But if not, it will have a chance to trigger a double-free bug. I
found phda is assigned in lpfc_bsg_mbox_cmd. But I am still not sure
about its value.
Appreciate if you can help me to understand more about the key condition :)

> So - the code is fine.  The SLI_CONFIG_HANDLED is a little weird, but
> the logic is fine. If the patch were added it would leak memory.
>
> I take it this was identified by some tool ?
>

Yes, I found it using Codeql. I didn't have a poc to verify.

Best Regards,
Zheng Wang