On Tue, 2022-02-22 at 10:03 -0500, John Pittman wrote:
> When scsi_dma_map() fails by returning a sges_left value less than
> zero, the amount of logging can be extremely high. In a recent
> end-user environment, 1200 messages per second were being sent to
> the log buffer. This eventually overwhelmed the system and it
> stalled. As the messages are almost all identical, use
> pr_err_ratelimited() instead of sdev_printk() to print the
> scsi_dma_map failure messages.
>
> Signed-off-by: John Pittman <jpittman@xxxxxxxxxx>
> ---
> drivers/scsi/mpt3sas/mpt3sas_base.c | 15 ++++++---------
> 1 file changed, 6 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/scsi/mpt3sas/mpt3sas_base.c
> b/drivers/scsi/mpt3sas/mpt3sas_base.c
> index 511726f92d9a..ac9ccde6f3f8 100644
> --- a/drivers/scsi/mpt3sas/mpt3sas_base.c
> +++ b/drivers/scsi/mpt3sas/mpt3sas_base.c
> @@ -2594,9 +2594,8 @@ _base_check_pcie_native_sgl(struct
> MPT3SAS_ADAPTER *ioc,
> /* Get the SG list pointer and info. */
> sges_left = scsi_dma_map(scmd);
> if (sges_left < 0) {
> - sdev_printk(KERN_ERR, scmd->device,
> - "scsi_dma_map failed: request for %d bytes!\n",
> - scsi_bufflen(scmd));
> + pr_err_ratelimited("sd %s: scsi_dma_map failed: request
> for %d bytes!\n",
> + dev_name(&scmd->device->sdev_gendev),
> scsi_bufflen(scmd));
> return 1;
> }
>
> @@ -2706,9 +2705,8 @@ _base_build_sg_scmd(struct MPT3SAS_ADAPTER
> *ioc,
> sg_scmd = scsi_sglist(scmd);
> sges_left = scsi_dma_map(scmd);
> if (sges_left < 0) {
> - sdev_printk(KERN_ERR, scmd->device,
> - "scsi_dma_map failed: request for %d bytes!\n",
> - scsi_bufflen(scmd));
> + pr_err_ratelimited("sd %s: scsi_dma_map failed: request
> for %d bytes!\n",
> + dev_name(&scmd->device->sdev_gendev),
> scsi_bufflen(scmd));
> return -ENOMEM;
> }
>
> @@ -2854,9 +2852,8 @@ _base_build_sg_scmd_ieee(struct MPT3SAS_ADAPTER
> *ioc,
> sg_scmd = scsi_sglist(scmd);
> sges_left = scsi_dma_map(scmd);
> if (sges_left < 0) {
> - sdev_printk(KERN_ERR, scmd->device,
> - "scsi_dma_map failed: request for %d bytes!\n",
> - scsi_bufflen(scmd));
> + pr_err_ratelimited("sd %s: scsi_dma_map failed: request
> for %d bytes!\n",
> + dev_name(&scmd->device->sdev_gendev),
> scsi_bufflen(scmd));
> return -ENOMEM;
> }
>
This high message rate triggered a double completion at a customer and
this patch will help prevent the issue of the latency causing the race
window for the double completion. This was triaged by David Jeffery
Reviewed-by: Laurence Oberman <loberman@xxxxxxxxxx>
