From: James Smart jsmart2021@gmail.com
commit a4c21acca2be6729ecbe72eda9b08092725b0a77 upstream.
Many of the sgl-per-hdwq paths are locking with spin_lock_irq() and spin_unlock_irq() and may unwittingly raising irq when it shouldn't. Hard deadlocks were seen around lpfc_scsi_prep_cmnd().
Fix by converting the locks to irqsave/irqrestore.
Fixes: d79c9e9d4b3d ("scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.") Link: https://lore.kernel.org/r/20190922035906.10977-16-jsmart2021@gmail.com Signed-off-by: Dick Kennedy dick.kennedy@broadcom.com Signed-off-by: James Smart jsmart2021@gmail.com Signed-off-by: Martin K. Petersen martin.petersen@oracle.com Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org
--- drivers/scsi/lpfc/lpfc_sli.c | 38 ++++++++++++++++++++++---------------- 1 file changed, 22 insertions(+), 16 deletions(-)
--- a/drivers/scsi/lpfc/lpfc_sli.c +++ b/drivers/scsi/lpfc/lpfc_sli.c @@ -20414,8 +20414,9 @@ lpfc_get_sgl_per_hdwq(struct lpfc_hba *p struct sli4_hybrid_sgl *allocated_sgl = NULL; struct lpfc_sli4_hdw_queue *hdwq = lpfc_buf->hdwq; struct list_head *buf_list = &hdwq->sgl_list; + unsigned long iflags;
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags);
if (likely(!list_empty(buf_list))) { /* break off 1 chunk from the sgl_list */ @@ -20427,7 +20428,7 @@ lpfc_get_sgl_per_hdwq(struct lpfc_hba *p } } else { /* allocate more */ - spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags); tmp = kmalloc_node(sizeof(*tmp), GFP_ATOMIC, cpu_to_node(smp_processor_id())); if (!tmp) { @@ -20449,7 +20450,7 @@ lpfc_get_sgl_per_hdwq(struct lpfc_hba *p return NULL; }
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags); list_add_tail(&tmp->list_node, &lpfc_buf->dma_sgl_xtra_list); }
@@ -20457,7 +20458,7 @@ lpfc_get_sgl_per_hdwq(struct lpfc_hba *p struct sli4_hybrid_sgl, list_node);
- spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags);
return allocated_sgl; } @@ -20481,8 +20482,9 @@ lpfc_put_sgl_per_hdwq(struct lpfc_hba *p struct sli4_hybrid_sgl *tmp = NULL; struct lpfc_sli4_hdw_queue *hdwq = lpfc_buf->hdwq; struct list_head *buf_list = &hdwq->sgl_list; + unsigned long iflags;
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags);
if (likely(!list_empty(&lpfc_buf->dma_sgl_xtra_list))) { list_for_each_entry_safe(list_entry, tmp, @@ -20495,7 +20497,7 @@ lpfc_put_sgl_per_hdwq(struct lpfc_hba *p rc = -EINVAL; }
- spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags); return rc; }
@@ -20516,8 +20518,9 @@ lpfc_free_sgl_per_hdwq(struct lpfc_hba * struct list_head *buf_list = &hdwq->sgl_list; struct sli4_hybrid_sgl *list_entry = NULL; struct sli4_hybrid_sgl *tmp = NULL; + unsigned long iflags;
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags);
/* Free sgl pool */ list_for_each_entry_safe(list_entry, tmp, @@ -20529,7 +20532,7 @@ lpfc_free_sgl_per_hdwq(struct lpfc_hba * kfree(list_entry); }
- spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags); }
/** @@ -20553,8 +20556,9 @@ lpfc_get_cmd_rsp_buf_per_hdwq(struct lpf struct fcp_cmd_rsp_buf *allocated_buf = NULL; struct lpfc_sli4_hdw_queue *hdwq = lpfc_buf->hdwq; struct list_head *buf_list = &hdwq->cmd_rsp_buf_list; + unsigned long iflags;
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags);
if (likely(!list_empty(buf_list))) { /* break off 1 chunk from the list */ @@ -20567,7 +20571,7 @@ lpfc_get_cmd_rsp_buf_per_hdwq(struct lpf } } else { /* allocate more */ - spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags); tmp = kmalloc_node(sizeof(*tmp), GFP_ATOMIC, cpu_to_node(smp_processor_id())); if (!tmp) { @@ -20594,7 +20598,7 @@ lpfc_get_cmd_rsp_buf_per_hdwq(struct lpf tmp->fcp_rsp = (struct fcp_rsp *)((uint8_t *)tmp->fcp_cmnd + sizeof(struct fcp_cmnd));
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags); list_add_tail(&tmp->list_node, &lpfc_buf->dma_cmd_rsp_list); }
@@ -20602,7 +20606,7 @@ lpfc_get_cmd_rsp_buf_per_hdwq(struct lpf struct fcp_cmd_rsp_buf, list_node);
- spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags);
return allocated_buf; } @@ -20627,8 +20631,9 @@ lpfc_put_cmd_rsp_buf_per_hdwq(struct lpf struct fcp_cmd_rsp_buf *tmp = NULL; struct lpfc_sli4_hdw_queue *hdwq = lpfc_buf->hdwq; struct list_head *buf_list = &hdwq->cmd_rsp_buf_list; + unsigned long iflags;
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags);
if (likely(!list_empty(&lpfc_buf->dma_cmd_rsp_list))) { list_for_each_entry_safe(list_entry, tmp, @@ -20641,7 +20646,7 @@ lpfc_put_cmd_rsp_buf_per_hdwq(struct lpf rc = -EINVAL; }
- spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags); return rc; }
@@ -20662,8 +20667,9 @@ lpfc_free_cmd_rsp_buf_per_hdwq(struct lp struct list_head *buf_list = &hdwq->cmd_rsp_buf_list; struct fcp_cmd_rsp_buf *list_entry = NULL; struct fcp_cmd_rsp_buf *tmp = NULL; + unsigned long iflags;
- spin_lock_irq(&hdwq->hdwq_lock); + spin_lock_irqsave(&hdwq->hdwq_lock, iflags);
/* Free cmd_rsp buf pool */ list_for_each_entry_safe(list_entry, tmp, @@ -20676,5 +20682,5 @@ lpfc_free_cmd_rsp_buf_per_hdwq(struct lp kfree(list_entry); }
- spin_unlock_irq(&hdwq->hdwq_lock); + spin_unlock_irqrestore(&hdwq->hdwq_lock, iflags); }