From: James Smart james.smart@broadcom.com
commit 62e3a931db60daf94fdb3159d685a5bc6ad4d0cf upstream.
The following calltrace was seen:
BUG: sleeping function called from invalid context at mm/slab.h:494 ... Call Trace: dump_stack+0x9a/0xf0 ___might_sleep.cold.63+0x13d/0x178 slab_pre_alloc_hook+0x6a/0x90 kmem_cache_alloc_trace+0x3a/0x2d0 lpfc_sli4_nvmet_alloc+0x4c/0x280 [lpfc] lpfc_post_rq_buffer+0x2e7/0xa60 [lpfc] lpfc_sli4_hba_setup+0x6b4c/0xa4b0 [lpfc] lpfc_pci_probe_one_s4.isra.15+0x14f8/0x2280 [lpfc] lpfc_pci_probe_one+0x260/0x2880 [lpfc] local_pci_probe+0xd4/0x180 work_for_cpu_fn+0x51/0xa0 process_one_work+0x8f0/0x17b0 worker_thread+0x536/0xb50 kthread+0x30c/0x3d0 ret_from_fork+0x3a/0x50
A prior patch introduced a spin_lock_irqsave(hbalock) in the lpfc_post_rq_buffer() routine. Call trace is seen as the hbalock is held with interrupts disabled during a GFP_KERNEL allocation in lpfc_sli4_nvmet_alloc().
Fix by reordering locking so that hbalock not held when calling sli4_nvmet_alloc() (aka rqb_buf_list()).
Link: https://lore.kernel.org/r/20201020202719.54726-2-james.smart@broadcom.com Fixes: 411de511c694 ("scsi: lpfc: Fix RQ empty firmware trap") Cc: stable@vger.kernel.org # v4.17+ Co-developed-by: Dick Kennedy dick.kennedy@broadcom.com Signed-off-by: Dick Kennedy dick.kennedy@broadcom.com Signed-off-by: James Smart james.smart@broadcom.com Signed-off-by: Martin K. Petersen martin.petersen@oracle.com Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org
--- drivers/scsi/lpfc/lpfc_mem.c | 4 +--- drivers/scsi/lpfc/lpfc_sli.c | 10 ++++++++-- 2 files changed, 9 insertions(+), 5 deletions(-)
--- a/drivers/scsi/lpfc/lpfc_mem.c +++ b/drivers/scsi/lpfc/lpfc_mem.c @@ -588,8 +588,6 @@ lpfc_sli4_rb_free(struct lpfc_hba *phba, * Description: Allocates a DMA-mapped receive buffer from the lpfc_hrb_pool PCI * pool along a non-DMA-mapped container for it. * - * Notes: Not interrupt-safe. Must be called with no locks held. - * * Returns: * pointer to HBQ on success * NULL on failure @@ -599,7 +597,7 @@ lpfc_sli4_nvmet_alloc(struct lpfc_hba *p { struct rqb_dmabuf *dma_buf;
- dma_buf = kzalloc(sizeof(struct rqb_dmabuf), GFP_KERNEL); + dma_buf = kzalloc(sizeof(*dma_buf), GFP_KERNEL); if (!dma_buf) return NULL;
--- a/drivers/scsi/lpfc/lpfc_sli.c +++ b/drivers/scsi/lpfc/lpfc_sli.c @@ -7248,12 +7248,16 @@ lpfc_post_rq_buffer(struct lpfc_hba *phb struct rqb_dmabuf *rqb_buffer; LIST_HEAD(rqb_buf_list);
- spin_lock_irqsave(&phba->hbalock, flags); rqbp = hrq->rqbp; for (i = 0; i < count; i++) { + spin_lock_irqsave(&phba->hbalock, flags); /* IF RQ is already full, don't bother */ - if (rqbp->buffer_count + i >= rqbp->entry_count - 1) + if (rqbp->buffer_count + i >= rqbp->entry_count - 1) { + spin_unlock_irqrestore(&phba->hbalock, flags); break; + } + spin_unlock_irqrestore(&phba->hbalock, flags); + rqb_buffer = rqbp->rqb_alloc_buffer(phba); if (!rqb_buffer) break; @@ -7262,6 +7266,8 @@ lpfc_post_rq_buffer(struct lpfc_hba *phb rqb_buffer->idx = idx; list_add_tail(&rqb_buffer->hbuf.list, &rqb_buf_list); } + + spin_lock_irqsave(&phba->hbalock, flags); while (!list_empty(&rqb_buf_list)) { list_remove_head(&rqb_buf_list, rqb_buffer, struct rqb_dmabuf, hbuf.list);