From d082dc1562a2ff0947b214796f12faaa87e816a9 Mon Sep 17 00:00:00 2001 From: James Smart Date: Mon, 16 Jul 2018 14:38:14 -0700 Subject: [PATCH 1/4] nvmet-fc: fix target sgl list on large transfers The existing code to carve up the sg list expected an sg element-per-page which can be very incorrect with iommu's remapping multiple memory pages to fewer bus addresses. To hit this error required a large io payload (greater than 256k) and a system that maps on a per-page basis. It's possible that large ios could get by fine if the system condensed the sgl list into the first 64 elements. This patch corrects the sg list handling by specifically walking the sg list element by element and attempting to divide the transfer up on a per-sg element boundary. While doing so, it still tries to keep sequences under 256k, but will exceed that rule if a single sg element is larger than 256k. Fixes: 48fa362b6c3f ("nvmet-fc: simplify sg list handling") Cc: # 4.14 Signed-off-by: James Smart Signed-off-by: Christoph Hellwig --- drivers/nvme/target/fc.c | 44 ++++++++++++++++++++++++++++++++-------- 1 file changed, 35 insertions(+), 9 deletions(-) diff --git a/drivers/nvme/target/fc.c b/drivers/nvme/target/fc.c index 408279cb6f2c..29b4b236afd8 100644 --- a/drivers/nvme/target/fc.c +++ b/drivers/nvme/target/fc.c @@ -58,8 +58,8 @@ struct nvmet_fc_ls_iod { struct work_struct work; } __aligned(sizeof(unsigned long long)); +/* desired maximum for a single sequence - if sg list allows it */ #define NVMET_FC_MAX_SEQ_LENGTH (256 * 1024) -#define NVMET_FC_MAX_XFR_SGENTS (NVMET_FC_MAX_SEQ_LENGTH / PAGE_SIZE) enum nvmet_fcp_datadir { NVMET_FCP_NODATA, @@ -74,6 +74,7 @@ struct nvmet_fc_fcp_iod { struct nvme_fc_cmd_iu cmdiubuf; struct nvme_fc_ersp_iu rspiubuf; dma_addr_t rspdma; + struct scatterlist *next_sg; struct scatterlist *data_sg; int data_sg_cnt; u32 offset; @@ -1025,8 +1026,7 @@ nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo, INIT_LIST_HEAD(&newrec->assoc_list); kref_init(&newrec->ref); ida_init(&newrec->assoc_cnt); - newrec->max_sg_cnt = min_t(u32, NVMET_FC_MAX_XFR_SGENTS, - template->max_sgl_segments); + newrec->max_sg_cnt = template->max_sgl_segments; ret = nvmet_fc_alloc_ls_iodlist(newrec); if (ret) { @@ -1722,6 +1722,7 @@ nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod) ((fod->io_dir == NVMET_FCP_WRITE) ? DMA_FROM_DEVICE : DMA_TO_DEVICE)); /* note: write from initiator perspective */ + fod->next_sg = fod->data_sg; return 0; @@ -1866,24 +1867,49 @@ nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport, struct nvmet_fc_fcp_iod *fod, u8 op) { struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq; + struct scatterlist *sg = fod->next_sg; unsigned long flags; - u32 tlen; + u32 remaininglen = fod->req.transfer_len - fod->offset; + u32 tlen = 0; int ret; fcpreq->op = op; fcpreq->offset = fod->offset; fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC; - tlen = min_t(u32, tgtport->max_sg_cnt * PAGE_SIZE, - (fod->req.transfer_len - fod->offset)); + /* + * for next sequence: + * break at a sg element boundary + * attempt to keep sequence length capped at + * NVMET_FC_MAX_SEQ_LENGTH but allow sequence to + * be longer if a single sg element is larger + * than that amount. This is done to avoid creating + * a new sg list to use for the tgtport api. + */ + fcpreq->sg = sg; + fcpreq->sg_cnt = 0; + while (tlen < remaininglen && + fcpreq->sg_cnt < tgtport->max_sg_cnt && + tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) { + fcpreq->sg_cnt++; + tlen += sg_dma_len(sg); + sg = sg_next(sg); + } + if (tlen < remaininglen && fcpreq->sg_cnt == 0) { + fcpreq->sg_cnt++; + tlen += min_t(u32, sg_dma_len(sg), remaininglen); + sg = sg_next(sg); + } + if (tlen < remaininglen) + fod->next_sg = sg; + else + fod->next_sg = NULL; + fcpreq->transfer_length = tlen; fcpreq->transferred_length = 0; fcpreq->fcp_error = 0; fcpreq->rsplen = 0; - fcpreq->sg = &fod->data_sg[fod->offset / PAGE_SIZE]; - fcpreq->sg_cnt = DIV_ROUND_UP(tlen, PAGE_SIZE); - /* * If the last READDATA request: check if LLDD supports * combined xfr with response. From 6cdefc6e2ad52170f89a8d0e8b1a1339f91834dc Mon Sep 17 00:00:00 2001 From: James Smart Date: Fri, 20 Jul 2018 15:49:48 -0700 Subject: [PATCH 2/4] nvme: if_ready checks to fail io to deleting controller The revised if_ready checks skipped over the case of returning error when the controller is being deleted. Instead it was returning BUSY, which caused the ios to retry, which caused the ns delete to hang waiting for the ios to drain. Stack trace of hang looks like: kworker/u64:2 D 0 74 2 0x80000000 Workqueue: nvme-delete-wq nvme_delete_ctrl_work [nvme_core] Call Trace: ? __schedule+0x26d/0x820 schedule+0x32/0x80 blk_mq_freeze_queue_wait+0x36/0x80 ? remove_wait_queue+0x60/0x60 blk_cleanup_queue+0x72/0x160 nvme_ns_remove+0x106/0x140 [nvme_core] nvme_remove_namespaces+0x7e/0xa0 [nvme_core] nvme_delete_ctrl_work+0x4d/0x80 [nvme_core] process_one_work+0x160/0x350 worker_thread+0x1c3/0x3d0 kthread+0xf5/0x130 ? process_one_work+0x350/0x350 ? kthread_bind+0x10/0x10 ret_from_fork+0x1f/0x30 Extend nvmf_fail_nonready_command() to supply the controller pointer so that the controller state can be looked at. Fail any io to a controller that is deleting. Fixes: 3bc32bb1186c ("nvme-fabrics: refactor queue ready check") Fixes: 35897b920c8a ("nvme-fabrics: fix and refine state checks in __nvmf_check_ready") Signed-off-by: James Smart Signed-off-by: Christoph Hellwig Tested-by: Ewan D. Milne Reviewed-by: Ewan D. Milne --- drivers/nvme/host/fabrics.c | 10 +++++++--- drivers/nvme/host/fabrics.h | 3 ++- drivers/nvme/host/fc.c | 2 +- drivers/nvme/host/rdma.c | 2 +- drivers/nvme/target/loop.c | 2 +- 5 files changed, 12 insertions(+), 7 deletions(-) diff --git a/drivers/nvme/host/fabrics.c b/drivers/nvme/host/fabrics.c index 903eb4545e26..f7efe5a58cc7 100644 --- a/drivers/nvme/host/fabrics.c +++ b/drivers/nvme/host/fabrics.c @@ -539,14 +539,18 @@ static struct nvmf_transport_ops *nvmf_lookup_transport( /* * For something we're not in a state to send to the device the default action * is to busy it and retry it after the controller state is recovered. However, - * anything marked for failfast or nvme multipath is immediately failed. + * if the controller is deleting or if anything is marked for failfast or + * nvme multipath it is immediately failed. * * Note: commands used to initialize the controller will be marked for failfast. * Note: nvme cli/ioctl commands are marked for failfast. */ -blk_status_t nvmf_fail_nonready_command(struct request *rq) +blk_status_t nvmf_fail_nonready_command(struct nvme_ctrl *ctrl, + struct request *rq) { - if (!blk_noretry_request(rq) && !(rq->cmd_flags & REQ_NVME_MPATH)) + if (ctrl->state != NVME_CTRL_DELETING && + ctrl->state != NVME_CTRL_DEAD && + !blk_noretry_request(rq) && !(rq->cmd_flags & REQ_NVME_MPATH)) return BLK_STS_RESOURCE; nvme_req(rq)->status = NVME_SC_ABORT_REQ; return BLK_STS_IOERR; diff --git a/drivers/nvme/host/fabrics.h b/drivers/nvme/host/fabrics.h index e1818a27aa2d..aa2fdb2a2e8f 100644 --- a/drivers/nvme/host/fabrics.h +++ b/drivers/nvme/host/fabrics.h @@ -162,7 +162,8 @@ void nvmf_unregister_transport(struct nvmf_transport_ops *ops); void nvmf_free_options(struct nvmf_ctrl_options *opts); int nvmf_get_address(struct nvme_ctrl *ctrl, char *buf, int size); bool nvmf_should_reconnect(struct nvme_ctrl *ctrl); -blk_status_t nvmf_fail_nonready_command(struct request *rq); +blk_status_t nvmf_fail_nonready_command(struct nvme_ctrl *ctrl, + struct request *rq); bool __nvmf_check_ready(struct nvme_ctrl *ctrl, struct request *rq, bool queue_live); diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c index 41d45a1b5c62..9bac912173ba 100644 --- a/drivers/nvme/host/fc.c +++ b/drivers/nvme/host/fc.c @@ -2272,7 +2272,7 @@ nvme_fc_queue_rq(struct blk_mq_hw_ctx *hctx, if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE || !nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready)) - return nvmf_fail_nonready_command(rq); + return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq); ret = nvme_setup_cmd(ns, rq, sqe); if (ret) diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c index 518c5b09038c..66ec5985c9f3 100644 --- a/drivers/nvme/host/rdma.c +++ b/drivers/nvme/host/rdma.c @@ -1639,7 +1639,7 @@ static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx, WARN_ON_ONCE(rq->tag < 0); if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready)) - return nvmf_fail_nonready_command(rq); + return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq); dev = queue->device->dev; ib_dma_sync_single_for_cpu(dev, sqe->dma, diff --git a/drivers/nvme/target/loop.c b/drivers/nvme/target/loop.c index d8d91f04bd7e..ae7586b8be07 100644 --- a/drivers/nvme/target/loop.c +++ b/drivers/nvme/target/loop.c @@ -162,7 +162,7 @@ static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx, blk_status_t ret; if (!nvmf_check_ready(&queue->ctrl->ctrl, req, queue_ready)) - return nvmf_fail_nonready_command(req); + return nvmf_fail_nonready_command(&queue->ctrl->ctrl, req); ret = nvme_setup_cmd(ns, req, &iod->cmd); if (ret) From 5613d31214eb4c5c04cdfce4966bb661c8b43191 Mon Sep 17 00:00:00 2001 From: Hannes Reinecke Date: Wed, 25 Jul 2018 08:35:17 +0200 Subject: [PATCH 3/4] nvmet: fixup crash on NULL device path When writing an empty string into the device_path attribute the kernel will crash with nvmet: failed to open block device (null): (-22) BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 This patch sanitizes the error handling for invalid device path settings. Fixes: a07b4970 ("nvmet: add a generic NVMe target") Signed-off-by: Hannes Reinecke Reviewed-by: Chaitanya Kulkarni Signed-off-by: Christoph Hellwig --- drivers/nvme/target/configfs.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/drivers/nvme/target/configfs.c b/drivers/nvme/target/configfs.c index d3f3b3ec4d1a..ebea1373d1b7 100644 --- a/drivers/nvme/target/configfs.c +++ b/drivers/nvme/target/configfs.c @@ -282,6 +282,7 @@ static ssize_t nvmet_ns_device_path_store(struct config_item *item, { struct nvmet_ns *ns = to_nvmet_ns(item); struct nvmet_subsys *subsys = ns->subsys; + size_t len; int ret; mutex_lock(&subsys->lock); @@ -289,10 +290,14 @@ static ssize_t nvmet_ns_device_path_store(struct config_item *item, if (ns->enabled) goto out_unlock; - kfree(ns->device_path); + ret = -EINVAL; + len = strcspn(page, "\n"); + if (!len) + goto out_unlock; + kfree(ns->device_path); ret = -ENOMEM; - ns->device_path = kstrndup(page, strcspn(page, "\n"), GFP_KERNEL); + ns->device_path = kstrndup(page, len, GFP_KERNEL); if (!ns->device_path) goto out_unlock; From 405a7519607e7a364114896264440c0f87b325c0 Mon Sep 17 00:00:00 2001 From: Hannes Reinecke Date: Wed, 25 Jul 2018 08:34:45 +0200 Subject: [PATCH 4/4] nvmet: only check for filebacking on -ENOTBLK We only need to check for a file-backed namespace if nvmet_bdev_ns_enable() returns -ENOTBLK. For any other error it's pointless as the open() error will remain the same. Fixes: d5eff33e ("nvmet: add simple file backed ns support") Signed-off-by: Hannes Reinecke Reviewed-by: Chaitanya Kulkarni Signed-off-by: Christoph Hellwig --- drivers/nvme/target/core.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/nvme/target/core.c b/drivers/nvme/target/core.c index 74d4b785d2da..9838103f2d62 100644 --- a/drivers/nvme/target/core.c +++ b/drivers/nvme/target/core.c @@ -339,7 +339,7 @@ int nvmet_ns_enable(struct nvmet_ns *ns) goto out_unlock; ret = nvmet_bdev_ns_enable(ns); - if (ret) + if (ret == -ENOTBLK) ret = nvmet_file_ns_enable(ns); if (ret) goto out_unlock;