Skip to contentSkip to conversation

nvme-pci: fix dma mapping leak on data setup error

PATCH In review Keith Busch 2 participants 2 messages May 19 – May 20 NVMe

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 nvme-pci: fix dma mapping leak on data setup error

From: Keith Busch <[email protected]>

We're leaking the initial DMA mapping during iteration if we fail to allocate the tracking descriptor for both PRP and SGL. Unmap the iterator directly; we can't use the existing unmap helper because it depends on the tracking descriptor being successfully allocated.

The mappings were also leaking when the driver detects an invalid bio_vec when mapping PRPs, so fix that too.

Fixes b8b7570a7ec87 ("nvme-pci: fix dma unmapping when using PRPs and not using the IOVA mapping")
Fixes 7ce3c1dd78fca ("nvme-pci: convert the data mapping to blk_rq_dma_map")

---
v1->v2:

Check the conditions that indicate we have to unmap. We know we don't have a coalesced iova in this path, so we can skip that, but need to check P2P and device dma requirements. And introduce a helper function for this so PRP and SGL don't duplicate the open coding for it.

Add fixes tags.

 drivers/nvme/host/pci.c | 22 ++++++++++++++++++++--
 1 file changed, 20 insertions(+), 2 deletions(-)
drivers/nvme/host/pci.c +20 -2
index 139a10cd687f9..c0de3b4afd072 100644
@@ -996,6 +996,18 @@ static bool nvme_pci_prp_iter_next(struct request *req, struct device *dma_dev,
 	return nvme_pci_prp_save_mapping(req, dma_dev, iter);
 }
 
+static void nvme_unmap_iter(struct request *req, struct blk_dma_iter *iter)
+{
+	struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
+	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
+
+	if (iod->flags & IOD_DATA_P2P || !dma_need_unmap(dev))
+		return;
+
+	dma_unmap_phys(nvmeq->dev->dev, iter->addr, iter->len, rq_dma_dir(req),
+		       iod->flags & IOD_DATA_MMIO ? DMA_ATTR_MMIO : 0);
+}
+
 static blk_status_t nvme_pci_setup_data_prp(struct request *req,
 		struct blk_dma_iter *iter)
 {
@@ -1006,8 +1018,10 @@ static blk_status_t nvme_pci_setup_data_prp(struct request *req,
 	unsigned int prp_len, i;
 	__le64 *prp_list;
 
-	if (!nvme_pci_prp_save_mapping(req, nvmeq->dev->dev, iter))
+	if (!nvme_pci_prp_save_mapping(req, nvmeq->dev->dev, iter)) {
+		nvme_unmap_iter(req, iter);
 		return iter->status;
+	}
 
 	/*
 	 * PRP1 always points to the start of the DMA transfers.
@@ -1112,6 +1126,7 @@ static blk_status_t nvme_pci_setup_data_prp(struct request *req,
 	dev_err_once(nvmeq->dev->dev,
 		"Incorrectly formed request for payload:%d nents:%d\n",
 		blk_rq_payload_bytes(req), blk_rq_nr_phys_segments(req));
+	nvme_unmap_data(req);
 	return BLK_STS_IOERR;
 }
 
@@ -1155,8 +1170,11 @@ static blk_status_t nvme_pci_setup_data_sgl(struct request *req,
 
 	sg_list = dma_pool_alloc(nvme_dma_pool(nvmeq, iod), GFP_ATOMIC,
 			&sgl_dma);
-	if (!sg_list)
+	if (!sg_list) {
+		nvme_unmap_iter(req, iter);
 		return BLK_STS_RESOURCE;
+	}
+
 	iod->descriptors[iod->nr_descriptors++] = sg_list;
 
 	do {
Signature · 1 line
2.53.0-Meta
maintainer
#1 lore raw
Quoted 4 lines · Keith Busch

We're leaking the initial DMA mapping during iteration if we fail to allocate the tracking descriptor for both PRP and SGL. Unmap the iterator directly; we can't use the existing unmap helper because it depends on the tracking descriptor being successfully allocated.

I hate to do another rapid respin here, but I gotta do it. The metadata sgl setup also can leak for the same reason, so need to close that one too. And there's also a leak for a P2P case that needs fixing. So v3 coming shortly.