nvme-pci: fix mempool alloc size

[ Upstream commit c89a529 ]

Convert the max size to bytes to match the units of the divisor that
calculates the worst-case number of PRP entries.

The result is used to determine how many PRP Lists are required. The
code was previously rounding this to 1 list, but we can require 2 in the
worst case. In that scenario, the driver would corrupt memory beyond the
size provided by the mempool.

While unlikely to occur (you'd need a 4MB in exactly 127 phys segments
on a queue that doesn't support SGLs), this memory corruption has been
observed by kfence.

Cc: Jens Axboe <axboe@kernel.dk>
Fixes: 943e942 ("nvme-pci: limit max IO size and segments to avoid high order allocations")
Signed-off-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Kanchan Joshi <joshi.k@samsung.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>

Author: keithbusch

drivers/nvme/host/pci.c

@@ -370,8 +370,8 @@ static bool nvme_dbbuf_update_and_check_event(u16 value, __le32 *dbbuf_db,
  */
 static int nvme_pci_npages_prp(void)
 {
-	unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
-				      NVME_CTRL_PAGE_SIZE);
+	unsigned max_bytes = (NVME_MAX_KB_SZ * 1024) + NVME_CTRL_PAGE_SIZE;
+	unsigned nprps = DIV_ROUND_UP(max_bytes, NVME_CTRL_PAGE_SIZE);
 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 }