[PATCH v2] nvmet-tcp: report a bounded MDTS instead of "no limit"

nvmet-tcp does not implement .get_mdts, so nvmet_ctrl_mdts() falls back
to the port value (0 by default) and identify-controller advertises "no
maximum data transfer size". The initiator believes it:
nvme_init_ctrl_finish() sets max_hw_sectors to UINT_MAX.

What the initiator then issues is decided by the block layer. The
generic cap is 4 MiB (BLK_DEF_MAX_SECTORS_CAP), but a namespace that
advertises a large NOWS raises it: blk_validate_limits() takes
max_sectors from io_opt once io_opt exceeds that cap. On the array we
measured, NOWS is 65535, so io_opt is 32 MiB and the initiator emitted
32 MiB commands -- from a controller that advertised no MDTS at all.

The target cannot serve those, in one of two ways depending on the
kernel.

Since 4a3f00262a04 ("nvmet-tcp: bound SGL data length before allocating
command buffers"), nvmet_tcp_map_data() rejects any len above
NVMET_TCP_MAXH2CDATA (4 MiB) with NVME_SC_SGL_INVALID_DATA | DNR, so the
command fails hard and deterministically. That is the right response to
an oversized command, but the initiator had no way to avoid sending it:
nothing advertised the limit it was exceeding.

Before that commit the command reached sgl_alloc(), which for 32 MiB
needs 8192 scatterlist entries -- 256 KiB, an order-6 kmalloc. Under
fragmentation that fails, and it surfaces as NVME_SC_INTERNAL, a generic
status, so NVMe multipath does not fail the command over to another
path. A 1 MiB command needs 256 entries, 8 KiB, served from the
kmalloc-8k slab, which does not fail. That failure mode is intermittent
and load-dependent rather than deterministic; on a vendor kernel
predating the bound it cost us roughly 98 GiB of a 2 TiB image copy,
with the copy tool exiting 0.

nvmet-rdma has advertised a bounded MDTS since "nvmet-rdma: Implement
get_mdts controller op" (Max Gurtovoy, Mar 2020), which left the other
transports untouched. Do the same for TCP, using the same 1 MiB value,
so initiators size their commands to something the target accepts
instead of discovering the limit by failing.

Signed-off-by: Alfonso Kuen <[email protected]>
---
Changes in v2:
 - Rewrite the rationale. v1 described the sgl_alloc() order-6 failure as
   current behaviour; since 4a3f00262a04 that path is unreachable and an
   oversized command is rejected earlier with SGL_INVALID_DATA | DNR. Both
   failure modes are now described, and which kernels see which.
 - v1 attributed the 32 MiB command size to a "generic 32 MiB ceiling" in
   the block layer. The generic cap is 4 MiB; the 32 MiB came from the
   target's own NOWS raising io_opt. Corrected -- and it sharpens the point,
   since the same controller advertises an optimal I/O size it cannot serve.
 - v1 called the 1 MiB SGL an order-0 allocation; it is 8 KiB.
 - No functional change: the diff is byte-identical to v1.
 drivers/nvme/target/tcp.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/drivers/nvme/target/tcp.c b/drivers/nvme/target/tcp.c
index e4f603b2a..414ba896e 100644
--- a/drivers/nvme/target/tcp.c
+++ b/drivers/nvme/target/tcp.c
@@ -23,6 +23,9 @@
 #include "nvmet.h"
 
 #define NVMET_TCP_DEF_INLINE_DATA_SIZE	(4 * PAGE_SIZE)
+
+/* Assume mpsmin == device_page_size == 4KB */
+#define NVMET_TCP_MAX_MDTS		8
 #define NVMET_TCP_MAXH2CDATA		0x400000 /* 16M arbitrary limit */
 #define NVMET_TCP_BACKLOG 128
 
@@ -2243,6 +2246,11 @@ static ssize_t nvmet_tcp_host_port_addr(struct nvmet_ctrl *ctrl,
 			(struct sockaddr *)&queue->sockaddr_peer);
 }
 
+static u8 nvmet_tcp_get_mdts(const struct nvmet_ctrl *ctrl)
+{
+	return NVMET_TCP_MAX_MDTS;
+}
+
 static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
 	.owner			= THIS_MODULE,
 	.type			= NVMF_TRTYPE_TCP,
@@ -2254,6 +2262,7 @@ static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
 	.install_queue		= nvmet_tcp_install_queue,
 	.disc_traddr		= nvmet_tcp_disc_port_addr,
 	.host_traddr		= nvmet_tcp_host_port_addr,
+	.get_mdts		= nvmet_tcp_get_mdts,
 };
 
 static int __init nvmet_tcp_init(void)
-- 
2.47.3