Add helpers that export a zoned null_blk device as a namespace of the default subsystem and connect to it, tear that setup down again, and issue a partial Zone Management Receive and check its completion status. The zone report tests added next share them. Assisted-by: LLM Signed-off-by: Andrea Parri --- tests/nvme/rc | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 61 insertions(+) diff --git a/tests/nvme/rc b/tests/nvme/rc index c415ca1234407..270fad19d61f1 100644 --- a/tests/nvme/rc +++ b/tests/nvme/rc @@ -408,6 +408,67 @@ _nvmet_passthru_target_cleanup() { _remove_nvmet_host "${def_hostnqn}" } +# Export a zoned null_blk device as a namespace of the default subsystem and +# connect to it. Callers must source common/null_blk. +_nvmet_zoned_target_setup() { + if ! _init_null_blk nr_devices=0; then + return 1 + fi + if ! _configure_null_blk nullb0 blocksize=512 size=1024 zoned=1 \ + zone_size=1 power=1 >>"$FULL" 2>&1; then + echo "could not configure zoned null_blk device" + return 1 + fi + + _nvmet_target_setup --blkdev none + _create_nvmet_ns --blkdev /dev/nullb0 --uuid "$def_subsys_uuid" \ + >/dev/null + _nvme_connect_subsys +} + +_nvmet_zoned_target_cleanup() { + _nvme_disconnect_subsys + _nvmet_target_cleanup + _exit_null_blk +} + +# Test cleanup handler to register with _nvmet_zoned_target_setup(). +_cleanup_nvmet_zoned_target() { + _cleanup_nvmet + _exit_null_blk +} + +# Issue a partial zone report from zone 0 of namespace device $1 with a +# $2-byte data buffer and NUMD $3, and check that it completes with status +# code $4, or succeeds if $4 is 0. +_check_zone_report() { + local dev=$1 data_len=$2 numd=$3 expected=$4 desc err rc status + + desc="zone report with $data_len bytes and NUMD $numd" + err=$(nvme io-passthru "$dev" --namespace-id=1 --opcode=0x7a --read \ + --data-len="$data_len" --cdw10=0 --cdw11=0 --cdw12="$numd" \ + --cdw13=0x10000 2>&1 >/dev/null) + rc=$? + status=$(sed -n \ + 's/^NVMe status:.*(\(0x[[:xdigit:]][[:xdigit:]]*\)).*/\1/p' \ + <<<"$err") + printf '%s: exit=%d status=%s\n' "$desc" "$rc" "${status:-none}" \ + >>"$FULL" + + if ((expected == 0)); then + if ((rc != 0)); then + echo "$desc failed" + fi + elif ((rc == 0)); then + echo "$desc was not rejected" + elif [[ -z "$status" ]]; then + echo "$desc returned no NVMe status" + elif (((status & 0x7ff) != expected)); then + printf '%s returned status %s, expected %#x\n' "$desc" \ + "$status" "$expected" + fi +} + _discovery_genctr() { local port=${1} -- 2.53.0 Issue a failing Identify into a buffer filled with a sentinel and verify that the target transfers no data, with a successful Identify and a namespace read as controls. The test runs on every configured transport; only nvmet-fc failed it before the fix. Add a small helper because nvme-cli cannot prefill the data buffer. The test fails over fc on the unpatched kernel and passes with the fix applied. Assisted-by: LLM Signed-off-by: Andrea Parri --- src/.gitignore | 1 + src/Makefile | 1 + src/nvme-sentinel-read.c | 89 ++++++++++++++++++++++++++++++++++++++++ tests/nvme/071 | 80 ++++++++++++++++++++++++++++++++++++ tests/nvme/071.out | 2 + 5 files changed, 173 insertions(+) create mode 100644 src/nvme-sentinel-read.c create mode 100755 tests/nvme/071 create mode 100644 tests/nvme/071.out diff --git a/src/.gitignore b/src/.gitignore index e9869e172b20f..2429da9d9cf44 100644 --- a/src/.gitignore +++ b/src/.gitignore @@ -18,3 +18,4 @@ /ioctl-lbmd-query /nvme-passthru-admin-uring /nvme-delay-ioctl +/nvme-sentinel-read diff --git a/src/Makefile b/src/Makefile index dd646940fa25e..78433f0762b4a 100644 --- a/src/Makefile +++ b/src/Makefile @@ -24,6 +24,7 @@ C_TARGETS := \ mount_clear_sock \ nvme-delay-ioctl \ nvme-passthrough-meta \ + nvme-sentinel-read \ ioctl-lbmd-query \ nbdsetsize \ openclose \ diff --git a/src/nvme-sentinel-read.c b/src/nvme-sentinel-read.c new file mode 100644 index 0000000000000..dd36f283f46e2 --- /dev/null +++ b/src/nvme-sentinel-read.c @@ -0,0 +1,89 @@ +// SPDX-License-Identifier: GPL-3.0+ +// Copyright (C) 2026 Andrea Parri + +/* + * Issue an NVMe admin read command into a buffer filled with a sentinel and + * print the NVMe status and the number of bytes the command overwrote. + */ +#define _GNU_SOURCE +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#define SENTINEL 0xa5 + +static int parse_u32(const char *s, uint32_t *val) +{ + unsigned long v; + char *end; + + errno = 0; + v = strtoul(s, &end, 0); + if (errno || end == s || *end || v > UINT32_MAX) + return -1; + *val = v; + return 0; +} + +int main(int argc, char **argv) +{ + struct nvme_passthru_cmd cmd = { }; + uint32_t opcode, cdw10, len, i, changed = 0; + unsigned char *buf; + void *p; + int fd, ret; + + if (argc != 5 || parse_u32(argv[2], &opcode) || opcode > 0xff || + parse_u32(argv[3], &cdw10) || parse_u32(argv[4], &len) || !len) { + fprintf(stderr, "usage: %s /dev/nvmeX opcode cdw10 data-len\n", + argv[0]); + return 2; + } + + /* + * Page-align the buffer so that the kernel maps it for the transfer + * instead of bouncing it: a bounce buffer is zero-filled and copied + * back even when the command fails, which would hide the sentinel. + */ + ret = posix_memalign(&p, 4096, len); + if (ret) { + fprintf(stderr, "posix_memalign: %s\n", strerror(ret)); + return 2; + } + buf = p; + memset(buf, SENTINEL, len); + + fd = open(argv[1], O_RDONLY); + if (fd < 0) { + perror(argv[1]); + free(buf); + return 2; + } + + cmd.opcode = opcode; + cmd.addr = (uintptr_t)buf; + cmd.data_len = len; + cmd.cdw10 = cdw10; + + ret = ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd); + close(fd); + if (ret < 0) { + perror("NVME_IOCTL_ADMIN_CMD"); + free(buf); + return 2; + } + + for (i = 0; i < len; i++) + if (buf[i] != SENTINEL) + changed++; + + printf("%#x %u\n", ret, changed); + free(buf); + return 0; +} diff --git a/tests/nvme/071 b/tests/nvme/071 new file mode 100755 index 0000000000000..5333bfac8db53 --- /dev/null +++ b/tests/nvme/071 @@ -0,0 +1,80 @@ +#!/bin/bash +# SPDX-License-Identifier: GPL-3.0+ +# Copyright (C) 2026 Andrea Parri +# +# Check that the NVMe target transfers no read data for a failed command. +# The host buffer is filled with a sentinel, so any data the target sends +# shows up as changed bytes. +# +# Regression test for patch "nvmet-fc: only send read data for successful +# commands". + +. tests/nvme/rc + +DESCRIPTION="verify nvmet does not transfer data for failed reads" +QUICK=1 + +requires() { + _nvme_requires + _have_loop + _require_kernel_nvme_target + _have_src_program nvme-sentinel-read +} + +set_conditions() { + _set_nvme_trtype "$@" +} + +# Issue an Identify with CNS $2 to controller $1 and print the NVMe status +# and the number of changed bytes in the 4096-byte data buffer. +_identify() { + src/nvme-sentinel-read "$1" 0x06 "$2" 4096 2>>"$FULL" +} + +test() { + echo "Running ${TEST_NAME}" + + local changed ctrl ns status + + _setup_nvmet + # File backing: the namespace read below fails over loop with the + # default loop device backing, for reasons unrelated to this test. + _nvmet_target_setup --blkdev file + _nvme_connect_subsys + + ctrl=$(_find_nvme_dev "$def_subsysnqn") + ns="/dev/${ctrl}n1" + + # Identify Controller succeeds and must overwrite the sentinel. + if ! read -r status changed < <(_identify "/dev/$ctrl" 0x01); then + echo "successful Identify could not be issued" + else + echo "cns=0x01 status=$status changed=$changed" >>"$FULL" + if ((status != 0 || changed == 0)); then + echo "successful Identify returned status $status with $changed bytes" + fi + fi + + # An unsupported CNS fails and must leave the sentinel intact. + if ! read -r status changed < <(_identify "/dev/$ctrl" 0xff); then + echo "failed Identify could not be issued" + else + echo "cns=0xff status=$status changed=$changed" >>"$FULL" + if (((status & 0x7ff) != 2)); then + echo "failed Identify returned status $status, expected 0x2" + fi + if ((changed != 0)); then + echo "failed Identify transferred data: $changed bytes changed" + fi + fi + + if ! dd if="$ns" of=/dev/null bs=4096 count=256 iflag=direct \ + status=none >>"$FULL" 2>&1; then + echo "namespace read failed" + fi + + _nvme_disconnect_subsys + _nvmet_target_cleanup + + echo "Test complete" +} diff --git a/tests/nvme/071.out b/tests/nvme/071.out new file mode 100644 index 0000000000000..146809b91e0a3 --- /dev/null +++ b/tests/nvme/071.out @@ -0,0 +1,2 @@ +Running nvme/071 +Test complete -- 2.53.0 Export a zoned null_blk device through the NVMe target and verify that Zone Management Receive zeroes reserved and unused report bytes and rejects a transfer length inconsistent with NUMD. The test runs on every configured transport, with RDMA restricted to 4 KiB page kernels so that the preparatory write exceeds the default inline-data limit. Add a small passthrough helper because nvme-cli does not expose the reserved report bytes. The test fails on the unpatched kernel and passes with the fix applied. Assisted-by: LLM Signed-off-by: Andrea Parri --- src/.gitignore | 1 + src/Makefile | 1 + src/nvme-zone-report.c | 165 +++++++++++++++++++++++++++++++++++++++++ tests/nvme/072 | 68 +++++++++++++++++ tests/nvme/072.out | 2 + 5 files changed, 237 insertions(+) create mode 100644 src/nvme-zone-report.c create mode 100755 tests/nvme/072 create mode 100644 tests/nvme/072.out diff --git a/src/.gitignore b/src/.gitignore index 2429da9d9cf44..8119476416937 100644 --- a/src/.gitignore +++ b/src/.gitignore @@ -19,3 +19,4 @@ /nvme-passthru-admin-uring /nvme-delay-ioctl /nvme-sentinel-read +/nvme-zone-report diff --git a/src/Makefile b/src/Makefile index 78433f0762b4a..7c4fd600ace1d 100644 --- a/src/Makefile +++ b/src/Makefile @@ -25,6 +25,7 @@ C_TARGETS := \ nvme-delay-ioctl \ nvme-passthrough-meta \ nvme-sentinel-read \ + nvme-zone-report \ ioctl-lbmd-query \ nbdsetsize \ openclose \ diff --git a/src/nvme-zone-report.c b/src/nvme-zone-report.c new file mode 100644 index 0000000000000..951b1d03d4a75 --- /dev/null +++ b/src/nvme-zone-report.c @@ -0,0 +1,165 @@ +// SPDX-License-Identifier: GPL-3.0+ +// Copyright (C) 2026 Andrea Parri + +/* Check that unused bytes in an NVMe zone report are zeroed. */ +#define _GNU_SOURCE +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#define REPORT_LEN 4096 +/* + * Larger than the default RDMA in-capsule data size on 4 KiB page kernels. + * nvme/072 skips RDMA on other page sizes, where this write could stay + * inline rather than use buffers allocated and freed per command. + */ +#define POISON_LEN 8192 +#define WRITE_OPCODE 0x01 +#define ZONE_REPORT_OPCODE 0x7a +#define PARTIAL_REPORT (1U << 16) +#define ZONE_REPORT_HEADER 64 +#define ZONE_DESC_LEN 64 +#define ZONE_START_OFFSET 16 +#define NR_ATTEMPTS 64 + +static int poison_target(int fd, unsigned int iteration) +{ + unsigned char data[POISON_LEN]; + struct nvme_passthru_cmd cmd = { + .opcode = WRITE_OPCODE, + .nsid = 1, + .addr = (uintptr_t)data, + .data_len = sizeof(data), + .cdw10 = iteration * (POISON_LEN / 512), + .cdw12 = POISON_LEN / 512 - 1, + }; + + memset(data, 0x5a, sizeof(data)); + return ioctl(fd, NVME_IOCTL_IO_CMD, &cmd); +} + +static uint64_t read_le64(const unsigned char *p) +{ + uint64_t value; + + memcpy(&value, p, sizeof(value)); + return le64toh(value); +} + +int main(int argc, char **argv) +{ + unsigned char *report; + unsigned long long slba; + char *end; + void *p; + int fd, i, j, ret, rc = 0; + + if (argc != 3) { + fprintf(stderr, "usage: %s /dev/nvmeXnY last-zone-slba\n", + argv[0]); + return 2; + } + + errno = 0; + slba = strtoull(argv[2], &end, 0); + if (errno || *end) { + fprintf(stderr, "invalid SLBA: %s\n", argv[2]); + return 2; + } + + /* + * Page-align the report so that the kernel maps it for the transfer + * instead of bouncing it: a bounce buffer is zero-filled, which would + * hide the sentinel and any bytes the target left unwritten. + */ + ret = posix_memalign(&p, 4096, REPORT_LEN); + if (ret) { + fprintf(stderr, "posix_memalign: %s\n", strerror(ret)); + return 2; + } + report = p; + + fd = open(argv[1], O_RDWR); + if (fd < 0) { + perror(argv[1]); + free(report); + return 2; + } + + for (i = 0; i < NR_ATTEMPTS; i++) { + struct nvme_passthru_cmd cmd = { + .opcode = ZONE_REPORT_OPCODE, + .nsid = 1, + .addr = (uintptr_t)report, + .data_len = REPORT_LEN, + .cdw10 = (uint32_t)slba, + .cdw11 = (uint32_t)(slba >> 32), + .cdw12 = REPORT_LEN / 4 - 1, + .cdw13 = PARTIAL_REPORT, + }; + unsigned int nonzero = 0, stale = 0, poisoned = 0; + unsigned char sentinel = i & 1 ? 0xa5 : 0x3c; + + ret = poison_target(fd, i); + if (ret) { + if (ret < 0) + perror("poison_target"); + else + fprintf(stderr, "poison write status: %#x\n", ret); + rc = 2; + goto out; + } + + memset(report, sentinel, REPORT_LEN); + ret = ioctl(fd, NVME_IOCTL_IO_CMD, &cmd); + if (ret) { + if (ret < 0) + perror("NVME_IOCTL_IO_CMD"); + else + fprintf(stderr, "NVMe status: %#x\n", ret); + rc = 2; + goto out; + } + + if (read_le64(report) != 1 || + read_le64(report + ZONE_REPORT_HEADER + + ZONE_START_OFFSET) != slba) { + fprintf(stderr, "unexpected zone report header or SLBA\n"); + rc = 1; + goto out; + } + + for (j = 8; j < REPORT_LEN; j++) { + if (j >= ZONE_REPORT_HEADER && + j < ZONE_REPORT_HEADER + ZONE_DESC_LEN) + continue; + if (!report[j]) + continue; + nonzero++; + if (report[j] == sentinel) + stale++; + if (report[j] == 0x5a) + poisoned++; + } + + fprintf(stderr, "attempt %d: %u nonzero bytes " + "(%u poison, %u sentinel)\n", i + 1, nonzero, + poisoned, stale); + if (nonzero) { + rc = 1; + goto out; + } + } + +out: + close(fd); + free(report); + return rc; +} diff --git a/tests/nvme/072 b/tests/nvme/072 new file mode 100755 index 0000000000000..c7bcf93ab4d20 --- /dev/null +++ b/tests/nvme/072 @@ -0,0 +1,68 @@ +#!/bin/bash +# SPDX-License-Identifier: GPL-3.0+ +# Copyright (C) 2026 Andrea Parri +# +# Check that NVMe target zone reports are fully initialized and that their +# transfer length matches NUMD. +# +# Regression test for patch "nvmet: zns: don't leak stale memory in zone +# management receive". + +. tests/nvme/rc +. common/null_blk + +DESCRIPTION="check NVMe target zone report initialization and length" +QUICK=1 + +requires() { + _nvme_requires + _have_loop + _require_kernel_nvme_target + _have_module null_blk + _have_kernel_options BLK_DEV_ZONED + _have_src_program nvme-zone-report + + # RDMA defaults to PAGE_SIZE bytes of inline data. The helper's + # 8 KiB preparatory write must exceed that limit. + if [[ "$nvme_trtype" == "rdma" ]]; then + _page_size_equals 4096 + fi +} + +set_conditions() { + _set_nvme_trtype "$@" +} + +test() { + echo "Running ${TEST_NAME}" + + local block_size dev last_sector ret sectors slba zone_sectors + + _setup_nvmet + _register_test_cleanup _cleanup_nvmet_zoned_target + if ! _nvmet_zoned_target_setup; then + return 1 + fi + dev="/dev/$(_find_nvme_ns "$def_subsys_uuid")" + + sectors=$(cat /sys/block/nullb0/size) + zone_sectors=$(cat /sys/block/nullb0/queue/chunk_sectors) + block_size=$(cat /sys/block/nullb0/queue/logical_block_size) + last_sector=$(((sectors - 1) / zone_sectors * zone_sectors)) + slba=$((last_sector * 512 / block_size)) + + src/nvme-zone-report "$dev" "$slba" >>"$FULL" 2>&1 + ret=$? + case $ret in + 0) ;; + 1) echo "zone report contained stale data" ;; + *) echo "zone report checker failed" ;; + esac + + _check_zone_report "$dev" 4096 1023 0 + _check_zone_report "$dev" 8192 1023 15 + + _nvmet_zoned_target_cleanup + + echo "Test complete" +} diff --git a/tests/nvme/072.out b/tests/nvme/072.out new file mode 100644 index 0000000000000..3f765b30a22b0 --- /dev/null +++ b/tests/nvme/072.out @@ -0,0 +1,2 @@ +Running nvme/072 +Test complete -- 2.53.0 Verify that Zone Management Receive rejects a NUMD whose byte count wraps a 32-bit calculation, while accepting an ordinary request. The test runs on every configured transport. The test fails on the unpatched kernel and passes with the fix applied. Assisted-by: LLM Signed-off-by: Andrea Parri --- tests/nvme/073 | 46 ++++++++++++++++++++++++++++++++++++++++++++++ tests/nvme/073.out | 2 ++ 2 files changed, 48 insertions(+) create mode 100755 tests/nvme/073 create mode 100644 tests/nvme/073.out diff --git a/tests/nvme/073 b/tests/nvme/073 new file mode 100755 index 0000000000000..c6d3d29f4b906 --- /dev/null +++ b/tests/nvme/073 @@ -0,0 +1,46 @@ +#!/bin/bash +# SPDX-License-Identifier: GPL-3.0+ +# Copyright (C) 2026 Andrea Parri +# +# Check that the NVMe target rejects a zone report NUMD whose byte count +# overflows a 32-bit value. +# +# Regression test for patch "nvmet: zns: reject a NUMD whose byte count wraps". + +. tests/nvme/rc +. common/null_blk + +DESCRIPTION="reject an NVMe target zone report NUMD that wraps" +QUICK=1 + +requires() { + _nvme_requires + _have_loop + _require_kernel_nvme_target + _have_module null_blk + _have_kernel_options BLK_DEV_ZONED +} + +set_conditions() { + _set_nvme_trtype "$@" +} + +test() { + echo "Running ${TEST_NAME}" + + local dev + + _setup_nvmet + _register_test_cleanup _cleanup_nvmet_zoned_target + if ! _nvmet_zoned_target_setup; then + return 1 + fi + dev="/dev/$(_find_nvme_ns "$def_subsys_uuid")" + + _check_zone_report "$dev" 64 15 0 + _check_zone_report "$dev" 64 0x4000000f 2 + + _nvmet_zoned_target_cleanup + + echo "Test complete" +} diff --git a/tests/nvme/073.out b/tests/nvme/073.out new file mode 100644 index 0000000000000..ed6690e578688 --- /dev/null +++ b/tests/nvme/073.out @@ -0,0 +1,2 @@ +Running nvme/073 +Test complete -- 2.53.0