BLKREPORTZONEV2 takes the flags field of struct blk_zone_report as an input. With BLK_ZONE_REP_CACHED the report is built from the zone information cached by the block layer. Without it, BLKREPORTZONEV2 is documented to behave like BLKREPORTZONE and to get the report from the device: this is what the changelog of the commit that added the ioctl and the comments in include/uapi/linux/blkzoned.h say. blkdev_report_zones_ioctl() only uses the flag to validate the input and calls blkdev_report_zones_cached() for every BLKREPORTZONEV2 request. A caller that passes flags == 0 gets the cached report, in which implicitly open, explicitly open and closed zones all have the condition BLK_ZONE_COND_ACTIVE, and an explicitly opened empty zone is reported as empty. Example with a zoned null_blk device (10 MiB, 4 MiB zones) after BLKOPENZONE on zone 0, and a write of 8 sectors to zone 1 followed by BLKCLOSEZONE: BLKREPORTZONE: zone 0 cond 0x3 (EXP_OPEN) zone 1 cond 0x4 (CLOSED) BLKREPORTZONEV2, flags 0: zone 0 cond 0x1 (EMPTY) zone 1 cond 0xff (ACTIVE) The uapi header marks BLKREPORTZONE as deprecated in favour of BLKREPORTZONEV2. A program that follows this and does not ask for cached information can no longer tell open zones from closed ones, and is handed BLK_ZONE_COND_ACTIVE, which is only defined for the cached report. Use the cached report only when BLK_ZONE_REP_CACHED is set and the device has a cached report; otherwise, and whenever the flag is not set, get the report from the device with blkdev_report_zones(). Guard the cached path with blkdev_has_cached_report_zones() so the redundant call is avoided when there is no cached report: in that case blkdev_report_zones_cached() already falls back to a device report, so the observable behaviour is unchanged. blkdev_has_cached_report_zones() is defined lower in the file, so move it, with its comment, above blkdev_report_zones_ioctl(). Tested with zoned null_blk devices in qemu: a program that compares the reports of BLKREPORTZONE and BLKREPORTZONEV2 with flags 0 after the sequence above finds the differences before this change and none with it. The report with BLK_ZONE_REP_CACHED is unchanged. Fixes: b30ffcdc0c15 ("block: introduce BLKREPORTZONESV2 ioctl") Cc: stable@vger.kernel.org Assisted-by: LLM Signed-off-by: Shashank Mohan Jain --- Prepared with Claude Code (Anthropic), model Claude Opus 4.8 (claude-opus-4-8). block/blk-zoned.c | 37 ++++++++++++++++++++++--------------- 1 file changed, 22 insertions(+), 15 deletions(-) diff --git a/block/blk-zoned.c b/block/blk-zoned.c index 475aa16bc..c9b66b8a7 100644 --- a/block/blk-zoned.c +++ b/block/blk-zoned.c @@ -361,6 +361,19 @@ static int blkdev_copy_zone_to_user(struct blk_zone *zone, unsigned int idx, return 0; } +/* + * For devices that natively support zone append operations, we do not use zone + * write plugging for zone append writes, which makes the zone condition + * tracking invalid once zone append was used. In that case fall back to a + * regular report zones to get correct information. + */ +static inline bool blkdev_has_cached_report_zones(struct block_device *bdev) +{ + return disk_need_zone_resources(bdev->bd_disk) && + (bdev_emulates_zone_append(bdev) || + !test_bit(GD_ZONE_APPEND_USED, &bdev->bd_disk->state)); +} + /* * Mask of valid input flags for BLKREPORTZONEV2 ioctl. */ @@ -400,8 +413,15 @@ int blkdev_report_zones_ioctl(struct block_device *bdev, unsigned int cmd, case BLKREPORTZONEV2: if (rep.flags & ~BLK_ZONE_REPV2_INPUT_FLAGS) return -EINVAL; - ret = blkdev_report_zones_cached(bdev, rep.sector, rep.nr_zones, - blkdev_copy_zone_to_user, &args); + if ((rep.flags & BLK_ZONE_REP_CACHED) && + blkdev_has_cached_report_zones(bdev)) + ret = blkdev_report_zones_cached(bdev, rep.sector, + rep.nr_zones, + blkdev_copy_zone_to_user, &args); + else + ret = blkdev_report_zones(bdev, rep.sector, + rep.nr_zones, + blkdev_copy_zone_to_user, &args); break; default: return -EINVAL; @@ -905,19 +925,6 @@ static int blkdev_report_zone_fallback(struct block_device *bdev, return 0; } -/* - * For devices that natively support zone append operations, we do not use zone - * write plugging for zone append writes, which makes the zone condition - * tracking invalid once zone append was used. In that case fall back to a - * regular report zones to get correct information. - */ -static inline bool blkdev_has_cached_report_zones(struct block_device *bdev) -{ - return disk_need_zone_resources(bdev->bd_disk) && - (bdev_emulates_zone_append(bdev) || - !test_bit(GD_ZONE_APPEND_USED, &bdev->bd_disk->state)); -} - /** * blkdev_get_zone_info - Get a single zone information from cached data * @bdev: Target block device -- 2.43.0 blkdev_get_zone_info() sets the length of every zone to the zone size of the device. The last zone of a zoned block device may be smaller than the others: blk_revalidate_zone_cb() accepts that and remembers the capacity of that zone in disk->last_zone_capacity. For such a zone the length reported from the cached information is too large, and start + len is beyond the end of the device. A conventional last zone is also given the capacity of a full zone. Example with a zoned null_blk device of 10 MiB (20480 sectors) with 4 MiB zones, last zone: BLKREPORTZONE: start 16384 len 4096 capacity 4096 BLKREPORTZONEV2 with BLK_ZONE_REP_CACHED: start 16384 len 8192 capacity 4096 Limit the length to what is left of the device. disk_zone_is_last() still recognises the last zone, as start + len is then equal to the capacity of the device. The capacity of a conventional zone is equal to its length, which blk_revalidate_conv_zone() checks, so use the length there. Tested with zoned null_blk devices in qemu (all zones sequential, and two conventional zones followed by sequential ones, both with a smaller last zone): the start, length and capacity of all zones in the cached report are the same as in the report from the device with this change. null_blk cannot create a device with a conventional last zone, so that case is not tested. Fixes: f2284eec5053 ("block: introduce blkdev_get_zone_info()") Cc: stable@vger.kernel.org Assisted-by: LLM Signed-off-by: Shashank Mohan Jain Reviewed-by: Damien Le Moal --- Prepared with Claude Code (Anthropic), model Claude Opus 4.8 (claude-opus-4-8). block/blk-zoned.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/block/blk-zoned.c b/block/blk-zoned.c index c9b66b8a7..4ebaba9aa 100644 --- a/block/blk-zoned.c +++ b/block/blk-zoned.c @@ -971,7 +971,7 @@ int blkdev_get_zone_info(struct block_device *bdev, sector_t sector, rcu_read_unlock(); zone->start = sector; - zone->len = zone_sectors; + zone->len = min(zone_sectors, get_capacity(disk) - sector); /* * If this is a conventional zone, we do not have a zone write plug and @@ -979,7 +979,7 @@ int blkdev_get_zone_info(struct block_device *bdev, sector_t sector, */ if (zone->cond == BLK_ZONE_COND_NOT_WP) { zone->type = BLK_ZONE_TYPE_CONVENTIONAL; - zone->capacity = zone_sectors; + zone->capacity = zone->len; zone->wp = ULLONG_MAX; return 0; } -- 2.43.0