dm table: add zoned block devices validation
authorDamien Le Moal <damien.lemoal@wdc.com>
Mon, 8 May 2017 23:40:43 +0000 (16:40 -0700)
committerMike Snitzer <snitzer@redhat.com>
Mon, 19 Jun 2017 15:03:50 +0000 (11:03 -0400)
1) Introduce DM_TARGET_ZONED_HM feature flag:

The target drivers currently available will not operate correctly if a
table target maps onto a host-managed zoned block device.

To avoid problems, introduce the new feature flag DM_TARGET_ZONED_HM to
allow a target to explicitly state that it supports host-managed zoned
block devices.  This feature is checked for all targets in a table if
any of the table's block devices are host-managed.

Note that as host-aware zoned block devices are backward compatible with
regular block devices, they can be used by any of the current target
types.  This new feature is thus restricted to host-managed zoned block
devices.

2) Check device area zone alignment:

If a target maps to a zoned block device, check that the device area is
aligned on zone boundaries to avoid problems with REQ_OP_ZONE_RESET
operations (resetting a partially mapped sequential zone would not be
possible).  This also facilitates the processing of zone report with
REQ_OP_ZONE_REPORT bios.

3) Check block devices zone model compatibility

When setting the DM device's queue limits, several possibilities exists
for zoned block devices:
1) The DM target driver may want to expose a different zone model
(e.g. host-managed device emulation or regular block device on top of
host-managed zoned block devices)
2) Expose the underlying zone model of the devices as-is

To allow both cases, the underlying block device zone model must be set
in the target limits in dm_set_device_limits() and the compatibility of
all devices checked similarly to the logical block size alignment.  For
this last check, introduce validate_hardware_zoned_model() to check that
all targets of a table have the same zone model and that the zone size
of the target devices are equal.

Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
[Mike Snitzer refactored Damien's original work to simplify the code]
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-table.c
include/linux/device-mapper.h

index 5f5eae41f8041ee3a70540f1fd78e01444abf247..a39bcd9b982a443cdfdac2013fcbf5ce3561ac4e 100644 (file)
@@ -319,6 +319,39 @@ static int device_area_is_invalid(struct dm_target *ti, struct dm_dev *dev,
                return 1;
        }
 
+       /*
+        * If the target is mapped to zoned block device(s), check
+        * that the zones are not partially mapped.
+        */
+       if (bdev_zoned_model(bdev) != BLK_ZONED_NONE) {
+               unsigned int zone_sectors = bdev_zone_sectors(bdev);
+
+               if (start & (zone_sectors - 1)) {
+                       DMWARN("%s: start=%llu not aligned to h/w zone size %u of %s",
+                              dm_device_name(ti->table->md),
+                              (unsigned long long)start,
+                              zone_sectors, bdevname(bdev, b));
+                       return 1;
+               }
+
+               /*
+                * Note: The last zone of a zoned block device may be smaller
+                * than other zones. So for a target mapping the end of a
+                * zoned block device with such a zone, len would not be zone
+                * aligned. We do not allow such last smaller zone to be part
+                * of the mapping here to ensure that mappings with multiple
+                * devices do not end up with a smaller zone in the middle of
+                * the sector range.
+                */
+               if (len & (zone_sectors - 1)) {
+                       DMWARN("%s: len=%llu not aligned to h/w zone size %u of %s",
+                              dm_device_name(ti->table->md),
+                              (unsigned long long)len,
+                              zone_sectors, bdevname(bdev, b));
+                       return 1;
+               }
+       }
+
        if (logical_block_size_sectors <= 1)
                return 0;
 
@@ -456,6 +489,8 @@ static int dm_set_device_limits(struct dm_target *ti, struct dm_dev *dev,
                       q->limits.alignment_offset,
                       (unsigned long long) start << SECTOR_SHIFT);
 
+       limits->zoned = blk_queue_zoned_model(q);
+
        return 0;
 }
 
@@ -1346,6 +1381,88 @@ bool dm_table_has_no_data_devices(struct dm_table *table)
        return true;
 }
 
+static int device_is_zoned_model(struct dm_target *ti, struct dm_dev *dev,
+                                sector_t start, sector_t len, void *data)
+{
+       struct request_queue *q = bdev_get_queue(dev->bdev);
+       enum blk_zoned_model *zoned_model = data;
+
+       return q && blk_queue_zoned_model(q) == *zoned_model;
+}
+
+static bool dm_table_supports_zoned_model(struct dm_table *t,
+                                         enum blk_zoned_model zoned_model)
+{
+       struct dm_target *ti;
+       unsigned i;
+
+       for (i = 0; i < dm_table_get_num_targets(t); i++) {
+               ti = dm_table_get_target(t, i);
+
+               if (zoned_model == BLK_ZONED_HM &&
+                   !dm_target_supports_zoned_hm(ti->type))
+                       return false;
+
+               if (!ti->type->iterate_devices ||
+                   !ti->type->iterate_devices(ti, device_is_zoned_model, &zoned_model))
+                       return false;
+       }
+
+       return true;
+}
+
+static int device_matches_zone_sectors(struct dm_target *ti, struct dm_dev *dev,
+                                      sector_t start, sector_t len, void *data)
+{
+       struct request_queue *q = bdev_get_queue(dev->bdev);
+       unsigned int *zone_sectors = data;
+
+       return q && blk_queue_zone_sectors(q) == *zone_sectors;
+}
+
+static bool dm_table_matches_zone_sectors(struct dm_table *t,
+                                         unsigned int zone_sectors)
+{
+       struct dm_target *ti;
+       unsigned i;
+
+       for (i = 0; i < dm_table_get_num_targets(t); i++) {
+               ti = dm_table_get_target(t, i);
+
+               if (!ti->type->iterate_devices ||
+                   !ti->type->iterate_devices(ti, device_matches_zone_sectors, &zone_sectors))
+                       return false;
+       }
+
+       return true;
+}
+
+static int validate_hardware_zoned_model(struct dm_table *table,
+                                        enum blk_zoned_model zoned_model,
+                                        unsigned int zone_sectors)
+{
+       if (zoned_model == BLK_ZONED_NONE)
+               return 0;
+
+       if (!dm_table_supports_zoned_model(table, zoned_model)) {
+               DMERR("%s: zoned model is not consistent across all devices",
+                     dm_device_name(table->md));
+               return -EINVAL;
+       }
+
+       /* Check zone size validity and compatibility */
+       if (!zone_sectors || !is_power_of_2(zone_sectors))
+               return -EINVAL;
+
+       if (!dm_table_matches_zone_sectors(table, zone_sectors)) {
+               DMERR("%s: zone sectors is not consistent across all devices",
+                     dm_device_name(table->md));
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
 /*
  * Establish the new table's queue_limits and validate them.
  */
@@ -1355,6 +1472,8 @@ int dm_calculate_queue_limits(struct dm_table *table,
        struct dm_target *ti;
        struct queue_limits ti_limits;
        unsigned i;
+       enum blk_zoned_model zoned_model = BLK_ZONED_NONE;
+       unsigned int zone_sectors = 0;
 
        blk_set_stacking_limits(limits);
 
@@ -1372,6 +1491,15 @@ int dm_calculate_queue_limits(struct dm_table *table,
                ti->type->iterate_devices(ti, dm_set_device_limits,
                                          &ti_limits);
 
+               if (zoned_model == BLK_ZONED_NONE && ti_limits.zoned != BLK_ZONED_NONE) {
+                       /*
+                        * After stacking all limits, validate all devices
+                        * in table support this zoned model and zone sectors.
+                        */
+                       zoned_model = ti_limits.zoned;
+                       zone_sectors = ti_limits.chunk_sectors;
+               }
+
                /* Set I/O hints portion of queue limits */
                if (ti->type->io_hints)
                        ti->type->io_hints(ti, &ti_limits);
@@ -1396,8 +1524,42 @@ combine_limits:
                               dm_device_name(table->md),
                               (unsigned long long) ti->begin,
                               (unsigned long long) ti->len);
+
+               /*
+                * FIXME: this should likely be moved to blk_stack_limits(), would
+                * also eliminate limits->zoned stacking hack in dm_set_device_limits()
+                */
+               if (limits->zoned == BLK_ZONED_NONE && ti_limits.zoned != BLK_ZONED_NONE) {
+                       /*
+                        * By default, the stacked limits zoned model is set to
+                        * BLK_ZONED_NONE in blk_set_stacking_limits(). Update
+                        * this model using the first target model reported
+                        * that is not BLK_ZONED_NONE. This will be either the
+                        * first target device zoned model or the model reported
+                        * by the target .io_hints.
+                        */
+                       limits->zoned = ti_limits.zoned;
+               }
        }
 
+       /*
+        * Verify that the zoned model and zone sectors, as determined before
+        * any .io_hints override, are the same across all devices in the table.
+        * - this is especially relevant if .io_hints is emulating a disk-managed
+        *   zoned model (aka BLK_ZONED_NONE) on host-managed zoned block devices.
+        * BUT...
+        */
+       if (limits->zoned != BLK_ZONED_NONE) {
+               /*
+                * ...IF the above limits stacking determined a zoned model
+                * validate that all of the table's devices conform to it.
+                */
+               zoned_model = limits->zoned;
+               zone_sectors = limits->chunk_sectors;
+       }
+       if (validate_hardware_zoned_model(table, zoned_model, zone_sectors))
+               return -EINVAL;
+
        return validate_hardware_logical_block_alignment(table, limits);
 }
 
index 19bc7fdfd6dabd53fc4ec1cb2df8c80096bb4dd2..186ef74009cb12a370e92a7a7bf3cd2739dca597 100644 (file)
@@ -237,6 +237,12 @@ typedef unsigned (*dm_num_write_bios_fn) (struct dm_target *ti, struct bio *bio)
 #define DM_TARGET_PASSES_INTEGRITY     0x00000020
 #define dm_target_passes_integrity(type) ((type)->features & DM_TARGET_PASSES_INTEGRITY)
 
+/*
+ * Indicates that a target supports host-managed zoned block devices.
+ */
+#define DM_TARGET_ZONED_HM             0x00000040
+#define dm_target_supports_zoned_hm(type) ((type)->features & DM_TARGET_ZONED_HM)
+
 struct dm_target {
        struct dm_table *table;
        struct target_type *type;