diff options
Diffstat (limited to 'Documentation/md')
| -rw-r--r-- | Documentation/md/md-cluster.txt | 325 | ||||
| -rw-r--r-- | Documentation/md/raid5-cache.txt | 109 | ||||
| -rw-r--r-- | Documentation/md/raid5-ppl.txt | 45 |
3 files changed, 479 insertions, 0 deletions
diff --git a/Documentation/md/md-cluster.txt b/Documentation/md/md-cluster.txt new file mode 100644 index 000000000..e1055f105 --- /dev/null +++ b/Documentation/md/md-cluster.txt @@ -0,0 +1,325 @@ +The cluster MD is a shared-device RAID for a cluster, it supports +two levels: raid1 and raid10 (limited support). + + +1. On-disk format + +Separate write-intent-bitmaps are used for each cluster node. +The bitmaps record all writes that may have been started on that node, +and may not yet have finished. The on-disk layout is: + +0 4k 8k 12k +------------------------------------------------------------------- +| idle | md super | bm super [0] + bits | +| bm bits[0, contd] | bm super[1] + bits | bm bits[1, contd] | +| bm super[2] + bits | bm bits [2, contd] | bm super[3] + bits | +| bm bits [3, contd] | | | + +During "normal" functioning we assume the filesystem ensures that only +one node writes to any given block at a time, so a write request will + + - set the appropriate bit (if not already set) + - commit the write to all mirrors + - schedule the bit to be cleared after a timeout. + +Reads are just handled normally. It is up to the filesystem to ensure +one node doesn't read from a location where another node (or the same +node) is writing. + + +2. DLM Locks for management + +There are three groups of locks for managing the device: + +2.1 Bitmap lock resource (bm_lockres) + + The bm_lockres protects individual node bitmaps. They are named in + the form bitmap000 for node 1, bitmap001 for node 2 and so on. When a + node joins the cluster, it acquires the lock in PW mode and it stays + so during the lifetime the node is part of the cluster. The lock + resource number is based on the slot number returned by the DLM + subsystem. Since DLM starts node count from one and bitmap slots + start from zero, one is subtracted from the DLM slot number to arrive + at the bitmap slot number. + + The LVB of the bitmap lock for a particular node records the range + of sectors that are being re-synced by that node. No other + node may write to those sectors. This is used when a new nodes + joins the cluster. + +2.2 Message passing locks + + Each node has to communicate with other nodes when starting or ending + resync, and for metadata superblock updates. This communication is + managed through three locks: "token", "message", and "ack", together + with the Lock Value Block (LVB) of one of the "message" lock. + +2.3 new-device management + + A single lock: "no-new-dev" is used to co-ordinate the addition of + new devices - this must be synchronized across the array. + Normally all nodes hold a concurrent-read lock on this device. + +3. Communication + + Messages can be broadcast to all nodes, and the sender waits for all + other nodes to acknowledge the message before proceeding. Only one + message can be processed at a time. + +3.1 Message Types + + There are six types of messages which are passed: + + 3.1.1 METADATA_UPDATED: informs other nodes that the metadata has + been updated, and the node must re-read the md superblock. This is + performed synchronously. It is primarily used to signal device + failure. + + 3.1.2 RESYNCING: informs other nodes that a resync is initiated or + ended so that each node may suspend or resume the region. Each + RESYNCING message identifies a range of the devices that the + sending node is about to resync. This overrides any previous + notification from that node: only one ranged can be resynced at a + time per-node. + + 3.1.3 NEWDISK: informs other nodes that a device is being added to + the array. Message contains an identifier for that device. See + below for further details. + + 3.1.4 REMOVE: A failed or spare device is being removed from the + array. The slot-number of the device is included in the message. + + 3.1.5 RE_ADD: A failed device is being re-activated - the assumption + is that it has been determined to be working again. + + 3.1.6 BITMAP_NEEDS_SYNC: if a node is stopped locally but the bitmap + isn't clean, then another node is informed to take the ownership of + resync. + +3.2 Communication mechanism + + The DLM LVB is used to communicate within nodes of the cluster. There + are three resources used for the purpose: + + 3.2.1 token: The resource which protects the entire communication + system. The node having the token resource is allowed to + communicate. + + 3.2.2 message: The lock resource which carries the data to + communicate. + + 3.2.3 ack: The resource, acquiring which means the message has been + acknowledged by all nodes in the cluster. The BAST of the resource + is used to inform the receiving node that a node wants to + communicate. + +The algorithm is: + + 1. receive status - all nodes have concurrent-reader lock on "ack". + + sender receiver receiver + "ack":CR "ack":CR "ack":CR + + 2. sender get EX on "token" + sender get EX on "message" + sender receiver receiver + "token":EX "ack":CR "ack":CR + "message":EX + "ack":CR + + Sender checks that it still needs to send a message. Messages + received or other events that happened while waiting for the + "token" may have made this message inappropriate or redundant. + + 3. sender writes LVB. + sender down-convert "message" from EX to CW + sender try to get EX of "ack" + [ wait until all receivers have *processed* the "message" ] + + [ triggered by bast of "ack" ] + receiver get CR on "message" + receiver read LVB + receiver processes the message + [ wait finish ] + receiver releases "ack" + receiver tries to get PR on "message" + + sender receiver receiver + "token":EX "message":CR "message":CR + "message":CW + "ack":EX + + 4. triggered by grant of EX on "ack" (indicating all receivers + have processed message) + sender down-converts "ack" from EX to CR + sender releases "message" + sender releases "token" + receiver upconvert to PR on "message" + receiver get CR of "ack" + receiver release "message" + + sender receiver receiver + "ack":CR "ack":CR "ack":CR + + +4. Handling Failures + +4.1 Node Failure + + When a node fails, the DLM informs the cluster with the slot + number. The node starts a cluster recovery thread. The cluster + recovery thread: + + - acquires the bitmap<number> lock of the failed node + - opens the bitmap + - reads the bitmap of the failed node + - copies the set bitmap to local node + - cleans the bitmap of the failed node + - releases bitmap<number> lock of the failed node + - initiates resync of the bitmap on the current node + md_check_recovery is invoked within recover_bitmaps, + then md_check_recovery -> metadata_update_start/finish, + it will lock the communication by lock_comm. + Which means when one node is resyncing it blocks all + other nodes from writing anywhere on the array. + + The resync process is the regular md resync. However, in a clustered + environment when a resync is performed, it needs to tell other nodes + of the areas which are suspended. Before a resync starts, the node + send out RESYNCING with the (lo,hi) range of the area which needs to + be suspended. Each node maintains a suspend_list, which contains the + list of ranges which are currently suspended. On receiving RESYNCING, + the node adds the range to the suspend_list. Similarly, when the node + performing resync finishes, it sends RESYNCING with an empty range to + other nodes and other nodes remove the corresponding entry from the + suspend_list. + + A helper function, ->area_resyncing() can be used to check if a + particular I/O range should be suspended or not. + +4.2 Device Failure + + Device failures are handled and communicated with the metadata update + routine. When a node detects a device failure it does not allow + any further writes to that device until the failure has been + acknowledged by all other nodes. + +5. Adding a new Device + + For adding a new device, it is necessary that all nodes "see" the new + device to be added. For this, the following algorithm is used: + + 1. Node 1 issues mdadm --manage /dev/mdX --add /dev/sdYY which issues + ioctl(ADD_NEW_DISK with disc.state set to MD_DISK_CLUSTER_ADD) + 2. Node 1 sends a NEWDISK message with uuid and slot number + 3. Other nodes issue kobject_uevent_env with uuid and slot number + (Steps 4,5 could be a udev rule) + 4. In userspace, the node searches for the disk, perhaps + using blkid -t SUB_UUID="" + 5. Other nodes issue either of the following depending on whether + the disk was found: + ioctl(ADD_NEW_DISK with disc.state set to MD_DISK_CANDIDATE and + disc.number set to slot number) + ioctl(CLUSTERED_DISK_NACK) + 6. Other nodes drop lock on "no-new-devs" (CR) if device is found + 7. Node 1 attempts EX lock on "no-new-dev" + 8. If node 1 gets the lock, it sends METADATA_UPDATED after + unmarking the disk as SpareLocal + 9. If not (get "no-new-dev" lock), it fails the operation and sends + METADATA_UPDATED. + 10. Other nodes get the information whether a disk is added or not + by the following METADATA_UPDATED. + +6. Module interface. + + There are 17 call-backs which the md core can make to the cluster + module. Understanding these can give a good overview of the whole + process. + +6.1 join(nodes) and leave() + + These are called when an array is started with a clustered bitmap, + and when the array is stopped. join() ensures the cluster is + available and initializes the various resources. + Only the first 'nodes' nodes in the cluster can use the array. + +6.2 slot_number() + + Reports the slot number advised by the cluster infrastructure. + Range is from 0 to nodes-1. + +6.3 resync_info_update() + + This updates the resync range that is stored in the bitmap lock. + The starting point is updated as the resync progresses. The + end point is always the end of the array. + It does *not* send a RESYNCING message. + +6.4 resync_start(), resync_finish() + + These are called when resync/recovery/reshape starts or stops. + They update the resyncing range in the bitmap lock and also + send a RESYNCING message. resync_start reports the whole + array as resyncing, resync_finish reports none of it. + + resync_finish() also sends a BITMAP_NEEDS_SYNC message which + allows some other node to take over. + +6.5 metadata_update_start(), metadata_update_finish(), + metadata_update_cancel(). + + metadata_update_start is used to get exclusive access to + the metadata. If a change is still needed once that access is + gained, metadata_update_finish() will send a METADATA_UPDATE + message to all other nodes, otherwise metadata_update_cancel() + can be used to release the lock. + +6.6 area_resyncing() + + This combines two elements of functionality. + + Firstly, it will check if any node is currently resyncing + anything in a given range of sectors. If any resync is found, + then the caller will avoid writing or read-balancing in that + range. + + Secondly, while node recovery is happening it reports that + all areas are resyncing for READ requests. This avoids races + between the cluster-filesystem and the cluster-RAID handling + a node failure. + +6.7 add_new_disk_start(), add_new_disk_finish(), new_disk_ack() + + These are used to manage the new-disk protocol described above. + When a new device is added, add_new_disk_start() is called before + it is bound to the array and, if that succeeds, add_new_disk_finish() + is called the device is fully added. + + When a device is added in acknowledgement to a previous + request, or when the device is declared "unavailable", + new_disk_ack() is called. + +6.8 remove_disk() + + This is called when a spare or failed device is removed from + the array. It causes a REMOVE message to be send to other nodes. + +6.9 gather_bitmaps() + + This sends a RE_ADD message to all other nodes and then + gathers bitmap information from all bitmaps. This combined + bitmap is then used to recovery the re-added device. + +6.10 lock_all_bitmaps() and unlock_all_bitmaps() + + These are called when change bitmap to none. If a node plans + to clear the cluster raid's bitmap, it need to make sure no other + nodes are using the raid which is achieved by lock all bitmap + locks within the cluster, and also those locks are unlocked + accordingly. + +7. Unsupported features + +There are somethings which are not supported by cluster MD yet. + +- change array_sectors. diff --git a/Documentation/md/raid5-cache.txt b/Documentation/md/raid5-cache.txt new file mode 100644 index 000000000..2b210f295 --- /dev/null +++ b/Documentation/md/raid5-cache.txt @@ -0,0 +1,109 @@ +RAID5 cache + +Raid 4/5/6 could include an extra disk for data cache besides normal RAID +disks. The role of RAID disks isn't changed with the cache disk. The cache disk +caches data to the RAID disks. The cache can be in write-through (supported +since 4.4) or write-back mode (supported since 4.10). mdadm (supported since +3.4) has a new option '--write-journal' to create array with cache. Please +refer to mdadm manual for details. By default (RAID array starts), the cache is +in write-through mode. A user can switch it to write-back mode by: + +echo "write-back" > /sys/block/md0/md/journal_mode + +And switch it back to write-through mode by: + +echo "write-through" > /sys/block/md0/md/journal_mode + +In both modes, all writes to the array will hit cache disk first. This means +the cache disk must be fast and sustainable. + +------------------------------------- +write-through mode: + +This mode mainly fixes the 'write hole' issue. For RAID 4/5/6 array, an unclean +shutdown can cause data in some stripes to not be in consistent state, eg, data +and parity don't match. The reason is that a stripe write involves several RAID +disks and it's possible the writes don't hit all RAID disks yet before the +unclean shutdown. We call an array degraded if it has inconsistent data. MD +tries to resync the array to bring it back to normal state. But before the +resync completes, any system crash will expose the chance of real data +corruption in the RAID array. This problem is called 'write hole'. + +The write-through cache will cache all data on cache disk first. After the data +is safe on the cache disk, the data will be flushed onto RAID disks. The +two-step write will guarantee MD can recover correct data after unclean +shutdown even the array is degraded. Thus the cache can close the 'write hole'. + +In write-through mode, MD reports IO completion to upper layer (usually +filesystems) after the data is safe on RAID disks, so cache disk failure +doesn't cause data loss. Of course cache disk failure means the array is +exposed to 'write hole' again. + +In write-through mode, the cache disk isn't required to be big. Several +hundreds megabytes are enough. + +-------------------------------------- +write-back mode: + +write-back mode fixes the 'write hole' issue too, since all write data is +cached on cache disk. But the main goal of 'write-back' cache is to speed up +write. If a write crosses all RAID disks of a stripe, we call it full-stripe +write. For non-full-stripe writes, MD must read old data before the new parity +can be calculated. These synchronous reads hurt write throughput. Some writes +which are sequential but not dispatched in the same time will suffer from this +overhead too. Write-back cache will aggregate the data and flush the data to +RAID disks only after the data becomes a full stripe write. This will +completely avoid the overhead, so it's very helpful for some workloads. A +typical workload which does sequential write followed by fsync is an example. + +In write-back mode, MD reports IO completion to upper layer (usually +filesystems) right after the data hits cache disk. The data is flushed to raid +disks later after specific conditions met. So cache disk failure will cause +data loss. + +In write-back mode, MD also caches data in memory. The memory cache includes +the same data stored on cache disk, so a power loss doesn't cause data loss. +The memory cache size has performance impact for the array. It's recommended +the size is big. A user can configure the size by: + +echo "2048" > /sys/block/md0/md/stripe_cache_size + +Too small cache disk will make the write aggregation less efficient in this +mode depending on the workloads. It's recommended to use a cache disk with at +least several gigabytes size in write-back mode. + +-------------------------------------- +The implementation: + +The write-through and write-back cache use the same disk format. The cache disk +is organized as a simple write log. The log consists of 'meta data' and 'data' +pairs. The meta data describes the data. It also includes checksum and sequence +ID for recovery identification. Data can be IO data and parity data. Data is +checksumed too. The checksum is stored in the meta data ahead of the data. The +checksum is an optimization because MD can write meta and data freely without +worry about the order. MD superblock has a field pointed to the valid meta data +of log head. + +The log implementation is pretty straightforward. The difficult part is the +order in which MD writes data to cache disk and RAID disks. Specifically, in +write-through mode, MD calculates parity for IO data, writes both IO data and +parity to the log, writes the data and parity to RAID disks after the data and +parity is settled down in log and finally the IO is finished. Read just reads +from raid disks as usual. + +In write-back mode, MD writes IO data to the log and reports IO completion. The +data is also fully cached in memory at that time, which means read must query +memory cache. If some conditions are met, MD will flush the data to RAID disks. +MD will calculate parity for the data and write parity into the log. After this +is finished, MD will write both data and parity into RAID disks, then MD can +release the memory cache. The flush conditions could be stripe becomes a full +stripe write, free cache disk space is low or free in-kernel memory cache space +is low. + +After an unclean shutdown, MD does recovery. MD reads all meta data and data +from the log. The sequence ID and checksum will help us detect corrupted meta +data and data. If MD finds a stripe with data and valid parities (1 parity for +raid4/5 and 2 for raid6), MD will write the data and parities to RAID disks. If +parities are incompleted, they are discarded. If part of data is corrupted, +they are discarded too. MD then loads valid data and writes them to RAID disks +in normal way. diff --git a/Documentation/md/raid5-ppl.txt b/Documentation/md/raid5-ppl.txt new file mode 100644 index 000000000..bfa092589 --- /dev/null +++ b/Documentation/md/raid5-ppl.txt @@ -0,0 +1,45 @@ +Partial Parity Log + +Partial Parity Log (PPL) is a feature available for RAID5 arrays. The issue +addressed by PPL is that after a dirty shutdown, parity of a particular stripe +may become inconsistent with data on other member disks. If the array is also +in degraded state, there is no way to recalculate parity, because one of the +disks is missing. This can lead to silent data corruption when rebuilding the +array or using it is as degraded - data calculated from parity for array blocks +that have not been touched by a write request during the unclean shutdown can +be incorrect. Such condition is known as the RAID5 Write Hole. Because of +this, md by default does not allow starting a dirty degraded array. + +Partial parity for a write operation is the XOR of stripe data chunks not +modified by this write. It is just enough data needed for recovering from the +write hole. XORing partial parity with the modified chunks produces parity for +the stripe, consistent with its state before the write operation, regardless of +which chunk writes have completed. If one of the not modified data disks of +this stripe is missing, this updated parity can be used to recover its +contents. PPL recovery is also performed when starting an array after an +unclean shutdown and all disks are available, eliminating the need to resync +the array. Because of this, using write-intent bitmap and PPL together is not +supported. + +When handling a write request PPL writes partial parity before new data and +parity are dispatched to disks. PPL is a distributed log - it is stored on +array member drives in the metadata area, on the parity drive of a particular +stripe. It does not require a dedicated journaling drive. Write performance is +reduced by up to 30%-40% but it scales with the number of drives in the array +and the journaling drive does not become a bottleneck or a single point of +failure. + +Unlike raid5-cache, the other solution in md for closing the write hole, PPL is +not a true journal. It does not protect from losing in-flight data, only from +silent data corruption. If a dirty disk of a stripe is lost, no PPL recovery is +performed for this stripe (parity is not updated). So it is possible to have +arbitrary data in the written part of a stripe if that disk is lost. In such +case the behavior is the same as in plain raid5. + +PPL is available for md version-1 metadata and external (specifically IMSM) +metadata arrays. It can be enabled using mdadm option --consistency-policy=ppl. + +There is a limitation of maximum 64 disks in the array for PPL. It allows to +keep data structures and implementation simple. RAID5 arrays with so many disks +are not likely due to high risk of multiple disks failure. Such restriction +should not be a real life limitation. |