1. Field of Invention
The invention generally relates to high reliability electronic data storage, and, more particularly, to an architecture for decomposing a data access request into a number of smaller tasks.
2. Description of Related Art
A file server is a computer that provides file service relating to the organization of information on writeable persistent storage devices, such as memories, tapes or disks of an array. The file server or filer may be embodied as a storage system including an operating system that implements a file system to logically organize the information as a hierarchical structure of directories and files on a storage devices (e.g., disks). Each “on-disk” file may be implemented as set of data structures, e.g., disk blocks, configured to store information, such as the actual data for the file. A directory, on the other hand, may be implemented as a specially formatted file in which information about other files and directories are stored.
A file server may be further configured to operate according to a client/server model of information delivery to thereby allow many clients to access files stored on the server. In this model, the client may comprise an application executing on a computer that “connects” to the storage system over a computer network, such as a point-to-point link, shared local area network, wide area network or virtual private network implemented over a public network, such as the Internet. Each client may request the services of the file system on the filer by issuing file system protocol messages (in the form of packets) to the system over the network. It should be noted, however, that the filer may alternatively be configured to operate as an assembly of storage devices that is directly attached to a (e.g., client or “host”) computer. Here, a user may request the services of the file system to access (i.e., read and/or write) data from/to the storage devices (e.g., data access request).
A common type of file system is a “write in-place” file system, an example of which is the conventional Berkeley fast file system. In a write in-place file system, the locations of the data structures, such as data blocks, on disk are typically fixed. Changes to the data blocks are made “in-place” in accordance with the write in-place file system. If an update to a file extends the quantity of data for the file, an additional data block is allocated.
In the operation of a disk array, it is fairly common that a disk, or other storage medium, such as tape, will fail. A goal of a high performance storage system is to make the mean time to data loss (MTTDL) as long as possible, preferably much longer than the expected service life of the system. Data can be lost when one or more storage devices fail, making it impossible to recover data from the device. Typical schemes to avoid loss of data include mirroring, backup and parity protection. Mirroring stores the same data on two or more disks, so that if one disk fails, the mirror disk can be used to read data. Backup periodically copies data on one disk to another disk assuming thereby that both disks are unlikely to fail simultaneously. Parity schemes are common because they provide a redundant encoding of the data that allows for the loss of one or more disks without the loss of data while only requiring a minimal number of additional disk drives in the storage system.
Parity protection is used in computer systems to protect against loss of data on a storage device, such as a disk. A parity value may be computed by summing (usually modulo 2) data of a particular word size (usually one bit) across a number of similar disks holding different data and then storing the results on the disks. That is, parity may be computed on 1-bit wide vectors, composed of bits in predetermined positions on each of the disks. Addition and subtraction on 1-bit vectors are an equivalent to an exclusive-OR (XOR) logical operation, and the addition and subtraction operations can be replaced by XOR operations. The data is then protected against the loss of any of the disks. If the disk storing the parity is lost, the parity can be regenerated from the data. If one of the data disks is lost, the data can be regenerated by adding the contents of the surviving data disks together and then subtracting the result from the stored parity.
In one embodiment, typically, the disks are divided into parity groups, each of which comprises one or more data disks and a parity disk. The disk space is divided into stripes, with each stripe containing one block from each disk. The blocks of a stripe are usually at the equivalent locations on each disk in the parity group. Within a stripe, all but one block are blocks containing data (“data blocks”) and one block is a block containing parity (“parity block”) computed by the XOR of all the data from all the disks. If the parity blocks are all stored on one disk, thereby providing a single disk that contains all (and only) parity information, the system is referred to as a RAID level four implementation. If the parity blocks are contained within different disks in each stripe, usually in a rotating pattern, then the implementation is RAID level five. In addition to RAID levels four and five, one skilled in the art knows that there are several other well-known RAID levels and hybrid combinations of those RAID levels.
In a known implementation, the file system operating on top of a RAID subsystem treats the RAID disk array as a large collection of blocks wherein each block is numbered sequentially across the RAID disk array. The data blocks of a file are scattered across the data disks to fill each stripe as fully as possible, thereby placing each data block in a stripe on a different disk. Once N data blocks of a first stripe are allocated to N data disks of the RAID array, remaining data blocks are allocated on subsequent stripes in the same fashion until the entire file is written in the RAID array. Thus, a file is written across the data disks of a RAID system in stripes comprising modulo N data blocks. As stripes are filled, they are sent to the RAID subsystem to be stored.
In a known implementation, the RAID subsystem performs locking and I/O tasks on the stripe level, with these tasks being implemented through a collection of dedicated stripe owner threads. Each thread performs synchronous I/O on one stripe at a time, with additional I/O requests on the same stripe being queued up on that stripe owner (providing mutual exclusion). The limited number of threads used for stripe I/O and XOR operations can lead to bottlenecks, particularly during reconstruction, affecting system response time.
In a known implementation, RAID state transitions due to disk failures and removals are sometimes not properly coordinated with the I/O path. This can result in buffers that refer to unusable disks, which can lead to errors, from which the system may not be able to recover. As stated above, I/O from the RAID stripe owners to the disk driver is synchronous. This, combined with the fact that the I/O path may handle state transitions, can lead to deadlock situations.
The resources used by these known systems (threads and memory buffers) are statically allocated during boot and a simple reservation mechanism exists to reserve buffers before performing an I/O. Such an allocation, typically accounting for worst-case error handling, results in a large allocation of resources that are never used, but nevertheless allocated and not available to other I/O threads in the system. This can reduce the system's ability to adapt to load and configuration changes.
The present invention implements an I/O task architecture in which a data access request requested by the file system, for example a stripe write, is decomposed into a number of lower-level asynchronous I/O tasks that can be scheduled independently. Resources needed by these lower-level I/O tasks are dynamically assigned, on an as-needed basis, to balance the load and use resources efficiently, achieving high scalability without the loss of performance. A hierarchical order is assigned to the I/O tasks to ensure that there is a forward progression of the higher-level I/O task and to ensure that resources do not become deadlocked. This architecture allows data blocks on the same disk to be processed efficiently, for example so that a disk write involves an optimal chain length.
In one aspect the invention relates to a method for avoiding runtime resource deadlock in a RAID system. The method includes defining a plurality of resources, the plurality of resources including a plurality of I/O tasks, assigning a hierarchical order to the plurality of resources, dynamically allocating a first resource for a first I/O task when the first I/O task is commenced and preventing the dynamic allocation from allocating a second resource to the first I/O task higher in the hierarchical order than the first resource. In another embodiment, the method includes dynamically allocating a third resource for the first I/O task, the third resource lower in the hierarchical order than the first resource and deallocating the first and third resources when a child I/O task of the first I/O task completes. In another embodiment, the method includes requesting a third resource for the first I/O task, the third resource lower in the hierarchical order than the first resource and suspending the first I/O task until the third resource is available for allocation to the first I/O task. In another embodiment, the method includes restarting the first I/O task when the third resource is available for allocation to the first I/O task. In another embodiment, the method includes selecting the first and second resources from at least one of stripe locks, memory data structures, data buffers and NVRAM slots.
In another embodiment, the method includes commencing the first I/O task responsive to a data access request. In another embodiment, the data access request is a user request. In another embodiment, the data access request is a system request. In another embodiment, the method includes defining a plurality of resources, each resource associated with one of a volume read, a volume write, a mirror read, a mirror write, a mirror read error, a plex read, a RAID group read, an array read, an array write, a disk read, a disk write, a plex read, a plex write, a parity reconstruction of a RAID group, a stripe parity reconstruction and a mirror verification. In another embodiment, the method includes assigning a hierarchical order in the plurality that is, from highest level to lowest level a volume read, a volume write, a mirror read, a mirror write, a mirror read error, a plex read, a RAID group read, an array read, an array write, a disk read, a disk write, a stripe read, a stripe write, a parity reconstruction of a RAID group, a stripe parity reconstruction and a mirror verification.
In another aspect, the invention relates to an article of manufacture having computer-readable program portions embodied therein for avoiding runtime resource deadlock in a RAID system, wherein the computer-readable program portions of the article perform the above-listed methods. In another aspect, the invention relates to a storage system for avoiding runtime resource deadlock in a RAID system. The storage system includes an instantiator module and a resource manager.
The above and further advantages of the invention may be better understood by referring to the following description taken in conjunction with the accompanying drawing, in which:
Please note that for clarity, the first digit of the reference number refers to the figure in which that reference appears.
In broad overview,
In the illustrative embodiment, the memory 125 includes an instantiator module 145 and an I/O manager 150 that implements an I/O task architecture for processing data access requests. The instantiator module 145 is the portion of the software and/or hardware that defines the data structures of an object, for example, an I/O task and manages the generation of each instance of an I/O task. The I/O manager 150 is the portion of software and/or hardware that implements an I/O threading mechanism. The memory 125 also includes a resource manager module 155 that implements a resource manager to manage the allocation of resources. In one embodiment, each instance of an I/O task is a type of resource, because, for example, a portion of memory is needed and CPU time is required to generate the instance of an I/O task. In such an embodiment, the resource manager 155 and the instantiator module 145 are the same entity.
The memory 125 also includes a storage operating system 160 that implements a storage manager to logically organize data on the disks 110. In one embodiment, the storage system 105 receives data in the form of files (e.g., where the storage system 105 is a file server) and the storage manager logically organizes the received data as a hierarchical structure of directories and files on the disks 110. In another embodiment, the storage system 105 receives data in the form of data blocks and the storage manager logically organizes the received data blocks within the storage locations on the disks 110.
This illustrative embodiment is configured as shown because the memory 125 comprises storage locations that are addressable by the processor 120 and adapters 130, 135 for storing software program code and data structures associated with the present invention. These blocks can be implemented, fully or partially, elsewhere without departing from the scope of the invention.
For example, the processor 120 and adapters 130, 135 may, in turn, comprise processing elements and/or logic circuitry configured to execute the software code and manipulate the data structures. Similarly, the modules 145, 150, 155 may be implemented in hardware using, for example, FPGA, and/or ASIC devices. In another embodiment, a distributed computing model is used. Other storage systems (not shown) including their own processors (not shown) are in communication with the storage system 105 and its processor 120 over the network 140 and/or the network 180. The I/O task processing is distributed over any of the storage systems and their processors using various optimization techniques, such as load balancing and the like. In some embodiments where the storage system 105 is implemented in a distributed environment, the file system and file system semantics are managed by one or more computing devices separate and in communication with another computing device comprising modules 145, 150 and 155. It will be apparent to those skilled in the art that other processing and memory means, including various computer readable media, may be used for storing and executing program instructions pertaining to the inventive techniques described herein.
The network adapter 135 comprises the mechanical, electrical and/or signaling circuitry needed to connect the storage system 105 to a client 175 over a computer network 180, which may comprise a point-to-point connection or a shared medium, such as a local area network. The client 175 may be a general-purpose computer configured to execute applications. Moreover, the client 175 may interact with the storage system 105 in accordance with a client/server model of information delivery. That is, the client 175 may request the services of the storage system 105, and the storage system 105 may return the results of the services requested by the client 175. In one embodiment, where the storage system 105 is a file server, this is accomplished by exchanging packets encapsulating file system protocol format (e.g., the Common Internet File System (CIFS) protocol or Network File System (NFS)) over the network 175.
The storage adapter 135 cooperates with the storage operating system 160 executing on the storage system 105 to access information requested by the client 175. The information may be stored on any type of attached array of writeable media such as video tape, optical, DVD, magnetic tape, bubble memory and any other similar media adapted to store information, including data and parity information. In the illustrative embodiment described herein, however, the information is preferably stored on the disks 110 of array 115. The storage adapter 135 includes input/output (I/O) interface circuitry that couples to the disks 110 over an I/O interconnect arrangement, such as a conventional high-performance, Fibre Channel serial link topology. The information is retrieved by the storage adapter 135 and, if necessary, processed by the processor 120 (or the adapter 135 itself) prior to being forwarded over the system bus 140 to the network adapter 130, where the information is formatted into a packet and returned to the client 175.
Storage of information on array 115 is preferably implemented as one or more storage “volumes” that comprise a cluster of physical storage disks 110, defining an overall logical arrangement of disk space. Each volume is generally associated with its own storage manager. The disks 110 within a volume are typically organized as one or more groups of Redundant Array of Independent (or Inexpensive) Disks (RAID), sometimes referred to as RAID groups. RAID implementations enhance the reliability/integrity of data storage through the redundant writing of data “stripes” across a given number of physical disks in the RAID group, and the appropriate storing of redundant information with respect to the striped data. The redundant information enables recovery of data lost when a storage device fails.
The illustrative embodiment represents a hybrid RAID level. The RAID group objects 220 represent a RAID level four implementation, striping data on two disks (e.g., 225a and 225b) with a third parity disk (e.g., 225c). The mirror object represents a RAID level one implementation where the data is mirrored, eventually down the tree, by two disk objects 225, one on each side of the mirror object 210. For example disk object 225g mirrors disk object 225a, meaning both will contain identical data. The illustrated embodiment depicts one possible arrangement of the set of objects 200 for this hybrid RAID level. Those skilled in the art understand that other arrangements and/or implementations of objects are also possible. Further, the inventive techniques are not limited to this specific RAID level implementation and other RAID levels and hybrid RAID levels can advantageously be used with the inventive techniques described herein. Likewise, symmetry between the plex objects 215 is not required. For example, disk-level mirroring can be represented as logical mirrored disk objects with physical disk objects below.
In some embodiments, the configuration information includes virtual block numbers (“VBNs”) that map to disk block numbers (“DBNs”).
The RAID group object 220 is associated with the aggregate of the VBNs of its disk objects 225. For example, the VBN range of the RAID group 220a is the aggregate of the disk objects 225a, 225b and 225c. In the above example using these disk objects, the RAID group object 220a has a VBN range of 0-50 and 80-90, which is the combination of the VBNs of the disk objects 225a, 225b and 225c. The plex object 215 is associated with the aggregate of the VBNs of its RAID group objects 220. For example, the VBN range of the plex object 215a is the aggregate of the RAID group objects 220a and 220b. In the above example using these RAID group objects, where the RAID group 220b has the same size range as RAID group 220a, the plex 215a has a VBN range of 0-120.
In this embodiment, because the mirror object 210 mirrors plex 215a onto 215b, or vice-versa, the VBN range 230 of the mirror object is not the aggregate of its plex objects 215. Instead, the VBN range of the mirror object 210 is the union of the VBNs of its plex objects 215. Thus, if the implementation of the disks 110 associated with each of the plex objects 215 is identical and the map of the ranges is identical, then VBN 230a is identical to 230b and the VBN range 230 of the mirror object 210 is identical to 230a. The VBN range 230 of the volume object 205 is the VBN range 230 of the mirror object 210. In one embodiment, the translation of DBNs to VBNs is performed at the mirror object 210 level. In this embodiment, those objects at levels below the mirror object 210 refer to the target storage space using its DBN reference and those objects above the mirror object refer to the target storage space using its VBN reference.
In another embodiment, the storage manager associates received data with each VBN range 230a and 230b separately. Even if asymmetrical, the mirrored data must go into the same blocks in each of the VBNs 230a and 230b. In another embodiment, the plex objects 215a and 215b are not symmetrical and one plex has a larger VBN range than the other does by delta VBNs. The size of VBN range 230 of the mirror object 210 is the mirrored range of VBNs plus the non-mirrored delta VBNs.
The association of a range or ranges of VBNs to objects at each level is sometimes referred to as a RAID map. By transmitting the RAID map of the set of objects 200 to the storage manager, the storage manager has configuration information regarding the disk array 115 topology and can associate data with the VBNs as arrays of stripes (“array”), optimizing for disk writes and/or disk reads.
Referring to
An I/O task implements an I/O operation on an object, for example one of the objects within the set of objects 200 (
A second parameter of the exemplary I/O task is a memory reference pointer to configuration information associated with the RAID configuration element (e.g., volume 205, plex 215, RAID group 220, disk 225) upon which the I/O task is operating. A third parameter of the exemplary I/O task is an enumerated value indicating the type of the I/O task. A fourth parameter of the exemplary I/O task is a value of the initial starting location (VBN) for the I/O operation. A fifth parameter of the exemplary I/O task is a value for the number of virtual blocks that are involved in the I/O operation. A sixth parameter of the exemplary I/O task is a memory pointer reference to a set of memory buffers that contain data that is involved in the I/O operation.
A seventh parameter of the exemplary I/O task is a memory pointer reference to a superior I/O task in a chained collection of I/O operations (e.g., a parent I/O task). An eighth parameter of the exemplary I/O task is a memory pointer reference to a set of standard operational vectors that define the specific functional behaviors associated with this type of I/O task. In other embodiments, an I/O task has more or less than the eight exemplary parameters. For example, the I/O task can include a memory pointer reference to a buffer that contains data that is specific to this type and instance of I/O task (e.g., private data).
In addition to the parameters, an I/O task includes one or more function vectors. The function vectors invoke a specific set of operations based on the type of the I/O task. Each type of I/O task (e.g., disk read, mirrored write, parity reconstruction, and the like) has its own type-specific operations vector, defining a specific set of functional behaviors. An exemplary I/O task includes four function vectors. A first function vector of the exemplary I/O task is a function vector that is invoked to start an I/O task or to restart a previously suspended I/O task. A second function vector of the exemplary I/O task is a function vector that is invoked to indicate that a given 110 task has completed all requested behaviors. For example, the disk driver uses this routine to notify a disk object 225 (
A third function vector of the exemplary I/O task is a function vector that is invoked to indicate that a given subordinate I/O task (e.g., child I/O task) has completed all requested behaviors. In one embodiment, the functions invoked by this vector check for child I/O task errors and try to recover from them. For example, a RAID group object 220 (
A fourth function vector of the exemplary I/O task is a function vector indicating that an exclusive-or operation has completed. Although included in the exemplary embodiment, this routine is not necessary if a RAID level is used that does not incorporate parity checking. In other embodiments, an I/O task has more or less than the four exemplary function vectors. In another embodiment, the I/O task structure uses a standard interface, so that each I/O task is preferably called the same way without regard to the object. For example, the storage manager can perform a stripe_write I/O on an object without knowing if it is a RAID level four or RAID level zero.
Once an I/O task is instantiated, it is queued to and processed by an I/O thread. In one embodiment, the I/O manager 150 (
In one embodiment, I/O tasks are queued to the I/O thread as messages and the I/O thread services the I/O task by invoking a specific set of operations using the function vectors of the I/O task. The I/O thread suspends the I/O task on the occurrence of an event, for example, an unavailable resource. If an I/O task is suspended, the I/O thread does not schedule or service that I/O task. In one embodiment, the I/O thread suspends an I/O task by queuing the suspended I/O task in a separate wait queue. When another event occurs, for example the resource becomes available, the I/O task is restarted by sending the message to the I/O thread again. The I/O thread receives I/O requests and then calls the appropriate object routine to service the request. Upon return from the I/O routine, the I/O thread services the next request. In one embodiment, the queue for the I/O thread is a FIFO queue. In other embodiments, I/O tasks have a priority value contained, for example, in a header element of the I/O task, and the priority value is based on certain characteristics of the I/O task. The characteristics can include, for example, the I/O task's level in the hierarchical order of objects 200, its status as a client requested data access request (e.g., store data) or a system requested data access request (e.g., a parity verification, mirror verification and the like), and the like.
As illustrated in the VBN array 302, VBN 28 is full and not available. The storage manager assigns the received data to the remaining VBNs 26, 27 and 29. The VBN array 302 is illustrative only and in other embodiments, different stripe and/or array of stripes VBN configurations can be used. For example, VBN 27 and VBN 28 can be switched, so VBNs go horizontally across a stripe. In another embodiment, the parity blocks P1 and P2 can be distributed among different disks, for example using a RAID level five implementation.
To fulfill the request of the storage manager to write data to the allocated VBNs of a particular volume, the storage manager calls an I/O task from volume object 205 (
The instantiator module 145 instantiates a mirror_write I/O task 308, containing the necessary information to accomplish this task. The volume_write I/O task 304 is suspended (e.g., put into a wait queue) until its child I/O task is complete. The mirror_write I/O task 308 is sent to the I/O thread and is scheduled. When the mirror_write I/O task 308 reaches the front of the queue, the I/O manager 150 starts the mirror_write I/O task 308 operation. In this embodiment, the mirror_write I/O task 308 calls (step 310) an I/O task from the master plex object 215a and calls (step 312) a write I/O task from the slave plex object 215b to perform the requested write.
The identifiers of master and slave for plex objects 215a and 215b are used because, in the illustrated embodiment, the implementation of the disk objects 225 (
In the illustrated embodiment, the instantiator module 145 instantiates a master array_write I/O task 314 and a slave array_write I/O task 316, both containing the necessary information to accomplish this task. The mirror_write I/O task 304 is suspended (e.g., put into a wait queue) until its child I/O tasks are complete. The array_write I/O tasks 314 and 316 use the term array because as illustrated they can handle the processing of multiple stripes at the same time, which allows the VBNs 302 to be processed as an entity, as opposed to processing the VBNs 302 as two separate stripes. This allows writes to disks to be optimized.
The master array_write I/O task 314 and the slave array_write I/O task 316 are both sent to the I/O thread and are scheduled independently. In other words, each I/O task is scheduled as an individual entity, without regard to each other. There is thus no need for the I/O thread to take into account the relationship of each I/O task to each other when scheduling. The I/O thread does not need to synchronize an I/O task with another, because each I/O task is a separate entity and its progress is dependent on its current state and the occurrence of a certain event to change that state, (e.g., completion of a child I/O task), as described in more detail below. The I/O thread processes the I/O tasks in its queue in order, or as determined by a priority algorithm, and if the I/O thread cannot service the I/O task because a certain event has not yet occurred (e.g., a child task is not complete, all needed resources are not available), the I/O task is suspended (e.g., moved to a wait queue) until the event occurs.
When the master array_write I/O task 314 reaches the front of the queue, the I/O manager 150 starts the master array_write I/O task 314 operation. When the slave array_write I/O task 316 reaches the front of the queue, the I/O manager 150 starts the slave array_write I/O task 316 operation. In this embodiment, the slave array_write I/O task 316 suspends until the master array_write I/O task 314 performs the parity calculation.
In one embodiment, the master array_write I/O task 314 determines that because VBN 28 is full, that VBN must be read so that the master array_write I/O task 314 can perform a parity calculation for the stripe corresponding to VBN 28. The master array_write I/O task 314 calls (step 318) an I/O task from the RAID group object 220a (
In the illustrated embodiment, the instantiator module 145 instantiates a group_read I/O task 320, containing the necessary information to accomplish this task. The array_write I/O task 314 is suspended until its child I/O task is complete. The group_read I/O task 320 is sent to the I/O thread and is scheduled. When the group_read I/O task 320 reaches the front of the queue, the I/O manager 150 starts the group_read I/O task 320 operation. In this embodiment, the group_read I/O task 320 calls (step 322) an I/O task from the disk object 225b (
The instantiator module 145 instantiates a disk_read I/O task 324, containing the necessary information to read VBN 28. The group_read I/O task 320 is suspended until its child I/O task is complete. The disk_read I/O task 324 is sent to the I/O thread and is scheduled. When the disk_read I/O task 324 reaches the front of the queue, the I/O manager 150 starts the disk_read I/O task 324 operation. With the disk_read I/O task 324 started, the disk_read I/O task 324 reads VBN 28 from the disk 110 and stores the read data in a predetermined buffer. The disk_read I/O task 324 is complete. The disk_read I/O task 324 notifies its parent I/O task, in this case, the group_read I/O task 320, that the requested I/O task is complete. In this embodiment, the disk_read I/O task 324 notifies the RAID group object 220a by invoking (step 326) the function vector indicating that the subordinate I/O task is complete (e.g., the third function vector of the exemplary I/O task) from the group_read I/O task 320. In another embodiment, the array_write I/O task 314 directly calls the disk_read I/O task 324 from the disk object 225b and a group_read I/O task is not used. In this embodiment, when the disk_read I/O task 324 is complete it notifies its parent I/O task, the array_write I/O task 314, that the requested I/O task is complete.
In the illustrated embodiment, upon completion of its child I/O task, the group_read I/O task 320 determines that no other child I/O tasks are needed to perform the requested group_read I/O task 320. The group_read I/O task 320 is complete and notifies its parent I/O task, array_write 314, that the requested I/O task is complete. This is accomplished by the group_read I/O task 320 notifying the array_write I/O task 314 by invoking (step 328) the function vector indicating that the subordinate I/O task is complete (e.g., the third function vector of the exemplary I/O task) from the array_write I/O task 314. Upon completion of its child I/O task group_read 320, the array_write I/O task 314 determines that it now has enough information to calculate parity.
The instantiator module 145 instantiates (step 330) an XOR calculation I/O task 332, containing the necessary information to calculate the parity for P1 and P2. In another embodiment, the XOR calculation I/O task for P2 is separate from P1 because no reads are necessary to calculate P2. If the P2 calculation is separate, the array_write I/O task 314 can call an XOR calculation I/O task to calculate P2 before or in parallel with the group_read I/O task 320.
As illustrated, the group_read I/O task 320 is suspended until its child I/O task, the XOR calculation I/O task 332, is complete. The XOR calculation I/O task 332 is sent to the I/O thread and is scheduled. When the XOR calculation I/O task 332 reaches the front of the queue, the I/O manager 150 starts the XOR calculation I/O task 332 operation. With the XOR calculation I/O task 332 started, a processing device in the storage system 105, for example processor 120, calculates the parity for P1 and P2 and stores (step 334) the calculated data in a buffer 326. In another embodiment, the storage system 105 has a separate XOR thread and the XOR calculation I/O task 332 is sent to the XOR thread and not the I/O thread.
Once the calculation is stored (step 334), the XOR calculation I/O task 332 notifies its parent I/O task, in this case, the array_write I/O task 314, that the requested I/O task is complete. In this embodiment, the XOR calculation I/O task 332 notifies the array_write I/O task 314 by invoking (step 348) the function vector indicating that the exclusive-or operation is complete (e.g., the fourth function vector of the exemplary I/O task) from the array_write I/O task 314. Upon completion of its child I/O task XOR calculation I/O task 332, the array_write I/O task 314 determines that it now has enough information to write the data to the disks 110 associated with the VBNs 302. The master array_write I/O task 314 calls (step 350) an I/O task from the RAID group object 220a to perform the requested writes.
The instantiator module 145 instantiates a group_write I/O task 352, containing the necessary information to accomplish this task. The array_write I/O task 314 is suspended until its child I/O task is complete. The group_write I/O task 352 is sent to the I/O thread and is scheduled. When the group_write I/O task 352 reaches the front of the queue, the I/O manager 150 starts the group_write I/O task 352 operation. In this embodiment, the group_write I/O task 352 calls (step 354) an I/O task from the disk object 225a (
In the illustrated embodiment, the instantiator module 145 instantiates a disk_write I/O task 360 corresponding to the call (step 354) for an I/O task to the disk object 225a, containing the necessary information to write VBNs 26 and 27. The instantiator module 145 also instantiates a disk_write I/O task 362 corresponding to the call (step 356) for an I/O task to the disk object 225b, containing the necessary information to write VBN 29. The instantiator module 145 also instantiates a disk_write I/O task 340 corresponding to the call (step 358) for an I/O task to the disk object 225c, containing the necessary information to write P1 and P2. The group_write I/O task 352 is suspended until all of its child I/O tasks 360, 362 and 364 are complete.
The disk_write I/O tasks 360, 362 and 364 are each individually sent to the I/O thread and are individually scheduled. When each of the disk_write I/O tasks 360, 362 and 364 reaches the front of the queue, the I/O manager 150 starts that particular disk_write I/O task operation. When each of the disk_write I/O tasks 360, 362 and 364 are started, the disk_write I/O tasks 360, 362 and 364 writes data to its disks corresponding to the VBNs.
The disk_write I/O tasks 360, 362 and 364 complete individually, as they are scheduled individually. The order of completion is dependent on the occurrence of events necessary to complete the disk_write I/O tasks 360, 362 and 364. Some events are, for example, the availability of all of the resources needed to complete a disk_write I/O task and the availability of the disk 110 on which the data is to be written. As each of the disk_write I/O tasks 360, 362 and 364 completes, each notifies its parent I/O task, in this case, the group_write I/O task 352, that the requested I/O task is complete. In this embodiment, each of the disk_write I/O tasks 360, 362 and 364 notifies the RAID group object 220a by invoking (steps 366, 368, 370 respectively) the function vector indicating that the subordinate I/O task is complete (e.g., the third function vector of the exemplary I/O task) from the group_write I/O task 320.
Upon completion of its child I/O tasks (i.e., 360, 362 and 364), the group_write 352 determines that no other child I/O tasks are needed to perform the requested group_write I/O task 352. The group_write I/O task 352 is complete and notifies its parent I/O task, array_write 314, that the requested I/O task is complete. The group_write I/O task 352 notifies the array_write I/O task 314 by invoking (step 372) the function vector indicating that the subordinate I/O task is complete (e.g., the third function vector of the exemplary I/O task) from the array_write I/O task 314. Upon completion of its child I/O task group_write 320, the array_write I/O task 314 determines that no other child I/O tasks are needed to perform the requested array_write I/O task 314. The array_write I/O task 314 is complete and notifies (step 374) its parent I/O task, mirror_write 308, that the requested I/O task is complete.
In the illustrated master/slave embodiment, the mirror_write I/O task 308 is complete when both of the array_write I/O tasks 314 and 316 are complete. As explained above each I/O task is scheduled individually, so the array_write I/O task 314 may complete before or after the slave_array_write I/O task 316, depending on, for example, the availability of resources and access to the disks 110. As described above, the slave_array_write I/O task 316 is identified as a slave task and thus waits for the master array_write I/O task 314 to perform the XOR calculation(s) necessary to write the VBN array 302. In an embodiment where there is no mirroring and/or no symmetry, the array_write I/O task 316 does not have to wait for this calculation and generates all of its needed child I/O tasks as resources allow.
In the illustrated embodiment, upon completion of the XOR calculation and storage (step 334) in the buffer 336, the slave_array_write I/O task 316 retrieves (step 378) the needed parity data from the buffer 336. The slave_array_write I/O task 316 proceeds to write (step 380) the data to the disks, similarly as shown for the array_write I/O task 314. The slave_array_write I/O task 316 generates a group_write I/O task, which in turn generates three disk_write I/O tasks that write the data to the disks. Upon completion of its child group_write I/O task, the group_write I/O task determines that no other child I/O tasks are needed to perform the requested slave_array_write I/O task 316. The slave_array_write I/O task 316 is complete and notifies (step 382) its parent I/O task, mirror_write 308, that the requested I/O task is complete. In another embodiment, the slave_array_write I/O task 316 directly calls the disk_write I/O tasks from their respective disk objects 225 and a group_write I/O task is not used. In this embodiment, when the disk_write I/O tasks are complete, they each notify their parent I/O task, the slave_array_write I/O task 316, that the requested child I/O task is complete.
Upon completion of its child I/O tasks (i.e., 314 and 316), the mirror object 210 determines that no other child I/O tasks are needed to perform the requested mirror_write I/O task 308. The mirror_write I/O task 308 is complete and notifies (step 376) its parent I/O task, volume_write 304, that the requested I/O task is complete. The volume_write I/O task 304 notifies the storage manager that the data access request has been successfully fulfilled.
A simple data access request may involve a large number of I/O task operations. Consider the example of writing VBNs in a system configured as shown in
In one embodiment, this process can be optimized. The storage manager, or any parent object can traverse through the set of objects 200 (
In an embodiment where bypass I/O tasks are not used, each I/O task has an immediate parent (except for the root). In that embodiment, when an I/O task is completed, the immediate parent I/O task checks the error status on the reply and handles any error that may have occurred. In an embodiment with bypass I/O tasks, since a bypass I/O task does not have an immediate parent object, when an error happens, the bypass I/O task calls an error handling operation from what would be the immediate parent object of the hierarchical order 200. The error handling I/O operation will normally allocate a child I/O task to handle such errors.
For example, a volume object 205 uses a bypass I/O task to make a call directly to a disk object 225 to obtain a read I/O task. The immediate parent of the disk object 225, a RAID group object 220, was not used in the bypass I/O task. However, if an error happens on a bypass read I/O task to the disk object 225 in the RAID group 220, the bypass I/O task will call the parent RAID group object's 225 error handling I/O task, even though the volume object 205 generated the parent I/O task. The RAID group 225 generates the error handling I/O task(s). Localized error handling allows a more optimal loading of resources, as described in more detail below.
An object can perform its associated I/O task using one of three types of exemplary procedures. In the first type of procedure, the object performs the necessary I/O task directly. For example, a disk object 225 performs a disk_read I/O task 324 by sending the appropriate read commands directly to the disk drive 110.
In the second type of procedure, the object cannot perform the necessary I/O task directly (e.g., as occurs when the object is a higher-level object in the set of objects 200). In order to complete the I/O task, the object generates (path 440) a child I/O task. The state of the I/O task (i.e., parent I/O task) changes to an I/O task suspended state 445. In this state, the I/O task waits, for example on a wait queue, until the occurrence of an event, for example, completion of the child I/O task. When the event occurs, a child I/O task finished operation (path 450) is called and the I/O task changes back to the perform I/O task state 415. If additional child I/O tasks are needed to perform the I/O task they are also generated (path 440) and the I/O task remains in the suspended state 445 until all of the child I/O tasks are completed.
When the object generates (path 440) a child I/O task, a separate and independent instance of the child I/O is generated in an initial state 405. Before the I/O thread can schedule the child I/O task, the child I/O task must obtain all of the resources needed by that child I/O task. If all of the necessary resources are available, the necessary resources are allocated to that child I/O task, a start/restart operation (path 410) is called, and that child I/O task changes to a perform I/O task state 415. In this state, the child I/O task is ready to initiate the necessary procedures to implement the requested I/O task. If all of the necessary resources are not available (path 420), the state of the child I/O task changes to a suspended state 425. In this state, the I/O task waits, for example on a wait queue, until the occurrence of an event, for example, all the necessary resources becoming available. When the event occurs, a start/restart operation (path 410) is called, the necessary resources are allocated to that I/O task, and that I/O task changes to the perform I/O task state 415. If the child I/O task cannot perform the necessary I/O task directly, the object generates (path 440) its own child I/O task (e.g., grandchild I/O task to the original I/O task) and the cycle continues recursively.
In the third type of procedure, the object cannot perform the necessary I/O task without performing a parity calculation (e.g., an XOR calculation). In order to complete the I/O task, the object generates (path 455) a request to perform a parity calculation. The state of the I/O task changes to an I/O task suspended state 460. In this state, the I/O task waits, for example on a wait queue, until the occurrence of an event, for example, completion of the parity calculation. When the event occurs, an XOR task finished operation (path 465) is called and the I/O task changes back to the perform I/O task state 415. If additional XOR calculation tasks are needed to perform the I/O task they are also generated (path 455) and the I/O task remains in the suspended state 460 until all of the XOR calculation tasks are completed. In one embodiment, only the portion associated with the XOR calculation is suspended. For example, if the I/O task is an array_write with multiple stripes, once an XOR calculation is complete for a stripe, the associated write I/O child tasks for that particular stripe can be generated, even though XOR calculations for other stripes in the array are not finished.
Once the I/O task is complete (e.g., the data is read from a disk and placed in a buffer, the child I/O tasks are completed, and/or the XOR calculation tasks are completed), an I/O task finished operation (path 430) is called. The necessary resources allocated to the I/O task are released, and the I/O task changes to an I/O task complete state 435. If this was a child I/O task, then the I/O task complete state 435 includes notifying the parent I/O task of completion (e.g., calling the exemplary third function vector of the parent I/O task to indicate that the subordinate I/O task has completed). In the illustrated embodiment, the I/O task can store intermediate state information in a persistent store 480 so that it may be reconstructed, for example after a power interrupt or system reboot.
If the resources are available, the storage system 105 determines (step 525) whether this child I/O task must perform any reads (i.e., reading data from a disk 110d to fulfill the request of the parent I/O task. If the parent I/O task is a request to read data from a disk, the storage system 105 determines (step 525) that a read is needed. If the parent I/O task is a request to write data to a disk, the storage system 105 determines (step 525) that a read is needed if any of the data blocks in a stripe already contain data and the value of that data needs to be read for a parity calculation. If the storage system 105 determines (step 525) that a read is necessary, the storage system 105 determines (step 530) whether the object can perform the read directly (e.g., is this a disk object 225). If the object cannot perform the read directly, the storage system 105 generates (step 510) another child I/O task (i.e., an I/O task at least one level lower than the prior I/O task). If the object can perform the read directly, the storage system 105 performs (step 535) the required read. The storage system 105 determines (step 525) whether any additional reads are needed by this child I/O task to fulfill the parent I/O task.
If no more reads are needed, the storage system 105 determines (step 540) whether any XOR calculations (i.e., parity calculations) are needed to fulfill the parent I/O task. If the storage system 105 determines that an XOR calculation is necessary, the storage system 105 sends (step 545) the request to an XOR thread and saves the results in a buffer (allocated in step 515). The storage system 105 determines (step 540) whether any additional XOR calculations are needed by this child I/O task to fulfill the parent I/O task.
If no more XOR calculations are needed, the storage system 105 determines (step 545) whether any writes (i.e., writing data to a disk 110) are needed by the child I/O task to fulfill the parent I/O task. If the data access request is a request to read data from a disk, the storage system 105 determines (step 545) that no writes are needed. If the data access request is a request to write data to a disk, the storage system 105 determines (step 545) that a write is needed. If the storage system 105 determines (step 545) that a write is necessary, the storage system 105 determines (step 550) whether the object can perform the write directly (e.g., is this a disk object 225). If the object cannot perform the write directly, the storage system 105 generates (step 510) another child I/O task (i.e., an I/O task at a level at least one lower than the prior I/O task). If the object can perform the write directly, the storage system 105 performs (step 555) the required write. The storage system 105 determines (step 545) whether any additional writes are needed by this child I/O task to fulfill the parent I/O task.
If any errors are encountered (step 560) at this child I/O task level, the storage system 105 initiates (step 565) error handling at the parent object level. Although the error determination (step 560) is located towards the end of the process 500, if at anytime during the process an error is encountered, the process will jump (connections omitted for clarity) to step 565. Once the child I/O task is complete, it notifies (step 570) its parent I/O task.
As described above, in one embodiment an I/O task has to obtain all necessary resources before the I/O task can begin performing its necessary actions. In one embodiment, the resource manager 155 manages the allocation of all resources available to the storage system 105. The resource manager 155 allocates and deallocates resources in response to requests from the I/O tasks. The resources managed by the resource manager 155 are, for example, stripe locks, memory data structures, data buffer memory allocations, NVRAM slots and the like. A stripe lock is a lock obtained on a range of DBNs at a particular object within the set of objects 200. A memory data structure is an allocation of memory sufficient to hold an instance of the I/O task that is generated. A data buffer memory allocation is an allocation of memory sufficient to perform and hold the read data so that parity calculations can be made and sufficient to hold the results of the parity calculations. A data buffer memory allocation is also an allocation of memory sufficient to hold read data and calculations to reconstruct a stripe in a degraded condition when one disk of a stripe is not working properly. A NVRAM slot is an allocation of non-volatile memory sized to hold task-specific intermediate state information. Use of this NVRAM slot for storage of data until it is written allows for recovery from a system crash or disk failure during a write I/O task.
As stated above, the I/O tasks are sent to the I/O stack individually and are scheduled and serviced without regard to their relationship to each other. The resource manager 155 allocates resources as the I/O tasks are generated and the I/O tasks send requests to the resource manager for required resources. In other words, the resource manager 155 assigns a specific resource to a specific instance of an I/O task, preventing that resource from being used by any other process except the specific instance of the I/O task. The resource manager 155 is configured to balance the loading on the resources, keep a forward progression of the I/O tasks and prevent the situation of a deadlock of I/O tasks waiting for each other's resources.
In one embodiment, the I/O tasks request a resource allocation through the resource manager 155. The resource manager 155 has no independent knowledge of the state of the I/O tasks, the ordering of requests that the I/O tasks will make or the future resource requirements of a particular I/O task. As described above, each I/O task has a series of states through which the I/O task progresses, one or more of which involve the allocation of resources (e.g., memory, NVRAM, child I/O tasks, and the like). At each point where an allocation of resources is required, the I/O task invokes a resource manager allocation function to satisfy the requirement. If the allocation cannot be satisfied, the resource manager 155 is responsible for suspending the requesting I/O task. In another embodiment, the invoked resource manager allocation function returns an indication back to the I/O task that the requested resource cannot be allocated. The I/O task is then responsible for performing the predefined actions in response to the indication (e.g., changing to a suspended state, scheduling a subsequent request for resources and the like).
In general, for a given top-level I/O (e.g., data access request, including a read or write ), resource allocations occur in the order mandated by the directed graph 600 in order to avoid deadlock. Once the resource manager 155 allocates a resource ‘A’ (e.g., 632), no other resource whose pool is at a higher level in the hierarchical order (e.g., 625, 630) from the allocated resource ‘A’can be allocated until resource ‘A’ is freed. Also, the resource manager 155 allocates all resources from a given pool at once. For example, the resource manager 155 allocates all resources requested for a mirror write I/O task 610 before allocating resources needed for buffer headers 615. Resources can be freed in any order. Not all resource pools need be involved in a given I/O (e.g., data access request, including read and/or write ). For example, resources associated with errors (e.g., 620, 625 and 645) are allocated only if errors actually occur.
Using the process illustrated in
While allocating resources, the resource manager 155 is prevented from allocating, or waiting for, resources assigned to an I/O task at a higher level. For example, if the memory data structure 610 for the mirror_write I/O task 308 is not available, the resource manager 155 does not allocate, or wait for resources assigned to the volume_write I/O task 304 (e.g., the memory data structure 605). In one embodiment, if all of the resources for a particular I/O task are not available, the resource manager 155 releases any resources already allocated to that I/O task until all of the needed I/O tasks are available. For example, if after allocating resources 610, 615 and 620 for the mirror_write I/O task 308 the resource manager 155 cannot allocate the necessary mirror lock 630, the resource manager deallocates the resources 610, 615 and 620 (e.g., releases the resources back to the storage system 105 by indicating they are available).
In another embodiment, an I/O task holds its allocated resources, even while suspended, and restarts once the I/O task obtains all necessary resources. In yet another embodiment, even if the resources are held, there cannot be any allocation of resources that is in the reverse order of graph 600. For example, for the mirror_write I/O task 308, after allocating resources 610 and 615 and then jumping to 630 because there are no error conditions, the resource manager 155 cannot go backwards and subsequently allocate mirror readerrorio resources 620 if an error is subsequently encountered. In this situation, the resource manager deallocates the resources 610, 615 and 630 (e.g., releases the resources back to the storage system 105 by indicating they are available) and starts the allocation process again.
The resource manager 155 also allocates a data buffer 625 of a predetermined size to hold data for error handling by the mirror object 210 when error handling is necessary. Error handling at this level is necessary when two disks 110 associated with the same RAID group 220 fail, thus preventing a reconstruction of the stripe and requiring the data to be read from the mirror storage devices. The predetermined size of the data buffer 625 can be based on the physical configuration of the RAID groups 220, the size of the VBN array (e.g., 302) to be written and the like. In the preferred embodiment, resources for error handling are only allocated when an error occurs.
Proceeding down the graph 600, when the array_write I/O task 314 is generated, the resource manager 155 allocates a memory data structure 635 sufficient, for example, to hold an instance of the array_write I/O task 314, including stripe data. If error handling at the mirror level is necessary, the resource manager 155 allocates the necessary resources for the plex readio 632 and the group readio 640. Regardless of the path, the resource manager 155 may also allocate resources to support a read I/O error handling task, referred to as a read error I/O task. The read error I/O task is used to handle errors during normal reads or reads of blocks that belong to the broken disk in degraded mode.
For the read error I/O error handling tasks, the resource manager 155 allocates a memory data structure 645 sufficient, for example, to hold an instance of the read error I/O error handling task. Along with this, the resource manager 155 also allocates a data buffer sufficient to hold a stripe of data for any parity calculation that may be necessary for the read.
Proceeding down the graph 600, the resource manager 155 also allocates all stripe locks 650 necessary to lock the DBN range being written to in the array_write I/O task 314. The resource manager 155 also allocates data buffers 655 of a predetermined size to hold data for processing by the array_write I/O task 314. For a write I/O task, the resource manager 155 also allocates NVRAM slots 660 sufficient to hold task-specific intermediate state information.
Again, while allocating resources, the resource manager 155 does not allocate, or wait for resources assigned to an I/O task at a higher level. For example, if the memory data structure 635 for the array_write I/O task 314 is not available, the resource manager 155 does not allocate, or wait for resources assigned to the volume_write I/O task 304 (e.g., the memory data structure 605) or to the mirror_write I/O task 308 (e.g., the memory data structure 610).
When the disk_write I/O task 360 is generated, the resource manager 155 allocates a memory data structure 675 sufficient to hold an instance of the disk_write I/O task 360. If the memory data structure 675 for the disk_write I/O task 360 is not available, the resource manager 155 does not allocate, or wait for resources assigned to the volume_write I/O task 304 (e.g., the memory data structure 605), to the mirror_write I/O task 308 (e.g., the memory data structure 610) or the array_write I/O task 314 (e.g., the memory data structure 635).
Resources for a read data access request are allocated in a very similar manner. Some of the differences will be highlighted. When a volume_read I/O task is generated, the resource manager 155 allocates a memory data structure 680 sufficient to hold an instance of the volume_read I/O task. When the mirror_read I/O task is generated, the resource manager 155 allocates a memory data structure 685 sufficient, for example, to hold an instance of the mirror_read I/O task. The other resources for the mirror_read I/O task are the same as for the mirror_write I/O task except that instead of a memory data structure 615, the resource manager 155 allocates data buffers 690 sufficient to hold data necessary to verify the read.
As with the write I/O tasks, while allocating resources, the resource manager 155 does not allocate, or wait for resources assigned to an I/O task at a higher level. For example, if the memory data structure 685 for the mirror_read I/O task is not available, the resource manager 155 does not allocate, or wait for resources assigned to the volume_read I/O task (e.g., the memory data structure 680).
Another difference is that for a read I/O task, the read I/O task does not need to allocate the NVRAM slots 660. For the read I/O task, once the resource manager 155 allocates the buffers 655, the read I/O task can generate a disk_read I/O task without regard to the NVRAM slot availability. When the disk_read I/O task is generated, the resource manager 155 allocates a memory data structure 695 sufficient to hold an instance of the disk_read I/O task. If the memory data structure 695 for the disk_read I/O task is not available, the resource manager 155 does not allocate, or wait for resources assigned to the volume_read I/O task (e.g., the memory data structure 680), or to the mirror_read I/O task (e.g., the memory data structure 685).
In another embodiment, the resource manager 155 uses a list to prioritize the allocation of resources and prevent any I/O task in the list from taking a resource from an I/O task higher in the list. Such a list is, for example, from highest priority to lowest priority, a volume read, a volume write, a mirror read, a mirror write, a mirror read error, a plex read, a RAID group read, an array read, an array write, a disk read, a disk write, a stripe read, a stripe write, a parity reconstruction of a RAID group, a stripe parity reconstruction and a mirror verification. While allocating resources to a particular I/O task in the list, the resource manager 155 is prevented from allocating, or waiting for, resources assigned to an I/O task at a higher priority in the list (i.e., an I/O task that comes before the particular I/O task in the list). For example, if the memory data structure for a mirror_write I/O task is not available, the resource manager 155 does not allocate, or wait for resources assigned to volume read I/O tasks, volume write I/O tasks, or mirror read I/O tasks.
The resource types are, for example, stripe locks, memory data structures, data buffer memory allocations, NVRAM slots and the like. To define a set of resources of a specific type (e.g., type A 710a, type B 715a and type C 720a), the resource manager 155 defines, for example, a quantity of that resource type and a location of that resource type associated with the set. The location of a resource type can be a physical location or it can be a logical/virtual location that is mapped to a physical location, allowing the resource manager 155 to dynamically change the mapping according to optimization algorithms.
In one embodiment, the sum of the sets of resources contained in the three resource pools 705a, 705b and 705c represent all of the resources available to the storage system 105. For example, the sum of the sets of resource type A, 710a, 710b and 710c, generally referred to as 710, represent the total set of available resources of type A 710 that the resource manager 155 can allocate. Similarly, the sum of the sets of resource type B, 715a and 715b, generally referred to as 715, represent the total set of available resources of type B 715 that the resource manager 155 can allocate. The sum of the sets of resource type C, 720a and 720b, generally referred to as 720, represent the total set of available resources of type C 720 that the resource manager 155 can allocate. The sum of the sets of resource type D, 725a and 725b, generally referred to as 725, represent the total set of available resources of type D 725 that the resource manager 155 can allocate.
The resource manager 155 associates each of the I/O tasks (e.g., volume_write 304, mirror_write 306, disk_read 324 and the like) with a resource pool 705. When an I/O task is generated, the resource manager 155 allocates the needed resource(s) from the resource pool 705 associated with that generated I/O task. When that generated I/O task is complete, the resource manager 155 returns the allocated resources (i.e., deallocates) to the pool 705 from which they came.
For example, the resource type A 710 is a memory data structure type resource. The resource manager 155 associates the volume_write I/O task with resource pool A 705a. The resource required when a volume_write I/O task is generated is a memory data structure 605 sufficient to hold an instance of the volume_write I/O task. When an instance of a volume_write I/O task is generated (e.g., volume_write I/O task 304), the resource manager 155 determines whether the unallocated amount of resource type A 710a of resource pool A 705a is large enough to allocate the required memory data structure.
In other words, a portion of the set of the resource type A 710a of resource pool A 705a may be allocated to one or more other volume_write I/O tasks currently being processed by the I/O thread. If the portion of the currently (i.e., when volume_write I/O task 304 is generated) unallocated resource type A 710a is not large enough, the resource manager cannot allocate the required resource. The I/O thread suspends the volume_write I/O task 304 until a resource from the associated resource pool A 705a is available. The resource manager 155 waits until the portion of unallocated resource type A 710a increases to be large enough to accommodate volume_write I/O task 304. In another embodiment, the I/O thread does not suspend the volume_write I/O task 304. Instead, when the invoked resource manager allocation function returns an indication back to the I/O task that the requested resource cannot be allocated, the I/O task performs the predefined actions in response to the indication (e.g., changing to a suspended state, scheduling a subsequent request for resources and the like).
The increase occurs when, for example, a different volume_write I/O task completes and its resource of type A 710a is returned to the pool A 705a. The increase can also occur if the resource manager 155 dynamically increases the set of the resource type A 710a by transferring a portion of the set of resource type A 710b from the resource pool B 705b. The resource manager 155 makes this dynamic allocation in response to, for example, statistical data, historical maximum usage and the like.
In one embodiment, the resource manager 155 defines pools according to the type of I/O tasks and associates any I/O tasks of the same type with that pool. For example, one resource pool 705a is defined as write I/O tasks. The resource manager 155 associates write I/O task types (e.g., volume_write, disk_write, and the like) with that resource pool 705. The sets of resources 710a, 715a, and 720a are bounded minimally such that, for example, a write to one stripe can be achieved with those defined resources. The sets of resources 710a, 715a, and 720a are bounded maximally such that, for example, there are other system resources available for read type I/O tasks, associated with another pool 705b, that can be allocated if needed during the write process to ensure forward progression.
In another embodiment, the resource manager 155 defines resource pools 705 according to the hierarchical representation (e.g., 200) of the disk array 115. For example, the resource manager 155 defines a first resource pool 705a for any I/O tasks associated with the volume object 205 (e.g., volume_write, volume_read). The resource manager 155 includes in the first resource pool 705a all of the resource types that are required by the I/O tasks associated with the volume object 205. The resource manager 155 defines a second resource pool 705b for any I/O tasks associated with the mirror object 210 (e.g., mirror_write, mirror_read). The resource manager 155 includes in that resource pool 705b all of the resource types that are required by the I/O tasks associated with the mirror object 210. The resource manager 155 similarly defines additional pools for the I/O tasks associated with each level (e.g., plex objects 215, RAID group objects 220 and disk objects 225).
For each defined resource pool 705, the resource manager 155 determines the set of resources that are associated with that respective resource pool 705. In addition to the set of resources being determined by the type of resources each I/O task of the hierarchy requires, the resource manager 155 determines the set of resources also based on configuration information. Configuration information represents the implementation of the storage devices (e.g., disks 110 of disk array 115).
In the illustrated embodiment of
Wherever located, however, the resource manager 155 determines that based on this configuration information, for each mirror_write I/O task (e.g., 308), the illustrated configuration requires two plex I/O tasks (e.g., 314 and 316). Similarly, for each RAID_group_write I/O task (e.g., 352), the illustrated configuration requires three disk_write I/O tasks (e.g., 360, 362, 364). The resource manager 155 determines that the set of resources for each resource pool 705 associated with these I/O tasks must be proportioned according to the configuration information.
For an illustrative example, referring to
The resource manager 155 associates disk_write I/O tasks with resource pool B 705b. One resource type a disk_write I/O task requires is a memory data structure resource type 710b. Because the predetermined proportion of RAID_group_write I/O tasks to disk_write I/O tasks is one to three, the resource manager 155 associates with resource pool B 705b a quantity of the memory data structure resource type 710a to enable allocation of this resource type to thirty disk_write I/O tasks at the same time. This sizing of the set of resources in each of the resource pools 705 ensures that there is forward progression of children I/O tasks. For each RAID_group_write I/O task that obtains its needed resource, its three child disk_write I/O tasks can also obtain their needed resources. The sizing of the resource pools 705 controls the flow of I/O tasks.
The predetermined proportion can change. For example, the addition or deletion of disks 110 changes the configuration information, which can affect the exemplary predetermined proportion of one to three. Also, an error can arise, requiring additional resources to process I/O tasks in the degraded condition. Continuing with the previous example, an error is encountered within the RAID group, requiring twice as many resource type A 710b for each disk_write I/O task. Thus, the error condition has modified the predetermined proportion to a ratio of one to six.
With the error condition, if ten RAID_group_write I/O tasks obtain resources from the resource pool 705a, only half of their child disk_write I/O tasks (i.e., fifteen disk_write I/O tasks) can obtain resources from resource pool B 705b, because of the doubling of needed resources by a disk_write I/O task in the degraded error condition. Half of the disk_write I/O tasks are suspended until other disk_write I/O tasks complete and resources from resource pool B 705b become unallocated. This is sometimes referred to as a livelock condition or starvation. In other words, livelock occurs when the system is processing I/O tasks, but only a portion (e.g., one or two) child disk_write I/O tasks can obtain resources, thus not allowing any single parent RAID_group_write I/O task to make forward progress or complete. In one embodiment, the pools 705 are sized to include at least the minimum resources necessary to write one stripe of the largest RAID group size in the system. The distribution of the minimum resources necessary among the pools 705 is dependent on which of the exemplary techniques are used to associate the I/O tasks with the pools 705.
In one embodiment, the I/O thread controls the flow of I/O tasks to eliminate the child I/O task suspension. In this embodiment, the I/O thread schedules only five concurrent RAID_group_write I/O tasks to ensure that the child disk_write I/O tasks can obtain their needed resources and forward progression is maintained. Once the error condition is eliminated, the I/O thread returns to scheduling ten concurrent RAID_group_write I/O tasks, enough to keep all resources fully utilized, but limited to ensure there is forward progression with child I/O tasks.
In another embodiment, the resource manager 155 defines a first pool 705a and associates with the first pool any I/O tasks associated with long-running behaviors. Long-running behaviors are those tasks that require access to the disks 110 many times to complete, require calculations to be made, are typically in the nature of data maintenance, and can be run as background tasks (e.g., should not have priority over other types of I/O tasks). I/O tasks associated with long-running behaviors include, for example, a parity verification I/O task, a mirror verification I/O task, a parity reconstruction of a RAID group, a mirror resynchronization and the like.
The resource manager 155 defines a second pool 705a and associates with the second pool any I/O tasks associated with time-critical behaviors. Time-critical behaviors are those tasks that require immediate access to the disks 110, are typically in the nature of fulfilling a request from the client 175, and cannot be run as background tasks (e.g., should have priority over other types of I/O tasks). I/O tasks associated with time-critical behaviors include, for example a disk read for non-degraded RAID groups, a disk write for non-degraded RAID groups and the like.
By splitting the pools in this manner, the resource manager 155 controls how much of the total resources of the storage system 105 are used by the long-running behaviors, ensuring there is enough resources available to process client 175 data access requests. The resource manager 155 can dynamically change the allocation of the resource types to the first pool 705a and the second pool 705b. The resource manager 155 makes this dynamic allocation in response to, for example, statistical data, historical maximum usage and the like.
In another embodiment, regardless of how the resource manager 155 defines pools 705 as exemplified above, the resource manager defines a separate error pool 705e (not shown). The resource manager 155 associates with an error pool any I/O tasks associated with error handling. To process the error handling, the resource manager 155 allocates resources to these error-handling I/O tasks from this error pool 705e. By allocating resources from the error pool 705e to these I/O tasks processing an error, the resource manager 155 ensures that no error I/O task consumes all of the available system resources and drastically slows down the response of the storage system 105 to a data access request. All I/O tasks that are not associated in the path where the error is located should be allowed to be processed quickly. The resource manager 155 can dynamically change the allocation of the resource types to the error pool 705e. The resource manager 155 makes this dynamic allocation in response to, for example, currently known disk errors, statistical data, historical maximum usage and the like. In another embodiment, there is an error pool 705 for each RAID group in the system. In this embodiment, the error handling I/O tasks for each RAID group are associated with their own error pool 705. This prevents all of the resources of the system from being dedicated to reconstruct data on writes to a specific RAID group by limiting the amount of error handling I/O tasks that may be allocated on behalf of the raid group.
In another embodiment, each resource group (i.e., the resources identified as 605, 610, 615, 620, 625, 630, 632, 635, 640, 645, 650, 655, 660, 665, 670, 680, 685, and 690) of the graph 600 (
The invention can be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The foregoing embodiments are therefore to be considered in all respects illustrative rather than limiting on the invention described herein. Scope of the invention is thus indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.
Number | Name | Date | Kind |
---|---|---|---|
3876978 | Bossen et al. | Apr 1975 | A |
4092732 | Ouchi | May 1978 | A |
4201976 | Patel | May 1980 | A |
4205324 | Patel | May 1980 | A |
4375100 | Tsuji et al. | Feb 1983 | A |
4467421 | White | Aug 1984 | A |
4517663 | Imazeki et al. | May 1985 | A |
4667326 | Young et al. | May 1987 | A |
4688221 | Nakamura et al. | Aug 1987 | A |
4722085 | Flora et al. | Jan 1988 | A |
4755978 | Takizawa et al. | Jul 1988 | A |
4761785 | Clark et al. | Aug 1988 | A |
4775978 | Hartness | Oct 1988 | A |
4791554 | Hirota et al. | Dec 1988 | A |
4796260 | Schilling et al. | Jan 1989 | A |
4817035 | Timsit | Mar 1989 | A |
4825403 | Gershenson et al. | Apr 1989 | A |
4837680 | Crockett et al. | Jun 1989 | A |
4847842 | Schilling | Jul 1989 | A |
4849929 | Timsit | Jul 1989 | A |
4849974 | Schilling et al. | Jul 1989 | A |
4849976 | Schillling et al. | Jul 1989 | A |
4870643 | Bultman et al. | Sep 1989 | A |
4899342 | Potter et al. | Feb 1990 | A |
4989205 | Dunphy, Jr. et al. | Jan 1991 | A |
4989206 | Dunphy, Jr. et al. | Jan 1991 | A |
5077736 | Dunphy, Jr. et al. | Dec 1991 | A |
5088081 | Farr | Feb 1992 | A |
5101492 | Schultz et al. | Mar 1992 | A |
5128810 | Halford | Jul 1992 | A |
5148432 | Gordon et al. | Sep 1992 | A |
RE34100 | Hartness | Oct 1992 | E |
5163131 | Row et al. | Nov 1992 | A |
5166936 | Ewert et al. | Nov 1992 | A |
5166939 | Jaffe et al. | Nov 1992 | A |
5179704 | Jibbe et al. | Jan 1993 | A |
5202979 | Hillis et al. | Apr 1993 | A |
5208813 | Stallmo | May 1993 | A |
5210860 | Pfeffer et al. | May 1993 | A |
5218689 | Hotle | Jun 1993 | A |
5233618 | Glider et al. | Aug 1993 | A |
5235601 | Stallmo et al. | Aug 1993 | A |
5237658 | Walker et al. | Aug 1993 | A |
5257367 | Goodlander et al. | Oct 1993 | A |
5274799 | Brant et al. | Dec 1993 | A |
5305326 | Solomon et al. | Apr 1994 | A |
5351246 | Blaum et al. | Sep 1994 | A |
5392244 | Jacobson et al. | Feb 1995 | A |
5410667 | Belsan et al. | Apr 1995 | A |
5537567 | Galbraith et al. | Jul 1996 | A |
5579475 | Blaum et al. | Nov 1996 | A |
5583995 | Gardner et al. | Dec 1996 | A |
5598549 | Rathunde | Jan 1997 | A |
5623595 | Bailey | Apr 1997 | A |
5682530 | Shimamura | Oct 1997 | A |
5768621 | Young | Jun 1998 | A |
5797034 | Young | Aug 1998 | A |
5802345 | Matsunami et al. | Sep 1998 | A |
5805788 | Johnson | Sep 1998 | A |
5812753 | Chiariotti | Sep 1998 | A |
5819292 | Hitz et al. | Oct 1998 | A |
5845117 | Fujita | Dec 1998 | A |
5848229 | Morita | Dec 1998 | A |
5860003 | Eidler et al. | Jan 1999 | A |
5862158 | Baylor et al. | Jan 1999 | A |
5884098 | Mason, Jr. | Mar 1999 | A |
5892969 | Young | Apr 1999 | A |
5913060 | Discavage | Jun 1999 | A |
5948110 | Hitz et al. | Sep 1999 | A |
5950225 | Kleiman | Sep 1999 | A |
5963962 | Hitz et al. | Oct 1999 | A |
5974544 | Jeffries et al. | Oct 1999 | A |
6038570 | Hitz et al. | Mar 2000 | A |
6073218 | DeKoning et al. | Jun 2000 | A |
6092215 | Hodges et al. | Jul 2000 | A |
6138126 | Hitz et al. | Oct 2000 | A |
6138201 | Rebalski | Oct 2000 | A |
6158017 | Han et al. | Dec 2000 | A |
6223300 | Gotoh | Apr 2001 | B1 |
6289356 | Hitz et al. | Sep 2001 | B1 |
6333936 | Johansson et al. | Dec 2001 | B1 |
6457098 | DeKoning et al. | Sep 2002 | B1 |
6532548 | Hughes | Mar 2003 | B1 |
6581185 | Hughes | Jun 2003 | B1 |
6594698 | Chow et al. | Jul 2003 | B1 |
6636940 | Hodges | Oct 2003 | B1 |
6671772 | Cousins | Dec 2003 | B1 |
6684313 | Voigt et al. | Jan 2004 | B2 |
6718486 | Roselli et al. | Apr 2004 | B1 |
6728832 | Yamamoto et al. | Apr 2004 | B2 |
6732166 | Woodruff | May 2004 | B1 |
6732232 | Krishnamurthy | May 2004 | B2 |
6918006 | Archibald et al. | Jul 2005 | B1 |
6928647 | Sager | Aug 2005 | B2 |
6961729 | Toohey et al. | Nov 2005 | B1 |
20020095602 | Pherson et al. | Jul 2002 | A1 |
20030061264 | Benhase et al. | Mar 2003 | A1 |
Number | Date | Country |
---|---|---|
0420142 | Sep 1990 | DE |
Number | Date | Country | |
---|---|---|---|
20030182348 A1 | Sep 2003 | US |