Parallel NFS (pNFS) RDMA Layout
hch@lst.de
Transport
NFSv4NFSv4
The Parallel Network File System (pNFS) allows a separation between
the metadata (onto a metadata server) and data (onto a storage device)
for a file. The RDMA Layout Type is defined in this document as an
extension to pNFS to allow the use of RDMA Verbs operations to access
remote storage, with a special focus on accessing byte addressable
persistent memory.
shows the overall
architecture of a Parallel NFS (pNFS) system:
The overall approach is that pNFS-enhanced clients obtain
sufficient information from the server to enable them to access
the underlying storage (on the storage systems) directly. See
the Section 12 of for more details.
RDMA () is a technique for moving data
efficiently between end nodes. By directing data into destination
buffers as it is sent on a network, and placing it via direct memory
access by hardware, the benefits of faster transfers and reduced host
overhead are obtained. Unlike the RPC RDMA transport
the pNFS RDMA layout does not transfer
remote procedural calls over RDMA networks, but instead uses
raw RDMA READ and WRITE operations to access a memory region exposed
on a storage device.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in .
The following definitions are provided for the purpose of providing
an appropriate context for the reader.
This document defines a byte as an octet, i.e., a datum exactly 8
bits in length.
The "client" is the entity that accesses the NFS server's
resources. The client may be an application that contains the
logic to access the NFS server directly. The client may also be
the traditional operating system client that provides remote file
system services for a set of applications.
The "server" is the entity responsible for coordinating client
access to a set of file systems and is identified by a server
owner.
The metadata server is a pNFS server which provides metadata
information for a file system object. It also is responsible for
generating layouts for file system objects. Note that
the MDS is also responsible for directory-based operations.
The external data representation (XDR) description and scripts
for extracting the XDR description are Code Components as
described in Section 4 of "Legal Provisions
Relating to IETF Documents". These Code Components are
licensed according to the terms of Section 4 of "Legal Provisions
Relating to IETF Documents".
This document contains the XDR description
of the NFSv4.1 RDMA layout protocol. The XDR description is
embedded in this document in a way that makes it simple for the
reader to extract into a ready-to-compile form. The reader can
feed this document into the following shell script to produce
the machine readable XDR description of the NFSv4.1 RDMA layout:
That is, if the above script is stored in a file called "extract.sh", and
this document is in a file called "spec.txt", then the reader can do:
The effect of the script is to remove leading white space from each
line, plus a sentinel sequence of "///".
The embedded XDR file header follows.
Subsequent XDR descriptions, with the sentinel sequence are
embedded throughout the document.
Note that the XDR code contained in this document depends on
types from the NFSv4.1 nfs4_prot.x file .
This includes both nfs types that end with a 4, such as
offset4, length4, etc., as well as more generic types such as
uint32_t and uint64_t.
A pNFS RDMA layout is responsible for mapping from an NFS file (or
portion of a file) to memory regions that contain the file.
These regions are expressed as extents with 64-bit offsets and lengths
using the existing NFSv4 offset4 and length4 types, and map to
memory regions that the servers registered, and for which it exposes
a handle (R_key or stag) that allows for RDMA READ and RDMA WRITE
operations from the client.
The pNFS operation for requesting a layout (LAYOUTGET) includes the
"layoutiomode4 loga_iomode" argument, which indicates whether the
requested layout is for read-only use or read-write use. A read-only
layout may contain holes that are read as zero, whereas a read-write
layout will contain allocated, but un-initialized storage in those
holes (read as zero, can be written by client). This document also
supports client participation in copy-on-write (e.g., for file
systems with snapshots) by providing both read-only and un-
initialized storage for the same extent in a layout. Reads are
initially performed on the read-only storage, with writes going to
the un-initialized storage. After the first write that initializes
the un-initialized storage, all reads are performed to that now-
initialized writable storage, and the corresponding read-only storage
is no longer used.
The layout4 type defined in
is extended with a new value as follows:
This document defines structure associated with the layouttype4
value LAYOUT4_RDMA. specifies
the loc_body structure as an XDR type "opaque". The opaque
layout is uninterpreted by the generic pNFS client layers, but
obviously must be interpreted by the Layout Type implementation.
Data operations to a storage device require the client to know the
network address of the storage device. The NFSv4.1+ GETDEVICEINFO
operation (Section 18.40 of ) is used by
the client to retrieve that information.
The "pnfs_rdma_device_addr4" data structure is returned by the
server as the storage-protocol-specific opaque field da_addr_body in
the "device_addr4" structure by a successful GETDEVICEINFO operation
.
It contains the network address of the storage device. The RDMA
Connection manager (RDMA/CM) shall be used to establish the queue
pair for the RDMA READ and RDMA WRITE operations used by the layout.
Details of connection establishment will be provided in future
versions of this document.
A pNFS RDMA layout is a list of extents within a flat array of data in
a device. The RDMA layout describes the individual byte ranges (extents)
on the device that make up the file. The offsets and length contained in
an extent are specified in units of bytes.
The RDMA layout consists of a list of extents that map the regions
of the file to locations on a device. The "re_storage_offset" field
within each extent identifies a location on the device specified by
the "re_device_id" field in the extent.
Each extent maps a region of the file onto a portion of the
specified device. The re_file_offset, re_length, and re_state fields for
an extent returned from the server are valid for all extents. In
contrast, the interpretation of the re_storage_offset field depends on
the value of re_state as follows (in increasing order):
means that re_storage_offset is valid, and points to
valid/initialized data that can be read and written.
means that re_storage_offset is valid and points to valid/initialized
data that can only be read. Write operations are prohibited.
means that re_storage_offset is valid, but points to invalid
un-initialized data. This data MUST not be read from the
device until it has been initialized. A read request for a
PNFS_RDMA_INVALID_DATA extent MUST fill the user buffer with zeros,
unless the extent is covered by a PNFS_RDMA_READ_DATA extent of a
copy-on-write file system.
Write requests MUST write whole
server-sized blocks to the device; bytes not initialized by the user
MUST be set to zero. Any write to parts of a device covered by a
PNFS_RDMA_INVALID_DATA extent changes the written portion of the
extent to PNFS_RDMA_READ_WRITE_DATA; the pNFS client is responsible
for reporting this change via LAYOUTCOMMIT.
means that re_storage_offset is not valid, and this extent MAY not
be used to satisfy write requests. Read requests MAY be satisfied
by zero-filling as for PNFS_RDMA_INVALID_DATA. PNFS_RDMA_NONE_DATA
extents MAY be returned by requests for readable extents; they are
never returned if the request was for a writable extent.
An extent list contains all relevant extents in increasing order of
the re_file_offset of each extent; any ties are broken by increasing
order of the extent state (re_state).
Each request for a layout specifies at least three parameters: file
offset, desired size, and minimum size. If the status of a request
indicates success, the extent list returned MUST meet the following
criteria:
A request for a readable (but not writable) layout MUST return
either PNFS_RDMA_READ_DATA or PNFS_RDMA_NONE_DATA extents. It
SHALL NOT return PNFS_RDMA_INVALID_DATA or
PNFS_RDMA_READ_WRITE_DATA extents.
A request for a writable layout MUST return
PNFS_RDMA_READ_WRITE_DATA or PNFS_RDMA_INVALID_DATA extents,
and it MAY return addition PNFS_RDMA_READ_DATA extents for
ranges covered by PNFS_RDMA_INVALID_DATA extents to allow
client side copy-on-write operations. A request for a writable
layout SHALL NOT return PNFS_RDMA_NONE_DATA extents.
The first extent in the list MUST contain the requested starting
offset.
The total size of extents within the requested range MUST cover at
least the minimum size. One exception is allowed: the total size
MAY be smaller if only readable extents were requested and EOF is
encountered.
Extents in the extent list MUST be logically contiguous for a
read-only layout. For a read-write layout, the set of writable
extents (i.e., excluding PNFS_RDMA_READ_DATA extents) MUST be
logically contiguous. Every PNFS_RDMA_READ_DATA extent in a
read-write layout MUST be covered by one or more
PNFS_RDMA_INVALID_DATA extents. This overlap of
PNFS_RDMA_READ_DATA and PNFS_RDMA_INVALID_DATA extents is the
only permitted extent overlap.
Extents MUST be ordered in the list by starting offset, with
PNFS_RDMA_READ_DATA extents preceding PNFS_RDMA_INVALID_DATA
extents in the case of equal re_file_offsets.
The server shall ensure that it has registered handles for the memory
regions that the extents in the layout refer to so that RDMA READ and/or
RDMA WRITE requests can be performed by the client. Multiple extents
may refer to the same handle. The handle shall be invalidated on
LAYOUTRETURN operation, including implicit layout returns as part
of CB_LAYOUTRECALL operations, or when a layout is revoked.
According to , if the minimum requested
size, loga_minlength, is zero, this is an indication to the
metadata server that the client desires any layout at offset
loga_offset or less that the metadata server has "readily
available". Given the lack of a clear definition of this phrase,
in the context of the RDMA layout type, when loga_minlength is
zero, the metadata server SHOULD:
when processing requests for readable layouts, return all
such, even if some extents are in the PNFS_RDMA_NONE_DATA
state.
when processing requests for writable layouts, return
extents which can be returned in the PNFS_RDMA_READ_WRITE_DATA
state.
The "pnfs_rdma_layoutupdate4" structure is used by the client as the
RDMA layout-specific argument in a LAYOUTCOMMIT operation. The
"rlu_commit_list" field is a list covering regions of the file layout
that were previously in the PNFS_RDMA_INVALID_DATA state, but have
been written by the client and SHOULD now be considered in the
PNFS_RDMA_READ_WRITE_DATA state. The extents in the commit list MUST
be disjoint and MUST be sorted by rr_file_offset. Implementors should
be aware that a server MAY be unable to commit regions at a granularity
smaller than a file-system block (typically 4 KB or 8 KB). As noted
above, the block-size that the server uses is available as an NFSv4
attribute, and any extents included in the "rlu_commit_list" MUST be
aligned to this granularity and have a size that is a multiple of
this granularity.
Since the block in question is in state PNFS_RDMA_INVALID_DATA,
byte ranges not written SHOULD be filled with zeros. This applies
even if it appears that the area being written is beyond what the
client believes to be the end of file.
A LAYOUTRETURN operation represents an explicit release of
resources by the client. This MAY be done in response to a
CB_LAYOUTRECALL or before any recall, in order to avoid a future
CB_LAYOUTRECALL. When the LAYOUTRETURN operation specifies a
LAYOUTRETURN4_FILE return type, then the layoutreturn_file4 data
structure specifies the region of the file layout that is no
longer needed by the client.
The LAYOUTRETURN operation is done without any RDMA layout
specific data. The opaque "lrf_body" field of the
"layoutreturn_file4" data structure MUST have length zero.
Layouts MAY be unilaterally revoked by the server, due to the
client's lease time expiring, or the client failing to return a
layout which has been recalled in a timely manner.
For the RDMA layout type this is accomplished by invalidating the
handle for the remote memory region exposed to the client.
Once the invalidation has completed the HCA will reject all
access from the client to the memory region.
Copy-on-write is a mechanism used to support file and/or file system
snapshots. When writing to unaligned regions, or to regions smaller
than a file system block, the writer MUST copy the portions of the
original file data to a new location on disk. This behavior can
either be implemented on the client or the server. The paragraphs
below describe how a pNFS RDMA layout client implements access to a
file that requires copy-on-write semantics.
Distinguishing the PNFS_RDMA_READ_WRITE_DATA and
PNFS_RDMA_READ_DATA extent types in combination with the allowed
overlap of PNFS_RDMA_READ_DATA extents with PNFS_RDMA_INVALID_DATA
extents allows copy-on-write processing to be done by pNFS clients.
In classic NFS, this operation would be done by the server. Since
pNFS enables clients to do direct block access, it is useful for
clients to participate in copy-on-write operations. All pNFS
RDMA layout clients MUST support this copy-on-write processing.
When a client wishes to write data covered by a PNFS_RDMA_READ_DATA
extent, it MUST have requested a writable layout from the server;
that layout will contain PNFS_RDMA_INVALID_DATA extents to cover all
the data ranges of that layout's PNFS_RDMA_READ_DATA extents. More
precisely, for any re_file_offset range covered by one or more
PNFS_RDMA_READ_DATA extents in a writable layout, the server MUST
include one or more PNFS_RDMA_INVALID_DATA extents in the layout
that cover the same re_file_offset range. When performing a write
to such an area of a layout, the client MUST effectively copy the
data from the PNFS_RDMA_READ_DATA extent for any partial blocks of
re_file_offset and range, merge in the changes to be written, and
write the result to the PNFS_RDMA_INVALID_DATA extent for the blocks
for that re_file_offset and range. That is, if entire blocks of
data are to be overwritten by an operation, the corresponding
PNFS_RDMA_READ_DATA blocks need not be fetched, but any partial-
block writes MUST be merged with data fetched via
PNFS_RDMA_READ_DATA extents before storing the result via
PNFS_RDMA_INVALID_DATA extents. For the purposes of this
discussion, "entire blocks" and "partial blocks" refer to the
server's file-system block size. Storing of data in a
PNFS_RDMA_INVALID_DATA extent converts the written portion of the
PNFS_RDMA_INVALID_DATA extent to a PNFS_RDMA_READ_WRITE_DATA
extent; all subsequent reads MUST be performed from this extent; the
corresponding portion of the PNFS_RDMA_READ_DATA extent MUST NOT be
used after storing data in a PNFS_RDMA_INVALID_DATA extent. If a
client writes only a portion of an extent, the extent MAY be split at
block aligned boundaries.
When a client wishes to write data to a PNFS_RDMA_INVALID_DATA
extent that is not covered by a PNFS_RDMA_READ_DATA extent, it MUST
treat this write identically to a write to a file not involved with
copy-on-write semantics. Thus, data MUST be written in at least
block-sized increments, aligned to multiples of block-sized offsets,
and unwritten portions of blocks MUST be zero filled.
Layout extents returned to pNFS clients grant permission to read or
write; PNFS_RDMA_READ_DATA and PNFS_RDMA_NONE_DATA are read-only
(PNFS_RDMA_NONE_DATA reads as zeroes), PNFS_RDMA_READ_WRITE_DATA
and PNFS_RDMA_INVALID_DATA are read/write, (PNFS_RDMA_INVALID_DATA
reads as zeros, any write converts it to PNFS_RDMA_READ_WRITE_DATA).
This is the only means a client has of obtaining permission to
perform direct I/O to storage devices; a pNFS client MUST NOT perform
direct I/O operations that are not permitted by an extent held by the
client. Client adherence to this rule places the pNFS server in
control of potentially conflicting storage device operations,
enabling the server to determine what does conflict and how to avoid
conflicts by granting and recalling extents to/from clients.
If a client makes a layout request that conflicts with an existing
layout delegation, the request will be rejected with the error
NFS4ERR_LAYOUTTRYLATER. This client is then expected to retry the
request after a short interval. During this interval, the server
SHOULD recall the conflicting portion of the layout delegation from
the client that currently holds it. This reject-and-retry approach
does not prevent client starvation when there is contention for the
layout of a particular file. For this reason, a pNFS server SHOULD
implement a mechanism to prevent starvation. One possibility is that
the server can maintain a queue of rejected layout requests. Each
new layout request can be checked to see if it conflicts with a
previous rejected request, and if so, the newer request can be
rejected. Once the original requesting client retries its request,
its entry in the rejected request queue can be cleared, or the entry
in the rejected request queue can be removed when it reaches a
certain age.
NFSv4 supports mandatory locks and share reservations. These are
mechanisms that clients can use to restrict the set of I/O operations
that are permissible to other clients. Since all I/O operations
ultimately arrive at the NFSv4 server for processing, the server is
in a position to enforce these restrictions. However, with pNFS
layouts, I/Os will be issued from the clients that hold the layouts
directly to the storage devices that host the data. These devices
have no knowledge of files, mandatory locks, or share reservations,
and are not in a position to enforce such restrictions. For this
reason the NFSv4 server MUST NOT grant layouts that conflict with
mandatory locks or share reservations. Further, if a conflicting
mandatory lock request or a conflicting open request arrives at the
server, the server MUST recall the part of the layout in conflict
with the request before granting the request.
The end-of-file location can be changed in two ways: implicitly as
the result of a WRITE or LAYOUTCOMMIT beyond the current end-of-file,
or explicitly as the result of a SETATTR request. Typically,
when a file is truncated by an NFSv4 client via the SETATTR call,
the server frees any disk blocks belonging to the file that are
beyond the new end-of-file byte, and MUST write zeros to the
portion of the new end-of-file block beyond the new end-of-file
byte. These actions render any pNFS layouts that refer to the
blocks that are freed or written semantically invalid. Therefore,
the server MUST recall from clients the portions of any pNFS
layouts that refer to blocks that will be freed or written by
the server before effecting the file truncation. These recalls
may take time to complete; as explained in ,
if the server cannot respond to the client SETATTR request
in a reasonable amount of time, it SHOULD reply to the client
with the error NFS4ERR_DELAY.
Blocks in the PNFS_RDMA_INVALID_DATA state that lie beyond the new
end-of-file block present a special case. The server has reserved
these blocks for use by a pNFS client with a writable layout for the
file, but the client has yet to commit the blocks, and they are not
yet a part of the file mapping on disk. The server MAY free these
blocks while processing the SETATTR request. If so, the server MUST
recall any layouts from pNFS clients that refer to the blocks before
processing the truncate. If the server does not free the
PNFS_RDMA_INVALID_DATA blocks while processing the SETATTR request,
it need not recall layouts that refer only to the
PNFS_RDMA_INVALID_DATA blocks.
When a file is extended implicitly by a WRITE or LAYOUTCOMMIT beyond
the current end-of-file, or extended explicitly by a SETATTR request,
the server need not recall any portions of any pNFS layouts.
The layout hint attribute specified in
is not supported by the RDMA layout, and the pNFS server MUST
reject setting a layout hint attribute with a loh_type value
of LAYOUT4_RDMA_VOLUME during OPEN or SETATTR operations. On a
file system only supporting the RDMA layout a server MUST NOT
report the layout_hint attribute in the supported_attrs attribute.
A critical requirement in crash recovery is that both the client and
the server know when the other has failed. Additionally, it is
required that a client sees a consistent view of data across
server restarts. These requirements and a full discussion of
crash recovery issues are covered in the "Crash Recovery" section
of the NFSv41 specification . This
document contains additional crash recovery material specific
only to the RDMA layout.
When the server crashes while the client holds a writable layout, and
the client has written data to blocks covered by the layout, and the
blocks are still in the PNFS_RDMA_INVALID_DATA state, the client has
two options for recovery. If the data that has been written to these
blocks is still cached by the client, the client can simply re-write
the data via NFSv4, once the server has come back online. However,
if the data is no longer in the client's cache, the client MUST NOT
attempt to source the data from the data servers. Instead, it SHOULD
attempt to commit the blocks in question to the server during the
server's recovery grace period, by sending a LAYOUTCOMMIT with the
"loca_reclaim" flag set to true. This process is described in detail
in Section 18.42.4 of .
The server may respond to LAYOUTGET with a variety of error statuses.
These errors can convey transient conditions or more permanent
conditions that are unlikely to be resolved soon.
The error NFS4ERR_RECALLCONFLICT indicates that the server has
recently issued a CB_LAYOUTRECALL to the requesting client, making it
necessary for the client to respond to the recall before processing
the layout request. A client can wait for that recall to be receive
and processe or it can retry as for NFS4ERR_TRYLATER, as described
below.
The error NFS4ERR_TRYLATER is used to indicate that the server cannot
immediately grant the layout to the client. This may be due to
constraints on writable sharing of blocks by multiple clients or to a
conflict with a recallable lock (e.g. a delegation). In either case, a
reasonable approach for the client is to wait several milliseconds
and retry the request. The client SHOULD track the number of retries,
and if forward progress is not made, the client SHOULD abandon the
attempt to get a layout and perform READ and WRITE operations by
sending them to the server
The error NFS4ERR_LAYOUTUNAVAILABLE MAY be returned by the server
if layouts are not supported for the requested file or its containing
file system. The server MAY also return this error code if the server
is the progress of migrating the file from secondary storage, there is
a conflicting lock that would prevent the layout from being granted,
or for any other reason that causes the server to be unable to supply
the layout. As a result of receiving NFS4ERR_LAYOUTUNAVAILABLE, the
client SHOULD abandon the attempt to get a layout and perform READ and
WRITE operations by sending them to the MDS.
It is expected that a client will not cache the file's layoutunavailable
state forever. In particular, when the file is closed or opened by the
client, issuing a new LAYOUTGET is appropriate.
The pNFS extension partitions the NFSv4.1+ file system protocol
into two parts, the control path and the data path (storage
protocol). The control path contains all the new operations
described by this extension; all existing NFSv4 security
mechanisms and features apply to the control path. The combination
of components in a pNFS system is required to preserve the
security properties of NFSv4.1+ with respect to an entity accessing
data via a client, including security countermeasures to defend
against threats that NFSv4.1+ provides defenses for in environments
where these threats are considered significant.
The metadata server enforces the file access-control policy at
LAYOUTGET time. The client should use suitable authorization
credentials for getting the layout for the requested iomode
(READ or RW) and the server verifies the permissions and ACL
for these credentials, possibly returning NFS4ERR_ACCESS if the
client is not allowed the requested iomode. If the LAYOUTGET
operation succeeds the client receives, as part of the layout,
a set of credentials allowing it I/O access to the specified
data files corresponding to the requested iomode. When the client
acts on I/O operations on behalf of its local users, it MUST
authenticate and authorize the user by issuing respective OPEN
and ACCESS calls to the metadata server, similar to having NFSv4
data delegations. If access is allowed, the client uses the
corresponding (READ or RW) credentials to perform the I/O
operations at the data file's storage devices. When the metadata
server receives a request to change a file's permissions or
ACL, it SHOULD recall all layouts for that file and it MUST
fence off the clients holding outstanding layouts for the
respective file by implicitly invalidating the outstanding
credentials on all data files comprising before committing
to the new permissions and ACL. Doing this will ensure that
clients re-authorize their layouts according to the modified
permissions and ACL by requesting new layouts. Recalling the
layouts in this case is courtesy of the server intended to
prevent clients from getting an error on I/Os done after the
client was fenced off.
IANA is requested to assign a new pNFS layout type in the pNFS Layout
Types Registry as follows (the value 5 is suggested):
Layout Type Name: LAYOUT4_RDMA
Value: 0x00000006
RFC: RFCTBD10
How: L (new layout type)
Minor Versions: 1
Key words for use in RFCs to Indicate Requirement LevelsHarvard University1350 Mass. Ave.CambridgeMA 02138- +1 617 495 3864sob@harvard.eduLegal Provisions Relating to IETF DocumentsIETF TrustXDR: External Data Representation StandardNetwork Appliance, Inc.Network File System (NFS) Version 4 Minor Version 1 ProtocolSun Microsystems, Inc.Network Appliance, Inc.Network Appliance, Inc.Network File System (NFS) Version 4 Minor Version 1 External Data Representation Standard (XDR) DescriptionSun Microsystems, Inc.Network Appliance, Inc.Network Appliance, Inc.Remote Direct Memory Access Transport for Remote Procedure Call Version 1Oracle CorporationRed HatMicrosoftInfiniBand Architecture Specification Volume 1 Release 1.3InfiniBand Trade AssociationA Remote Direct Memory Access Protocol SpecificationIBM Corp.IBM Research GmbH Zurich Research LaboratoryHewlett-Packard CompanyHewlett-Packard CompanyDirect Data Placement over Reliable TransportsBroadcom CorporationMicrosoft CorporationIBM Corp.Hewlett-Packard Company
[RFC Editor: please remove this section prior to publishing
this document as an RFC]
[RFC Editor: prior to publishing this document as an RFC, please replace
all occurrences of RFCTBD10 with RFCxxxx where xxxx is the RFC number of
this document]