| NFSv4 | D. Noveck, Ed. |
| Internet-Draft | EMC |
| Intended status: Informational | P. Shivam |
| Expires: January 06, 2013 | C. Lever |
| B. Baker | |
| ORACLE | |
| July 07, 2012 |
NFSv4 migration: Implementation experience and spec issues to resolve
draft-ietf-nfsv4-migration-issues-01
The migration feature of NFSv4 provides for moving responsibility for a single filesystem from one server to another, without disruption to clients. Recent implementation experience has shown problems in the existing specification for this feature. This document discusses the issues which have arisen and explores the options available for curing the issues via clarification and correction of the NFSv4.0 and NFSv4.1 specifications.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on January 06, 2013.
Copyright (c) 2012 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
This document is in the informational category, and while the facts it reports may have normative implications, any such normative significance reflects the readers' preferences. For example, we may report that the reboot of a client with migrated state results in state not being promptly cleared and that this will prevent granting of conflicting lock requests at least for the lease time, which is a fact. While it is to be expected that client and server implementers will judge this to be a situation that is best avoided, the judgment as to how pressing this issue should be considered is a judgment for the reader, and eventually the nfsv4 working group to make.
We do explore possible ways in which such issues can be avoided, with minimal negative effects, in the expectation that the working group will choose to address these issues, but the choice of exactly how to address these is best given effect in one or more standards-track documents and/or errata.
This document focuses on NFSv4.0, since that is where the majority of implementation experience has been. Nevertheless, there is some discussion of the implications of the NFSv4.0 experience for migration in NFSv4.1.
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 [RFC2119].
In the context of this informational document, these normative keywords will always occur in the context of a quotation, most often direct but sometimes indirect. The context will make it clear whether the quotation is from:
Note that the examples below reflect current experience which arises from clients implementing the recommendation to use different nfs_client_id4 id strings for different server addresses, i.e. using what is later referred to herein as the "non-uniform client-string approach"
This is simply because that is the experience implementers have had. The reader should not assume that in all cases, this practice is the source of the difficulty. It may be so in some cases but clearly it is not in all cases.
The following sort of situation has proved troublesome:
Note here that while it seems clear to us in this example that C-XYZ and C-ABC are from the same client, the server has no way to determine the structure of the "opaque" id string. In the protocol, it really is treated as opaque. Only the client knows which nfs_client_id4 values designate the same client on a different server.
Further problems arise from scenarios like the following.
Note that if the client used "C" (rather than "C-ABC") as the nfs_client_id4 id string, the exact same situation would arise.
One of the first cases in which this sort of situation has resulted in difficulties is in connection with doing a SETCLIENTID for callback update.
The SETCLIENTID for callback update only includes the nfs_client_id4, assuming there can only be one such with a given nfs_client_id4 value. If there were multiple, confirmed client records with identical nfs_client_id4 id string values, there would be no way to map the callback update request to the correct client record. Apart from the migration handling specified in [RFC3530], such a situation cannot arise.
One possible accommodation for this particular issue that has been used is to add a RENEW operation along with SETCLIENTID (on a callback update) to disambiguate the client.
When the client updates the callback info to the destination, the client would, by convention, send a compound like this:
{ RENEW clientid4, SETCLIENTID nfs_client_id4,verf,cb }
The presence of the clientid4 in the compound would allow the server to differentiate among the various leases that it knows of, all with the same nfs_client_id4 value.
While this would be a reasonable patch for an isolated protocol weakness, interoperable clients and servers would require that the protocol truly be updated to allow such a situation, specifically that of multiple clientid4's with the same nfs_client_id4 value. The protocol is currently designed and implemented assuming this can't happen. We need to either prevent the situation from happening, or fully adapt to the possibilities which can arise. See Section 4 for a discussion of such issues.
Consider the following situation:
Now, instead of a clientid4 identifying a client-server pair, we have many more entities for the client to deal with. In addition, it isn't clear how new state is to be incorporated in this structure.
The limitations of the migrated state (inability to be freed on reboot) would argue against adding more such state but trying to avoid that would run into its own difficulties. For example, a single lockowner string presented under two different clientids would appear as two different entities.
Thus we have to choose between:
In any case, we have gone (in adding migration as it was described) from a situation in which
To one in which
This sort of additional client complexity is troublesome and needs to be eliminated.
The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Client ID" says:
There are two possible interpretations of the phrase "uniquely defines" in the above:
The first interpretation would make these client-strings like phone numbers (a single person can have several) while the second would make them like social security numbers.
Endless debate about the true meaning of "uniquely defines" in this context is quite possible but not very helpful. The following points should be noted though:
Given the need for the server to be aware of client identity with regard to migrated state, either client-string construction rules will have to change or there will be a need to get around current issues, or perhaps a combination of these two will be required. Later sections will examine the options and propose a solution.
One consideration that may indicate that this cannot remain exactly as it is today has to do with the fact that the current explanation for this behavior is not correct. The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Client ID" says:
In point of fact, a "SETCLIENTID with the same id string" sent to multiple network addresses will be treated as all from the same client but will not "cause the server to begin the process of removing the client's previous leased state" unless the server believes it is a different instance of the same client, i.e. if the id string is the same and there is a different boot verifier. If the client does not reboot, the verifier should not change. If it does reboot, the verifier will change, and the server should "begin the process of removing the client's previous leased state.
The situation of multiple SETCLIENTID requests received by a server on multiple network addresses is exactly the same, from the protocol design point of view, as when multiple (i.e. duplicate) SETCLIENTID requests are received by the server on a single network address. The same protocol mechanisms that prevent erroneous state deletion in the latter case prevent it in the former case. There is no reason for special handling of the multiple-network-appearance case, in this regard.
It is often felt that this is a consequence of the client-string construction issues, and it is certainly the case that the two are closely connected in that non-uniform client-strings make it impossible for the server to appropriately combine leases from the same client. See Section 5.2.1 for a discussion of non-uniform client-strings.
However, even where the server could combine leases from the same client, it needs to be clear how and when it will do so, so that the client will be prepared. These issues will have to be addressed at various places in the spec.
This could be enough only if we are prepared to do away with the "should" recommending non-uniform client-strings and replace it with a "should not" or even a "SHOULD NOT". Current client implementation patterns make this an unpalatable choice for use as a general solution, but it is reasonable to "RECOMMEND" this choice for a well-defined subset of clients. One alternative would be to create a way for the server to infer from client behavior which leases are held by the same client and use this information to do appropriate lease mergers. Prototyping and detailed specification work has shown that this could be done but the resulting complexity is such that a better choice is to "RECOMMEND" use of the uniform approach for clients supporting the migration feature.
Because of the discussion of client-string construction in [RFC3530], most existing clients implement the non-uniform client-string approach. As a result, existing servers may not have been tested with clients implementing uniform client-strings. As a consequence, care must be taken to preserve interoperability between UCS-capable clients and servers that don't tolerate uniform client strings for one reason or another. See Section 5.2.3 for details.
The fact that the reason given in client-string-BP3 is not valid makes the existing "should" insupportable. We can't either
What are often presented as reasons that motivate use of the non-uniform approach always turn out to be cases in which, if the uniform approach were used, the server will treat a client which accesses that server via two different IP addresses as part of a single client, as it in fact is. This may be disconcerting to a client unaware that the two IP addresses connect to the same server. This is thus not a reason to use the non-uniform approach but rather an illustration of the fact that those using the uniform approach must use server behavior to determine whether any trunking of IP addresses exists, as is described in Section 5.2.2.
It is always possible that a valid new reason will be found, but so far none has been proposed. Given the history, the burden of proof should be on those asserting the validity of a proposed new reason.
So we will assume for now that the "should" will have to go. The question is what to replace it with.
Given the difficulties caused by having different nfs_client_id4 client-string values for the same client, we have two choices:
There are a number of issues where the existing text is unclear and/or wrong and needs to be fixed in some way.
There are a number of cases in which certain sections, not specifically related to migration, require additional clarification. This is generally because text that is clear in a context in which leases and clientids are created in one place and live there forever may need further refinement in the more dynamic environment that arises as part of migration.
Some examples:
We propose replacing client-string-BP3 with the following text and adding the following proposed Section 5.2 to provide implementation guidance.
One particular aspect of the construction of the nfs4_client_id4 string has proved recurrently troublesome. The client has a choice of:
Note that implementation considerations, including compatibility with existing servers, may make it desirable for a client to use both approaches, based on configuration information, such as mount options. This issue will be discussed in Section 5.2.3.
Construction of the client-string has been a troublesome issue because of the way in which the NFS protocols have evolved.
NFSv4.0 is unfortunately halfway between these two. The two client-string approaches have arisen in attempts to deal with the changing requirements of the protocol as implementation has proceeded and features that were not very substantial in [RFC3530], got more substantial.
Both approaches have to deal with the asymmetry in client and server identity information between client and server. Each seeks to make the client's and the server's views match. In the process, each encounters some combination of inelegant protocol features and/or implementation difficulties. The choice of which to use is up to the client implementer and the sections below try to give some useful guidance.
The non-uniform client-string approach is an attempt to handle these matters in NFSv4.0 client implementations in as NFSv3-like a way as possible.
For a client using the non-uniform approach, all internal recording of clientid4 values is to include, whether explicitly or implicitly, the server IP address so that one always has an (IP-address, clientid4) pair. Two such pairs from different servers are always distinct even when the clientid4 values are the same, as they may occasionally be. In this approach, such equality is always treated as simple happenstance.
Making the client-string different on different servers means that a server has no way of tying together information from the same client and so will treat a single client as multiple clients with multiple leases for each server network address. Since there is no way in the protocol for the client to determine if two network addresses are connected to the same server, the resulting lack of knowledge is symmetrical and can result in simpler client implementations in which there is a single clientid/lease per server network addresses.
Support for migration, particularly with transparent state migration, is more complex in the case of non-uniform client-strings. For example, migration of a lease can result in multiple leases for the same client accessing the same server addresses, vitiating many of the advantages of this approach. Therefore, client implementations that support migration with transparent state migration SHOULD NOT use the non-uniform client-string approach, except where it is necessary for compatibility with existing server implementations (For details of arranging use of multiple client-string approaches, see Section 5.2.3).
When the client-string is kept uniform, the server has the basis to have a single clientid4/lease for each distinct client. The problem that has to be addressed is the lack of explicit server identity information, which is made available in NFSv4.1.
When the same client-string is given to multiple IP addresses, the client can determine whether two IP addresses correspond to a single server, based on the server's behavior. This is the inverse of the strategy adopted for the non-uniform approach in which different server IP addresses are told about different clients, simply to prevent a server from manifesting behavior that is inconsistent with there being a single server for each IP address, in line with the traditions of NFS. So, to compare:
The uniform client-string approach makes it necessary to exercise more care in the definition of the nfs_client_id4 boot verifier:
The following are advantages for the implementation of using the uniform client-string approach:
The following implementation considerations might cause issues for client implementations.
How to balance these considerations depends on implementation goals.
As noted above, a client which needs to use the uniform client-string approach (e.g. to support migration), may also need to support existing servers with implementations that do not work properly in this case.
Some examples of such server issues include:
In order to support such servers, the client can use different approaches for different mounts, as long as:
One effective way for clients to handle this is to support the uniform client-string approach as the default, but allow a mount option to specify use of the non-uniform client-string approach for particular mount points, as long as such mount points are not used when migration is to be supported.
In the case in which the same server has multiple mounts, and both approaches are specified for the same server, the client could have multiple clientids corresponding to the same server, one for each approach and would then have to keep these separate.
This section provides an example of how trunking determination could be done by a client following the uniform client-string approach (whether this is used for all mounts or not). Clients need not follow this procedure but implementers should make sure that the issues dealt with by this procedure are all properly addressed.
We need to clarify the various possible purposes of trunking determination and the corresponding requirements as to server behavior. The following points should be noted:
For a client using the uniform approach, clientid4 values are treated as important information in determining server trunking patterns. For two different IP addresses to return the same clientid4 value is a necessary, though not a sufficient condition for them to be considered as connected to the same server. As a result, when two different IP addresses return the same clientid4, the client needs to determine, using the procedure given below or otherwise, whether the IP addresses are connected to the same server. For such clients, all internal recording of clientid4 values needs to include, whether explicitly or implicitly, identification of the server from which the clientid4 was received so that one always has a (server, clientid4) pair. Two such pairs from different servers are always considered distinct even when the clientid4 values are the same, as they may occasionally be.
In order to make this approach work, the client must have accessible, for each nfs4_client_id4 used by the uniform approach (only one in general) a list of all server IP addresses, together with the associated clientid4 values and authentication flavors. As a part of the associated data structures, there should be the ability to mark a server IP structure as having the same server as another and to mark an IP-address as currently unresolved. One way to do this is to a allow each such entry to point to another with the pointer value being one of:
In order to keep the above information current, in the interests of the most effective trunking determination, RENEWs should be periodically done on each server. However, even if this is not done, the primary purpose of the trunking determination algorithm, to prevent confusion due to trunking hidden from the client, will be achieved.
Given this apparatus, when a SETCLIENTID is done and a clientid4 returned, the data structure can be searched for a matching clientid4 and if such is found, further processing can be done to determine whether the clientid4 match is accidental, or the result of trunking.
In this algorithm, when SETCLIENTID is done it will use the common nfs_client_id4 and specify the current target IP address as part of the callback parameters. We call the clientid4 and SETCLIENTID verifier returned by this operation XC and XV.
Note that when the client has done previous SETCLIENTID's, to any IP addresses, with more than one authentication flavor, we have the possibility of receiving NFS4ERR_CLID_INUSE, since we do not yet know which of our connections with existing IP addresses might be trunked with our current one. In the event that the SETCLIENID fails with NFS4ERR_CLID_INUSE, one must try all other authentication flavors currently in use and eventually one will be correct and not return NFS4ERR_CLID_INUSE.
Note that at this point, no SETCLIENTID_CONFIRM has yet been done. This is because our SETCLIENTID has either established a new clientid4 on a previously unknown server or changed the callback parameters on a clientid4 associated with some already known server. Given that we don't want to confirm something that we are not sure we want to happen, what is to be done next depends on information about existing clientid4's.
So for each lead IP address IPn with a clientid4 matching XC, the following steps are done.
Note here that we may set a number of possible values for the callback parameters to be used for XC, one for the possibility that X is untrunked, and others for each potential match with an existing IPn. Although there are multiple such updates at most one will be confirmed and, if X is untrunked, its original callback parameters will be put in effect by its SETCLIENID_CONFIRM.
The procedure above has made no explicit mention of the possibility that server reboot can occur at any time. To address this possibility the client should periodically use the clientid4 XC in RENEW operations, directed to both the IP address X and the current lead IP address that is currently being tested for identity.
If we have run out of IPn's without finding a matching server, X is considered as having no existing known IP addresses trunked with it. The IP address is marked as a lead IP address for a new server. A SETCLIENTID_CONFIRM is done using XC and XV.
The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Migration and State" says:
There are a number of problems with this and any resolution of our difficulties must address them somehow.
To avoid client complexity, we need to have no more than one lease between a single client and a single server. This requires merger of leases since there is no real help from synchronizing them at a single instant.
For the uniform approach, the destination server would simply merge leases as part of state transfer, since two leases with the same nfs_client_id4 values must be for the same client.
We have made the following decisions as far as proposed normative statements regarding for state merger. They reflect the facts that we want to support fully migration support in the simplest way possible and that we can't say MUST since we have older clients and servers to deal with.
If the clients and the servers obey the SHOULD's, having more than a single lease for a given client-server pair will be a transient situation, cleaned up as part of adapting to use of migrated state.
Since clients and servers will be a mixture of old and new and because nothing is a MUST we have to ensure that no combination will show worse behavior than is exhibited by current (i.e. old) clients and servers.
The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Migration and State" says:
This poses some difficulties, mostly because the part about "client ID" is not clear:
We have decided to address this issue as follows, with the relevant changes all reflected in Section 5.6.
The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Migration and State" says:
The above will need to be fixed to reflect the possibility of merging of leases and the text to do this appears as part of Section 5.6.
The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Notification of Migrated Lease" says:
There is a lack of clarity that is prompted by ambiguity about what exactly probing is and what the interlock between client and server must be. This has led to some worry about the scalability of the probing process, and although the time required does scale linearly with the number of fs's that the client may have state for with respect to a given server, the actual process can be done efficiently.
To address these issues we propose replacing the above with the text addressing NFS4RR_LEASE_MOVED as given in Section 5.6.3.
Some changes are necessary to reduce confusion about the process of callback information update and in particular to make it clear that no state is freed as a result:
To address both of the clientid4-related issues mentioned in Section 4.4, we propose replacing the last three paragraphs of the section entitled "Client ID" with the following:
It appears to be the intention that only a single authentication flavor be used for client establishment between any client-server pair. However:
As a result, servers exist which reject a SETCLIENTID simply because there already exists a clientid for the same client, established using a different IP address. Although this is generally understood to be erroneous, such servers still exist and the spec should make the correct behavior clear.
Although the error name cannot be changed, the following changes should be made to avoid confusion:
When responsibility for handling a given filesystem is transferred to a new server (migration) or the client chooses to use an alternate server (e.g., in response to server unresponsiveness) in the context of filesystem replication, the appropriate handling of state shared between the client and server (i.e., locks, leases, stateids, and client IDs) is as described below. The handling differs between migration and replication.
If a server replica or a server immigrating a filesystem agrees to, or is expected to, accept opaque values from the client that originated from another server, then it is a wise implementation practice for the servers to encode the "opaque" values in network byte order. When doing so, servers acting as replicas or immigrating filesystems will be able to parse values like stateids, directory cookies, filehandles, etc. even if their native byte order is different from that of other servers cooperating in the replication and migration of the filesystem.
In the case of migration, the servers involved in the migration of a filesystem SHOULD transfer all server state from the original to the new server. This must be done in a way that is transparent to the client. This state transfer will ease the client's transition when a filesystem migration occurs. If the servers are successful in transferring all state, the client will continue to use stateids assigned by the original server. Therefore the new server must recognize these stateids as valid.
If transferring stateids from server to server would result in a conflict for an existing stateid for the destination server with the existing client, transparent state migration MUST NOT happen for that client. Servers participating in using transparent state migration should co-ordinate their stateid assignment policies to make this situation unlikely or impossible. The means by which this might be done, like all of the inter-server interactions for migration, are not specified by the NFS version 4.0 protocol.
Handling of clientid values is similar but not identical. The clientid4 and nfs_client_id4 information (id string and boot verifier) will be transferred with the rest of the state information and the destination server should use that information to determine appropriate clientid4 handling. Although the destination server may make state stored under an existing lease available under the clientid4 used on the source server, the client should not assume that this is always so. In particular,
When leases are not merged, the transfer of state should result in creation of a confirmed client record with empty callback information but matching the {v, x, c} for the transferred client information. This should enable establishment of new callback information using SETCLIENTID and SETCLIENTID_CONFIRM.
A client may determine the disposition of migrated state by using a stateid associated with the migrated state and in an operation on the new server and using the associated clientid4 in a RENEW on the new server.
Since responsibility for an entire filesystem is transferred with a migration event, there is no possibility that conflicts will arise on the new server as a result of the transfer of locks.
The servers may choose not to transfer the state information upon migration. However, this choice is discouraged, except where specific issues such as stateid conflicts make it necessary. In the case of migration without state transfer, when the client presents state information from the original server (e.g. in a RENEW op or a READ op of zero length), the client must be prepared to receive either NFS4ERR_STALE_CLIENTID or NFS4ERR_STALE_STATEID from the new server. The client should then recover its state information as it normally would in response to a server failure. The new server must take care to allow for the recovery of state information as it would in the event of server restart.
When a lease is transferred to a new server (as opposed to being merged with a lease already on the new server), a client SHOULD re-establish new callback information with the new server as soon as possible, according to sequences described in sections "Operation 35: SETCLIENTID - Negotiate Client ID" and "Operation 36: SETCLIENTID_CONFIRM - Confirm Client ID". This ensures that server operations are not blocked by the inability to recall delegations.
In those situation in which state has not been transferred, as shown by a return of NFS4ERR_BAD_STATEID, the client may attempt to reclaim the locks in order to take advantage of cases in which destination server has set up a file-system-specific grace period in support of the migration.
Since client switch-over in the case of replication is not under server control, the handling of state is different. In this case, leases, stateids and client IDs do not have validity across a transition from one server to another. The client must re-establish its locks on the new server. This can be compared to the re-establishment of locks by means of reclaim-type requests after a server reboot. The difference is that the server has no provision to distinguish requests reclaiming locks from those obtaining new locks or to defer the latter. Thus, a client re-establishing a lock on the new server (by means of a LOCK or OPEN request), may have the requests denied due to a conflicting lock. Since replication is intended for read-only use of filesystems, such denial of locks should not pose large difficulties in practice. When an attempt to re-establish a lock on a new server is denied, the client should treat the situation as if its original lock had been revoked.
In the case of lease renewal, the client may not be submitting requests for a filesystem that has been migrated to another server. This can occur because of the implicit lease renewal mechanism. The client renews a lease containing state of multiple filesystems when submitting a request to any one filesystem at the server.
In order for the client to schedule renewal of leases that may have been relocated to the new server, the client must find out about lease relocation before those leases expire. Similarly, when migration occurs but there has not been transparent state migration, the client needs to find out about the change soon enough to be able to reclaim the lock within the destination server's grace period. To accomplish this, all operations which implicitly renew leases for a client (such as OPEN, CLOSE, READ, WRITE, RENEW, LOCK, and others), will return the error NFS4ERR_LEASE_MOVED if responsibility for any of the leases to be renewed has been transferred to a new server. Note that when the transfer of responsibility leaves remaining state for that lease on the source server, the lease is renewed just as it would have been in the NFS4ERR_OK case, despite returning the error. The transfer of responsibility happens when the server receives a GETATTR(fs_locations) from the client for each filesystem for which a lease has been moved to a new server. Normally it does this after receiving an NFS4ERR_MOVED for an access to the filesystem but the server is not required to verify that this happens in order to terminate the return of NFS4ERR_LEASE_MOVED. By convention, the compounds containing GETATTR(fs_locations) SHOULD include an appended RENEW operation to permit the server to identify the client getting the information.
Note that the NFS4ERR_LEASE_MOVED error is only required when responsibility for at least one stateid has been affected. In the case of a null lease, where the only associated state is a clientid, no NFS4ERR_LEASE_MOVED error need be generated.
Upon receiving the NFS4ERR_LEASE_MOVED error, a client that supports filesystem migration MUST perform the necessary GETATTR operation for each of the filesystems containing state that have been migrated and so give the server evidence that it is aware of the migration of the filesystem. Once the client has done this for all migrated filesystems on which the client holds state, the server MUST resume normal handling of stateful requests from that client.
One way in which clients can do this efficiently in the presence of large numbers of filesystems is described below. This approach divides the process into two phases, one devoted to finding the migrated filesystems and the second devoted to doing the necessary GETATTRs.
The client can find the migrated filesystems by building and issuing one or more COMPOUND requests, each consisting of a set of PUTFH/GETFH pairs, each pair using an fh in one of the filesystems in question. All such COMPOUND requests can be done in parallel. The successful completion of such a request indicates that none of the fs's interrogated have been migrated while termination with NFS4ERR_MOVED indicates that the filesystem getting the error has migrated while those interrogated before it in the same COMPOUND have not. Those whose interrogation follows the error remain in an uncertain state and can be interrogated by restarting the requests from after the point at which NFS4ERR_MOVED was returned or by issuing a new set of COMPOUND requests for the filesystems which remain in an uncertain state.
Once the migrated filesystems have been found, all that is needed is for the client to give evidence to the server that it is aware of the migrated status of filesystems found by this process, by interrogating the fs_locations attribute for an fh within each of the migrated filesystems. The client can do this by building and issuing one or more COMPOUND requests, each of which consists of a set of PUTFH operations, each followed by a GETATTR of the fs_locations attribute. A RENEW follows to help tie the operations to the lease returning NFS4ERR_LEASE_MOVED. Once the client has done this for all migrated filesystems on which the client holds state, the server will resume normal handling of stateful requests from that client.
In order to support legacy clients that do not handle the NFS4ERR_LEASE_MOVED error correctly, the server SHOULD time out after a wait of at least two lease periods, at which time it will resume normal handling of stateful requests from all clients. If a client attempts to access the migrated files, the server MUST reply NFS4ERR_MOVED.
When the client receives an NFS4ERR_MOVED error, the client can follow the normal process to obtain the new server information (through the fs_locations attribute) and perform renewal of those leases on the new server. If the server has not had state transferred to it transparently, the client will receive either NFS4ERR_STALE_CLIENTID or NFS4ERR_STALE_STATEID from the new server, as described above. The client can then recover state information as it does in the event of server failure.
Aside from recovering from a migration, there are other reasons a client may wish to retrieve fs_locations information from a server. When a server becomes unresponsive, for example, a client may use cached fs_locations data to discover an alternate server hosting the same fs data. A client may periodically request fs_locations data from a server in order to keep its cache of fs_locations data fresh.
Since a GETATTR(fs_locations) operation would be used for refreshing cached fs_locations data, a server could mistake such a request as indicating recognition of an NFS4ERR_LEASE_MOVED condition. Therefore a compound which is not intended to signal that a client has recognized a migrated lease SHOULD be prefixed with a guard operation which fails with NFS4ERR_MOVED if the file handle being queried is no longer present on the server. The guard can be as simple as a GETFH operation.
Though unlikely, it is possible that the target of such a compound could be migrated in the time after the guard operation is executed on the server but before the GETATTR(fs_locations) operation is encountered. When a client issues a GETATTR(fs_locations) operation as part of a compound not intended to signal recognition of a migrated lease, it SHOULD be prepared to process fs_locations data in the reply that shows the current location of the fs is gone.
In order that the client may appropriately manage its leases in the case of migration, the destination server must establish proper values for the lease_time attribute.
When state is transferred transparently, that state should include the correct value of the lease_time attribute. The lease_time attribute on the destination server must never be less than that on the source since this would result in premature expiration of leases granted by the source server. Upon migration in which state is transferred transparently, the client is under no obligation to re-fetch the lease_time attribute and may continue to use the value previously fetched (on the source server).
In the case in which lease merger occurs as part of state transfer, the lease_time attribute of the destination lease remains in effect. The client can simply renew that lease with its existing lease_time attribute. State in the source lease is renewed at the time of transfer so that it cannot expire, as long as the destination lease is appropriately renewed.
If state has not been transferred transparently (i.e., the client need to reclaim or re-obtain its locks), the client should fetch the value of lease_time on the new (i.e., destination) server, and use it for subsequent locking requests. However the server must respect a grace period at least as long as the lease_time on the source server, in order to ensure that clients have ample time to reclaim their locks before potentially conflicting non-reclaimed locks are granted. The means by which the new server obtains the value of lease_time on the old server is left to the server implementations. It is not specified by the NFS version 4.0 protocol.
The purpose of this section is to examine the troubling results reported in Section 3.1. We will look at the scenarios as they would be handled within the proposal.
Because the choice of uniform vs. non-uniform nfs_client_id4 id strings is a "SHOULD" in these cases, we will designate clients that follow this recommendation by SHOULD-UF-CID.
We will also have to take account of any merger-related "SHOULD" clauses to better understand how they have addressed the issues seen. We abbreviate as follows:
Let's look at the troublesome situation cited in Section 3.1.1. We have already seen what happens when SHOULD-UF-CID does not hold. Now let's look at the situation in which SHOULD-UF-CID holds, whether SHOULD-SVR-AM is in effect or not.
The correctness signature for this issue is
so if you have clients and servers that obey the SHOULD clauses, the problem is gone regardless of the choice on the MAY.
Now let's consider the scenario given in Section 3.1.2. We have already seen what happens when SHOULD-UF-CID does not hold . Now let's look at the situation in which SHOULD-UF-CID holds and SHOULD-SVR-AM holds as well.
Now let's consider the same scenario in the situation in which SHOULD-UF-CID holds and SHOULD-SVR-AM holds as well.
The correctness signature for this issue is
so if you have clients and servers that obey the SHOULD clauses, the problem is gone regardless of the choice on the MAY.
Consider the following situation:
Now look what will happen under various scenarios:
The correctness signature for this issue is
so if you have clients and servers that obey the SHOULD clauses, the problem is gone regardless of the choice on the MAY.
We have seen that (SHOULD-SVR-AM & SHOULD-UF-CID) are sufficient to solve the problems people have experienced.
Because NFSv4.1 embraces the uniform client-string approach, addressing migration issues is simpler. In the terms of Section 6, we already have SHOULD-UF-CID, for NFSv4.1, as advised by section 2.4 of [RFC5661], simplifying the work to be done.
Nevertheless, there are some issues that will have to be addressed. Some examples:
Discussion of how to resolve these issues will appear in the sections below.
The existing treatment of state transfer in [RFC5661], has similar problems to that in [RFC3530] in that it assumes that the state for multiple fs's on different servers will not be merged to so that it appears under a single common clientid. We've already seen the reasons that this is a problem, with regard to NFSv4.0.
Although we don't have the problems stemming from the non-uniform client-string approach, there are a number of complexities in the existing treatment of state management in the section entitled "Lock State and File System Transitions" in [RFC5661] that make this non-trivial to address:
This is made difficult because, within the PNFS framework, migration might mean any of several things:
Migration needs to support both the first and last of these models.
Section 2.10.5 of [RFC5661] states the following.
While this paragraph is literally true in that such reconfiguration events can happen and clients have to deal with them, it is confusing in that it can be read as suggesting that clients have to deal with them without disruption, which in general is impossible.
A clearer alternative would be:
In dealing with file system transitions, the client needs to handle cases in which the two servers have cooperated in state management and cases in which they have not.
The primary means by which a client finds out about state management co-operation is by comparing eir_server_scope values returned by each server. If the scope values do not match, then any co-operation of the servers in state management, is limited to transferring state in event of migration and making arrangements for the safe reclamation of locking state. If the scope values match, then this indicates the servers have cooperated in assigning client IDs and stateids to the point that the same id will not refer to different things on different servers. Servers may reject client IDs that refer to state they do not know about. See the section entitled "Server Scope" for more information about the use of server scope.
How the client needs to deal with locking state with regard to these situations will depend upon:
We will divide the basic description of these possibilities into three sections
In the case of migration, the servers involved in the migration of a file system SHOULD transfer all server state relevant to the migrating file system from the original to the new server. When this is done, it needs to be done in a way that is maximally transparent to the client in that all stateids used by the client to access state on the filesystem in question can be used on the new server, albeit possibly under different client IDs.
When layouts are active for a migrated file system, layout state SHOULD be included as part of the state transferred. Even if it is the case that there are circumstances preventing the layout from being supported on the new server, this should be dealt with by recalling layouts either before or after the transition. Where this cannot be done, layout revocation is possible but any such revocation should appear to the client just as any other layout revocation would.
With replication, such a degree of common state is typically not the case. Clients, however, should use the information provided by the eir_server_scope returned by EXCHANGE_ID (as modified by the validation procedures described in the section entitled "Server Scope") to determine whether such sharing may be in effect in non-migration cases, rather than making assumptions based solely on the reason for the transition.
This state transfer will reduce disruption to the client when a file system transition occurs. If the servers are successful in transferring all state, the client can access existing stateids, using either existing or new sessions between the client and the new server instance. If the server accepts such a transferred stateid as valid, then the client may use that stateid to access the same state that it represented on the old server.
When the two servers belong to the same server scope, it does not mean that when dealing with the transition, the client will not have to reclaim or otherwise reobtain state. However, it does mean that the client may proceed using its current stateids when communicating with the new server, and the new server will either recognize the stateids as valid or reject them, in which case locking state must be reobtained by the client.
File systems cooperating in state management may actually share state or simply divide the identifier space so as to recognize (and reject as stale) each other's stateids and client IDs. Servers that do share state may not do so under all conditions or at all times. If the server cannot be sure when accepting a stateid that it reflects the locks the client was given, the server must treat the state as stale and report it as such to the client.
When the two file system instances are on servers that do not share a server scope value, the client must establish a new client ID on the destination, if it does not have one already, to obtain access to its locks. Depending on the type of file system transition and facilities provided by the server, it may re-establish its connection to locking and layout state in a number of ways.
In the case of migration, the servers may have transferred stateids, making it possible for the client to access his state on the new server, simply by using the existing stateid. The server may transfer all state or a subset and the client can use TEST_STATEID to determine what state has been transferred and what needs to be reclaimed or otherwise reobtained as described in Section 8.3.
Lock reclaim may be used by the client for any sort of file system transition, but the server is not required to support it in any particular case.
Note that in this case, lock reclaim may be attempted even when the servers involved in the transfer have different server scope values (see Section 8.4.2.1 for the contrary case of reclaim after server reboot). Servers with different server scope values may cooperate to allow reclaim for locks associated with the transfer of a file system even if they do not cooperate sufficiently to share a server scope.
In either case, when actual locks are not known to be maintained, the destination server may establish a grace period specific to the given file system, with non-reclaim locks being rejected for that file system, even though normal locks are being granted for other file systems. Clients should not infer the absence of a grace period for file systems being transitioned to a server from responses to requests for other file systems.
In the case of lock reclamation for a given file system after a file system transition, edge conditions can arise similar to those for reclaim after server restart (although in the case of the planned state transfer associated with migration, these can be avoided by securely recording lock state as part of state migration). Unless the destination server can guarantee that locks will not be incorrectly granted, the destination server should not allow lock reclaims and should avoid establishing a grace period.
Once all locks have been reclaimed, or there were no locks to reclaim, the client indicates that there are no more reclaims to be done for the file system in question by sending a RECLAIM_COMPLETE operation with the rca_one_fs parameter set to true. Once this has been done, non-reclaim locking operations may be done, and any subsequent request to do a reclaim will be rejected with the error NFS4ERR_NO_GRACE.
Information about client identity may be propagated between servers in the form of a client_owner4 and associated verifiers, under the assumption that the client presents the same values to all the servers with which it deals.
Servers are encouraged to provide facilities to allow locks to be reclaimed on the new server after a file system transition. Often, however, in cases in which the two servers do not share a server scope value, such facilities may not be available and the client should be prepared to re-obtain locks, even though it is possible that the client may have its LOCK or OPEN request denied due to a conflicting lock.
Layouts may be reobtained when necessary even without special facilities for lock reclamation. However, the client MUST NOT depend on being able to obtain such layout since pNFS or the desired mapping type might not be supported on the new server.
The consequences of having no facilities available to reclaim locks on the new server will depend on the type of environment. In some environments, such as the transition between read-only file systems, such denial of locks should not pose large difficulties in practice. When an attempt to re-establish a lock on a new server is denied, the client should treat the situation as if its original lock had been revoked. Note that when the lock is granted, the client cannot assume that no conflicting lock could have been granted in the interim. Where change attribute continuity is present, the client may check the change attribute to check for unwanted file modifications. Where even this is not available, and the file system is not read-only, a client may reasonably treat all pending locks as having been revoked.
The current definitive definition of the NFSv4.0 protocol [RFC3530], and the current pending draft of RFC3530bis [cur-v4.0-bis] both agree. The section entitled "Security Considerations" encourages that clients protect the integrity of the SECINFO operation, any GETATTR operation for the fs_locations attribute, and the operations SETCLIENTID/SETCLIENTID_CONFIRM. A migration recovery event can use any or all of these operations. We do not recommend any change here.
This document does not require actions by IANA.
The editor and authors of this document gratefully acknowledge the contributions of Trond Myklebust of NetApp and Robert Thurlow of Oracle. We also thank Tom Haynes of NetApp and Spencer Shepler of Microsoft for their guidance and suggestions.
Special thanks go to members of the Oracle Solaris NFS team, especially Rick Mesta and James Wahlig, for their work implementing an NFSv4.0 migration prototype and identifying many of the issues documented here.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. |
| [RFC3530] | Shepler, S., Callaghan, B., Robinson, D., Thurlow, R., Beame, C., Eisler, M. and D. Noveck, "Network File System (NFS) version 4 Protocol", RFC 3530, April 2003. |
| [RFC5661] | Shepler, S., Eisler, M. and D. Noveck, "Network File System (NFS) Version 4 Minor Version 1 Protocol", RFC 5661, January 2010. |
| [cur-v4.0-bis] |
Haynes, T. and D. Noveck, "Network File System (NFS) Version 4 Protocol", 2011. Work in progress. |