linux-cifs-client Digest, Vol 70, Issue 25
Subsystems (2)
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 11 lines
Quoted 11 lines
This patchset is still preliminary and is just an RFC...
First, some background. When I was at Connectathon this year, Trond mentioned an interesting idea to me. He said (paraphrasing):
"Why doesn't CIFS just use the RPC layer for transport? It's very efficient at shoveling bits out onto the wire. You'd just need to abstract out the XDR/RPC specific bits."
My first reaction is that if you abstract out XDR/RPC specific parts of SunRPC it isn't SunRPC,
just a scheduler on top of tcp (not a bad thing in theory). Pulling out the two key pieces from
SunRPC:
- asynchronous event handling and scheduling
- upcall for credentials
could be useful, but does add a lot of complexity. If there is a way to use just the async
scheduling (and perhaps upcall) out of SunRPC, that part sounds fine as long as it can skip the encoding/decoding and just pass in a raw kvec containing the SMB header and data.
Quoted 5 lines
Quoted 5 lines
CIFS in particular is also designed around synchronous ops, which seriously limits throughput. Retrofitting it for asynchronous operation will be adding even more kludges.
There are only three operations that we can send asynchronous today, all of which require special case handling in the VFS already:
- readpages
- writepages
- blocking locks
(and also directory change notification which we and nfs don't do). I think the "slow_work"
mechanism is probably sufficient for these cases already.
Quoted 7 lines
Quoted 7 lines
works in our favor...
------------------------------------------------------------------------
Q: can we hook up cifs or smbfs to use this as a transport?A: Not trivially. CIFS in particular is not designed with each call having discrete encode and decode functions. They're sort of mashed
We certainly don't want to move to an abstract encoding mechanism, especially for SMB2 where there is only one encoding of wire operations, and no duplicate requests due
to 20 years of dialects. I can see an argument for abstract encoding for requests
like SMB open, vs. SMB OpenX vs. SMB NTCreateX but this would be harder or to abstract and has to be done case by case anyway due to differences in field length, missing fields, different compensations. It is not
like the simpler NFS case where encoding involves endian conversion etc.
Quoted 11 lines
Quoted 11 lines
------------------------------------------------------------------------
Q: could we use this as a transport layer for a smb2fs ?A: Yes, I think so. This particular prototype is build around SMB1, but SMB2 could be supported with only minor modifications. One of the reasons for sending this patchset now before I've built a filesystem on top of it is because I know that SMB2 work is in progress. I'd like to see it based around a more asynchronous transport model, or at least built with cleaner layering so that we can eventually bolt on a different transport layer if we so choose.
Amost all the ops use "send_receive" already - so there is no need to change the code much above
that if you want to experiment with changing the transport. I like the idea of the
abtraction of async operations, and creating completion routines (and an async send abstraction) for readpages, writepages and directory change notification would make sense. but in both cifs and smb2, the 95% of the operations that must be synchronous in the VFS (open, lookup, unlink, create etc.) can already be hooked up to any transport as long as it can send a kvec contain fs data and return a response (like the "send_receive" and equivalent).
The idea of doing abstract translation and encoding of SMB protocol frames does seem overengineered and probably would make it harder to read/understand the setup of certain complex request frames which are quite different from
Samba to Windows. As another example, generalized, abstract SMB frame
conversion isn't being done in Samba 3 for example, and with only
19 requests in SMB2 it makes even less sense. On the client, since
we have control over which types of requests we send, our case
is simpler than for the server for sending requests, but in response processing since we have to work around server bugs, xdr like decoding of SMB responses could get harder still.
I like the idea of the way SunRPC keeps task information, and it may make it easier to carry credentials around (although I think using Dave Howell's key management code
might be ok instead to access Winbind). I am not sure how easy it would be to tie
SunRPC credential mapping to Winbind but that could probably be done. I like the async scheduling capability of SunRPC although I suspect that it is a factor in a number of the (nfs client) performance problems we have seen so may need more work. I don't like adding (in effect) an extra transport and "encoding layer" though to
protocols (cifs and smb2). NFS since it is built on SunRPC on the wire, required
such a layer, and it makes sense for NFS to layer the code, like their protocol,
over SunRPC. CIFS and SMB2 don't require (or even allow) XDR translation,
variable encodings, and SunRPC encapsulation so the idea of abstracting the encoding of something that has a single defined encoding seems wrong.
Quoted 11 lines · On Mon, 28 Sep 2009 09:41:08 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 11 lines
Quoted 11 lines
This patchset is still preliminary and is just an RFC...
First, some background. When I was at Connectathon this year, Trond mentioned an interesting idea to me. He said (paraphrasing):
"Why doesn't CIFS just use the RPC layer for transport? It's very efficient at shoveling bits out onto the wire. You'd just need to abstract out the XDR/RPC specific bits."
Quoted 15 lines
My first reaction is that if you abstract out XDR/RPC specific parts of SunRPC it isn't SunRPC,
just a scheduler on top of tcp (not a bad thing in theory). Pulling
out the two key pieces from
SunRPC:
- asynchronous event handling and scheduling
- upcall for credentials
could be useful, but does add a lot of complexity. If there is a way
to use just the async
scheduling (and perhaps upcall) out of SunRPC, that part sounds fine as long as it
can skip the encoding/decoding and just pass in a raw kvec containing the SMB
header and data.
Well, the sunrpc layer currently contains a lot of pieces:
1) client side call/response handling (clnt routines)
2) server side call/response handling (svc routines)
3) XDR encoding and decoding routines (including crypto signatures, etc)
...the idea is to hook up new encoding and decoding routines and to add a new "transport class" which will make the client-side scheduler handle SMB/SMB2 properly.
We'll also eventually have to add new authentication/credential "classes" too. I haven't researched that yet in any real depth, so I can't state much about how difficult it'll be.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
CIFS in particular is also designed around synchronous ops, which seriously limits throughput. Retrofitting it for asynchronous operation will be adding even more kludges.
Quoted 10 lines
There are only three operations that we can send asynchronous today, all of which require
special case handling in the VFS already:
- readpages
- writepages
- blocking locks
(and also directory change notification which we and nfs don't do). I
think the "slow_work"
mechanism is probably sufficient for these cases already.
The problem is that rolling a mechanism to handle asynchronous ops is difficult to get right. I think it makes a lot of sense to reuse a proven engine here. It also makes a lot of sense to implement synchronous ops on top of an asynchronous infrastructure. RPC does this under the hood, and so did smbfs.
What you're proposing, in effect, is to do this in reverse -- implement an asynchronous transport engine using synchronous ops and offloading the background parts onto threads. That's possible I suppose, but it means you have a lot of tasks sleeping in the kernel and waiting for stuff to happen.
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
works in our favor...
------------------------------------------------------------------------
Q: can we hook up cifs or smbfs to use this as a transport?A: Not trivially. CIFS in particular is not designed with each call having discrete encode and decode functions. They're sort of mashed
Quoted 11 lines
We certainly don't want to move to an abstract encoding mechanism, especially for SMB2
where there is only one encoding of wire operations, and no duplicate requests due
to 20 years of dialects. I can see an argument for abstract encoding for requests
like SMB open, vs. SMB OpenX vs. SMB NTCreateX but this would be harder or to abstract and has to be done case by case anyway due to differences in field length, missing fields, different compensations. It is not like the simpler NFS case where encoding involves endian conversion etc.
I'm not sure what you mean by this. Assembling an SMB header and call is very similar to assembling an RPC header and call. There are differences of course, but they aren't that substantial.
SMB does introduce some more interesting wrinkles. For instance, since state is tied up with the actual socket connection, we'll probably need callbacks into the fs for socket state changes. That doesn't have much to do with how you abstract out the encoding and decoding though.
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
------------------------------------------------------------------------
Q: could we use this as a transport layer for a smb2fs ?A: Yes, I think so. This particular prototype is build around SMB1, but SMB2 could be supported with only minor modifications. One of the reasons for sending this patchset now before I've built a filesystem on top of it is because I know that SMB2 work is in progress. I'd like to see it based around a more asynchronous transport model, or at least built with cleaner layering so that we can eventually bolt on a different transport layer if we so choose.
Quoted 16 lines
Amost all the ops use "
send_receive" already - so there is no need to change the code much above
that if you want to experiment with changing the transport. I like the
idea of the
abtraction of async operations, and creating completion routines (and an async send
abstraction) for readpages, writepages and directory change notification would make sense.
but in both cifs and smb2, the 95% of the operations that must be synchronous in
the VFS (open, lookup, unlink, create etc.) can already be hooked up to any transport
as long as it can send a kvec contain fs data and return a response (like the "send_receive"
and equivalent).
The problem with the send_receive interface is that it assumes that the encoding, send and decoding will be done by the same task. I think that assumption will greatly limit this code later and force you to rely on workarounds (like slow_work) to get asynchronous behavior.
At the very least, I suggest splitting off the decode portions into separate functions. That at least should allow you the ability later to offload that part to other tasks (similar to how async tasks get offloaded to rpciod).
Quoted 12 lines
The idea of doing abstract translation and encoding of SMB protocol frames does seem overengineered and probably would make it harder to read/understand
the setup of certain complex request frames which are quite different from
Samba to Windows. As another example, generalized, abstract SMB frame
conversion isn't being done in Samba 3 for example, and with only
19 requests in SMB2 it makes even less sense. On the client, since
we have control over which types of requests we send, our case is simpler than for the server for sending requests, but in response processing since we have to work around server bugs, xdr like decoding of SMB responses could get harder still.
Again, I don't see SMB as being that different from NFS in this regard. You have a transport header (similar to the fraghdr with NFS TCP transport), then a protocol header (the SMB/SMB2 header), and then call-specific information. RPC/NFS works exactly the same way.
The code I've proposed more or less has abstractions along those lines. There's:
transport class -- (net/sunrpc/xprtsmb.c)
header encoding/decoding -- (net/sunrpc/smb.c)
...the other parts will be implemented by the filesystem (similar to how fs/nfs/nfs?xdr.c work).
Quoted 13 lines
I like the idea of the way SunRPC keeps task information, and it may make it easier
to carry credentials around (although I think using Dave Howell's key management code
might be ok instead to access Winbind). I am not sure how easy it
would be to tie
SunRPC credential mapping to Winbind but that could probably be done. I like the
async scheduling capability of SunRPC although I suspect that it is a factor in
a number of the (nfs client) performance problems we have seen so may need more work.
I don't like adding (in effect) an extra transport and "encoding layer"
Quoted 8 lines
though to
protocols (cifs and smb2). NFS since it is built on SunRPC on the
wire, required
such a layer, and it makes sense for NFS to layer the code, like their protocol,
over SunRPC. CIFS and SMB2 don't require (or even allow) XDR translation,
variable encodings, and SunRPC encapsulation so the idea of abstracting the encoding of something that has a single defined encoding seems wrong.
I'm not sure I understand this last comment. CIFS/SMB2 and NFS are just not that different in this regard. Either way you have to marshal up the buffer correctly before you send it, and decode it properly.
Is it possible to go too far with layering and abstraction? Sure. I don't think I've done that here though.
Signature · 1 line
Jeff Layton <[email protected]>
A message that is not in the archive (sent off-list or not yet received): 20090928115427.4103826e-9yPaYZwiELC+kQycOl6kW4xkIHaj4LzF@public.gmane.org
Re: [RFC PATCH 0/9] sunrpc: teach the SUNRPC layer how to speak SMB
Quoted 63 lines · Jeff Layton
Quoted 11 lines · On Mon, 28 Sep 2009 09:41:08 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 11 lines
Quoted 11 lines
This patchset is still preliminary and is just an RFC...
First, some background. When I was at Connectathon this year, Trond mentioned an interesting idea to me. He said (paraphrasing):
"Why doesn't CIFS just use the RPC layer for transport? It's very efficient at shoveling bits out onto the wire. You'd just need to abstract out the XDR/RPC specific bits."
Quoted 11 lines
My first reaction is that if you abstract out XDR/RPC specific parts of SunRPC it isn't SunRPC,
just a scheduler on top of tcp (not a bad thing in theory). Pulling out the two key pieces from
SunRPC:
- asynchronous event handling and scheduling
- upcall for credentials
could be useful, but does add a lot of complexity. If there is a way to use just the async
scheduling (and perhaps upcall) out of SunRPC, that part sounds fine as long as it can skip the encoding/decoding and just pass in a raw kvec containing the SMB header and data.Well, the sunrpc layer currently contains a lot of pieces:
1) client side call/response handling (clnt routines)
2) server side call/response handling (svc routines)
3) XDR encoding and decoding routines (including crypto signatures, etc)...the idea is to hook up new encoding and decoding routines and to add a new "transport class" which will make the client-side scheduler handle SMB/SMB2 properly.
We'll also eventually have to add new authentication/credential "classes" too. I haven't researched that yet in any real depth, so I can't state much about how difficult it'll be.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
CIFS in particular is also designed around synchronous ops, which seriously limits throughput. Retrofitting it for asynchronous operation will be adding even more kludges.
Quoted 9 lines
There are only three operations that we can send asynchronous today, all of which require special case handling in the VFS already:
- readpages
- writepages
- blocking locks
(and also directory change notification which we and nfs don't do). I think the "slow_work"
mechanism is probably sufficient for these cases already.The problem is that rolling a mechanism to handle asynchronous ops is difficult to get right. I think it makes a lot of sense to reuse a proven engine here. It also makes a lot of sense to implement synchronous ops on top of an asynchronous infrastructure. RPC does this under the hood, and so did smbfs.
Quoted 7 lines
What you're proposing, in effect, is to do this in reverse -- implement an asynchronous transport engine using synchronous ops and offloading the background parts onto threads. That's possible I suppose, but it means you have a lot of tasks sleeping in the kernel and waiting for stuff to happen.
I don't think it changes the number of tasks sleeping. For sync operations it doesn't
change how many tasks that are sleeping. For async operations, you no longer have a
task sleeping in cifs_writepages or cifs_readpages, but do have the ability to dispatch a decode routine when the SMB Write or Read response is returned (may or
may not have a pool to do this). Seems like about the same number of task (possibly
less). Moving to all asynchronous operations under open, unlink, close doesn't reduce
the number of sleeping tasks (it may increase it or stay the same).
Quoted 56 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
works in our favor...
------------------------------------------------------------------------
Q: can we hook up cifs or smbfs to use this as a transport?A: Not trivially. CIFS in particular is not designed with each call having discrete encode and decode functions. They're sort of mashed
Quoted 9 lines
We certainly don't want to move to an abstract encoding mechanism, especially for SMB2 where there is only one encoding of wire operations, and no duplicate requests due to 20 years of dialects. I can see an argument for abstract encoding for requests like SMB open, vs. SMB OpenX vs. SMB NTCreateX but this would be harder or to abstract and has to be done case by case anyway due to differences in field length, missing fields, different compensations. It is not
like the simpler NFS case where encoding involves endian conversion etc.I'm not sure what you mean by this. Assembling an SMB header and call is very similar to assembling an RPC header and call. There are differences of course, but they aren't that substantial.
SMB does introduce some more interesting wrinkles. For instance, since state is tied up with the actual socket connection, we'll probably need callbacks into the fs for socket state changes. That doesn't have much to do with how you abstract out the encoding and decoding though.
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
------------------------------------------------------------------------
Q: could we use this as a transport layer for a smb2fs ?A: Yes, I think so. This particular prototype is build around SMB1, but SMB2 could be supported with only minor modifications. One of the reasons for sending this patchset now before I've built a filesystem on top of it is because I know that SMB2 work is in progress. I'd like to see it based around a more asynchronous transport model, or at least built with cleaner layering so that we can eventually bolt on a different transport layer if we so choose.
Quoted 10 lines
Amost all the ops use "
send_receive" already - so there is no need to change the code much above
that if you want to experiment with changing the transport. I like the idea of the
abtraction of async operations, and creating completion routines (and an async send abstraction) for readpages, writepages and directory change notification would make sense. but in both cifs and smb2, the 95% of the operations that must be synchronous in the VFS (open, lookup, unlink, create etc.) can already be hooked up to any transport as long as it can send a kvec contain fs data and return a response (like the "send_receive" and equivalent).The problem with the
send_receiveinterface is that it assumes that the encoding, send and decoding will be done by the same task. I think that assumption will greatly limit this code later and force you to rely on workarounds (likeslow_work) to get asynchronous behavior.
Quoted 6 lines
At the very least, I suggest splitting off the decode portions into separate functions. That at least should allow you the ability later to offload that part to other tasks (similar to how async tasks get offloaded to rpciod).
That (splitting decode into a distinct helper) makes sense (at least for async capable ops, in particular write, read and directory change notification and byte range locks). The "smb_decode_write_response" case is a particularly simple and useful one to do and would
be an easy one to do as a test. I think the prototype patches for async write that someone did
for cifs a few years ago did that.
Quoted 13 lines
Quoted 12 lines
The idea of doing abstract translation and encoding of SMB protocol frames does seem overengineered and probably would make it harder to read/understand the setup of certain complex request frames which are quite different from
Samba to Windows. As another example, generalized, abstract SMB frame
conversion isn't being done in Samba 3 for example, and with only
19 requests in SMB2 it makes even less sense. On the client, since
we have control over which types of requests we send, our case
is simpler than for the server for sending requests, but in response processing since we have to work around server bugs, xdr like decoding of SMB responses could get harder still.
Quoted 6 lines
Again, I don't see SMB as being that different from NFS in this regard. You have a transport header (similar to the fraghdr with NFS TCP transport), then a protocol header (the SMB/SMB2 header), and then call-specific information. RPC/NFS works exactly the same way.
NFS and CIFS/SMB2 seem pretty different to me in this regard. CIFS and SMB2
are much simpler - for these once protocols unlike nfs, after you assemble the CIFS or SMB2 header (possibly with a data area as in SMB Write) you simply add a 4 byte length (actually 3 byte length, 0 byte 0) and
you send it - no endian conversions, xdr, no adding 80 bytes or so rpc prefix. SunRPC adds a whole layer (not just a length field) with credentials. SunRPC
typically adds 80 bytes (or more depending on auth flavor) before the nfs frame (cifs and smb2 frame don't need this, so the net/sunrpc code which handles the 80 bytes or so of rpc header before the network fs frame
is not used for the cifs code)
In SMB2 the pacing is part of the SMB packet, not the transport packet, and in SMB2 any reconnection code which would be built into the proposed modified SunRPC transport would probably have to be aware of the new handle types and locking operations and state that have been added in SMB2.1 which may break the abstraction between network fs and SunRPC transport
Quoted 24 lines
The code I've proposed more or less has abstractions along those lines. There's:
transport class -- (
net/sunrpc/xprtsmb.c)
header encoding/decoding -- (net/sunrpc/smb.c)...the other parts will be implemented by the filesystem (similar to how fs/nfs/nfs?xdr.c work).
Quoted 4 lines
I like the idea of the way SunRPC keeps task information, and it may make it easier to carry credentials around (although I think using Dave Howell's key management code
might be ok instead to access Winbind). I am not sure how easy it would be to tie
SunRPC credential mapping to Winbind but that could probably be done. I like theQuoted 9 lines
async scheduling capability of SunRPC although I suspect that it is a factor in a number of the (nfs client) performance problems we have seen so may need more work. I don't like adding (in effect) an extra transport and "encoding layer" though to
protocols (cifs and smb2). NFS since it is built on SunRPC on the wire, required
such a layer, and it makes sense for NFS to layer the code, like their protocol,
over SunRPC. CIFS and SMB2 don't require (or even allow) XDR translation,
variable encodings, and SunRPC encapsulation so the idea of abstracting the encoding of something that has a single defined encoding seems wrong.
Quoted 5 lines
I'm not sure I understand this last comment. CIFS/SMB2 and NFS are just not that different in this regard. Either way you have to marshal up the buffer correctly before you send it, and decode it properly.
As an example of one of the more complicated cases for cifs (setting a time field). The code looks something like this and is very straightforward to see where the "linux field" is being put into the packet - and to catch errors in size or
endianness. Most cases are simpler for cifs.
struct file_basic_info buf; /* the wire format of the file basic info SMB info level */
buf->LastAccessTime = 0; /* we can't set access time to Windows */ buf->LastWriteTime = cpu_to_le64(cifs_UnixTimeToNT(attrs->ia_atime)); (followed by a few similar assignment statements for the remaining fields)
then send the SMB header followed by the buffer. There is no marshalling
or xdr conversion needed. One thing I like about this approach is that
"sparse" (make modules C=1) immediately catches any mismatch between the wire format (size or endianness) and the vfs data structure element being put in the frame.
Looking at nfs4xdr.c as an example for comparison, it is much harder to see the actual line where a bigendian (or littlenedian) 64 bit quantity is being put into the request frame.
Splitting encode and decode routines probably would make code more reasonable - but converting from a "linux file system thing" to an abstract structure and then to one of many wire possibilities for a frame - doesn't make much sense for the case where there is only one wire encoding (SMB2) or for cases where a single operation maps to more than one frame in some dialects but not others (as in SMB) and so can't be encoded abstractly.
Re: [RFC PATCH 0/9] sunrpc: teach the SUNRPC layer how to speak SMB
Quoted 78 lines · On Mon, 28 Sep 2009 13:40:23 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 70 lines · Jeff Layton
Quoted 11 lines · On Mon, 28 Sep 2009 09:41:08 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 11 lines
Quoted 11 lines
This patchset is still preliminary and is just an RFC...
First, some background. When I was at Connectathon this year, Trond mentioned an interesting idea to me. He said (paraphrasing):
"Why doesn't CIFS just use the RPC layer for transport? It's very efficient at shoveling bits out onto the wire. You'd just need to abstract out the XDR/RPC specific bits."
Quoted 16 lines
My first reaction is that if you abstract out XDR/RPC specific parts of SunRPC it isn't SunRPC,
just a scheduler on top of tcp (not a bad thing in theory). Pulling
out the two key pieces from
SunRPC:
- asynchronous event handling and scheduling
- upcall for credentials
could be useful, but does add a lot of complexity. If there is a way
to use just the async
scheduling (and perhaps upcall) out of SunRPC, that part sounds fine as long as it
can skip the encoding/decoding and just pass in a raw kvec containing the SMB
header and data.Well, the sunrpc layer currently contains a lot of pieces:
1) client side call/response handling (clnt routines)
2) server side call/response handling (svc routines)
3) XDR encoding and decoding routines (including crypto signatures, etc)...the idea is to hook up new encoding and decoding routines and to add a new "transport class" which will make the client-side scheduler handle SMB/SMB2 properly.
We'll also eventually have to add new authentication/credential "classes" too. I haven't researched that yet in any real depth, so I can't state much about how difficult it'll be.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
CIFS in particular is also designed around synchronous ops, which seriously limits throughput. Retrofitting it for asynchronous operation will be adding even more kludges.
Quoted 11 lines
There are only three operations that we can send asynchronous today, all of which require
special case handling in the VFS already:
- readpages
- writepages
- blocking locks
(and also directory change notification which we and nfs don't do). I
think the "slow_work"
mechanism is probably sufficient for these cases already.The problem is that rolling a mechanism to handle asynchronous ops is difficult to get right. I think it makes a lot of sense to reuse a proven engine here. It also makes a lot of sense to implement synchronous ops on top of an asynchronous infrastructure. RPC does this under the hood, and so did smbfs.
Quoted 7 lines
What you're proposing, in effect, is to do this in reverse -- implement an asynchronous transport engine using synchronous ops and offloading the background parts onto threads. That's possible I suppose, but it means you have a lot of tasks sleeping in the kernel and waiting for stuff to happen.
Quoted 12 lines
I don't think it changes the number of tasks sleeping. For sync
operations it doesn't
change how many tasks that are sleeping. For async operations, you no
longer have a
task sleeping incifs_writepagesorcifs_readpages, but do have the ability to dispatch
a decode routine when the SMB Write or Read response is returned (may or
may not have a pool to do this). Seems like about the same number of
task (possibly
less). Moving to all asynchronous operations under open, unlink, close
doesn't reduce
the number of sleeping tasks (it may increase it or stay the same).
I think it does. Take the example of asynchronous writepages requests. You want to send a bunch of write calls in quick succession and then process the replies as they come in. If you rely on a SendReceive-type interface, you have no choice but to spawn a separate slow_work task for each call on the wire. No guarantee they'll run in parallel though (depends on the rules for spawning new kslowd threads).
Now, you could hack up a routine that just sends the calls, sprinkle in a callback routine that's run by cifsd or something when the writes come back (though you'll need to be wary of deadlock, of course).
...or you could just start with a transport layer that's designed for this from the beginning. That's what I'm suggesting that we do.
Heck, samba does the exact same thing. It's transport layer is fundamentally asynchronous too (built around the tevent lib).
Quoted 72 lines
Quoted 66 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
works in our favor...
------------------------------------------------------------------------
Q: can we hook up cifs or smbfs to use this as a transport?A: Not trivially. CIFS in particular is not designed with each call having discrete encode and decode functions. They're sort of mashed
Quoted 12 lines
We certainly don't want to move to an abstract encoding mechanism, especially for SMB2
where there is only one encoding of wire operations, and no duplicate requests due
to 20 years of dialects. I can see an argument for abstract encoding for requests
like SMB open, vs. SMB OpenX vs. SMB NTCreateX but this would be harder or to abstract and has to be done case by case anyway due to differences in field length, missing fields, different compensations. It is not like the simpler NFS case where encoding involves endian conversion etc.I'm not sure what you mean by this. Assembling an SMB header and call is very similar to assembling an RPC header and call. There are differences of course, but they aren't that substantial.
SMB does introduce some more interesting wrinkles. For instance, since state is tied up with the actual socket connection, we'll probably need callbacks into the fs for socket state changes. That doesn't have much to do with how you abstract out the encoding and decoding though.
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
------------------------------------------------------------------------
Q: could we use this as a transport layer for a smb2fs ?A: Yes, I think so. This particular prototype is build around SMB1, but SMB2 could be supported with only minor modifications. One of the reasons for sending this patchset now before I've built a filesystem on top of it is because I know that SMB2 work is in progress. I'd like to see it based around a more asynchronous transport model, or at least built with cleaner layering so that we can eventually bolt on a different transport layer if we so choose.
Quoted 17 lines
Amost all the ops use "
send_receive" already - so there is no need to change the code much above
that if you want to experiment with changing the transport. I like the
idea of the
abtraction of async operations, and creating completion routines (and an async send
abstraction) for readpages, writepages and directory change notification would make sense.
but in both cifs and smb2, the 95% of the operations that must be synchronous in
the VFS (open, lookup, unlink, create etc.) can already be hooked up to any transport
as long as it can send a kvec contain fs data and return a response (like the "send_receive"
and equivalent).The problem with the
send_receiveinterface is that it assumes that the encoding, send and decoding will be done by the same task. I think that assumption will greatly limit this code later and force you to rely on workarounds (likeslow_work) to get asynchronous behavior.Quoted 6 lines
At the very least, I suggest splitting off the decode portions into separate functions. That at least should allow you the ability later to offload that part to other tasks (similar to how async tasks get offloaded to rpciod).
Quoted 9 lines
That (splitting decode into a distinct helper) makes sense (at least for async capable ops,
in particular write, read and directory change notification and byte range locks). The
"smb_decode_write_response" case is a particularly simple and useful one to do and would
be an easy one to do as a test. I think the prototype patches for
async write that someone did
for cifs a few years ago did that.
Even for ops that are fundamentally synchronous, it makes sense to split the encoding and decoding into separate routines. Maybe it's just my personal bias, but I think it encourages a cleaner, more flexible design.
Quoted 36 lines
Quoted 14 lines
Quoted 13 lines
The idea of doing abstract translation and encoding of SMB protocol frames does seem overengineered and probably would make it harder to read/understand
the setup of certain complex request frames which are quite different from
Samba to Windows. As another example, generalized, abstract SMB frame
conversion isn't being done in Samba 3 for example, and with only
19 requests in SMB2 it makes even less sense. On the client, since
we have control over which types of requests we send, our case is simpler than for the server for sending requests, but in response processing since we have to work around server bugs, xdr like decoding of SMB responses could get harder still.Quoted 6 lines
Again, I don't see SMB as being that different from NFS in this regard. You have a transport header (similar to the fraghdr with NFS TCP transport), then a protocol header (the SMB/SMB2 header), and then call-specific information. RPC/NFS works exactly the same way.
NFS and CIFS/SMB2 seem pretty different to me in this regard. CIFS and
SMB2
are much simpler - for these once protocols unlike nfs, after you assemble the CIFS or SMB2 header (possibly with a data area as in SMB Write) you simply add a 4 byte length (actually 3 byte length, 0 byte 0) and you send it - no endian conversions, xdr, no adding 80 bytes or so rpc prefix.
SunRPC adds a whole layer (not just a length field) with credentials. SunRPC
typically adds 80 bytes (or more depending on auth flavor) before the nfs frame
(cifs and smb2 frame don't need this, so the net/sunrpc code which handles the
80 bytes or so of rpc header before the network fs frame
is not used for the cifs code)
Quoted 7 lines
In SMB2 the pacing is part of the SMB packet, not the transport packet, and in
SMB2 any reconnection code which would be built into the proposed modified SunRPC transport would probably have to be aware of the new handle types and locking operations and state that have been added in SMB2.1 which may break the abstraction between network fs and SunRPC transport
CIFS/SMB most definitely need to do endianness conversions -- BE arches have to convert nearly everything.
Though honestly, you're splitting hairs here. All of these details are meaningless. In NFS there are areas that are simpler to handle than SMB (like the header parsing of the response)...and there are parts that are more difficult (fragheader handling, authentication).
The fact of that matter is that they are just not that different. There are distinct parallels to many of the important pieces:
SMB transport header is 32 bits and contains a length with a zeroed out upper byte. The RPC over TCP transport header is 32 bits and contains a 31 bit length + a "last fragment" flag.
SMB has MID's, RPC has XID's
RPC has the RPC auth info, SMB has the session UID and TID
...the parallels go on...
We'll of course need a different set of arguments to build a SMB header vs. a RPC one. That information comes from the filesystem anyway and it'll know what sort of arguments it needs to pass in.
Quoted 59 lines
Quoted 33 lines
The code I've proposed more or less has abstractions along those lines. There's:
transport class -- (
net/sunrpc/xprtsmb.c)
header encoding/decoding -- (net/sunrpc/smb.c)...the other parts will be implemented by the filesystem (similar to how fs/nfs/nfs?xdr.c work).
Quoted 13 lines
I like the idea of the way SunRPC keeps task information, and it may make it easier
to carry credentials around (although I think using Dave Howell's key management code
might be ok instead to access Winbind). I am not sure how easy it
would be to tie
SunRPC credential mapping to Winbind but that could probably be done. I like the
async scheduling capability of SunRPC although I suspect that it is a factor in
a number of the (nfs client) performance problems we have seen so may need more work.
I don't like adding (in effect) an extra transport and "encoding layer"Quoted 9 lines
though to
protocols (cifs and smb2). NFS since it is built on SunRPC on the
wire, required
such a layer, and it makes sense for NFS to layer the code, like their protocol,
over SunRPC. CIFS and SMB2 don't require (or even allow) XDR translation,
variable encodings, and SunRPC encapsulation so the idea of abstracting the encoding of something that has a single defined encoding seems wrong.Quoted 5 lines
I'm not sure I understand this last comment. CIFS/SMB2 and NFS are just not that different in this regard. Either way you have to marshal up the buffer correctly before you send it, and decode it properly.
As an example of one of the more complicated cases for cifs (setting a time field).
The code looks something like this and is very straightforward to see where the "linux field" is being put into the packet - and to catch errors in size or
endianness. Most cases are simpler for cifs.struct
file_basic_infobuf; /* the wire format of the file basic info SMB info level */
buf->LastAccessTime= 0; /* we can't set access time to Windows */buf->LastWriteTime= cpu_to_le64(cifs_UnixTimeToNT(attrs->ia_atime)); (followed by a few similar assignment statements for the remaining fields)then send the SMB header followed by the buffer. There is no marshalling
or xdr conversion needed. One thing I like about this approach is that
"sparse" (make modules C=1) immediately catches any mismatch between the wire format (size or endianness) and the vfs data structure element being put in the frame.
Quoted 5 lines
Looking at
nfs4xdr.cas an example for comparison, it is much harder to see the actual line
where a bigendian (or littlenedian) 64 bit quantity is being put into the request frame.
Woah...I think we just have a difference in terminology here. When I say "encoding" the packet, I just mean that we're stuffing the data into the buffer in the format that the wire expects.
Now, whether you do that with packed structures (like CIFS does and which I tend to prefer), or via "hand-marshaling" (like the WRITE/READ macros in the nfs?xdr.c files) is simply an implementation detail.
In fact, if you take a look at the sample "smbtest" module, you'll notice that the negotiate protocol packets are "encoded" and "decoded" using packed structures.
In truth, we could probably utilize packed structures in most of the NFS xdr code too, it just hasn't traditionally been done that way...
Quoted 8 lines
Splitting encode and decode routines probably would make code more reasonable - but converting from a "linux file system thing" to an abstract structure
and then to one of many wire possibilities for a frame - doesn't make much sense
for the case where there is only one wire encoding (SMB2) or for cases where a single operation maps to more than one frame in some dialects but not others (as in SMB) and so can't be encoded abstractly.
The transport engine doesn't (and shouldn't) care about any of this. All it needs to know is how to encode the packet given a set of arguments, how to send it, and what to do with the reply. The rest is all the job of the actual filesystem.
The only difference here from what CIFS does today is that this is handled via a discrete set of interfaces to the transport engine rather than just calling SendReceive with an already layed-out buffer.
Signature · 1 line
Jeff Layton <[email protected]>
A message that is not in the archive (sent off-list or not yet received): 20090928153737.65c1ba3e-9yPaYZwiELC+kQycOl6kW4xkIHaj4LzF@public.gmane.org
Re: [RFC PATCH 0/9] sunrpc: teach the SUNRPC layer how to speak SMB
Quoted 29 lines · Jeff Layton
Quoted 9 lines · On Mon, 28 Sep 2009 13:40:23 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 7 lines · Jeff Layton
On Mon, 28 Sep 2009 09:41:08 -0500
"Steve French (smfltc)" <[email protected]> wrote:Quoted 8 lines
I don't think it changes the number of tasks sleeping. For sync operations it doesn't
change how many tasks that are sleeping. For async operations, you no longer have a
task sleeping incifs_writepagesorcifs_readpages, but do have the ability to dispatch a decode routine when the SMB Write or Read response is returned (may or
may not have a pool to do this). Seems like about the same number of task (possibly
less). Moving to all asynchronous operations under open, unlink, close doesn't reduce
the number of sleeping tasks (it may increase it or stay the same).I think it does. Take the example of asynchronous writepages requests. You want to send a bunch of write calls in quick succession and then process the replies as they come in. If you rely on a SendReceive-type interface, you have no choice but to spawn a separate
slow_worktask for each call on the wire. No guarantee they'll run in parallel though (depends on the rules for spawning new kslowd threads).
Quoted 5 lines
Now, you could hack up a routine that just sends the calls, sprinkle in a callback routine that's run by cifsd or something when the writes come back (though you'll need to be wary of deadlock, of course).
I was assuming the latter ... ie that cifsd processes the completion (or spawns a slow
work task to process the response). For the case of write (actually writepages)
there is not processing to be done (checking the rc, updating the status of
the corresponding pages in the page cache). For read there are various choices
(launching a slow work thread in readpages - but you only need one for what is potentially a very large number of reads from that readpages call) or processing in cifs or launching a slow work thread to process (my current preference since this is just cache updates/readahead anyway).
Quoted 8 lines
Quoted 7 lines
That (splitting decode into a distinct helper) makes sense (at least for async capable ops, in particular write, read and directory change notification and byte range locks). The "
smb_decode_write_response" case is a particularly simple and useful one to do and would
be an easy one to do as a test. I think the prototype patches for async write that someone did
for cifs a few years ago did that.
Quoted 5 lines
Even for ops that are fundamentally synchronous, it makes sense to split the encoding and decoding into separate routines. Maybe it's just my personal bias, but I think it encourages a cleaner, more flexible design.
I don't have a strong preference either way. There are a large number of SMBs which
are basically "responseless" (we just look at the rc, but they don't include parameters)
so they would not have interesting decode routines. Even for SMB2 although we
probably have to look at the credits, the generic SMB response processing can
probably deal with those. For those SMBs which have response processing,
if it makes for smaller more readable functions to break out the decode routines, that seems fine.
Quoted 10 lines
SMB transport header is 32 bits and contains a length with a zeroed out upper byte. The RPC over TCP transport header is 32 bits and contains a 31 bit length + a "last fragment" flag.
SMB has MID's, RPC has XID's
RPC has the RPC auth info, SMB has the session UID and TID
There are some similar concepts but at different levels of the network stack. Having net/sunrpc assemble SMB headers (uid, mids) would be needed if you really want to leverage the similar concepts, and it seems like a strange idea to move much of fs/cifs into net/sunrpc helpers
Quoted 52 lines
Quoted 41 lines
Quoted 25 lines
The code I've proposed more or less has abstractions along those lines. There's:
transport class -- (
net/sunrpc/xprtsmb.c)
header encoding/decoding -- (net/sunrpc/smb.c)...the other parts will be implemented by the filesystem (similar to how fs/nfs/nfs?xdr.c work).
Quoted 4 lines
I like the idea of the way SunRPC keeps task information, and it may make it easier to carry credentials around (although I think using Dave Howell's key management code
might be ok instead to access Winbind). I am not sure how easy it would be to tie
SunRPC credential mapping to Winbind but that could probably be done. I like theQuoted 10 lines
async scheduling capability of SunRPC although I suspect that it is a factor in a number of the (nfs client) performance problems we have seen so may need more work. I don't like adding (in effect) an extra transport and "encoding layer" though to
protocols (cifs and smb2). NFS since it is built on SunRPC on the wire, required
such a layer, and it makes sense for NFS to layer the code, like their protocol,
over SunRPC. CIFS and SMB2 don't require (or even allow) XDR translation,
variable encodings, and SunRPC encapsulation so the idea of abstracting the encoding of something that has a single defined encoding seems wrong.Quoted 6 lines
I'm not sure I understand this last comment. CIFS/SMB2 and NFS are just not that different in this regard. Either way you have to marshal up the buffer correctly before you send it, and decode it properly.
As an example of one of the more complicated cases for cifs (setting a time field). The code looks something like this and is very straightforward to see where the "linux field" is being put into the packet - and to catch errors in size or
endianness. Most cases are simpler for cifs.struct
file_basic_infobuf; /* the wire format of the file basic info SMB info level */
buf->LastAccessTime= 0; /* we can't set access time to Windows */buf->LastWriteTime= cpu_to_le64(cifs_UnixTimeToNT(attrs->ia_atime)); (followed by a few similar assignment statements for the remaining fields)Quoted 10 lines
then send the SMB header followed by the buffer. There is no marshalling
or xdr conversion needed. One thing I like about this approach is that
"sparse" (make modules C=1) immediately catches any mismatch between the wire format (size or endianness) and the vfs data structure element being put in the frame.Looking at
nfs4xdr.cas an example for comparison, it is much harder to see the actual line where a bigendian (or littlenedian) 64 bit quantity is being put into the request frame.
Quoted 8 lines
Woah...I think we just have a difference in terminology here. When I say "encoding" the packet, I just mean that we're stuffing the data into the buffer in the format that the wire expects.
Now, whether you do that with packed structures (like CIFS does and which I tend to prefer), or via "hand-marshaling" (like the WRITE/READ macros in the nfs?xdr.c files) is simply an implementation detail.
If we are stuffing the data into the buffer in the format the wire expects ie a series of assignment statements more or less as we do today (but broken into distinct encode and decode helpers)
pSMB->SomeField = cpu_to_le64(inode->some_attribute_field);
and simply using SunRPC as a convenient wrapper around the socket API for async dispatch that is probably fine (not sure if it is slower or faster ... and RPC has various performance limitations like the RPC_SLOT_COUNT) although not sure it is worth the trouble.
Re: [RFC PATCH 0/9] sunrpc: teach the SUNRPC layer how to speak SMB
Quoted 40 lines · On Mon, 28 Sep 2009 16:40:35 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 34 lines · Jeff Layton
Quoted 9 lines · On Mon, 28 Sep 2009 13:40:23 -0500 "Steve French (smfltc)" <[email protected]>…
Quoted 7 lines · Jeff Layton
On Mon, 28 Sep 2009 09:41:08 -0500
"Steve French (smfltc)" <[email protected]> wrote:Quoted 13 lines
I don't think it changes the number of tasks sleeping. For sync
operations it doesn't
change how many tasks that are sleeping. For async operations, you no
longer have a
task sleeping incifs_writepagesorcifs_readpages, but do have the ability to dispatch
a decode routine when the SMB Write or Read response is returned (may or
may not have a pool to do this). Seems like about the same number of
task (possibly
less). Moving to all asynchronous operations under open, unlink, close
doesn't reduce
the number of sleeping tasks (it may increase it or stay the same).I think it does. Take the example of asynchronous writepages requests. You want to send a bunch of write calls in quick succession and then process the replies as they come in. If you rely on a SendReceive-type interface, you have no choice but to spawn a separate
slow_worktask for each call on the wire. No guarantee they'll run in parallel though (depends on the rules for spawning new kslowd threads).Quoted 5 lines
Now, you could hack up a routine that just sends the calls, sprinkle in a callback routine that's run by cifsd or something when the writes come back (though you'll need to be wary of deadlock, of course).
Quoted 14 lines
I was assuming the latter ... ie that cifsd processes the completion (or spawns a slow
work task to process the response). For the case of write (actually
writepages)
there is not processing to be done (checking the rc, updating the status of
the corresponding pages in the page cache). For read there are various
choices
(launching a slow work thread in readpages - but you only need one for what is potentially a very large number of reads from that readpages call) or processing
in cifs or launching a slow work thread to process (my current preference since
this is just cache updates/readahead anyway).
Fair enough. In fact, I have a patch in my for-2.6.33 branch that generalizes the callback routine when cifsd matches a response to a MID. That may help get that moving.
Still, the fact of the matter is that cifs carries a thread (cifsd) that's wouldn't truly be needed were the responses to be processed in softirq context as sunrpc does.
Quoted 17 lines
Quoted 12 lines
Quoted 11 lines
That (splitting decode into a distinct helper) makes sense (at least for async capable ops,
in particular write, read and directory change notification and byte range locks). The
"smb_decode_write_response" case is a particularly simple and useful one to do and would
be an easy one to do as a test. I think the prototype patches for
async write that someone did
for cifs a few years ago did that.Quoted 5 lines
Even for ops that are fundamentally synchronous, it makes sense to split the encoding and decoding into separate routines. Maybe it's just my personal bias, but I think it encourages a cleaner, more flexible design.
Quoted 13 lines
I don't have a strong preference either way. There are a large number
of SMBs which
are basically "responseless" (we just look at the rc, but they don't include parameters)
so they would not have interesting decode routines. Even for SMB2
although we
probably have to look at the credits, the generic SMB response processing can
probably deal with those. For those SMBs which have response processing,
if it makes for smaller more readable functions to break out the decode routines, that seems
fine.
Sure, it's the same way with NFS too. For those we can probably just use a standard "decode" routine that just looks at the error code. Code consolidation is a good thing.
Quoted 10 lines
Quoted 10 lines
SMB transport header is 32 bits and contains a length with a zeroed out upper byte. The RPC over TCP transport header is 32 bits and contains a 31 bit length + a "last fragment" flag.
SMB has MID's, RPC has XID's
RPC has the RPC auth info, SMB has the session UID and TID
Quoted 8 lines
There are some similar concepts but at different levels of the network stack. Having
net/sunrpc assemble SMB headers (uid, mids) would be needed if you really want
to leverage the similar concepts, and it seems like a strange idea to move much of fs/cifs
into net/sunrpc helpers
Are they really at different levels of the network stack? SMB/SMB2 carries a little more info within the protocol header than we need at the transport layer, but it's not really that big a deal. The code I have just copies off the data until it gets to the MID.
Assembling UID's/TID's etc for a packet I see as more a function for the credential/auth management code in the RPC layer. Essentially I think I'll have to write a "RPC_AUTH_SMB" rpc_auth plugin for it and probably also a corresponding credential piece.
It'll be quite a bit different from how it's handled with "real" RPC protocols, but should still fit within the basic framework.
Quoted 80 lines
Quoted 67 lines
Quoted 60 lines
Quoted 34 lines
The code I've proposed more or less has abstractions along those lines. There's:
transport class -- (
net/sunrpc/xprtsmb.c)
header encoding/decoding -- (net/sunrpc/smb.c)...the other parts will be implemented by the filesystem (similar to how fs/nfs/nfs?xdr.c work).
Quoted 13 lines
I like the idea of the way SunRPC keeps task information, and it may make it easier
to carry credentials around (although I think using Dave Howell's key management code
might be ok instead to access Winbind). I am not sure how easy it
would be to tie
SunRPC credential mapping to Winbind but that could probably be done. I like the
async scheduling capability of SunRPC although I suspect that it is a factor in
a number of the (nfs client) performance problems we have seen so may need more work.
I don't like adding (in effect) an extra transport and "encoding layer"Quoted 10 lines
though to
protocols (cifs and smb2). NFS since it is built on SunRPC on the
wire, required
such a layer, and it makes sense for NFS to layer the code, like their protocol,
over SunRPC. CIFS and SMB2 don't require (or even allow) XDR translation,
variable encodings, and SunRPC encapsulation so the idea of abstracting the encoding of something that has a single defined encoding seems wrong.Quoted 6 lines
I'm not sure I understand this last comment. CIFS/SMB2 and NFS are just not that different in this regard. Either way you have to marshal up the buffer correctly before you send it, and decode it properly.
As an example of one of the more complicated cases for cifs (setting a time field).
The code looks something like this and is very straightforward to see where the "linux field" is being put into the packet - and to catch errors in size or
endianness. Most cases are simpler for cifs.struct
file_basic_infobuf; /* the wire format of the file basic info SMB info level */
buf->LastAccessTime= 0; /* we can't set access time to Windows */buf->LastWriteTime= cpu_to_le64(cifs_UnixTimeToNT(attrs->ia_atime)); (followed by a few similar assignment statements for the remaining fields)then send the SMB header followed by the buffer. There is no marshalling
or xdr conversion needed. One thing I like about this approach is that
"sparse" (make modules C=1) immediately catches any mismatch between the wire format (size or endianness) and the vfs data structure element being put in the frame.Quoted 6 lines
Looking at
nfs4xdr.cas an example for comparison, it is much harder to see the actual line
where a bigendian (or littlenedian) 64 bit quantity is being put into the request frame.Quoted 8 lines
Woah...I think we just have a difference in terminology here. When I say "encoding" the packet, I just mean that we're stuffing the data into the buffer in the format that the wire expects.
Now, whether you do that with packed structures (like CIFS does and which I tend to prefer), or via "hand-marshaling" (like the WRITE/READ macros in the nfs?xdr.c files) is simply an implementation detail.
If we are stuffing the data into the buffer in the format the wire expects ie a series of assignment
statements more or less as we do today (but broken into distinct encode and decode helpers)
Quoted 7 lines
pSMB->SomeField=cpu_to_le64(inode->some_attribute_field);and simply using SunRPC as a convenient wrapper around the socket API for async dispatch that is probably fine (not sure if it is slower or faster ... and
RPC has various performance limitations like theRPC_SLOT_COUNT) although not sure it is worth the trouble.
Why would it be any extra trouble when you're designing something new in the first place? As I said above, trying to bolt CIFS onto this is probably going to be more trouble than its worth. SMB2 is another matter though.
When it comes to encoding, all of the header and call encoding routines have a defined set of args and return values. How you actually do the encoding is up to the person writing the implementation. I too think that packed structures are easier to deal with and would probably prefer those over hand-marshaling.
CIFS has a "slot count" limitation too (we default to 50 outstanding calls). In fact, for the prototype here, I made the slot count default to 50 because of that precedent.
Signature · 1 line
Jeff Layton <[email protected]>