Date: Mon, 15 Dec 2014 18:42:11 -0500 (EST) From: Rick Macklem <rmacklem@uoguelph.ca> To: Mehmet Erol Sanliturk <m.e.sanliturk@gmail.com> Cc: freebsd-net@freebsd.org, Gerrit =?utf-8?B?S8O8aG4=?= <gerrit.kuehn@aei.mpg.de> Subject: Re: compiling on nfs directories Message-ID: <306385509.13293746.1418686931068.JavaMail.root@uoguelph.ca> In-Reply-To: <CAOgwaMvAN1zLhZYSuPxa5n8V9=QAj9Y7hmsmCYSYwqjy0fg-6w@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Mehmet Erol Sanilturk wrote: > On Mon, Dec 15, 2014 at 12:59 PM, Rick Macklem <rmacklem@uoguelph.ca> > wrote: > > > > Mehmet Erol Sanliturk wrote: > > > On Mon, Dec 15, 2014 at 1:24 AM, Gerrit K=C3=BChn > > > <gerrit.kuehn@aei.mpg.de> > > > wrote: > > > > > > > > Hi all, > > > > > > > > I ran into some weird issue here last week: > > > > I have an NFS-Server for storage and diskless booting (pxe / > > > > nfs > > > > root) > > > > running under FreeBSD. The clients are running Gentoo Linux. > > > > Some > > > > time > > > > ago, I replaced the server, going from a HDD-based storage > > > > array > > > > (ZFS) > > > > under FreeBSD 8.3 to an SSD-based array under FreeBSD 10-stable > > > > (as > > > > of > > > > February this year - I know this needs updates). > > > > > > > > Only now I recognized that this somehow appears to have broken > > > > some > > > > of my > > > > Gentoo ebuilds that do not install cleanly anymore. They > > > > complain > > > > about > > > > "soiled libtool library files found" and "insecure RUNPATHs" in > > > > the > > > > installation stage of shared libs. > > > > > > > > I was not able to find any useful solution for this in the Net > > > > so > > > > far. > > > > However, I was able to verify that this is somehow an issue > > > > with > > > > the nfs > > > > server by plugging in a USB-drive into the diskless clients and > > > > mounting > > > > this as /var/tmp/portage (the directory structure where > > > > Gentoo's > > > > ebuilds > > > > are compiled). This makes the error messages go away, and > > > > everything works > > > > again (like it did before the server update). > > > > > > > > Are there any suggestions what might be causing this and how to > > > > fix > > > > it? > > > > > > > > > > > > cu > > > > Gerrit > > > > > > > > > > > > > > > > > With respect to information given in your message , may pure > > > guess is > > > the > > > following : > > > > > > > > > When a client generates a file in NFS server , it assumes that > > > everything > > > is written into the file . > > > The next step ( reading the generated file ) starts , BUT the > > > file is > > > NOT > > > completely written into disk yet , > > > therefore it reads an incomplete file which causes errors in the > > > client . > > > > > Well, not exactly. The NFS client chooses whether or not the > > written > > data must be committed to stable storage (disk) right away via a > > flag > > argument on the write. It is up to the client to keep track of what > > has > > been written and if the FILE_STABLE flag wasn't set, must do a > > separate > > Commit RPC to force the data to stable storage on the server. > > It is also up to the NFS client to keep track of the file's size > > while > > it is being grown, since the NFS server's size may be smaller until > > the data gets written to the server. > > Also, note that he didn't see the problem with FreeBSD8.3, which > > would > > have been following the same rules on the server as 10.1. > > > > What I suspect might cause this is one of two things: > > 1 - The modify time of the file is now changing at a time the Linux > > client doesn't expect, due to changes in ZFS or maybe TOD clock > > resolution. (At one time, the TOD clock was only at a > > resolution > > of 1sec, so the client wouldn't see the modify time change > > often. > > I think it is now at a much higher resolution, but would have > > to > > look at the code/test to be sure.) > > 2 - I think you mention this one later in your message, in that the > > build might be depending on file locking. If this is the case, > > trying NFSv4, which does better file locking, might fix the > > problem. > > > > Gerrit, I would suggest that you do "nfsstat -m" on the Linux > > client, > > to see what the mount options are. The Linux client might be using > > NFSv4 > > already. > > Also, avoid "soft, intr" especially if you are using NFSv4, since > > these > > can cause slow server response to result in a failure of a > > read/write > > when it shouldn't fail, due to timeout or interruption by a signal. > > > > If you could find out more about what causes the specific build > > failure > > on the Linux side, that might help. > > If you can reproduce a build failure quickly/easily, you can > > capture > > packets via "tcpdump -s 0 -w <file> host <client-hostname>" on the > > server and then look at it in wireshark to see what the server is > > replying > > when the build failure occurs. (I don't mind looking at a packet > > trace if > > it is relatively small, if you email it to me as an attachment.) > > > > Good luck with it, rick > > ps: I am not familiar with the Linux mount options, but if it has > > stuff like "nocto", you could try those. > > > > > In FreeBSD NFS server , there is NOT ( or I could NOT be able to > > > find > > > ) a > > > facility to store written data immediately into disk . > > > > > > NFS server is collecting data up to a point ( number of bytes ) > > > and > > > then > > > writing it to disk , during this phase ( whether the NFS server > > > is > > > busy or > > > not ) is not important ) . With this structure , > > > the tasks which a program writes a small number of bytes to be > > > read > > > by > > > another program can not be > > > processed by a NFS server only . > > > > > > I did not try "locking in NFS server" : If this route is taken , > > > then > > > it is > > > necessary to adjust the clients for such periods to wait that NFS > > > server > > > has removed the lock which themselves can continue ( Each such > > > read > > > requires a waiting loop without generating an error message about > > > unavailable data and termination . ) . > > > > > > In Linux NFS server , there is an option to immediately write the > > > received > > > data into disk . This is improving the above situation > > > considerable > > > but not > > > completely solving the problem ( because during reads of data , > > > data > > > in > > > cache is NOT concatenated to the data in disk ) . > > > > > > > > > Another MAJOR problem is that , the NFS server is NOT > > > concatenating > > > data in > > > cache to data in disk during reads : This defect is making NFS > > > server > > > useless for , let's say "real time" , applications used > > > concurrently > > > or as > > > a single one by the clients without using another "Server" within > > > NFS > > > server . > > > > > > > > > > > > In your case , during software builds , a step is using the > > > previously > > > generated files : In local disk , writing and reading are > > > sequential > > > , in > > > the sense that written data is found during reading . In NFS > > > server > > > this is > > > not the case . > > > > > > > > > With respect to my knowledge obtained from messages in FreeBSD > > > mailing > > > lists about making a possibility to read data immediately after > > > it is > > > written into NFS server is NOT available . > > > > > > > > > > > > Thank you very much . > > > > > > Mehmet Erol Sanliturk > > > _______________________________________________ > > > >=20 >=20 >=20 >=20 >=20 > When a C program is written to be used in an NFS environment , some > possibilities may be used to synchronize write and reads from the > programs > with the unsolved "cached data" problem . >=20 > When ready programs are used , such as "make" , "ld" , there is no > choice . >=20 > I am using Pascal programs , then there is no such facilities . >=20 > The solution may be to improve the NFS Server and Client modules to > use > cached data during reads : >=20 > If end of file is reached : Before sending EOF signal , check whether > there > is data in cache or not . > If there is data in cache : continue reading from the cache > up to > end , > else send an EOF signal . >=20 > ( For random access files , also there is a need to look at the > cached > values . ) >=20 Well, the FreeBSD NFS client (and most others) do extensive data caching and will read data from the client cache whenever possible. NFS performance without client caching is pretty terrible. The problem (which has existed since NFS was first developed in about 1985) is that NFS does not provide a cache coherency protocol, so when multiple clients write data to a file concurrently, there is no guarantee that the c= lient read will get the most up-to-date data. There has been something called close-to-open (cto) consistency adopted, which says that a client will read= data written by another client after the writing client has closed the file. (Most NFS clients only implement this "approximately", since they depend on seeing th= e modify time change to determine this. This may not happen when multiple mod= ifications occur in the same time of day clock tick or when clients cache the file's attributes and use a stale cached modify time. Turning off client attribute caching improves this, but also results in a performance hit, due to the extra Getattr RPCs done.) The current consensus within the NFS community (driven by the Linux client implementation) is to only provide data consistency among multiple clients when byte range locking is used on the file. I'm not sure if this was what you were referring to. (It is true that NFS is not and cannot be a POSIX compliant file system, due to it's design.) "make" can often be confused when the modify time isn't updated when expect= ed. If an application running on FreeBSD wants to ensure that data is written t= o stable storage on the server, the application can call fsync(2). > Since the above modification requires knowledge of internal structure > of > NFS Server , and perhaps NFS Client ,I am not able to supply any > patch . > Also I am not able to understand its implementation difficulty . >=20 > My opinion is that the above modification would be a wonderful > improvement > for NFS system in FreeBSD , because it will behave just like a local > data > store usable as "real time" data processing tasks . In the present > structure , this is NOT possible with NFS Client and Server only . >=20 Many years ago, I implemented a cache coherency protocol for NFS called NQNFS. No one used it (at least not much) and it never caught on. Most care about NFS performance and data coherency has never been a priority with most users, from what I've seen. rick >=20 > Thank you very much . >=20 >=20 > Mehmet Erol Sanliturk > _______________________________________________ > freebsd-net@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-net > To unsubscribe, send any mail to > "freebsd-net-unsubscribe@freebsd.org"
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?306385509.13293746.1418686931068.JavaMail.root>