Date: Mon, 19 Dec 2016 23:32:32 +0000 From: Colin Percival <cperciva@tarsnap.com> To: Rick Macklem <rmacklem@uoguelph.ca> Cc: "freebsd-fs@freebsd.org" <freebsd-fs@freebsd.org> Subject: Re: ESTALE after cwd deleted by same NFS client Message-ID: <01000159196dee7c-b727d595-c8e6-4761-b2a7-197d27f1f21a-000000@email.amazonses.com> In-Reply-To: <YTXPR01MB0189B6B6869327FD70210018DD910@YTXPR01MB0189.CANPRD01.PROD.OUTLOOK.COM> References: <01000158f023675b-41b35a73-4428-4937-853b-62db4fb9b984-000000@email.amazonses.com> <20161212054233.GU8460@kduck.kaduk.org> <01000158f1abc081-d4eddc58-3b4b-41dd-aa1e-0157d2fad812-000000@email.amazonses.com> <20161212163603.GV8460@kduck.kaduk.org> <YQBPR01MB018054EE62DEFDC73784AD9BDD9B0@YQBPR01MB0180.CANPRD01.PROD.OUTLOOK.COM> <01000158fc3da2c5-c13da088-e7b9-4ac0-ac01-ec49a275dd24-000000@email.amazonses.com> <YTXPR01MB0189ACD940B7D399A6855CB8DD9A0@YTXPR01MB0189.CANPRD01.PROD.OUTLOOK.COM> <YTXPR01MB01891CB02D5D7BA46579F1BFDD9D0@YTXPR01MB0189.CANPRD01.PROD.OUTLOOK.COM> <010001590945e9b3-015a4d05-2646-44ba-9db9-415e8b9119dd-000000@email.amazonses.com> <0100015915a5ee96-49de6100-5050-4a0a-a3c9-1bd4215bc6a0-000000@email.amazonses.com> <YTXPR01MB0189B6B6869327FD70210018DD910@YTXPR01MB0189.CANPRD01.PROD.OUTLOOK.COM>
next in thread | previous in thread | raw e-mail | index | archive | help
On 12/19/16 13:59, Rick Macklem wrote: > Colin Percival wrote: >> Further information: In addition to the "lookup relative to a directory which >> has been deleted out from underneath us" case which causes ESTALE to land in >> nfs_lookup, the cleandir step of buildworld results in ESTALE being returned >> by nfsrpc_getattr into nfs_getattr (landing ultimately in getcwd), and ESTALE >> being returned by nfsrpc_accessrpc into nfs34_access_otw (landing ultimately >> in stat and lstat). >> >> In UFS there are checks for effnlink == 0 which result in e.g. ufs_lookup >> returning ENOENT; would it make sense to add NREMOVED to struct nfsnode.n_flag >> and check this in the appropriate nfs_* calls? > To be honest, I can't think of a reason why userland would ever want to see ESTALE? > The function you see above "nfscl_maperr()" could easily map all ESTALEs to ENOENTs? I was wondering about that. I hesitated to suggest it since it seemed like doing this could mask bugs and/or throw away useful information -- I mean, I assume there was a reason ESTALE existed in the first place... > - The question is: "Would returning ENOENT for stat(2) and access(2) actually make the > buildworld happy? I think buildworld would need s/ESTALE/ENOENT/ in access, lstat, rmdir, stat, and unlink. But with those I'm 99% confident that buildworld will complete. > --> The cheat for regular files is "sillyrename". This could be done for directories, > but there are multiple comments in the code (not put there by me) that say > "no sillyrename for directories". > #1 Does this imply something breaks when you do sillyrename for dirs? Yes. You'd run into this scenario: $ mkdir /nfs/foo $ cd /nfs/foo $ rmdir /nfs/foo # still in use, sillyrename happens $ touch bar # creates /nfs/foo.sillyname/bar $ cd / # directory no longer in use, time to delete it... Whereas keeping track of "this nfsnode refers to a directory which has been deleted" would allow us to return ENOENT to file-creation attempt, just like UFS does if you try to create a file inside a removed directory. -- Colin Percival Security Officer Emeritus, FreeBSD | The power to serve Founder, Tarsnap | www.tarsnap.com | Online backups for the truly paranoid
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?01000159196dee7c-b727d595-c8e6-4761-b2a7-197d27f1f21a-000000>