Re: [PATCH] [RFC] nfs4: inject process namespace into COMPOUND tag

["Mkrtchyan, Tigran" <tigran.mkrtchyan@desy.de>]

[-- Attachment #1: Type: text/plain, Size: 7228 bytes --]



----- Original Message -----
> From: "Trond Myklebust" <trondmy@kernel.org>
> To: "Tigran Mkrtchyan" <tigran.mkrtchyan@desy.de>, "chuck lever" <chuck.lever@oracle.com>, "jlayton"
> <jlayton@kernel.org>, "anna" <anna@kernel.org>
> Cc: "linux-nfs" <linux-nfs@vger.kernel.org>
> Sent: Sunday, 28 June, 2026 23:15:42
> Subject: Re: [PATCH] [RFC] nfs4: inject process namespace into COMPOUND tag

> On Fri, 2026-06-26 at 17:10 +0200, Tigran Mkrtchyan wrote:
>> On large shared machines often multiple jobs of a same user run in
>> parallel. For debugging, it's usually impossible to identify requests
>> coming from different processes.
>> 
>> The batch systems like HTCondor or SLURM start every job in it's own
>> namespace, thus passing namespace info to the server will help by
>> debugging.
>> 
>> 192.168.122.150 → 192.168.178.69 NFS 260 V4 Call GETATTR FH:
>> 0xd5ffb2cb
>> 192.168.178.69 → 192.168.122.150 NFS 324 V4 Reply (Call In 89)
>> GETATTR
>> 192.168.122.150 → 192.168.178.69 NFS 260 V4 Call GETATTR FH:
>> 0xd0b0a44e
>> 192.168.178.69 → 192.168.122.150 NFS 324 V4 Reply (Call In 95)
>> GETATTR
>> 192.168.122.150 → 192.168.178.69 NFS 268 V4 Call ACCESS FH:
>> 0xd0b0a44e, [Check: RD LU MD XT DL XAR XAW XAL]
>> 192.168.178.69 → 192.168.122.150 NFS 240 V4 Reply (Call In 101)
>> ACCESS, [Allowed: RD LU MD XT DL XAR XAW XAL]
>> 192.168.122.150 → 192.168.178.69 NFS 284 V4 Call READDIR FH:
>> 0xd0b0a44e
>> 192.168.178.69 → 192.168.122.150 NFS 664 V4 Reply (Call In 105)
>> READDIR
>> 192.168.122.150 → 192.168.178.69 NFS 284 V4 Call ns:4026532507
>> GETATTR FH: 0xd67b66a5
>> 192.168.178.69 → 192.168.122.150 NFS 340 V4 Reply (Call In 111)
>> ns:4026532507 GETATTR
>> 192.168.122.150 → 192.168.178.69 NFS 292 V4 Call ns:4026532507 ACCESS
>> FH: 0xd67b66a5, [Check: RD LU MD XT DL XAR XAW XAL]
>> 192.168.178.69 → 192.168.122.150 NFS 256 V4 Reply (Call In 117)
>> ns:4026532507 ACCESS, [Access Denied: MD XT DL XAW], [Allowed: RD LU
>> XAR XAL]
>> 192.168.122.150 → 192.168.178.69 NFS 308 V4 Call ns:4026532507
>> READDIR FH: 0xd67b66a5
>> 192.168.178.69 → 192.168.122.150 NFS 200 V4 Reply (Call In 121)
>> ns:4026532507 READDIR
>> 
>> Suggested-by: Chuck Lever <chuck.lever@oracle.com>
>> Signed-off-by: Tigran Mkrtchyan <tigran.mkrtchyan@desy.de>
>> ---
>>  fs/nfs/nfs4xdr.c             | 24 +++++++++++++++++++-----
>>  include/linux/sunrpc/sched.h |  2 ++
>>  net/sunrpc/sched.c           |  6 ++++++
>>  3 files changed, 27 insertions(+), 5 deletions(-)
>> 
>> diff --git a/fs/nfs/nfs4xdr.c b/fs/nfs/nfs4xdr.c
>> index c23c2eee1b5c..9c035c74a3b5 100644
>> --- a/fs/nfs/nfs4xdr.c
>> +++ b/fs/nfs/nfs4xdr.c
>> @@ -46,6 +46,7 @@
>>  #include <linux/kdev_t.h>
>>  #include <linux/module.h>
>>  #include <linux/utsname.h>
>> +#include <linux/pid_namespace.h>
>>  #include <linux/sunrpc/clnt.h>
>>  #include <linux/sunrpc/msg_prot.h>
>>  #include <linux/sunrpc/gss_api.h>
>> @@ -71,12 +72,8 @@ static void encode_layoutget(struct xdr_stream
>> *xdr,
>>  static int decode_layoutget(struct xdr_stream *xdr, struct rpc_rqst
>> *req,
>>  			     struct nfs4_layoutget_res *res);
>>  
>> -/* NFSv4 COMPOUND tags are only wanted for debugging purposes */
>> -#ifdef DEBUG
>> +/* Enable compound tags to include namespace information */
>>  #define NFS4_MAXTAGLEN		20
>> -#else
>> -#define NFS4_MAXTAGLEN		0
>> -#endif
>>  
>>  /* lock,open owner id:
>>   * we currently use size 2 (u64) out of (NFS4_OPAQUE_LIMIT  >> 2)
>> @@ -1034,6 +1031,23 @@ static void encode_compound_hdr(struct
>> xdr_stream *xdr,
>>  {
>>  	__be32 *p;
>>  
>> +	/* Inject namespace info into compound tag if not already
>> set */
>> +	if (hdr->taglen == 0 && req->rq_task != NULL) {
>> +		/* Use namespace info captured at task creation time
>> */
>> +		struct rpc_task *task = req->rq_task;
>> +
>> +		if (taks->tk_ns_inum != 0) {
> 
> Hmm.... This has not been compile tested.

I am pretty sure this is the code that I run on VM right now :)

> 
>> +			char ns_tag[NFS4_MAXTAGLEN + 1];
>> +
>> +			hdr->taglen = snprintf(ns_tag,
>> sizeof(ns_tag), "ns:%u", taks->tk_ns_inum);
>> +			if (hdr->taglen > NFS4_MAXTAGLEN) {
>> +				hdr->taglen = NFS4_MAXTAGLEN;
>> +				ns_tag[NFS4_MAXTAGLEN] = '\0';
>> +			}
>> +			hdr->tag = ns_tag;
> 
> ns_tag is only scoped to this block. I suggest that instead of
> assigning to hdr->taglen and hdr->tag, you just do the assignment to
> hdr->replen + call to encode_string() here, so you don't have to assign
> a scope limited buffer to an externally visible struct.
> 
> Note also that there is no need to NUL terminate ns_tag[] when hdr-
>>taglen > NFS4_MAXTAGLEN, since encode_string() does not require a nul
> terminated string. In addition, snprintf() always guarantees that the
> string is nul terminated, even when truncated by the buffer size :-).
> 
>> +		}
>> +	}
>> +
>>  	/* initialize running count of expected bytes in reply.
>>  	 * NOTE: the replied tag SHOULD be the same is the one sent,
>>  	 * but this is not required as a MUST for the server to do
>> so. */
>> diff --git a/include/linux/sunrpc/sched.h
>> b/include/linux/sunrpc/sched.h
>> index 0dbdf3722537..d376b52a72a1 100644
>> --- a/include/linux/sunrpc/sched.h
>> +++ b/include/linux/sunrpc/sched.h
>> @@ -92,6 +92,8 @@ struct rpc_task {
>>  
>>  	pid_t			tk_owner;	/* Process id for
>> batching tasks */
>>  
>> +	unsigned int		tk_ns_inum;	/* PID namespace
>> inum for namespace tracking */
>> +
>>  	int			tk_rpc_status;	/* Result of last
>> RPC operation */
>>  	unsigned short		tk_flags;	/* misc flags */
>>  	unsigned short		tk_timeouts;	/* maj timeouts */
>> diff --git a/net/sunrpc/sched.c b/net/sunrpc/sched.c
>> index 016f16ca5779..4e8e7fa849d5 100644
>> --- a/net/sunrpc/sched.c
>> +++ b/net/sunrpc/sched.c
>> @@ -21,6 +21,7 @@
>>  #include <linux/mutex.h>
>>  #include <linux/freezer.h>
>>  #include <linux/sched/mm.h>
>> +#include <linux/pid_namespace.h>
>>  
>>  #include <linux/sunrpc/clnt.h>
>>  #include <linux/sunrpc/metrics.h>
>> @@ -1110,6 +1111,11 @@ static void rpc_init_task(struct rpc_task
>> *task, const struct rpc_task_setup *ta
>>  	task->tk_priority = task_setup_data->priority -
>> RPC_PRIORITY_LOW;
>>  	task->tk_owner = current->tgid;
>>  
>> +	struct pid_namespace *pid_ns = task_active_pid_ns(current);
>> +	/* Keep track on namespace id */
>> +	if (pid_ns != &init_pid_ns)
>> +		task->tk_ns_inum = pid_ns->ns.inum;
> 
> For buffered writes, this will tell you the pid namespace of the
> process that is flushing the data, not that of the process that wrote
> the data into the page cache. Is that what you expected?


In general, with read-ahead, delegations, and other caching mechanisms,
is it possible to match NFS I/O to an application?

Best regards,
   Tigran.


> 
>> +
>>  	/* Initialize workqueue for async tasks */
>>  	task->tk_workqueue = task_setup_data->workqueue;
>>  
> 
> --
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> trondmy@kernel.org, trond.myklebust@hammerspace.com

[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 2309 bytes --]