On Wed, 2024-03-27 at 14:40 +0800, Chen Hanxiao wrote:
> Hi, Jeff
>
> I wrote a POC patch, use name_to_handle_at to get nfs root filehandle
> instead of /proc/fs/nfsd/filehandle.
>
> I did some simple tests and it works.
>
> Pls check the POC patch below:
> ---
> support/export/cache.c | 95 +++++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 93 insertions(+), 2 deletions(-)
>
Very cool!
I do think this is the basic direction we want to go with revising how
mountd gets its filehandles.
> diff --git a/support/export/cache.c b/support/export/cache.c
> index 6c0a44a3..adecac2e 100644
> --- a/support/export/cache.c
> +++ b/support/export/cache.c
> @@ -43,6 +43,25 @@
> #include "blkid/blkid.h"
> #endif
>
> +#define NFS4_FHSIZE 128
> +
> +struct knfsd_fh {
> + unsigned int fh_size; /*
> + * Points to the current size while
> + * building a new file handle.
> + */
> + union {
> + char fh_raw[NFS4_FHSIZE];
> + struct {
> + uint8_t fh_version; /* == 1 */
> + uint8_t fh_auth_type; /* deprecated */
> + uint8_t fh_fsid_type;
> + uint8_t fh_fileid_type;
> + uint32_t fh_fsid[]; /* flexible-array member */
> + };
> + };
> +};
> +
This is a large departure from how this has worked in the past. Before
this, userland could always treat a knfsd_fh as opaque. With this change
though, knfsd_fh will need to be part of the kernel's ABI, and we'll
need to be very careful about future changes to the FH format (not that
we're planning any).
I think we probably ought to start by moving the definitions of knfsd_fh
and its component parts into a UAPI header that nfs-utils could then
use. For now, we can add duplicate definitions like above for when the
case where they aren't defined in UAPI headers.
> enum nfsd_fsid {
> FSID_DEV = 0,
> FSID_NUM,
Side note: I wonder if some of the nfsd_fsid values can (effectively) be
deprecated these days? I don't think we really use FSID_DEV anymore at
all, do we?
> @@ -1827,8 +1846,8 @@ int cache_export(nfs_export *exp, char *path)
> * read filehandle <&0
> * } <> /proc/fs/nfsd/filehandle
> */
> -struct nfs_fh_len *
> -cache_get_filehandle(nfs_export *exp, int len, char *p)
> +static struct nfs_fh_len *
> +cache_get_filehandle_by_proc(nfs_export *exp, int len, char *p)
> {
> static struct nfs_fh_len fh;
> char buf[RPC_CHAN_BUF_SIZE], *bp;
> @@ -1862,6 +1881,78 @@ cache_get_filehandle(nfs_export *exp, int len, char *p)
> return &fh;
> }
>
> +static struct nfs_fh_len *
> +cache_get_filehandle_by_name(nfs_export *exp, char *name)
> +{
> + static struct nfs_fh_len fh;
> + struct {
> + struct file_handle fh;
> + unsigned char handle[128];
> + } file_fh;
> + char buf[RPC_CHAN_BUF_SIZE] = {0};
> + char *mesg = buf;
> + int len, mnt_id;
> + unsigned int e_fsid;
> + struct knfsd_fh kfh;
> + char u[16];
> +
> + memset(fh.fh_handle, 0, sizeof(fh.fh_handle));
> +
> + file_fh.fh.handle_bytes = 128;
> + if (name_to_handle_at(AT_FDCWD, name, &file_fh.fh, &mnt_id, 0) < 0)
> + return NULL;
> +
> + memset(fh.fh_handle, 0, sizeof(fh.fh_handle));
> + memset(&kfh, 0, sizeof(struct knfsd_fh));
> + kfh.fh_version = 1;
> + kfh.fh_auth_type = 0;
> + e_fsid = exp->m_export.e_fsid;
> +
> + if (e_fsid > 0) {
> + len = 12;
> + fh.fh_size = 8;
> + kfh.fh_size = 12;
> + kfh.fh_fsid_type = 1;
> + kfh.fh_fsid[0] = e_fsid;
> + } else {
> + len = file_fh.fh.handle_bytes + 8;
> + fh.fh_size = file_fh.fh.handle_bytes;
> + kfh.fh_size = file_fh.fh.handle_bytes + sizeof(kfh.fh_size);
> + kfh.fh_fsid_type = FSID_UUID16_INUM;
Note that we will need to deal with NFSv2 here too. It has 32-byte
filehandles, and so we likely can't use FSID_UUID16_INUM for those.
> + if (file_fh.fh.handle_bytes <= 12) {
> + kfh.fh_fsid[0] = *(uint32_t *)file_fh.fh.f_handle;
> + kfh.fh_fsid[1] = 0;
> + } else {
> + kfh.fh_fsid[0] = *(uint32_t *)file_fh.fh.f_handle;
> + kfh.fh_fsid[1] = *((uint32_t *)file_fh.fh.f_handle + 1);
> + }
> + }
> + kfh.fh_fileid_type = 0; // FILEID_ROOT
> +
> + qword_addhex(&mesg, &len, kfh.fh_raw, kfh.fh_size);
> + mesg = buf;
> + len = qword_get(&mesg, (char *)fh.fh_handle, NFS3_FHSIZE);
We may need to pass to this function what version of NFS MNT request was
sent so that we know what sort of FH is needed here.
> + if (e_fsid == 0) {
> + len = 16;
> + uuid_by_path(name, 0, 16, u);
> + memcpy((char *)fh.fh_handle+12, u, 16);
> + fh.fh_size += 16;
> + }
> +
> + return &fh;
> +}
> +
> +struct nfs_fh_len *
> +cache_get_filehandle(nfs_export *exp, int len, char *p)
> +{
> + struct nfs_fh_len *fh;
> + fh = cache_get_filehandle_by_name(exp, p);
> + if (!fh)
> + fh = cache_get_filehandle_by_proc(exp, len, p);
> +
> + return fh;
> +}
> +
> /* Wait for all worker child processes to exit and reap them */
> void
> cache_wait_for_workers(char *prog)
This looks like a great start -- nice work! There are a number of more
obscure corner-cases that we'll need to deal with, but this does look
like a promising direction.
--
Jeff Layton <jlayton@xxxxxxxxxx>
