| Message ID | CAH2r5msdUQ=CVM6s7ENeH7SP-teYAOioOGq7zY5sDXZFrFYiCA@mail.gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | convert revalidate of directories to using directory metadata cache timeout | expand |
Updated version incorporates Ronnie's suggestion of leaving the default (for directory caching) the same as it is today, 1 second, to avoid unnecessary risk. Most users can safely improve performance by mounting with acdirmax to a higher value (e.g. 60 seconds as NFS defaults to). nfs and cifs on Linux currently have a mount parameter "actimeo" to control metadata (attribute) caching but cifs does not have additional mount parameters to allow distinguishing between caching directory metadata (e.g. needed to revalidate paths) and that for files. Add new mount parameter "acdirmax" to allow caching metadata for directories more loosely than file data. NFS adjusts metadata caching from acdirmin to acdirmax (and another two mount parms for files) but to reduce complexity, it is safer to just introduce the one mount parm to allow caching directories longer. The defaults for acdirmax and actimeo (for cifs.ko) are conservative, 1 second (NFS defaults acdirmax to 60 seconds). For many workloads, setting acdirmax to a higher value is safe and will improve performance. This patch leaves unchanged the default values for caching metadata for files and directories but gives the user more flexibility in adjusting them safely for their workload via the new mount parm. Signed-off-by: Steve French <stfrench@microsoft.com> Reviewed-by: Ronnie Sahlberg <lsahlber@redhat.com> --- fs/cifs/cifsfs.c | 3 ++- fs/cifs/cifsglob.h | 8 +++++++- fs/cifs/connect.c | 2 ++ fs/cifs/fs_context.c | 9 +++++++++ fs/cifs/fs_context.h | 4 +++- 5 files changed, 23 insertions(+), 3 deletions(-) diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c index 6f33ff3f625f..4e0b0b26e844 100644 --- a/fs/cifs/cifsfs.c +++ b/fs/cifs/cifsfs.c @@ -637,8 +637,9 @@ cifs_show_options(struct seq_file *s, struct dentry *root) seq_printf(s, ",snapshot=%llu", tcon->snapshot_time); if (tcon->handle_timeout) seq_printf(s, ",handletimeout=%u", tcon->handle_timeout); - /* convert actimeo and display it in seconds */ + /* convert actimeo and directory attribute timeout and display in seconds */ seq_printf(s, ",actimeo=%lu", cifs_sb->ctx->actimeo / HZ); + seq_printf(s, ",acdirmax=%lu", cifs_sb->ctx->acdirmax / HZ); if (tcon->ses->chan_max > 1) seq_printf(s, ",multichannel,max_channels=%zu", diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c index cd6dbeaf2166..a9dc39aee9f4 100644 --- a/fs/cifs/connect.c +++ b/fs/cifs/connect.c @@ -2278,6 +2278,8 @@ compare_mount_options(struct super_block *sb, struct cifs_mnt_data *mnt_data) if (old->ctx->actimeo != new->ctx->actimeo) return 0; + if (old->ctx->acdirmax != new->ctx->acdirmax) + return 0; return 1; } diff --git a/fs/cifs/fs_context.c b/fs/cifs/fs_context.c index 7d04f2255624..555969c8d586 100644 --- a/fs/cifs/fs_context.c +++ b/fs/cifs/fs_context.c @@ -140,6 +140,7 @@ const struct fs_parameter_spec smb3_fs_parameters[] = { fsparam_u32("rsize", Opt_rsize), fsparam_u32("wsize", Opt_wsize), fsparam_u32("actimeo", Opt_actimeo), + fsparam_u32("acdirmax", Opt_acdirmax), fsparam_u32("echo_interval", Opt_echo_interval), fsparam_u32("max_credits", Opt_max_credits), fsparam_u32("handletimeout", Opt_handletimeout), @@ -936,6 +937,13 @@ static int smb3_fs_context_parse_param(struct fs_context *fc, goto cifs_parse_mount_err; } break; + case Opt_acdirmax: + ctx->acdirmax = HZ * result.uint_32; + if (ctx->acdirmax > CIFS_MAX_ACTIMEO) { + cifs_dbg(VFS, "acdirmax too large\n"); + goto cifs_parse_mount_err; + } + break; case Opt_echo_interval: ctx->echo_interval = result.uint_32; break; @@ -1362,6 +1370,7 @@ int smb3_init_fs_context(struct fs_context *fc) ctx->strict_io = true; ctx->actimeo = CIFS_DEF_ACTIMEO; + ctx->acdirmax = CIFS_DEF_ACTIMEO; /* Most clients set timeout to 0, allows server to use its default */ ctx->handle_timeout = 0; /* See MS-SMB2 spec section 2.2.14.2.12 */ diff --git a/fs/cifs/fs_context.h b/fs/cifs/fs_context.h index 1c44a460e2c0..472372fec4e9 100644 --- a/fs/cifs/fs_context.h +++ b/fs/cifs/fs_context.h @@ -118,6 +118,7 @@ enum cifs_param { Opt_rsize, Opt_wsize, Opt_actimeo, + Opt_acdirmax, Opt_echo_interval, Opt_max_credits, Opt_snapshot, @@ -232,7 +233,8 @@ struct smb3_fs_context { unsigned int wsize; unsigned int min_offload; bool sockopt_tcp_nodelay:1; - unsigned long actimeo; /* attribute cache timeout (jiffies) */ + unsigned long actimeo; /* attribute cache timeout for files (jiffies) */ + unsigned long acdirmax; /* attribute cache timeout for directories (jiffies) */ struct smb_version_operations *ops; struct smb_version_values *vals; char *prepath; On Tue, Feb 23, 2021 at 4:22 PM Steve French <smfrench@gmail.com> wrote: > > nfs and cifs on Linux currently have a mount parameter "actimeo" to > control metadata (attribute) caching but cifs does not have additional > mount parameters to allow distinguishing between caching directory > metadata (e.g. needed to revalidate paths) and that for files. > > Add new mount parameter "acdirmax" to allow caching metadata for > directories more loosely than file data. NFS adjusts metadata > caching from acdirmin to acdirmax (and another two mount parms > for files) but to reduce complexity, it is safer to just introduce > the one mount parm to allow caching directories longer (30 seconds > vs. the 1 second default for file metadata) which is still more > conservative than other Linux filesystems (e.g. NFS sets acdirmax > to 60 seconds) > > -- > Thanks, > > Steve
On 2/23/2021 8:03 PM, Steve French wrote: > Updated version incorporates Ronnie's suggestion of leaving the > default (for directory caching) the same as it is today, 1 second, to > avoid > unnecessary risk. Most users can safely improve performance by > mounting with acdirmax to a higher value (e.g. 60 seconds as NFS > defaults to). > > nfs and cifs on Linux currently have a mount parameter "actimeo" to control > metadata (attribute) caching but cifs does not have additional mount > parameters to allow distinguishing between caching directory metadata > (e.g. needed to revalidate paths) and that for files. The behaviors seem to be slightly different with this change. With NFS, the actimeo option overrides the four min/max options, and by default the directory ac timers range between 30 and 60. The CIFS code I see below seems to completely separate actimeo and acdirmax, and if not set, uses the historic 1 second value. That's fine, but it's completely different from NFS. Shouldn't we use a different mount option, to avoid confusing the admin? > + /* > + * depending on inode type, check if attribute caching disabled for > + * files or directories > + */ > + if (S_ISDIR(inode->i_mode)) { > + if (!cifs_sb->ctx->acdirmax) > + return true; > + if (!time_in_range(jiffies, cifs_i->time, > + cifs_i->time + cifs_sb->ctx->acdirmax)) > + return true; > + } else { /* file */ > + if (!cifs_sb->ctx->actimeo) > + return true; > + if (!time_in_range(jiffies, cifs_i->time, > + cifs_i->time + cifs_sb->ctx->actimeo)) > + return true; > + } > Add new mount parameter "acdirmax" to allow caching metadata for > directories more loosely than file data. NFS adjusts metadata > caching from acdirmin to acdirmax (and another two mount parms > for files) but to reduce complexity, it is safer to just introduce > the one mount parm to allow caching directories longer. The > defaults for acdirmax and actimeo (for cifs.ko) are conservative, > 1 second (NFS defaults acdirmax to 60 seconds). For many workloads, > setting acdirmax to a higher value is safe and will improve > performance. This patch leaves unchanged the default values > for caching metadata for files and directories but gives the > user more flexibility in adjusting them safely for their workload > via the new mount parm. > > Signed-off-by: Steve French <stfrench@microsoft.com> > Reviewed-by: Ronnie Sahlberg <lsahlber@redhat.com> > --- > fs/cifs/cifsfs.c | 3 ++- > fs/cifs/cifsglob.h | 8 +++++++- > fs/cifs/connect.c | 2 ++ > fs/cifs/fs_context.c | 9 +++++++++ > fs/cifs/fs_context.h | 4 +++- > 5 files changed, 23 insertions(+), 3 deletions(-) > > diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c > index 6f33ff3f625f..4e0b0b26e844 100644 > --- a/fs/cifs/cifsfs.c > +++ b/fs/cifs/cifsfs.c > @@ -637,8 +637,9 @@ cifs_show_options(struct seq_file *s, struct dentry *root) > seq_printf(s, ",snapshot=%llu", tcon->snapshot_time); > if (tcon->handle_timeout) > seq_printf(s, ",handletimeout=%u", tcon->handle_timeout); > - /* convert actimeo and display it in seconds */ > + /* convert actimeo and directory attribute timeout and display in seconds */ > seq_printf(s, ",actimeo=%lu", cifs_sb->ctx->actimeo / HZ); > + seq_printf(s, ",acdirmax=%lu", cifs_sb->ctx->acdirmax / HZ); > > if (tcon->ses->chan_max > 1) > seq_printf(s, ",multichannel,max_channels=%zu", > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c > index cd6dbeaf2166..a9dc39aee9f4 100644 > --- a/fs/cifs/connect.c > +++ b/fs/cifs/connect.c > @@ -2278,6 +2278,8 @@ compare_mount_options(struct super_block *sb, > struct cifs_mnt_data *mnt_data) > > if (old->ctx->actimeo != new->ctx->actimeo) > return 0; > + if (old->ctx->acdirmax != new->ctx->acdirmax) > + return 0; > > return 1; > } > diff --git a/fs/cifs/fs_context.c b/fs/cifs/fs_context.c > index 7d04f2255624..555969c8d586 100644 > --- a/fs/cifs/fs_context.c > +++ b/fs/cifs/fs_context.c > @@ -140,6 +140,7 @@ const struct fs_parameter_spec smb3_fs_parameters[] = { > fsparam_u32("rsize", Opt_rsize), > fsparam_u32("wsize", Opt_wsize), > fsparam_u32("actimeo", Opt_actimeo), > + fsparam_u32("acdirmax", Opt_acdirmax), > fsparam_u32("echo_interval", Opt_echo_interval), > fsparam_u32("max_credits", Opt_max_credits), > fsparam_u32("handletimeout", Opt_handletimeout), > @@ -936,6 +937,13 @@ static int smb3_fs_context_parse_param(struct > fs_context *fc, > goto cifs_parse_mount_err; > } > break; > + case Opt_acdirmax: > + ctx->acdirmax = HZ * result.uint_32; > + if (ctx->acdirmax > CIFS_MAX_ACTIMEO) { > + cifs_dbg(VFS, "acdirmax too large\n"); > + goto cifs_parse_mount_err; > + } > + break; > case Opt_echo_interval: > ctx->echo_interval = result.uint_32; > break; > @@ -1362,6 +1370,7 @@ int smb3_init_fs_context(struct fs_context *fc) > ctx->strict_io = true; > > ctx->actimeo = CIFS_DEF_ACTIMEO; > + ctx->acdirmax = CIFS_DEF_ACTIMEO; > > /* Most clients set timeout to 0, allows server to use its default */ > ctx->handle_timeout = 0; /* See MS-SMB2 spec section 2.2.14.2.12 */ > diff --git a/fs/cifs/fs_context.h b/fs/cifs/fs_context.h > index 1c44a460e2c0..472372fec4e9 100644 > --- a/fs/cifs/fs_context.h > +++ b/fs/cifs/fs_context.h > @@ -118,6 +118,7 @@ enum cifs_param { > Opt_rsize, > Opt_wsize, > Opt_actimeo, > + Opt_acdirmax, > Opt_echo_interval, > Opt_max_credits, > Opt_snapshot, > @@ -232,7 +233,8 @@ struct smb3_fs_context { > unsigned int wsize; > unsigned int min_offload; > bool sockopt_tcp_nodelay:1; > - unsigned long actimeo; /* attribute cache timeout (jiffies) */ > + unsigned long actimeo; /* attribute cache timeout for files (jiffies) */ > + unsigned long acdirmax; /* attribute cache timeout for directories > (jiffies) */ > struct smb_version_operations *ops; > struct smb_version_values *vals; > char *prepath; > > On Tue, Feb 23, 2021 at 4:22 PM Steve French <smfrench@gmail.com> wrote: >> >> nfs and cifs on Linux currently have a mount parameter "actimeo" to >> control metadata (attribute) caching but cifs does not have additional >> mount parameters to allow distinguishing between caching directory >> metadata (e.g. needed to revalidate paths) and that for files. >> >> Add new mount parameter "acdirmax" to allow caching metadata for >> directories more loosely than file data. NFS adjusts metadata >> caching from acdirmin to acdirmax (and another two mount parms >> for files) but to reduce complexity, it is safer to just introduce >> the one mount parm to allow caching directories longer (30 seconds >> vs. the 1 second default for file metadata) which is still more >> conservative than other Linux filesystems (e.g. NFS sets acdirmax >> to 60 seconds) >> >> -- >> Thanks, >> >> Steve > > >
On Wed, Feb 24, 2021 at 10:11 AM Tom Talpey <tom@talpey.com> wrote: > > On 2/23/2021 8:03 PM, Steve French wrote: > > Updated version incorporates Ronnie's suggestion of leaving the > > default (for directory caching) the same as it is today, 1 second, to > > avoid > > unnecessary risk. Most users can safely improve performance by > > mounting with acdirmax to a higher value (e.g. 60 seconds as NFS > > defaults to). > > > > nfs and cifs on Linux currently have a mount parameter "actimeo" to control > > metadata (attribute) caching but cifs does not have additional mount > > parameters to allow distinguishing between caching directory metadata > > (e.g. needed to revalidate paths) and that for files. > > The behaviors seem to be slightly different with this change. > With NFS, the actimeo option overrides the four min/max options, > and by default the directory ac timers range between 30 and 60. > > The CIFS code I see below seems to completely separate actimeo > and acdirmax, and if not set, uses the historic 1 second value. > That's fine, but it's completely different from NFS. Shouldn't we > use a different mount option, to avoid confusing the admin? Ugh ... You are probably right. I was trying to avoid two problems: 1) (a minor one) adding a second mount option rather than just one (to solve the same problem). But reducing confusion is worth an extra mount option 2) how to avoid the user specifying *both* actimeo and acregmax - which one 'wins' (presumably the last one in the mount line) We could check for this and warn the user in mount.cifs so maybe not important to worry about in the kernel though. I will add the acregmax mount option and change actimeo to mean if (actimeo is set) acregmax = acdirmax = actimeo
Added additional patch to add "acregmax" so now behavior is more similar to nfs. "acregmax" changes file metadata caching timeout "acdirmax" changes directory metadata caching "actimeo" does what it did before - and changes both. On Wed, Feb 24, 2021 at 11:31 AM Steve French <smfrench@gmail.com> wrote: > > On Wed, Feb 24, 2021 at 10:11 AM Tom Talpey <tom@talpey.com> wrote: > > > > On 2/23/2021 8:03 PM, Steve French wrote: > > > Updated version incorporates Ronnie's suggestion of leaving the > > > default (for directory caching) the same as it is today, 1 second, to > > > avoid > > > unnecessary risk. Most users can safely improve performance by > > > mounting with acdirmax to a higher value (e.g. 60 seconds as NFS > > > defaults to). > > > > > > nfs and cifs on Linux currently have a mount parameter "actimeo" to control > > > metadata (attribute) caching but cifs does not have additional mount > > > parameters to allow distinguishing between caching directory metadata > > > (e.g. needed to revalidate paths) and that for files. > > > > The behaviors seem to be slightly different with this change. > > With NFS, the actimeo option overrides the four min/max options, > > and by default the directory ac timers range between 30 and 60. > > > > The CIFS code I see below seems to completely separate actimeo > > and acdirmax, and if not set, uses the historic 1 second value. > > That's fine, but it's completely different from NFS. Shouldn't we > > use a different mount option, to avoid confusing the admin? > > Ugh ... You are probably right. I was trying to avoid two problems: > 1) (a minor one) adding a second mount option rather than just one (to > solve the same problem). But reducing confusion is worth an extra > mount option > > 2) how to avoid the user specifying *both* actimeo and acregmax - > which one 'wins' (presumably the last one in the mount line) > We could check for this and warn the user in mount.cifs so maybe not > important to worry about in the kernel though. > > I will add the acregmax mount option and change actimeo to mean > if (actimeo is set) > acregmax = acdirmax = actimeo > > > -- > Thanks, > > Steve
On 2/24/2021 1:20 PM, Steve French wrote: > Added additional patch to add "acregmax" so now behavior is more similar to nfs. > > "acregmax" changes file metadata caching timeout > "acdirmax" changes directory metadata caching > "actimeo" does what it did before - and changes both. Clever. It's a little weird that specifying either max with actimeo will kick a warning, and maybe surprising to someone who sets both maxes to the same value will see actimeo instead. But they'll get over that. :) You can add to all three, my Reviewed-By: Tom Talpey <tom@talpey.com> > On Wed, Feb 24, 2021 at 11:31 AM Steve French <smfrench@gmail.com> wrote: >> >> On Wed, Feb 24, 2021 at 10:11 AM Tom Talpey <tom@talpey.com> wrote: >>> >>> On 2/23/2021 8:03 PM, Steve French wrote: >>>> Updated version incorporates Ronnie's suggestion of leaving the >>>> default (for directory caching) the same as it is today, 1 second, to >>>> avoid >>>> unnecessary risk. Most users can safely improve performance by >>>> mounting with acdirmax to a higher value (e.g. 60 seconds as NFS >>>> defaults to). >>>> >>>> nfs and cifs on Linux currently have a mount parameter "actimeo" to control >>>> metadata (attribute) caching but cifs does not have additional mount >>>> parameters to allow distinguishing between caching directory metadata >>>> (e.g. needed to revalidate paths) and that for files. >>> >>> The behaviors seem to be slightly different with this change. >>> With NFS, the actimeo option overrides the four min/max options, >>> and by default the directory ac timers range between 30 and 60. >>> >>> The CIFS code I see below seems to completely separate actimeo >>> and acdirmax, and if not set, uses the historic 1 second value. >>> That's fine, but it's completely different from NFS. Shouldn't we >>> use a different mount option, to avoid confusing the admin? >> >> Ugh ... You are probably right. I was trying to avoid two problems: >> 1) (a minor one) adding a second mount option rather than just one (to >> solve the same problem). But reducing confusion is worth an extra >> mount option >> >> 2) how to avoid the user specifying *both* actimeo and acregmax - >> which one 'wins' (presumably the last one in the mount line) >> We could check for this and warn the user in mount.cifs so maybe not >> important to worry about in the kernel though. >> >> I will add the acregmax mount option and change actimeo to mean >> if (actimeo is set) >> acregmax = acdirmax = actimeo >> >> >> -- >> Thanks, >> >> Steve > > >
Looks good to me. I'd prefer it if the default value of acdirmax is much higher, and not the same as acregmax value. Otherwise, this becomes something that we ask the users to do only after they start complaining about performance. Maybe we can change it once we document these new options in the man page? Regards, Shyam On Thu, Feb 25, 2021 at 9:20 AM Tom Talpey <tom@talpey.com> wrote: > > On 2/24/2021 1:20 PM, Steve French wrote: > > Added additional patch to add "acregmax" so now behavior is more similar to nfs. > > > > "acregmax" changes file metadata caching timeout > > "acdirmax" changes directory metadata caching > > "actimeo" does what it did before - and changes both. > > Clever. It's a little weird that specifying either max with > actimeo will kick a warning, and maybe surprising to someone > who sets both maxes to the same value will see actimeo instead. > But they'll get over that. :) > > You can add to all three, my > Reviewed-By: Tom Talpey <tom@talpey.com> > > > On Wed, Feb 24, 2021 at 11:31 AM Steve French <smfrench@gmail.com> wrote: > >> > >> On Wed, Feb 24, 2021 at 10:11 AM Tom Talpey <tom@talpey.com> wrote: > >>> > >>> On 2/23/2021 8:03 PM, Steve French wrote: > >>>> Updated version incorporates Ronnie's suggestion of leaving the > >>>> default (for directory caching) the same as it is today, 1 second, to > >>>> avoid > >>>> unnecessary risk. Most users can safely improve performance by > >>>> mounting with acdirmax to a higher value (e.g. 60 seconds as NFS > >>>> defaults to). > >>>> > >>>> nfs and cifs on Linux currently have a mount parameter "actimeo" to control > >>>> metadata (attribute) caching but cifs does not have additional mount > >>>> parameters to allow distinguishing between caching directory metadata > >>>> (e.g. needed to revalidate paths) and that for files. > >>> > >>> The behaviors seem to be slightly different with this change. > >>> With NFS, the actimeo option overrides the four min/max options, > >>> and by default the directory ac timers range between 30 and 60. > >>> > >>> The CIFS code I see below seems to completely separate actimeo > >>> and acdirmax, and if not set, uses the historic 1 second value. > >>> That's fine, but it's completely different from NFS. Shouldn't we > >>> use a different mount option, to avoid confusing the admin? > >> > >> Ugh ... You are probably right. I was trying to avoid two problems: > >> 1) (a minor one) adding a second mount option rather than just one (to > >> solve the same problem). But reducing confusion is worth an extra > >> mount option > >> > >> 2) how to avoid the user specifying *both* actimeo and acregmax - > >> which one 'wins' (presumably the last one in the mount line) > >> We could check for this and warn the user in mount.cifs so maybe not > >> important to worry about in the kernel though. > >> > >> I will add the acregmax mount option and change actimeo to mean > >> if (actimeo is set) > >> acregmax = acdirmax = actimeo > >> > >> > >> -- > >> Thanks, > >> > >> Steve > > > > > >
From 79f731e14559be57d48647d22283b93600caf192 Mon Sep 17 00:00:00 2001 From: Steve French <stfrench@microsoft.com> Date: Tue, 23 Feb 2021 16:16:09 -0600 Subject: [PATCH 2/2] cifs: convert revalidate of directories to using directory metadata cache timeout The new optional mount parm, "acdirmax" allows caching the metadata for a directory longer than file metadata, which can be very helpful for performance. Convert cifs_inode_needs_reval to check acdirmax for revalidating directory metadata (actimeo is checked for files) Signed-off-by: Steve French <stfrench@microsoft.com> --- fs/cifs/inode.c | 23 +++++++++++++++++------ 1 file changed, 17 insertions(+), 6 deletions(-) diff --git a/fs/cifs/inode.c b/fs/cifs/inode.c index a83b3a8ffaac..cfd31cc4520f 100644 --- a/fs/cifs/inode.c +++ b/fs/cifs/inode.c @@ -2198,12 +2198,23 @@ cifs_inode_needs_reval(struct inode *inode) if (!lookupCacheEnabled) return true; - if (!cifs_sb->ctx->actimeo) - return true; - - if (!time_in_range(jiffies, cifs_i->time, - cifs_i->time + cifs_sb->ctx->actimeo)) - return true; + /* + * depending on inode type, check if attribute caching disabled for + * files or directories + */ + if (S_ISDIR(inode->i_mode)) { + if (!cifs_sb->ctx->acdirmax) + return true; + if (!time_in_range(jiffies, cifs_i->time, + cifs_i->time + cifs_sb->ctx->acdirmax)) + return true; + } else { /* file */ + if (!cifs_sb->ctx->actimeo) + return true; + if (!time_in_range(jiffies, cifs_i->time, + cifs_i->time + cifs_sb->ctx->actimeo)) + return true; + } /* hardlinked files w/ noserverino get "special" treatment */ if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_SERVER_INUM) && -- 2.27.0