ceph: add new "nopagecache" option
RFC PATCH 1/1 ceph: add new "nopagecache" option
CephFS is a bit unlike most other filesystems in that it only conditionally does buffered I/O based on the caps that it gets from the MDS. In most cases, unless there is contended access for an inode the MDS does give Fbc caps to the client, so the unbuffered codepaths are only infrequently traveled and are difficult to test.
At one time, the "-o sync" mount option would give you this behavior, but that was removed in commit 7ab9b3807097 ("ceph: Don't use ceph-sync-mode for synchronous-fs.").
Add a new mount option to tell the client to ignore Fbc caps when doing I/O, and to use the synchronous codepaths exclusively, even on non-O_DIRECT file descriptors. We already have an ioctl that forces this behavior on a per-inode basis, so we can just always set the CEPH_F_SYNC flag on such mounts.
Additionally, this patch also changes the client to not request Fbc when doing direct I/O. We aren't using the cache with O_DIRECT so we don't have any need for those caps.
--- fs/ceph/file.c | 24 +++++++++++++++--------- fs/ceph/super.c | 11 +++++++++++ fs/ceph/super.h | 1 + 3 files changed, 27 insertions(+), 9 deletions(-)
I've been working with this patch in order to test the synchronous codepaths for content encryption. I think though that this might make sense to take into mainline, as it could be helpful for troubleshooting, and ensuring that these codepaths are regularly tested.
Either way, I think the cap handling changes probably make sense on their own. I can split that out if we don't want to take the mount option in as well.
Thoughts?
fs/ceph/file.c +15 -9
@@ -204,6 +204,8 @@ static int ceph_init_file_info(struct inode *inode, struct file *file, int fmode, bool isdir) {struct ceph_inode_info *ci =ceph_inode (inode); +struct ceph_mount_options *opt = +ceph_inode_to_client (&ci->vfs_inode)->mount_options;struct ceph_file_info *fi; dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
@@ -225,6 +227,9 @@ static int ceph_init_file_info(struct inode *inode, struct file *file,if (!fi)return -ENOMEM ; +if (opt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) + fi->flags |=CEPH_F_SYNC ; + file->private_data = fi; }
@@ -1536,7 +1541,7 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)struct ceph_inode_info *ci =ceph_inode (inode);bool direct_lock = iocb->ki_flags &IOCB_DIRECT ;ssize_t ret; -int want, got =0 ; +int want =0 , got =0 ;int retry_op =0 , read =0 ; again:
@@ -1551,13 +1556,14 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)else ceph_start_io_read (inode); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_CACHE ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_CACHE |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_CACHE ; + want |=CEPH_CAP_FILE_LAZYIO ; + ret =ceph_get_caps (filp,CEPH_CAP_FILE_RD , want,-1 , &got);if (ret <0 ) { -if (iocb->ki_flags &IOCB_DIRECT ) +if (direct_lock)ceph_end_io_direct (inode);else ceph_end_io_read (inode);
@@ -1691,7 +1697,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)struct ceph_osd_client *osdc = &fsc->client->osdc;struct ceph_cap_flush *prealloc_cf;ssize_t count, written =0 ; -int err, want, got; +int err, want =0 , got;bool direct_lock = false;u32 map_flags;u64 pool_flags;
@@ -1766,10 +1772,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)dout ("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n" , inode,ceph_vinop (inode), pos, count,i_size_read (inode)); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_BUFFER ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_BUFFER |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_BUFFER ; + want |=CEPH_CAP_FILE_LAZYIO ; got =0 ; err =ceph_get_caps (file,CEPH_CAP_FILE_WR , want, pos + count, &got);if (err <0 )
fs/ceph/super.c +11 -0
@@ -160,6 +160,7 @@ enum { Opt_quotadf, Opt_copyfrom, Opt_wsync, + Opt_pagecache, }; enum ceph_recover_session_mode {
@@ -201,6 +202,7 @@ static const struct fs_parameter_spec ceph_mount_parameters[] = {fsparam_string ("mon_addr" , Opt_mon_addr),fsparam_u32 ("wsize" , Opt_wsize),fsparam_flag_no ("wsync" , Opt_wsync), + fsparam_flag_no ("pagecache", Opt_pagecache), {} };
@@ -567,6 +569,12 @@ static int ceph_parse_mount_param(struct fs_context *fc,else fsopt->flags |=CEPH_MOUNT_OPT_ASYNC_DIROPS ;break ; +case Opt_pagecache: +if (result.negated) + fsopt->flags |=CEPH_MOUNT_OPT_NOPAGECACHE ; +else + fsopt->flags &= ~CEPH_MOUNT_OPT_NOPAGECACHE ; +break ;default :BUG (); }
@@ -702,6 +710,9 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)if (!(fsopt->flags &CEPH_MOUNT_OPT_ASYNC_DIROPS ))seq_puts (m,",wsync" ); +if (fsopt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) +seq_puts (m,",nopagecache" ); +if (fsopt->wsize !=CEPH_MAX_WRITE_SIZE )seq_printf (m,",wsize=%u" , fsopt->wsize);if (fsopt->rsize !=CEPH_MAX_READ_SIZE )
fs/ceph/super.h +1 -0
@@ -46,6 +46,7 @@#define CEPH_MOUNT_OPT_NOQUOTADF (1<<13)/* no root dir quota in statfs */ #define CEPH_MOUNT_OPT_NOCOPYFROM (1<<14)/* don't use RADOS 'copy-from' op */ #define CEPH_MOUNT_OPT_ASYNC_DIROPS (1<<15)/* allow async directory ops */ +#define CEPH_MOUNT_OPT_NOPAGECACHE (1<<16)/* bypass pagecache altogether */ #define CEPH_MOUNT_OPT_DEFAULT \ (CEPH_MOUNT_OPT_DCACHE | \
Signature · 1 line
2.33.1
This seems fine to me. Do you have a teuthology fragment to exercise this?
Quoted 172 lines · Jeff Layton
CephFS is a bit unlike most other filesystems in that it only conditionally does buffered I/O based on the caps that it gets from the MDS. In most cases, unless there is contended access for an inode the MDS does give Fbc caps to the client, so the unbuffered codepaths are only infrequently traveled and are difficult to test.
At one time, the "-o sync" mount option would give you this behavior, but that was removed in commit
7ab9b3807097("ceph: Don't use ceph-sync-mode for synchronous-fs.").Add a new mount option to tell the client to ignore Fbc caps when doing I/O, and to use the synchronous codepaths exclusively, even on non-O_DIRECT file descriptors. We already have an ioctl that forces this behavior on a per-inode basis, so we can just always set the
CEPH_F_SYNCflag on such mounts.Additionally, this patch also changes the client to not request Fbc when doing direct I/O. We aren't using the cache with
O_DIRECTso we don't have any need for those caps.Cc majianpeng <[email protected]>Signed-off-by Jeff Layton <[email protected]>--- fs/ceph/file.c | 24 +++++++++++++++--------- fs/ceph/super.c | 11 +++++++++++ fs/ceph/super.h | 1 + 3 files changed, 27 insertions(+), 9 deletions(-)I've been working with this patch in order to test the synchronous codepaths for content encryption. I think though that this might make sense to take into mainline, as it could be helpful for troubleshooting, and ensuring that these codepaths are regularly tested.
Either way, I think the cap handling changes probably make sense on their own. I can split that out if we don't want to take the mount option in as well.
Thoughts?
fs/ceph/file.c +15
-9@@ -204,6 +204,8 @@ static int ceph_init_file_info(struct inode *inode, struct file *file, int fmode, bool isdir) {struct ceph_inode_info *ci =ceph_inode (inode); +struct ceph_mount_options *opt = +ceph_inode_to_client (&ci->vfs_inode)->mount_options;struct ceph_file_info *fi; dout("%s %p %p 0%o (%s)\n", __func__, inode, file,@@ -225,6 +227,9 @@ static int ceph_init_file_info(struct inode *inode, struct file *file,if (!fi)return -ENOMEM ; +if (opt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) + fi->flags |=CEPH_F_SYNC ; + file->private_data = fi; }@@ -1536,7 +1541,7 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)struct ceph_inode_info *ci =ceph_inode (inode);bool direct_lock = iocb->ki_flags &IOCB_DIRECT ;ssize_t ret; -int want, got =0 ; +int want =0 , got =0 ;int retry_op =0 , read =0 ; again:@@ -1551,13 +1556,14 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)else ceph_start_io_read (inode); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_CACHE ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_CACHE |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_CACHE ; + want |=CEPH_CAP_FILE_LAZYIO ; + ret =ceph_get_caps (filp,CEPH_CAP_FILE_RD , want,-1 , &got);if (ret <0 ) { -if (iocb->ki_flags &IOCB_DIRECT ) +if (direct_lock)ceph_end_io_direct (inode);else ceph_end_io_read (inode);@@ -1691,7 +1697,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)struct ceph_osd_client *osdc = &fsc->client->osdc;struct ceph_cap_flush *prealloc_cf;ssize_t count, written =0 ; -int err, want, got; +int err, want =0 , got;bool direct_lock = false;u32 map_flags;u64 pool_flags;@@ -1766,10 +1772,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)dout ("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n" , inode,ceph_vinop (inode), pos, count,i_size_read (inode)); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_BUFFER ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_BUFFER |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_BUFFER ; + want |=CEPH_CAP_FILE_LAZYIO ; got =0 ; err =ceph_get_caps (file,CEPH_CAP_FILE_WR , want, pos + count, &got);if (err <0 )fs/ceph/super.c +11
-0@@ -160,6 +160,7 @@ enum { Opt_quotadf, Opt_copyfrom, Opt_wsync, + Opt_pagecache, }; enum ceph_recover_session_mode {@@ -201,6 +202,7 @@ static const struct fs_parameter_spec ceph_mount_parameters[] = {fsparam_string ("mon_addr" , Opt_mon_addr),fsparam_u32 ("wsize" , Opt_wsize),fsparam_flag_no ("wsync" , Opt_wsync), + fsparam_flag_no ("pagecache", Opt_pagecache), {} };@@ -567,6 +569,12 @@ static int ceph_parse_mount_param(struct fs_context *fc,else fsopt->flags |=CEPH_MOUNT_OPT_ASYNC_DIROPS ;break ; +case Opt_pagecache: +if (result.negated) + fsopt->flags |=CEPH_MOUNT_OPT_NOPAGECACHE ; +else + fsopt->flags &= ~CEPH_MOUNT_OPT_NOPAGECACHE ; +break ;default :BUG (); }@@ -702,6 +710,9 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)if (!(fsopt->flags &CEPH_MOUNT_OPT_ASYNC_DIROPS ))seq_puts (m,",wsync" ); +if (fsopt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) +seq_puts (m,",nopagecache" ); +if (fsopt->wsize !=CEPH_MAX_WRITE_SIZE )seq_printf (m,",wsize=%u" , fsopt->wsize);if (fsopt->rsize !=CEPH_MAX_READ_SIZE )fs/ceph/super.h +1
-1@@ -46,6 +46,7 @@#define CEPH_MOUNT_OPT_NOQUOTADF (1<<13)/* no root dir quota in statfs */ #define CEPH_MOUNT_OPT_NOCOPYFROM (1<<14)/* don't use RADOS 'copy-from' op */ #define CEPH_MOUNT_OPT_ASYNC_DIROPS (1<<15)/* allow async directory ops */ +#define CEPH_MOUNT_OPT_NOPAGECACHE (1<<16)/* bypass pagecache altogether */ #define CEPH_MOUNT_OPT_DEFAULT \ (CEPH_MOUNT_OPT_DCACHE | \ --2.33.1
No, I've only used this for local testing so far. Perhaps you can handhold me through adding such a patch to teuthology for it?
Note that this _really_ slows down I/O, especially if you're doing a lot of small reads or writes, so it may not be appropriate for throughput- heavy testing.
Quoted 178 lines · Gregory Farnum
This seems fine to me. Do you have a teuthology fragment to exercise this?
Acked-by Greg Farnum <[email protected]>Quoted 172 lines · Jeff Layton
CephFS is a bit unlike most other filesystems in that it only conditionally does buffered I/O based on the caps that it gets from the MDS. In most cases, unless there is contended access for an inode the MDS does give Fbc caps to the client, so the unbuffered codepaths are only infrequently traveled and are difficult to test.
At one time, the "-o sync" mount option would give you this behavior, but that was removed in commit
7ab9b3807097("ceph: Don't use ceph-sync-mode for synchronous-fs.").Add a new mount option to tell the client to ignore Fbc caps when doing I/O, and to use the synchronous codepaths exclusively, even on non-O_DIRECT file descriptors. We already have an ioctl that forces this behavior on a per-inode basis, so we can just always set the
CEPH_F_SYNCflag on such mounts.Additionally, this patch also changes the client to not request Fbc when doing direct I/O. We aren't using the cache with
O_DIRECTso we don't have any need for those caps.Cc majianpeng <[email protected]>Signed-off-by Jeff Layton <[email protected]>--- fs/ceph/file.c | 24 +++++++++++++++--------- fs/ceph/super.c | 11 +++++++++++ fs/ceph/super.h | 1 + 3 files changed, 27 insertions(+), 9 deletions(-)I've been working with this patch in order to test the synchronous codepaths for content encryption. I think though that this might make sense to take into mainline, as it could be helpful for troubleshooting, and ensuring that these codepaths are regularly tested.
Either way, I think the cap handling changes probably make sense on their own. I can split that out if we don't want to take the mount option in as well.
Thoughts?
fs/ceph/file.c +15
-9@@ -204,6 +204,8 @@ static int ceph_init_file_info(struct inode *inode, struct file *file, int fmode, bool isdir) {struct ceph_inode_info *ci =ceph_inode (inode); +struct ceph_mount_options *opt = +ceph_inode_to_client (&ci->vfs_inode)->mount_options;struct ceph_file_info *fi; dout("%s %p %p 0%o (%s)\n", __func__, inode, file,@@ -225,6 +227,9 @@ static int ceph_init_file_info(struct inode *inode, struct file *file,if (!fi)return -ENOMEM ; +if (opt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) + fi->flags |=CEPH_F_SYNC ; + file->private_data = fi; }@@ -1536,7 +1541,7 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)struct ceph_inode_info *ci =ceph_inode (inode);bool direct_lock = iocb->ki_flags &IOCB_DIRECT ;ssize_t ret; -int want, got =0 ; +int want =0 , got =0 ;int retry_op =0 , read =0 ; again:@@ -1551,13 +1556,14 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)else ceph_start_io_read (inode); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_CACHE ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_CACHE |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_CACHE ; + want |=CEPH_CAP_FILE_LAZYIO ; + ret =ceph_get_caps (filp,CEPH_CAP_FILE_RD , want,-1 , &got);if (ret <0 ) { -if (iocb->ki_flags &IOCB_DIRECT ) +if (direct_lock)ceph_end_io_direct (inode);else ceph_end_io_read (inode);@@ -1691,7 +1697,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)struct ceph_osd_client *osdc = &fsc->client->osdc;struct ceph_cap_flush *prealloc_cf;ssize_t count, written =0 ; -int err, want, got; +int err, want =0 , got;bool direct_lock = false;u32 map_flags;u64 pool_flags;@@ -1766,10 +1772,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)dout ("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n" , inode,ceph_vinop (inode), pos, count,i_size_read (inode)); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_BUFFER ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_BUFFER |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_BUFFER ; + want |=CEPH_CAP_FILE_LAZYIO ; got =0 ; err =ceph_get_caps (file,CEPH_CAP_FILE_WR , want, pos + count, &got);if (err <0 )fs/ceph/super.c +11
-0@@ -160,6 +160,7 @@ enum { Opt_quotadf, Opt_copyfrom, Opt_wsync, + Opt_pagecache, }; enum ceph_recover_session_mode {@@ -201,6 +202,7 @@ static const struct fs_parameter_spec ceph_mount_parameters[] = {fsparam_string ("mon_addr" , Opt_mon_addr),fsparam_u32 ("wsize" , Opt_wsize),fsparam_flag_no ("wsync" , Opt_wsync), + fsparam_flag_no ("pagecache", Opt_pagecache), {} };@@ -567,6 +569,12 @@ static int ceph_parse_mount_param(struct fs_context *fc,else fsopt->flags |=CEPH_MOUNT_OPT_ASYNC_DIROPS ;break ; +case Opt_pagecache: +if (result.negated) + fsopt->flags |=CEPH_MOUNT_OPT_NOPAGECACHE ; +else + fsopt->flags &= ~CEPH_MOUNT_OPT_NOPAGECACHE ; +break ;default :BUG (); }@@ -702,6 +710,9 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)if (!(fsopt->flags &CEPH_MOUNT_OPT_ASYNC_DIROPS ))seq_puts (m,",wsync" ); +if (fsopt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) +seq_puts (m,",nopagecache" ); +if (fsopt->wsize !=CEPH_MAX_WRITE_SIZE )seq_printf (m,",wsize=%u" , fsopt->wsize);if (fsopt->rsize !=CEPH_MAX_READ_SIZE )fs/ceph/super.h +1
-1@@ -46,6 +46,7 @@#define CEPH_MOUNT_OPT_NOQUOTADF (1<<13)/* no root dir quota in statfs */ #define CEPH_MOUNT_OPT_NOCOPYFROM (1<<14)/* don't use RADOS 'copy-from' op */ #define CEPH_MOUNT_OPT_ASYNC_DIROPS (1<<15)/* allow async directory ops */ +#define CEPH_MOUNT_OPT_NOPAGECACHE (1<<16)/* bypass pagecache altogether */ #define CEPH_MOUNT_OPT_DEFAULT \ (CEPH_MOUNT_OPT_DCACHE | \ --2.33.1
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 3 lines · Jeff Layton
No, I've only used this for local testing so far. Perhaps you can handhold me through adding such a patch to teuthology for it?
Well, we mount with the "kclient" task:
https://github.com/ceph/ceph/blob/master/qa/tasks/kclient.py
And it looks like that already supports passing arbitrary mount
options: https://github.com/ceph/ceph/blob/master/qa/tasks/kclient.py#L51
So I think that wherever we use the kclient, we look at if that subsuite/test is appropriate to test without the page cache. If it is, split up the kclient yaml fragment to specify both with and without the new mount option.
Right now it looks like the tree is organized so we just include
https://github.com/ceph/ceph/tree/master/qa/cephfs/mount anywhere we use CephFS mounts, and that will generate tests against both ceph-fuse and the kclient. So we could update that by replacing the single "kclient/mount.yaml" file with a folder "kclient_mount" that contains the existing mount.yaml, and a no_page_cache_mount.yaml that looks like
tasks:
- kclient:
- mntopts: ["nopagecache"]
Quoted 188 lines
Note that this _really_ slows down I/O, especially if you're doing a lot of small reads or writes, so it may not be appropriate for throughput- heavy testing.
Quoted 172 lines · Gregory Farnum
This seems fine to me. Do you have a teuthology fragment to exercise this?
Acked-by Greg Farnum <[email protected]>Quoted 167 lines · Jeff Layton
CephFS is a bit unlike most other filesystems in that it only conditionally does buffered I/O based on the caps that it gets from the MDS. In most cases, unless there is contended access for an inode the MDS does give Fbc caps to the client, so the unbuffered codepaths are only infrequently traveled and are difficult to test.
At one time, the "-o sync" mount option would give you this behavior, but that was removed in commit
7ab9b3807097("ceph: Don't use ceph-sync-mode for synchronous-fs.").Add a new mount option to tell the client to ignore Fbc caps when doing I/O, and to use the synchronous codepaths exclusively, even on non-O_DIRECT file descriptors. We already have an ioctl that forces this behavior on a per-inode basis, so we can just always set the
CEPH_F_SYNCflag on such mounts.Additionally, this patch also changes the client to not request Fbc when doing direct I/O. We aren't using the cache with
O_DIRECTso we don't have any need for those caps.Cc majianpeng <[email protected]>Signed-off-by Jeff Layton <[email protected]>--- fs/ceph/file.c | 24 +++++++++++++++--------- fs/ceph/super.c | 11 +++++++++++ fs/ceph/super.h | 1 + 3 files changed, 27 insertions(+), 9 deletions(-)I've been working with this patch in order to test the synchronous codepaths for content encryption. I think though that this might make sense to take into mainline, as it could be helpful for troubleshooting, and ensuring that these codepaths are regularly tested.
Either way, I think the cap handling changes probably make sense on their own. I can split that out if we don't want to take the mount option in as well.
Thoughts?
fs/ceph/file.c +15
-9@@ -204,6 +204,8 @@ static int ceph_init_file_info(struct inode *inode, struct file *file, int fmode, bool isdir) {struct ceph_inode_info *ci =ceph_inode (inode); +struct ceph_mount_options *opt = +ceph_inode_to_client (&ci->vfs_inode)->mount_options;struct ceph_file_info *fi; dout("%s %p %p 0%o (%s)\n", __func__, inode, file,@@ -225,6 +227,9 @@ static int ceph_init_file_info(struct inode *inode, struct file *file,if (!fi)return -ENOMEM ; +if (opt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) + fi->flags |=CEPH_F_SYNC ; + file->private_data = fi; }@@ -1536,7 +1541,7 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)struct ceph_inode_info *ci =ceph_inode (inode);bool direct_lock = iocb->ki_flags &IOCB_DIRECT ;ssize_t ret; -int want, got =0 ; +int want =0 , got =0 ;int retry_op =0 , read =0 ; again:@@ -1551,13 +1556,14 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)else ceph_start_io_read (inode); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_CACHE ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_CACHE |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_CACHE ; + want |=CEPH_CAP_FILE_LAZYIO ; + ret =ceph_get_caps (filp,CEPH_CAP_FILE_RD , want,-1 , &got);if (ret <0 ) { -if (iocb->ki_flags &IOCB_DIRECT ) +if (direct_lock)ceph_end_io_direct (inode);else ceph_end_io_read (inode);@@ -1691,7 +1697,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)struct ceph_osd_client *osdc = &fsc->client->osdc;struct ceph_cap_flush *prealloc_cf;ssize_t count, written =0 ; -int err, want, got; +int err, want =0 , got;bool direct_lock = false;u32 map_flags;u64 pool_flags;@@ -1766,10 +1772,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)dout ("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n" , inode,ceph_vinop (inode), pos, count,i_size_read (inode)); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_BUFFER ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_BUFFER |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_BUFFER ; + want |=CEPH_CAP_FILE_LAZYIO ; got =0 ; err =ceph_get_caps (file,CEPH_CAP_FILE_WR , want, pos + count, &got);if (err <0 )fs/ceph/super.c +11
-0@@ -160,6 +160,7 @@ enum { Opt_quotadf, Opt_copyfrom, Opt_wsync, + Opt_pagecache, }; enum ceph_recover_session_mode {@@ -201,6 +202,7 @@ static const struct fs_parameter_spec ceph_mount_parameters[] = {fsparam_string ("mon_addr" , Opt_mon_addr),fsparam_u32 ("wsize" , Opt_wsize),fsparam_flag_no ("wsync" , Opt_wsync), + fsparam_flag_no ("pagecache", Opt_pagecache), {} };@@ -567,6 +569,12 @@ static int ceph_parse_mount_param(struct fs_context *fc,else fsopt->flags |=CEPH_MOUNT_OPT_ASYNC_DIROPS ;break ; +case Opt_pagecache: +if (result.negated) + fsopt->flags |=CEPH_MOUNT_OPT_NOPAGECACHE ; +else + fsopt->flags &= ~CEPH_MOUNT_OPT_NOPAGECACHE ; +break ;default :BUG (); }@@ -702,6 +710,9 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)if (!(fsopt->flags &CEPH_MOUNT_OPT_ASYNC_DIROPS ))seq_puts (m,",wsync" ); +if (fsopt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) +seq_puts (m,",nopagecache" ); +if (fsopt->wsize !=CEPH_MAX_WRITE_SIZE )seq_printf (m,",wsize=%u" , fsopt->wsize);if (fsopt->rsize !=CEPH_MAX_READ_SIZE )fs/ceph/super.h +1
-0@@ -46,6 +46,7 @@#define CEPH_MOUNT_OPT_NOQUOTADF (1<<13)/* no root dir quota in statfs */ #define CEPH_MOUNT_OPT_NOCOPYFROM (1<<14)/* don't use RADOS 'copy-from' op */ #define CEPH_MOUNT_OPT_ASYNC_DIROPS (1<<15)/* allow async directory ops */ +#define CEPH_MOUNT_OPT_NOPAGECACHE (1<<16)/* bypass pagecache altogether */ Quoted 6 lines
Quoted 5 lines
fs/ceph/super.h +0
-1#define CEPH_MOUNT_OPT_DEFAULT \ (CEPH_MOUNT_OPT_DCACHE | \ --2.33.1
--
Jeff Layton <[email protected]>
Quoted 24 lines · Gregory Farnum
Quoted 3 lines · Jeff Layton
No, I've only used this for local testing so far. Perhaps you can handhold me through adding such a patch to teuthology for it?
Well, we mount with the "kclient" task:
https://github.com/ceph/ceph/blob/master/qa/tasks/kclient.pyAnd it looks like that already supports passing arbitrary mount
options: https://github.com/ceph/ceph/blob/master/qa/tasks/kclient.py#L51So I think that wherever we use the kclient, we look at if that subsuite/test is appropriate to test without the page cache. If it is, split up the kclient yaml fragment to specify both with and without the new mount option.
Right now it looks like the tree is organized so we just include
https://github.com/ceph/ceph/tree/master/qa/cephfs/mount anywhere we use CephFS mounts, and that will generate tests against both ceph-fuse and the kclient. So we could update that by replacing the single "kclient/mount.yaml" file with a folder "kclient_mount" that contains the existingmount.yaml, and ano_page_cache_mount.yamlthat looks like
Quoted 4 lines
tasks:
- kclient:
- mntopts: ["nopagecache"]
Do I also need one of the magic %, +, or $ files in this directory?
Also, what will be effect of doing this? Will the new mount option be used randomly, or will I need to specify a test run that includes this file?
Quoted 189 lines
Quoted 188 lines
Note that this _really_ slows down I/O, especially if you're doing a lot of small reads or writes, so it may not be appropriate for throughput- heavy testing.
Quoted 172 lines · Gregory Farnum
This seems fine to me. Do you have a teuthology fragment to exercise this?
Acked-by Greg Farnum <[email protected]>Quoted 167 lines · Jeff Layton
CephFS is a bit unlike most other filesystems in that it only conditionally does buffered I/O based on the caps that it gets from the MDS. In most cases, unless there is contended access for an inode the MDS does give Fbc caps to the client, so the unbuffered codepaths are only infrequently traveled and are difficult to test.
At one time, the "-o sync" mount option would give you this behavior, but that was removed in commit
7ab9b3807097("ceph: Don't use ceph-sync-mode for synchronous-fs.").Add a new mount option to tell the client to ignore Fbc caps when doing I/O, and to use the synchronous codepaths exclusively, even on non-O_DIRECT file descriptors. We already have an ioctl that forces this behavior on a per-inode basis, so we can just always set the
CEPH_F_SYNCflag on such mounts.Additionally, this patch also changes the client to not request Fbc when doing direct I/O. We aren't using the cache with
O_DIRECTso we don't have any need for those caps.Cc majianpeng <[email protected]>Signed-off-by Jeff Layton <[email protected]>--- fs/ceph/file.c | 24 +++++++++++++++--------- fs/ceph/super.c | 11 +++++++++++ fs/ceph/super.h | 1 + 3 files changed, 27 insertions(+), 9 deletions(-)I've been working with this patch in order to test the synchronous codepaths for content encryption. I think though that this might make sense to take into mainline, as it could be helpful for troubleshooting, and ensuring that these codepaths are regularly tested.
Either way, I think the cap handling changes probably make sense on their own. I can split that out if we don't want to take the mount option in as well.
Thoughts?
fs/ceph/file.c +15
-9@@ -204,6 +204,8 @@ static int ceph_init_file_info(struct inode *inode, struct file *file, int fmode, bool isdir) {struct ceph_inode_info *ci =ceph_inode (inode); +struct ceph_mount_options *opt = +ceph_inode_to_client (&ci->vfs_inode)->mount_options;struct ceph_file_info *fi; dout("%s %p %p 0%o (%s)\n", __func__, inode, file,@@ -225,6 +227,9 @@ static int ceph_init_file_info(struct inode *inode, struct file *file,if (!fi)return -ENOMEM ; +if (opt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) + fi->flags |=CEPH_F_SYNC ; + file->private_data = fi; }@@ -1536,7 +1541,7 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)struct ceph_inode_info *ci =ceph_inode (inode);bool direct_lock = iocb->ki_flags &IOCB_DIRECT ;ssize_t ret; -int want, got =0 ; +int want =0 , got =0 ;int retry_op =0 , read =0 ; again:@@ -1551,13 +1556,14 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)else ceph_start_io_read (inode); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_CACHE ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_CACHE |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_CACHE ; + want |=CEPH_CAP_FILE_LAZYIO ; + ret =ceph_get_caps (filp,CEPH_CAP_FILE_RD , want,-1 , &got);if (ret <0 ) { -if (iocb->ki_flags &IOCB_DIRECT ) +if (direct_lock)ceph_end_io_direct (inode);else ceph_end_io_read (inode);@@ -1691,7 +1697,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)struct ceph_osd_client *osdc = &fsc->client->osdc;struct ceph_cap_flush *prealloc_cf;ssize_t count, written =0 ; -int err, want, got; +int err, want =0 , got;bool direct_lock = false;u32 map_flags;u64 pool_flags;@@ -1766,10 +1772,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)dout ("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n" , inode,ceph_vinop (inode), pos, count,i_size_read (inode)); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_BUFFER ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_BUFFER |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_BUFFER ; + want |=CEPH_CAP_FILE_LAZYIO ; got =0 ; err =ceph_get_caps (file,CEPH_CAP_FILE_WR , want, pos + count, &got);if (err <0 )fs/ceph/super.c +11
-0@@ -160,6 +160,7 @@ enum { Opt_quotadf, Opt_copyfrom, Opt_wsync, + Opt_pagecache, }; enum ceph_recover_session_mode {@@ -201,6 +202,7 @@ static const struct fs_parameter_spec ceph_mount_parameters[] = {fsparam_string ("mon_addr" , Opt_mon_addr),fsparam_u32 ("wsize" , Opt_wsize),fsparam_flag_no ("wsync" , Opt_wsync), + fsparam_flag_no ("pagecache", Opt_pagecache), {} };@@ -567,6 +569,12 @@ static int ceph_parse_mount_param(struct fs_context *fc,else fsopt->flags |=CEPH_MOUNT_OPT_ASYNC_DIROPS ;break ; +case Opt_pagecache: +if (result.negated) + fsopt->flags |=CEPH_MOUNT_OPT_NOPAGECACHE ; +else + fsopt->flags &= ~CEPH_MOUNT_OPT_NOPAGECACHE ; +break ;default :BUG (); }@@ -702,6 +710,9 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)if (!(fsopt->flags &CEPH_MOUNT_OPT_ASYNC_DIROPS ))seq_puts (m,",wsync" ); +if (fsopt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) +seq_puts (m,",nopagecache" ); +if (fsopt->wsize !=CEPH_MAX_WRITE_SIZE )seq_printf (m,",wsize=%u" , fsopt->wsize);if (fsopt->rsize !=CEPH_MAX_READ_SIZE )fs/ceph/super.h +1
-0@@ -46,6 +46,7 @@#define CEPH_MOUNT_OPT_NOQUOTADF (1<<13)/* no root dir quota in statfs */ #define CEPH_MOUNT_OPT_NOCOPYFROM (1<<14)/* don't use RADOS 'copy-from' op */ #define CEPH_MOUNT_OPT_ASYNC_DIROPS (1<<15)/* allow async directory ops */ +#define CEPH_MOUNT_OPT_NOPAGECACHE (1<<16)/* bypass pagecache altogether */ Quoted 6 lines
Quoted 5 lines
fs/ceph/super.h +0
-1#define CEPH_MOUNT_OPT_DEFAULT \ (CEPH_MOUNT_OPT_DCACHE | \ --2.33.1
--
Jeff Layton <[email protected]>
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 17 lines · Jeff Layton
CephFS is a bit unlike most other filesystems in that it only conditionally does buffered I/O based on the caps that it gets from the MDS. In most cases, unless there is contended access for an inode the MDS does give Fbc caps to the client, so the unbuffered codepaths are only infrequently traveled and are difficult to test.
At one time, the "-o sync" mount option would give you this behavior, but that was removed in commit
7ab9b3807097("ceph: Don't use ceph-sync-mode for synchronous-fs.").Add a new mount option to tell the client to ignore Fbc caps when doing I/O, and to use the synchronous codepaths exclusively, even on non-O_DIRECT file descriptors. We already have an ioctl that forces this behavior on a per-inode basis, so we can just always set the
CEPH_F_SYNCflag on such mounts.
Quoted 3 lines
Additionally, this patch also changes the client to not request Fbc when doing direct I/O. We aren't using the cache with
O_DIRECTso we don't have any need for those caps.
This looks good.
Quoted 152 lines
Cc majianpeng <[email protected]>Signed-off-by Jeff Layton <[email protected]>--- fs/ceph/file.c | 24 +++++++++++++++--------- fs/ceph/super.c | 11 +++++++++++ fs/ceph/super.h | 1 + 3 files changed, 27 insertions(+), 9 deletions(-)I've been working with this patch in order to test the synchronous codepaths for content encryption. I think though that this might make sense to take into mainline, as it could be helpful for troubleshooting, and ensuring that these codepaths are regularly tested.
Either way, I think the cap handling changes probably make sense on their own. I can split that out if we don't want to take the mount option in as well.
Thoughts?
fs/ceph/file.c +15
-9@@ -204,6 +204,8 @@ static int ceph_init_file_info(struct inode *inode, struct file *file, int fmode, bool isdir) {struct ceph_inode_info *ci =ceph_inode (inode); +struct ceph_mount_options *opt = +ceph_inode_to_client (&ci->vfs_inode)->mount_options;struct ceph_file_info *fi; dout("%s %p %p 0%o (%s)\n", __func__, inode, file,@@ -225,6 +227,9 @@ static int ceph_init_file_info(struct inode *inode, struct file *file,if (!fi)return -ENOMEM ; +if (opt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) + fi->flags |=CEPH_F_SYNC ; + file->private_data = fi; }@@ -1536,7 +1541,7 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)struct ceph_inode_info *ci =ceph_inode (inode);bool direct_lock = iocb->ki_flags &IOCB_DIRECT ;ssize_t ret; -int want, got =0 ; +int want =0 , got =0 ;int retry_op =0 , read =0 ; again:@@ -1551,13 +1556,14 @@ static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)else ceph_start_io_read (inode); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_CACHE ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_CACHE |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_CACHE ; + want |=CEPH_CAP_FILE_LAZYIO ; + ret =ceph_get_caps (filp,CEPH_CAP_FILE_RD , want,-1 , &got);if (ret <0 ) { -if (iocb->ki_flags &IOCB_DIRECT ) +if (direct_lock)ceph_end_io_direct (inode);else ceph_end_io_read (inode);@@ -1691,7 +1697,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)struct ceph_osd_client *osdc = &fsc->client->osdc;struct ceph_cap_flush *prealloc_cf;ssize_t count, written =0 ; -int err, want, got; +int err, want =0 , got;bool direct_lock = false;u32 map_flags;u64 pool_flags;@@ -1766,10 +1772,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)dout ("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n" , inode,ceph_vinop (inode), pos, count,i_size_read (inode)); +if (!(fi->flags &CEPH_F_SYNC ) && !direct_lock) + want |=CEPH_CAP_FILE_BUFFER ;if (fi->fmode &CEPH_FILE_MODE_LAZY ) - want =CEPH_CAP_FILE_BUFFER |CEPH_CAP_FILE_LAZYIO ; -else - want =CEPH_CAP_FILE_BUFFER ; + want |=CEPH_CAP_FILE_LAZYIO ; got =0 ; err =ceph_get_caps (file,CEPH_CAP_FILE_WR , want, pos + count, &got);if (err <0 )fs/ceph/super.c +11
-0@@ -160,6 +160,7 @@ enum { Opt_quotadf, Opt_copyfrom, Opt_wsync, + Opt_pagecache, }; enum ceph_recover_session_mode {@@ -201,6 +202,7 @@ static const struct fs_parameter_spec ceph_mount_parameters[] = {fsparam_string ("mon_addr" , Opt_mon_addr),fsparam_u32 ("wsize" , Opt_wsize),fsparam_flag_no ("wsync" , Opt_wsync), + fsparam_flag_no ("pagecache", Opt_pagecache), {} };@@ -567,6 +569,12 @@ static int ceph_parse_mount_param(struct fs_context *fc,else fsopt->flags |=CEPH_MOUNT_OPT_ASYNC_DIROPS ;break ; +case Opt_pagecache: +if (result.negated) + fsopt->flags |=CEPH_MOUNT_OPT_NOPAGECACHE ; +else + fsopt->flags &= ~CEPH_MOUNT_OPT_NOPAGECACHE ; +break ;default :BUG (); }@@ -702,6 +710,9 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)if (!(fsopt->flags &CEPH_MOUNT_OPT_ASYNC_DIROPS ))seq_puts (m,",wsync" ); +if (fsopt->flags &CEPH_MOUNT_OPT_NOPAGECACHE ) +seq_puts (m,",nopagecache" ); +if (fsopt->wsize !=CEPH_MAX_WRITE_SIZE )seq_printf (m,",wsize=%u" , fsopt->wsize);if (fsopt->rsize !=CEPH_MAX_READ_SIZE )fs/ceph/super.h +1
-1@@ -46,6 +46,7 @@#define CEPH_MOUNT_OPT_NOQUOTADF (1<<13)/* no root dir quota in statfs */ #define CEPH_MOUNT_OPT_NOCOPYFROM (1<<14)/* don't use RADOS 'copy-from' op */ #define CEPH_MOUNT_OPT_ASYNC_DIROPS (1<<15)/* allow async directory ops */ +#define CEPH_MOUNT_OPT_NOPAGECACHE (1<<16)/* bypass pagecache altogether */ #define CEPH_MOUNT_OPT_DEFAULT \ (CEPH_MOUNT_OPT_DCACHE | \ --2.33.1
Signature · 2 lines
Cheers, Venky