| Message ID | 163244687436.2701674.5377184817013946444.stgit@magnolia (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | xfs: separate btree cursor slab zones | expand |
On Thu, Sep 23, 2021 at 06:27:54PM -0700, Darrick J. Wong wrote: > From: Darrick J. Wong <djwong@kernel.org> > > Add code for all five btree types so that we can compute the absolute > maximum possible btree height for each btree type, and then check that > none of them exceed XFS_BTREE_CUR_ZONE_MAXLEVELS. The code to do the > actual checking is a little excessive, but it sets us up for per-type > cursor zones in the next patch. Ok, I think the cursor "zone" array is the wrong approach here. First of all - can we stop using the term "zone" for new code? That's the old irix terminolgy for slab caches, and we have been moving away from that to the Linux "kmem_cache" terminology and types for quite some time. AFAICT, the only reason for having the zone array is so that xfs_btree_alloc_cursor() can do a lookup via btnum into the array to get the maxlevels and kmem cache pointer to allocate from. Given that we've just called into xfs_btree_alloc_cursor() from the specific btree type we are allocating the cursor for (that's where we got btnum from!), we should just be passing these type specific variables directly from the caller like we do for btnum. That gets rid of the need for the zone array completely.... i.e. I don't see why the per-type cache information needs to be global information. The individual max-level calculations could just be individual kmem_cache_alloc() calls to set locally defined (i.e. static global) cache pointers and max size variables. > diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c > index c8fea6a464d5..ce428c98e7c4 100644 > --- a/fs/xfs/libxfs/xfs_ialloc_btree.c > +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c > @@ -541,6 +541,17 @@ xfs_inobt_maxrecs( > return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > } > > +unsigned int > +xfs_inobt_absolute_maxlevels(void) > +{ > + unsigned int minrecs[2]; > + > + xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > + sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > + > + return xfs_btree_compute_maxlevels(minrecs, XFS_MAX_AG_INODES); > +} i.e. rather than returning the size here, we do: static int xfs_inobt_maxlevels; static struct kmem_cache xfs_inobt_cursor_cache; int __init xfs_inobt_create_cursor_cache(void) { unsigned int minrecs[2]; xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); xfs_inobt_maxlevels = xfs_btree_compute_maxlevels(minrecs, XFS_MAX_AG_INODES); xfs_inobt_cursor_cache = kmem_cache_alloc("xfs_inobt_cur", xfs_btree_cur_sizeof(xfs_inobt_maxlevels), 0, 0, NULL); if (!xfs_inobt_cursor_cache) return -ENOMEM; return 0; } void xfs_inobt_destroy_cursor_cache(void) { kmem_cache_destroy(xfs_inobt_cursor_cache); } and nothing outside fs/xfs/libxfs/xfs_ialloc_btree.c ever needs to know about these variables as they only ever feed into xfs_btree_alloc_cursor() from xfs_inobt_init_common(). Cheers, Dave.
On Sun, Sep 26, 2021 at 10:43:43AM +1000, Dave Chinner wrote: > On Thu, Sep 23, 2021 at 06:27:54PM -0700, Darrick J. Wong wrote: > > From: Darrick J. Wong <djwong@kernel.org> > > > > Add code for all five btree types so that we can compute the absolute > > maximum possible btree height for each btree type, and then check that > > none of them exceed XFS_BTREE_CUR_ZONE_MAXLEVELS. The code to do the > > actual checking is a little excessive, but it sets us up for per-type > > cursor zones in the next patch. > > Ok, I think the cursor "zone" array is the wrong approach here. > > First of all - can we stop using the term "zone" for new code? > That's the old irix terminolgy for slab caches, and we have been > moving away from that to the Linux "kmem_cache" terminology and > types for quite some time. > > AFAICT, the only reason for having the zone array is so that > xfs_btree_alloc_cursor() can do a lookup via btnum into the array to > get the maxlevels and kmem cache pointer to allocate from. > > Given that we've just called into xfs_btree_alloc_cursor() from the > specific btree type we are allocating the cursor for (that's where > we got btnum from!), we should just be passing these type specific > variables directly from the caller like we do for btnum. That gets > rid of the need for the zone array completely.... > > i.e. I don't see why the per-type cache information needs to be > global information. The individual max-level calculations could just > be individual kmem_cache_alloc() calls to set locally defined (i.e. > static global) cache pointers and max size variables. If the cache is a static variable inside xfs_fubar_btree.c, how do you know which cache to pass to kmem_cache_free in xfs_btree_del_cursor? Does this imply adding per-btree del_cursor functions and refactoring the entire codebase to use them? I was /trying/ to get a dependent patchset ready so that Chandan could submit the extent counters patchset for 5.16, not trigger a refactoring of a whole ton of btree code. If you want to hide the information that badly, please take over this patchset and solve both the above problem and then one below. > > diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c > > index c8fea6a464d5..ce428c98e7c4 100644 > > --- a/fs/xfs/libxfs/xfs_ialloc_btree.c > > +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c > > @@ -541,6 +541,17 @@ xfs_inobt_maxrecs( > > return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > } > > > > +unsigned int > > +xfs_inobt_absolute_maxlevels(void) > > +{ > > + unsigned int minrecs[2]; > > + > > + xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > > + sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > + > > + return xfs_btree_compute_maxlevels(minrecs, XFS_MAX_AG_INODES); > > +} > > i.e. rather than returning the size here, we do: > > static int xfs_inobt_maxlevels; > static struct kmem_cache xfs_inobt_cursor_cache; > > int __init > xfs_inobt_create_cursor_cache(void) > { > unsigned int minrecs[2]; > > xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > xfs_inobt_maxlevels = xfs_btree_compute_maxlevels(minrecs, > XFS_MAX_AG_INODES); Something you couldn't have seen here is that the xfsprogs port contains an addition to the xfs_db btheight switch to print these absolute maxima so that we won't have to compute them by hand anymore. Maybe I should have noted both of these points in the commit message? Though I've also been chided for submitting excessive comments in the past, which is why I didn't. --D > xfs_inobt_cursor_cache = kmem_cache_alloc("xfs_inobt_cur", > xfs_btree_cur_sizeof(xfs_inobt_maxlevels), > 0, 0, NULL); > if (!xfs_inobt_cursor_cache) > return -ENOMEM; > return 0; > } > > void > xfs_inobt_destroy_cursor_cache(void) > { > kmem_cache_destroy(xfs_inobt_cursor_cache); > } > > and nothing outside fs/xfs/libxfs/xfs_ialloc_btree.c ever needs to > know about these variables as they only ever feed into > xfs_btree_alloc_cursor() from xfs_inobt_init_common(). > > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com
On Mon, Sep 27, 2021 at 11:17:51AM -0700, Darrick J. Wong wrote: > On Sun, Sep 26, 2021 at 10:43:43AM +1000, Dave Chinner wrote: > > On Thu, Sep 23, 2021 at 06:27:54PM -0700, Darrick J. Wong wrote: > > > From: Darrick J. Wong <djwong@kernel.org> > > > > > > Add code for all five btree types so that we can compute the absolute > > > maximum possible btree height for each btree type, and then check that > > > none of them exceed XFS_BTREE_CUR_ZONE_MAXLEVELS. The code to do the > > > actual checking is a little excessive, but it sets us up for per-type > > > cursor zones in the next patch. > > > > Ok, I think the cursor "zone" array is the wrong approach here. > > > > First of all - can we stop using the term "zone" for new code? > > That's the old irix terminolgy for slab caches, and we have been > > moving away from that to the Linux "kmem_cache" terminology and > > types for quite some time. > > > > AFAICT, the only reason for having the zone array is so that > > xfs_btree_alloc_cursor() can do a lookup via btnum into the array to > > get the maxlevels and kmem cache pointer to allocate from. > > > > Given that we've just called into xfs_btree_alloc_cursor() from the > > specific btree type we are allocating the cursor for (that's where > > we got btnum from!), we should just be passing these type specific > > variables directly from the caller like we do for btnum. That gets > > rid of the need for the zone array completely.... > > > > i.e. I don't see why the per-type cache information needs to be > > global information. The individual max-level calculations could just > > be individual kmem_cache_alloc() calls to set locally defined (i.e. > > static global) cache pointers and max size variables. > > If the cache is a static variable inside xfs_fubar_btree.c, how do you > know which cache to pass to kmem_cache_free in xfs_btree_del_cursor? > Does this imply adding per-btree del_cursor functions and refactoring > the entire codebase to use them? > > I was /trying/ to get a dependent patchset ready so that Chandan could > submit the extent counters patchset for 5.16, not trigger a refactoring > of a whole ton of btree code. If you want to hide the information that > badly, please take over this patchset and solve both the above problem > and then one below. So of course 5 minutes after sending this grouchy message I notice that everthing you asked for can be done pretty easily by having each btree type call a generic btree function that does: int __init xfs_btree_create_cursor_cache(xfs_btnum_t btnum, const char *name, unsigned int maxlevels) { struct xfs_btree_cur_cache *cc; cc = &xfs_btree_cur_caches[btnum]; cc->name = name; cc->maxlevels = maxlevels; cc->alias = false; return 0; } and a similar destructor function to null all that out. Then the xfs_db callout becomes: unsigned int xfs_btree_absolute_maxlevels(xfs_btnum_t btnum) { return xfs_btree_cur_caches[btnum].maxlevels; } printf("rmap maxlevels %u\n", libxfs_btree_absolute_maxlevels(XFS_BTNUM_RMAP)); So yeah, I'll do that. --D > > > diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c > > > index c8fea6a464d5..ce428c98e7c4 100644 > > > --- a/fs/xfs/libxfs/xfs_ialloc_btree.c > > > +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c > > > @@ -541,6 +541,17 @@ xfs_inobt_maxrecs( > > > return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > > } > > > > > > +unsigned int > > > +xfs_inobt_absolute_maxlevels(void) > > > +{ > > > + unsigned int minrecs[2]; > > > + > > > + xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > > > + sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > > + > > > + return xfs_btree_compute_maxlevels(minrecs, XFS_MAX_AG_INODES); > > > +} > > > > i.e. rather than returning the size here, we do: > > > > static int xfs_inobt_maxlevels; > > static struct kmem_cache xfs_inobt_cursor_cache; > > > > int __init > > xfs_inobt_create_cursor_cache(void) > > { > > unsigned int minrecs[2]; > > > > xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > > sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > xfs_inobt_maxlevels = xfs_btree_compute_maxlevels(minrecs, > > XFS_MAX_AG_INODES); > > Something you couldn't have seen here is that the xfsprogs port contains > an addition to the xfs_db btheight switch to print these absolute maxima > so that we won't have to compute them by hand anymore. > > Maybe I should have noted both of these points in the commit message? > Though I've also been chided for submitting excessive comments in the > past, which is why I didn't. > > --D > > > xfs_inobt_cursor_cache = kmem_cache_alloc("xfs_inobt_cur", > > xfs_btree_cur_sizeof(xfs_inobt_maxlevels), > > 0, 0, NULL); > > if (!xfs_inobt_cursor_cache) > > return -ENOMEM; > > return 0; > > } > > > > void > > xfs_inobt_destroy_cursor_cache(void) > > { > > kmem_cache_destroy(xfs_inobt_cursor_cache); > > } > > > > and nothing outside fs/xfs/libxfs/xfs_ialloc_btree.c ever needs to > > know about these variables as they only ever feed into > > xfs_btree_alloc_cursor() from xfs_inobt_init_common(). > > > > Cheers, > > > > Dave. > > -- > > Dave Chinner > > david@fromorbit.com
On Mon, Sep 27, 2021 at 11:17:51AM -0700, Darrick J. Wong wrote: > On Sun, Sep 26, 2021 at 10:43:43AM +1000, Dave Chinner wrote: > > On Thu, Sep 23, 2021 at 06:27:54PM -0700, Darrick J. Wong wrote: > > > From: Darrick J. Wong <djwong@kernel.org> > > > > > > Add code for all five btree types so that we can compute the absolute > > > maximum possible btree height for each btree type, and then check that > > > none of them exceed XFS_BTREE_CUR_ZONE_MAXLEVELS. The code to do the > > > actual checking is a little excessive, but it sets us up for per-type > > > cursor zones in the next patch. > > > > Ok, I think the cursor "zone" array is the wrong approach here. > > > > First of all - can we stop using the term "zone" for new code? > > That's the old irix terminolgy for slab caches, and we have been > > moving away from that to the Linux "kmem_cache" terminology and > > types for quite some time. > > > > AFAICT, the only reason for having the zone array is so that > > xfs_btree_alloc_cursor() can do a lookup via btnum into the array to > > get the maxlevels and kmem cache pointer to allocate from. > > > > Given that we've just called into xfs_btree_alloc_cursor() from the > > specific btree type we are allocating the cursor for (that's where > > we got btnum from!), we should just be passing these type specific > > variables directly from the caller like we do for btnum. That gets > > rid of the need for the zone array completely.... > > > > i.e. I don't see why the per-type cache information needs to be > > global information. The individual max-level calculations could just > > be individual kmem_cache_alloc() calls to set locally defined (i.e. > > static global) cache pointers and max size variables. > > If the cache is a static variable inside xfs_fubar_btree.c, how do you > know which cache to pass to kmem_cache_free in xfs_btree_del_cursor? > Does this imply adding per-btree del_cursor functions and refactoring > the entire codebase to use them? Use a switch statement on cur->bc_btnum and a function call to free the cursor? That's about the same CPU cost as a cacheline miss when looking up the global array for the cache pointer. The other obvious way of doing it is putting the cache pointer in the cursor itself. Cost is 8 bytes per cursor, but you've just reduced the size of most cursors by more than that so I see no big deal in using some of that space. It also means we don't take a cacheline miss just to look up the array that contains the pointer - the cursor will be hot at this point in time.... <looks at kmem cache code> Huh. Both slab and slub keep the cache pointer in page->slab_cache. They don't actually need (or use!) the cache pointer to be supplied - kfree() would work just fine on kmem_cache allocated objects. The problem is SLOB - it doesn't do this, and the kmem_cache implementation is subtly different to the kmalloc implementation and it doesn't store the slab cache on page->slab_cache. So to support SLOB, we have to use kmem_cache_free(). How's this for a triple negative that makes you scratch your head, but demonstrates that SLAB/SLUB can handle null cache pointers correctly: /* * Caller must not use kfree_bulk() on memory not originally allocated * by kmalloc(), because the SLOB allocator cannot handle this. */ static __always_inline void kfree_bulk(size_t size, void **p) { kmem_cache_free_bulk(NULL, size, p); } Ok, so we can't completely optimise the cache pointer way (which would be the best solution), so I'd be leaning towards a pointer within the cursor.. > I was /trying/ to get a dependent patchset ready so that Chandan could > submit the extent counters patchset for 5.16, not trigger a refactoring > of a whole ton of btree code. If you want to hide the information that > badly, please take over this patchset and solve both the above problem > and then one below. I'm not trying to hide anything, just trying to keep the implementation modular and not reliant on an explosion of one-off global constructs that have to be repeatedly looked up on every operation. Just keep a pointer to the zone in the cursor.... > > > diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c > > > index c8fea6a464d5..ce428c98e7c4 100644 > > > --- a/fs/xfs/libxfs/xfs_ialloc_btree.c > > > +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c > > > @@ -541,6 +541,17 @@ xfs_inobt_maxrecs( > > > return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > > } > > > > > > +unsigned int > > > +xfs_inobt_absolute_maxlevels(void) > > > +{ > > > + unsigned int minrecs[2]; > > > + > > > + xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > > > + sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > > + > > > + return xfs_btree_compute_maxlevels(minrecs, XFS_MAX_AG_INODES); > > > +} > > > > i.e. rather than returning the size here, we do: > > > > static int xfs_inobt_maxlevels; > > static struct kmem_cache xfs_inobt_cursor_cache; > > > > int __init > > xfs_inobt_create_cursor_cache(void) > > { > > unsigned int minrecs[2]; > > > > xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), > > sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); > > xfs_inobt_maxlevels = xfs_btree_compute_maxlevels(minrecs, > > XFS_MAX_AG_INODES); > > Something you couldn't have seen here is that the xfsprogs port contains > an addition to the xfs_db btheight switch to print these absolute maxima > so that we won't have to compute them by hand anymore. > > Maybe I should have noted both of these points in the commit message? > Though I've also been chided for submitting excessive comments in the > past, which is why I didn't. Please err on the side of verbosity in commit messages - it's better to document the obvious than it is to omit important design criteria. Anyone reading the code is not going to know what your design constraints were if you don't document them. Making sure everyone knows you've created lots of little functions to build a global table in the kernel code "because userspace" needs to know of that "because userspace" requirement... I still don't like global tables like this we have to look up on every cursor del operation. I'd much prefer the object carries the state needed for generic code to perform operations on it efficiently... Cheers, Dave.
diff --git a/fs/xfs/libxfs/xfs_alloc_btree.c b/fs/xfs/libxfs/xfs_alloc_btree.c index c644b11132f6..7f8612e383ed 100644 --- a/fs/xfs/libxfs/xfs_alloc_btree.c +++ b/fs/xfs/libxfs/xfs_alloc_btree.c @@ -582,6 +582,18 @@ xfs_allocbt_maxrecs( return blocklen / (sizeof(xfs_alloc_key_t) + sizeof(xfs_alloc_ptr_t)); } +unsigned int +xfs_allocbt_absolute_maxlevels(void) +{ + unsigned int minrecs[2]; + + xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_alloc_rec_t), + sizeof(xfs_alloc_key_t) + sizeof(xfs_alloc_ptr_t)); + + return xfs_btree_compute_maxlevels(minrecs, + (XFS_MAX_AG_BLOCKS + 1) / 2); +} + /* Calculate the freespace btree size for some records. */ xfs_extlen_t xfs_allocbt_calc_size( diff --git a/fs/xfs/libxfs/xfs_alloc_btree.h b/fs/xfs/libxfs/xfs_alloc_btree.h index 2f6b816aaf9f..0d5b9aeb78b2 100644 --- a/fs/xfs/libxfs/xfs_alloc_btree.h +++ b/fs/xfs/libxfs/xfs_alloc_btree.h @@ -59,5 +59,6 @@ extern xfs_extlen_t xfs_allocbt_calc_size(struct xfs_mount *mp, void xfs_allocbt_commit_staged_btree(struct xfs_btree_cur *cur, struct xfs_trans *tp, struct xfs_buf *agbp); +unsigned int xfs_allocbt_absolute_maxlevels(void); #endif /* __XFS_ALLOC_BTREE_H__ */ diff --git a/fs/xfs/libxfs/xfs_bmap_btree.c b/fs/xfs/libxfs/xfs_bmap_btree.c index a06987e36db5..976a324460b4 100644 --- a/fs/xfs/libxfs/xfs_bmap_btree.c +++ b/fs/xfs/libxfs/xfs_bmap_btree.c @@ -586,6 +586,19 @@ xfs_bmbt_maxrecs( return blocklen / (sizeof(xfs_bmbt_key_t) + sizeof(xfs_bmbt_ptr_t)); } +unsigned int +xfs_bmbt_absolute_maxlevels(void) +{ + unsigned int minrecs[2]; + + xfs_btree_absolute_minrecs(minrecs, XFS_BTREE_LONG_PTRS, + sizeof(struct xfs_bmbt_rec), + sizeof(struct xfs_bmbt_key) + + sizeof(xfs_bmbt_ptr_t)); + + return xfs_btree_compute_maxlevels(minrecs, MAXEXTNUM) + 1; +} + /* * Calculate number of records in a bmap btree inode root. */ diff --git a/fs/xfs/libxfs/xfs_bmap_btree.h b/fs/xfs/libxfs/xfs_bmap_btree.h index 729e3bc569be..1c5168947b96 100644 --- a/fs/xfs/libxfs/xfs_bmap_btree.h +++ b/fs/xfs/libxfs/xfs_bmap_btree.h @@ -109,5 +109,6 @@ extern struct xfs_btree_cur *xfs_bmbt_init_cursor(struct xfs_mount *, extern unsigned long long xfs_bmbt_calc_size(struct xfs_mount *mp, unsigned long long len); +unsigned int xfs_bmbt_absolute_maxlevels(void); #endif /* __XFS_BMAP_BTREE_H__ */ diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c index 619319ff41e5..120280c998f8 100644 --- a/fs/xfs/libxfs/xfs_btree.c +++ b/fs/xfs/libxfs/xfs_btree.c @@ -27,6 +27,13 @@ * Cursor allocation zone. */ kmem_zone_t *xfs_btree_cur_zone; +struct xfs_btree_cur_zone xfs_btree_cur_zones[XFS_BTNUM_MAX] = { + [XFS_BTNUM_BNO] = { .name = "xfs_alloc_btree_cur" }, + [XFS_BTNUM_INO] = { .name = "xfs_ialloc_btree_cur" }, + [XFS_BTNUM_RMAP] = { .name = "xfs_rmap_btree_cur" }, + [XFS_BTNUM_REFC] = { .name = "xfs_refc_btree_cur" }, + [XFS_BTNUM_BMAP] = { .name = "xfs_bmap_btree_cur" }, +}; /* * Btree magic numbers. @@ -5027,3 +5034,44 @@ xfs_btree_alloc_cursor( return cur; } + +/* + * Compute absolute minrecs for leaf and node btree blocks. Callers should set + * BTREE_LONG_PTRS and BTREE_OVERLAPPING as they would for regular cursors. + * Set BTREE_CRC_BLOCKS if the btree type is supported on V5 or newer + * filesystems. + */ +void +xfs_btree_absolute_minrecs( + unsigned int *minrecs, + unsigned int bc_flags, + unsigned int leaf_recbytes, + unsigned int node_recbytes) +{ + unsigned int min_recbytes; + + /* + * If this btree type is supported on V4, we use the smaller V4 min + * block size along with the V4 header size. If the btree type is only + * supported on V5, use the (twice as large) V5 min block size along + * with the V5 header size. + */ + if (!(bc_flags & XFS_BTREE_CRC_BLOCKS)) { + if (bc_flags & XFS_BTREE_LONG_PTRS) + min_recbytes = XFS_MIN_BLOCKSIZE - + XFS_BTREE_LBLOCK_LEN; + else + min_recbytes = XFS_MIN_BLOCKSIZE - + XFS_BTREE_SBLOCK_LEN; + } else if (bc_flags & XFS_BTREE_LONG_PTRS) { + min_recbytes = XFS_MIN_CRC_BLOCKSIZE - XFS_BTREE_LBLOCK_CRC_LEN; + } else { + min_recbytes = XFS_MIN_CRC_BLOCKSIZE - XFS_BTREE_SBLOCK_CRC_LEN; + } + + if (bc_flags & XFS_BTREE_OVERLAPPING) + node_recbytes <<= 1; + + minrecs[0] = min_recbytes / (2 * leaf_recbytes); + minrecs[1] = min_recbytes / (2 * node_recbytes); +} diff --git a/fs/xfs/libxfs/xfs_btree.h b/fs/xfs/libxfs/xfs_btree.h index 57b7aa3f6366..5bebd26f8b2c 100644 --- a/fs/xfs/libxfs/xfs_btree.h +++ b/fs/xfs/libxfs/xfs_btree.h @@ -583,4 +583,14 @@ struct xfs_btree_cur *xfs_btree_alloc_cursor(struct xfs_mount *mp, struct xfs_trans *tp, xfs_btnum_t btnum); unsigned int xfs_btree_maxlevels(struct xfs_mount *mp, xfs_btnum_t btnum); +void xfs_btree_absolute_minrecs(unsigned int *minrecs, unsigned int bc_flags, + unsigned int leaf_recbytes, unsigned int node_recbytes); + +struct xfs_btree_cur_zone { + const char *name; + unsigned int maxlevels; +}; + +extern struct xfs_btree_cur_zone xfs_btree_cur_zones[XFS_BTNUM_MAX]; + #endif /* __XFS_BTREE_H__ */ diff --git a/fs/xfs/libxfs/xfs_fs.h b/fs/xfs/libxfs/xfs_fs.h index bde2b4c64dbe..33d323ddf6da 100644 --- a/fs/xfs/libxfs/xfs_fs.h +++ b/fs/xfs/libxfs/xfs_fs.h @@ -268,6 +268,9 @@ typedef struct xfs_fsop_resblks { */ #define XFS_MIN_AG_BYTES (1ULL << 24) /* 16 MB */ #define XFS_MAX_AG_BYTES (1ULL << 40) /* 1 TB */ +#define XFS_MAX_AG_BLOCKS (XFS_MAX_AG_BYTES / XFS_MIN_BLOCKSIZE) +#define XFS_MAX_CRC_AG_BLOCKS (XFS_MAX_AG_BYTES / XFS_MIN_CRC_BLOCKSIZE) +#define XFS_MAX_AG_INODES (XFS_MAX_AG_BYTES / XFS_DINODE_MIN_SIZE) /* keep the maximum size under 2^31 by a small amount */ #define XFS_MAX_LOG_BYTES \ diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c index c8fea6a464d5..ce428c98e7c4 100644 --- a/fs/xfs/libxfs/xfs_ialloc_btree.c +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c @@ -541,6 +541,17 @@ xfs_inobt_maxrecs( return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); } +unsigned int +xfs_inobt_absolute_maxlevels(void) +{ + unsigned int minrecs[2]; + + xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t), + sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t)); + + return xfs_btree_compute_maxlevels(minrecs, XFS_MAX_AG_INODES); +} + /* * Convert the inode record holemask to an inode allocation bitmap. The inode * allocation bitmap is inode granularity and specifies whether an inode is diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.h b/fs/xfs/libxfs/xfs_ialloc_btree.h index 8a322d402e61..e29570b4d3b6 100644 --- a/fs/xfs/libxfs/xfs_ialloc_btree.h +++ b/fs/xfs/libxfs/xfs_ialloc_btree.h @@ -74,5 +74,6 @@ int xfs_inobt_cur(struct xfs_mount *mp, struct xfs_trans *tp, void xfs_inobt_commit_staged_btree(struct xfs_btree_cur *cur, struct xfs_trans *tp, struct xfs_buf *agbp); +unsigned int xfs_inobt_absolute_maxlevels(void); #endif /* __XFS_IALLOC_BTREE_H__ */ diff --git a/fs/xfs/libxfs/xfs_refcount_btree.c b/fs/xfs/libxfs/xfs_refcount_btree.c index 48c45e31d897..9d428c3d4d90 100644 --- a/fs/xfs/libxfs/xfs_refcount_btree.c +++ b/fs/xfs/libxfs/xfs_refcount_btree.c @@ -408,6 +408,19 @@ xfs_refcountbt_maxrecs( sizeof(xfs_refcount_ptr_t)); } +unsigned int +xfs_refcountbt_absolute_maxlevels(void) +{ + unsigned int minrecs[2]; + + xfs_btree_absolute_minrecs(minrecs, XFS_BTREE_CRC_BLOCKS, + sizeof(struct xfs_refcount_rec), + sizeof(struct xfs_refcount_key) + + sizeof(xfs_refcount_ptr_t)); + + return xfs_btree_compute_maxlevels(minrecs, XFS_MAX_CRC_AG_BLOCKS); +} + /* Compute the maximum height of a refcount btree. */ void xfs_refcountbt_compute_maxlevels( diff --git a/fs/xfs/libxfs/xfs_refcount_btree.h b/fs/xfs/libxfs/xfs_refcount_btree.h index bd9ed9e1e41f..d7dad2dbac09 100644 --- a/fs/xfs/libxfs/xfs_refcount_btree.h +++ b/fs/xfs/libxfs/xfs_refcount_btree.h @@ -64,5 +64,6 @@ extern int xfs_refcountbt_calc_reserves(struct xfs_mount *mp, void xfs_refcountbt_commit_staged_btree(struct xfs_btree_cur *cur, struct xfs_trans *tp, struct xfs_buf *agbp); +unsigned int xfs_refcountbt_absolute_maxlevels(void); #endif /* __XFS_REFCOUNT_BTREE_H__ */ diff --git a/fs/xfs/libxfs/xfs_rmap_btree.c b/fs/xfs/libxfs/xfs_rmap_btree.c index 85caeb14e4db..d73768c7546f 100644 --- a/fs/xfs/libxfs/xfs_rmap_btree.c +++ b/fs/xfs/libxfs/xfs_rmap_btree.c @@ -534,6 +534,31 @@ xfs_rmapbt_maxrecs( (2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t)); } +unsigned int +xfs_rmapbt_absolute_maxlevels(void) +{ + unsigned int minrecs[2]; + + xfs_btree_absolute_minrecs(minrecs, + XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING, + sizeof(struct xfs_rmap_rec), + sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t)); + + /* + * Compute the asymptotic maxlevels for an rmapbt on any reflink fs. + * + * On a reflink filesystem, each AG block can have up to 2^32 (per the + * refcount record format) owners, which means that theoretically we + * could face up to 2^64 rmap records. However, we're likely to run + * out of blocks in the AG long before that happens, which means that + * we must compute the max height based on what the btree will look + * like if it consumes almost all the blocks in the AG due to maximal + * sharing factor. + */ + return xfs_btree_compute_maxlevels_size(XFS_MAX_CRC_AG_BLOCKS, + minrecs[1]); +} + /* Compute the maximum height of an rmap btree. */ unsigned int xfs_rmapbt_compute_maxlevels( diff --git a/fs/xfs/libxfs/xfs_rmap_btree.h b/fs/xfs/libxfs/xfs_rmap_btree.h index 5aaecf755abd..6d5001e92966 100644 --- a/fs/xfs/libxfs/xfs_rmap_btree.h +++ b/fs/xfs/libxfs/xfs_rmap_btree.h @@ -58,5 +58,6 @@ extern xfs_extlen_t xfs_rmapbt_max_size(struct xfs_mount *mp, extern int xfs_rmapbt_calc_reserves(struct xfs_mount *mp, struct xfs_trans *tp, struct xfs_perag *pag, xfs_extlen_t *ask, xfs_extlen_t *used); +unsigned int xfs_rmapbt_absolute_maxlevels(void); #endif /* __XFS_RMAP_BTREE_H__ */ diff --git a/fs/xfs/libxfs/xfs_types.h b/fs/xfs/libxfs/xfs_types.h index b6da06b40989..6ce15a4f9f47 100644 --- a/fs/xfs/libxfs/xfs_types.h +++ b/fs/xfs/libxfs/xfs_types.h @@ -126,6 +126,9 @@ typedef enum { XFS_BTNUM_INOi, XFS_BTNUM_FINOi, XFS_BTNUM_REFCi, XFS_BTNUM_MAX } xfs_btnum_t; +#define for_each_xfs_btnum(btnum) \ + for((btnum) = XFS_BTNUM_BNOi; (btnum) < XFS_BTNUM_MAX; (btnum)++) + #define XFS_BTNUM_STRINGS \ { XFS_BTNUM_BNOi, "bnobt" }, \ { XFS_BTNUM_CNTi, "cntbt" }, \ diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c index 90c92a6a49e0..83abe9bfe3ef 100644 --- a/fs/xfs/xfs_super.c +++ b/fs/xfs/xfs_super.c @@ -37,6 +37,12 @@ #include "xfs_reflink.h" #include "xfs_pwork.h" #include "xfs_ag.h" +#include "xfs_btree.h" +#include "xfs_alloc_btree.h" +#include "xfs_ialloc_btree.h" +#include "xfs_bmap_btree.h" +#include "xfs_rmap_btree.h" +#include "xfs_refcount_btree.h" #include <linux/magic.h> #include <linux/fs_context.h> @@ -1950,9 +1956,34 @@ static struct file_system_type xfs_fs_type = { }; MODULE_ALIAS_FS("xfs"); +STATIC int __init +xfs_init_btree_zones(void) +{ + struct xfs_btree_cur_zone *z = xfs_btree_cur_zones; + xfs_btnum_t btnum; + + z[XFS_BTNUM_BNO].maxlevels = xfs_allocbt_absolute_maxlevels(); + z[XFS_BTNUM_RMAP].maxlevels = xfs_rmapbt_absolute_maxlevels(); + z[XFS_BTNUM_BMAP].maxlevels = xfs_bmbt_absolute_maxlevels(); + z[XFS_BTNUM_INO].maxlevels = xfs_inobt_absolute_maxlevels(); + z[XFS_BTNUM_REFC].maxlevels = xfs_refcountbt_absolute_maxlevels(); + + for_each_xfs_btnum(btnum) { + ASSERT(z[btnum].maxlevels <= XFS_BTREE_CUR_ZONE_MAXLEVELS); + } + + /* struct copies for btree types that share zones */ + z[XFS_BTNUM_CNT] = z[XFS_BTNUM_BNO]; + z[XFS_BTNUM_FINO] = z[XFS_BTNUM_INO]; + + return 0; +} + STATIC int __init xfs_init_zones(void) { + int error; + xfs_log_ticket_zone = kmem_cache_create("xfs_log_ticket", sizeof(struct xlog_ticket), 0, 0, NULL); @@ -1965,6 +1996,10 @@ xfs_init_zones(void) if (!xfs_bmap_free_item_zone) goto out_destroy_log_ticket_zone; + error = xfs_init_btree_zones(); + if (error) + goto out_destroy_bmap_free_item_zone; + xfs_btree_cur_zone = kmem_cache_create("xfs_btree_cur", xfs_btree_cur_sizeof(XFS_BTREE_CUR_ZONE_MAXLEVELS), 0, 0, NULL);