Commit 66ee620f06 ("idr: Permit any valid kernel pointer to be stored")
changed the radix tree lookup so that it stops when reaching the bottom
of the tree. However, the condition was added in the wrong place,
making it possible to return retry entries to the caller. Reorder the
tests to check for the retry entry before checking whether we're at the
bottom of the tree. The retry entry should never be found in the tree
root, so it's safe to defer the check until the end of the loop.
Add a regression test to the test-suite to be sure this doesn't come
back.
Fixes: 66ee620f06 ("idr: Permit any valid kernel pointer to be stored")
Reported-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Matthew Wilcox <willy@infradead.org>
All users have now been converted to the XArray. Removing the support
reduces code size and ensures new users will use the XArray instead.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
The tag_tagged_items() function is supposed to test the page-writeback
tagging code. Since that has been converted to the XArray, there's
not much point in testing the radix tree's tagging code. This requires
using the pthread mutex embedded in the xarray instead of an external
lock, so remove the pthread mutexes which protect xarrays/radix trees.
Also remove radix_tree_iter_tag_set() as this was the last user.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
The page cache was the only user of this interface and it has now
been converted to the XArray. Transform the test into a test of
xas_init_marks().
Signed-off-by: Matthew Wilcox <willy@infradead.org>
radix_tree_split and radix_tree_join were never used upstream. Remove
them; if they're needed in future they will be replaced by XArray
equivalents.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
The only user of this functionality was the workingset code, and it's
now been converted to the XArray. Remove __radix_tree_delete_node()
entirely as it was also only used by the workingset code.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
xa_find() is a slightly easier API to use than
radix_tree_gang_lookup_slot() because it contains its own RCU locking.
This commit removes the last user of radix_tree_gang_lookup_slot()
so remove the function too.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Use the XArray APIs to add and replace pages in the page cache. This
removes two uses of the radix tree preload API and is significantly
shorter code. It also removes the last user of __radix_tree_create()
outside radix-tree.c itself, so make it static.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Use the XA_TRACK_FREE ability to track which entries have a free bit,
similarly to how it uses the radix tree's IDR_FREE tag. This eliminates
the per-cpu ida_bitmap preload, and fixes the memory consumption
regression I introduced when making the IDR able to store any pointer.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
xa_store() differs from radix_tree_insert() in that it will overwrite an
existing element in the array rather than returning an error. This is
the behaviour which most users want, and those that want more complex
behaviour generally want to use the xas family of routines anyway.
For memory allocation, xa_store() will first attempt to request memory
from the slab allocator; if memory is not immediately available, it will
drop the xa_lock and allocate memory, keeping a pointer in the xa_state.
It does not use the per-CPU cache, although those will continue to exist
until all radix tree users are converted to the xarray.
This patch also includes xa_erase() and __xa_erase() for a streamlined
way to store NULL. Since there is no need to allocate memory in order
to store a NULL in the XArray, we do not need to trouble the user with
deciding what memory allocation flags to use.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
The xa_load function brings with it a lot of infrastructure; xa_empty(),
xa_is_err(), and large chunks of the XArray advanced API that are used
to implement xa_load.
As the test-suite demonstrates, it is possible to use the XArray functions
on a radix tree. The radix tree functions depend on the GFP flags being
stored in the root of the tree, so it's not possible to use the radix
tree functions on an XArray.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
This is a direct replacement for struct radix_tree_node. A couple of
struct members have changed name, so convert those. Use a #define so
that radix tree users continue to work without change.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
This is a direct replacement for struct radix_tree_root. Some of the
struct members have changed name; convert those, and use a #define so
that radix_tree users continue to work without change.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Instead of storing a pointer to the slot containing the canonical entry,
store the offset of the slot. Produces slightly more efficient code
(~300 bytes) and simplifies the implementation.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Introduce xarray value entries and tagged pointers to replace radix
tree exceptional entries. This is a slight change in encoding to allow
the use of an extra bit (we can now store BITS_PER_LONG - 1 bits in a
value entry). It is also a change in emphasis; exceptional entries are
intimidating and different. As the comment explains, you can choose
to store values or pointers in the xarray and they are both first-class
citizens.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
An upcoming change to the encoding of internal entries will set the bottom
two bits to 0b10. Unfortunately, m68k only aligns some data structures
to 2 bytes, so the IDR will interpret them as internal entries and things
will go badly wrong.
Change the radix tree so that it stops either when the node indicates
that it's the bottom of the tree (shift == 0) or when the entry is not an
internal entry. This means we cannot insert an arbitrary kernel pointer
as a multiorder entry, but the IDR does not permit multiorder entries.
Annoyingly, this means the IDR can no longer take advantage of the radix
tree's ability to store a single entry at offset 0 without allocating
memory. A pointer which is 2-byte aligned cannot be stored directly in
the root as it would be indistinguishable from a node, so we must allocate
a node in order to store a 2-byte pointer at index 0. The idr_replace()
function does not take a GFP flags argument, so cannot allocate memory.
If a user inserts a 4-byte aligned pointer at index 0 and then replaces
it with a 2-byte aligned pointer, we must be able to store it.
Arbitrary pointer values are still not permitted; pointers of the
form 2 + (i * 4) for values of i between 0 and 1023 are reserved for
the implementation. These are not valid kernel pointers as they would
point into the zero page.
This change does cause a runtime memory consumption regression for
the IDA. I will recover that later.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Tested-by: Guenter Roeck <linux@roeck-us.net>
Delete ida_pre_get(), ida_get_new(), ida_get_new_above() and ida_remove()
from the public API. Some of these functions still exist as internal
helpers, but they should not be called by consumers.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
get_slot_offset() can be called with a NULL 'parent' argument.
In this case, the calculated value will not be used, but calculating
it is undefined. Rather than fixing the caller (__radix_tree_delete)
to not call get_slot_offset(), make get_slot_offset() robust against
being called with a NULL parent.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
If the radix tree underlying the IDR happens to be full and we attempt
to remove an id which is larger than any id in the IDR, we will call
__radix_tree_delete() with an uninitialised 'slot' pointer, at which
point anything could happen. This was easiest to hit with a single
entry at id 0 and attempting to remove a non-0 id, but it could have
happened with 64 entries and attempting to remove an id >= 64.
Roman said:
The syzcaller test boils down to opening /dev/kvm, creating an
eventfd, and calling a couple of KVM ioctls. None of this requires
superuser. And the result is dereferencing an uninitialized pointer
which is likely a crash. The specific path caught by syzbot is via
KVM_HYPERV_EVENTD ioctl which is new in 4.17. But I guess there are
other user-triggerable paths, so cc:stable is probably justified.
Matthew added:
We have around 250 calls to idr_remove() in the kernel today. Many of
them pass an ID which is embedded in the object they're removing, so
they're safe. Picking a few likely candidates:
drivers/firewire/core-cdev.c looks unsafe; the ID comes from an ioctl.
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c is similar
drivers/atm/nicstar.c could be taken down by a handcrafted packet
Link: http://lkml.kernel.org/r/20180518175025.GD6361@bombadil.infradead.org
Fixes: 0a835c4f09 ("Reimplement IDR and IDA using the radix tree")
Reported-by: <syzbot+35666cba7f0a337e2e79@syzkaller.appspotmail.com>
Debugged-by: Roman Kagan <rkagan@virtuozzo.com>
Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix a race in the multi-order iteration code which causes the kernel to
hit a GP fault. This was first seen with a production v4.15 based
kernel (4.15.6-300.fc27.x86_64) utilizing a DAX workload which used
order 9 PMD DAX entries.
The race has to do with how we tear down multi-order sibling entries
when we are removing an item from the tree. Remember for example that
an order 2 entry looks like this:
struct radix_tree_node.slots[] = [entry][sibling][sibling][sibling]
where 'entry' is in some slot in the struct radix_tree_node, and the
three slots following 'entry' contain sibling pointers which point back
to 'entry.'
When we delete 'entry' from the tree, we call :
radix_tree_delete()
radix_tree_delete_item()
__radix_tree_delete()
replace_slot()
replace_slot() first removes the siblings in order from the first to the
last, then at then replaces 'entry' with NULL. This means that for a
brief period of time we end up with one or more of the siblings removed,
so:
struct radix_tree_node.slots[] = [entry][NULL][sibling][sibling]
This causes an issue if you have a reader iterating over the slots in
the tree via radix_tree_for_each_slot() while only under
rcu_read_lock()/rcu_read_unlock() protection. This is a common case in
mm/filemap.c.
The issue is that when __radix_tree_next_slot() => skip_siblings() tries
to skip over the sibling entries in the slots, it currently does so with
an exact match on the slot directly preceding our current slot.
Normally this works:
V preceding slot
struct radix_tree_node.slots[] = [entry][sibling][sibling][sibling]
^ current slot
This lets you find the first sibling, and you skip them all in order.
But in the case where one of the siblings is NULL, that slot is skipped
and then our sibling detection is interrupted:
V preceding slot
struct radix_tree_node.slots[] = [entry][NULL][sibling][sibling]
^ current slot
This means that the sibling pointers aren't recognized since they point
all the way back to 'entry', so we think that they are normal internal
radix tree pointers. This causes us to think we need to walk down to a
struct radix_tree_node starting at the address of 'entry'.
In a real running kernel this will crash the thread with a GP fault when
you try and dereference the slots in your broken node starting at
'entry'.
We fix this race by fixing the way that skip_siblings() detects sibling
nodes. Instead of testing against the preceding slot we instead look
for siblings via is_sibling_entry() which compares against the position
of the struct radix_tree_node.slots[] array. This ensures that sibling
entries are properly identified, even if they are no longer contiguous
with the 'entry' they point to.
Link: http://lkml.kernel.org/r/20180503192430.7582-6-ross.zwisler@linux.intel.com
Fixes: 148deab223 ("radix-tree: improve multiorder iterators")
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reported-by: CR, Sapthagirish <sapthagirish.cr@intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "XArray", v9. (First part thereof).
This patchset is, I believe, appropriate for merging for 4.17. It
contains the XArray implementation, to eventually replace the radix
tree, and converts the page cache to use it.
This conversion keeps the radix tree and XArray data structures in sync
at all times. That allows us to convert the page cache one function at
a time and should allow for easier bisection. Other than renaming some
elements of the structures, the data structures are fundamentally
unchanged; a radix tree walk and an XArray walk will touch the same
number of cachelines. I have changes planned to the XArray data
structure, but those will happen in future patches.
Improvements the XArray has over the radix tree:
- The radix tree provides operations like other trees do; 'insert' and
'delete'. But what most users really want is an automatically
resizing array, and so it makes more sense to give users an API that
is like an array -- 'load' and 'store'. We still have an 'insert'
operation for users that really want that semantic.
- The XArray considers locking as part of its API. This simplifies a
lot of users who formerly had to manage their own locking just for
the radix tree. It also improves code generation as we can now tell
RCU that we're holding a lock and it doesn't need to generate as much
fencing code. The other advantage is that tree nodes can be moved
(not yet implemented).
- GFP flags are now parameters to calls which may need to allocate
memory. The radix tree forced users to decide what the allocation
flags would be at creation time. It's much clearer to specify them at
allocation time.
- Memory is not preloaded; we don't tie up dozens of pages on the off
chance that the slab allocator fails. Instead, we drop the lock,
allocate a new node and retry the operation. We have to convert all
the radix tree, IDA and IDR preload users before we can realise this
benefit, but I have not yet found a user which cannot be converted.
- The XArray provides a cmpxchg operation. The radix tree forces users
to roll their own (and at least four have).
- Iterators take a 'max' parameter. That simplifies many users and will
reduce the amount of iteration done.
- Iteration can proceed backwards. We only have one user for this, but
since it's called as part of the pagefault readahead algorithm, that
seemed worth mentioning.
- RCU-protected pointers are not exposed as part of the API. There are
some fun bugs where the page cache forgets to use rcu_dereference()
in the current codebase.
- Value entries gain an extra bit compared to radix tree exceptional
entries. That gives us the extra bit we need to put huge page swap
entries in the page cache.
- Some iterators now take a 'filter' argument instead of having
separate iterators for tagged/untagged iterations.
The page cache is improved by this:
- Shorter, easier to read code
- More efficient iterations
- Reduction in size of struct address_space
- Fewer walks from the top of the data structure; the XArray API
encourages staying at the leaf node and conducting operations there.
This patch (of 8):
None of these bits may be used for slab allocations, so we can use them
as radix tree flags as long as we mask them off before passing them to
the slab allocator. Move the IDR flag from the high bits to the
GFP_ZONEMASK bits.
Link: http://lkml.kernel.org/r/20180313132639.17387-3-willy@infradead.org
Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com>
Acked-by: Jeff Layton <jlayton@kernel.org>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As far as I can tell, the only place the per-cpu ida_bitmap is populated
is in ida_pre_get. The pre-allocated element is stolen in two places in
ida_get_new_above, in both cases immediately followed by a memset(0).
Since ida_get_new_above is called with locks held, do the zeroing in
ida_pre_get, or rather let kmalloc() do it. Also, apparently gcc
generates ~44 bytes of code to do a memset(, 0, 128):
$ scripts/bloat-o-meter vmlinux.{0,1}
add/remove: 0/0 grow/shrink: 2/1 up/down: 5/-88 (-83)
Function old new delta
ida_pre_get 115 119 +4
vermagic 27 28 +1
ida_get_new_above 715 627 -88
Link: http://lkml.kernel.org/r/20180108225634.15340-1-linux@rasmusvillemoes.dk
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Acked-by: Matthew Wilcox <mawilcox@microsoft.com>
Cc: Eric Biggers <ebiggers@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It has no more users, so remove it. Move idr_alloc() back into idr.c,
move the guts of idr_alloc_cmn() into idr_alloc_u32(), remove the
wrappers around idr_get_free_cmn() and rename it to idr_get_free().
While there is now no interface to allocate IDs larger than a u32,
the IDR internals remain ready to handle a larger ID should a need arise.
These changes make it possible to provide the guarantee that, if the
nextid pointer points into the object, the object's ID will be initialised
before a concurrent lookup can find the object.
Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com>
During truncation, the mapping has already been checked for shmem and
dax so it's known that workingset_update_node is required.
This patch avoids the checks on mapping for each page being truncated.
In all other cases, a lookup helper is used to determine if
workingset_update_node() needs to be called. The one danger is that the
API is slightly harder to use as calling workingset_update_node directly
without checking for dax or shmem mappings could lead to surprises.
However, the API rarely needs to be used and hopefully the comment is
enough to give people the hint.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
oneirq-v1r1 pickhelper-v1r1
Min Time 141.00 ( 0.00%) 140.00 ( 0.71%)
1st-qrtle Time 142.00 ( 0.00%) 141.00 ( 0.70%)
2nd-qrtle Time 142.00 ( 0.00%) 142.00 ( 0.00%)
3rd-qrtle Time 143.00 ( 0.00%) 143.00 ( 0.00%)
Max-90% Time 144.00 ( 0.00%) 144.00 ( 0.00%)
Max-95% Time 147.00 ( 0.00%) 145.00 ( 1.36%)
Max-99% Time 195.00 ( 0.00%) 191.00 ( 2.05%)
Max Time 230.00 ( 0.00%) 205.00 ( 10.87%)
Amean Time 144.37 ( 0.00%) 143.82 ( 0.38%)
Stddev Time 10.44 ( 0.00%) 9.00 ( 13.74%)
Coeff Time 7.23 ( 0.00%) 6.26 ( 13.41%)
Best99%Amean Time 143.72 ( 0.00%) 143.34 ( 0.26%)
Best95%Amean Time 142.37 ( 0.00%) 142.00 ( 0.26%)
Best90%Amean Time 142.19 ( 0.00%) 141.85 ( 0.24%)
Best75%Amean Time 141.92 ( 0.00%) 141.58 ( 0.24%)
Best50%Amean Time 141.69 ( 0.00%) 141.31 ( 0.27%)
Best25%Amean Time 141.38 ( 0.00%) 140.97 ( 0.29%)
As you'd expect, the gain is marginal but it can be detected. The
differences in bonnie are all within the noise which is not surprising
given the impact on the microbenchmark.
radix_tree_update_node_t is a callback for some radix operations that
optionally passes in a private field. The only user of the callback is
workingset_update_node and as it no longer requires a mapping, the
private field is removed.
Link: http://lkml.kernel.org/r/20171018075952.10627-3-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Matthew Wilcox
Matthew Wilcox
Ross Zwisler
Rasmus Villemoes
Mel Gorman