Sander reports following splat after netfilter nat bysrc table got
converted to rhashtable:
swapper/0: page allocation failure: order:3, mode:0x2084020(GFP_ATOMIC|__GFP_COMP)
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.8.0-rc1 [..]
[<ffffffff811633ed>] warn_alloc_failed+0xdd/0x140
[<ffffffff811638b1>] __alloc_pages_nodemask+0x3e1/0xcf0
[<ffffffff811a72ed>] alloc_pages_current+0x8d/0x110
[<ffffffff8117cb7f>] kmalloc_order+0x1f/0x70
[<ffffffff811aec19>] __kmalloc+0x129/0x140
[<ffffffff8146d561>] bucket_table_alloc+0xc1/0x1d0
[<ffffffff8146da1d>] rhashtable_insert_rehash+0x5d/0xe0
[<ffffffff819fcfff>] nf_nat_setup_info+0x2ef/0x400
The failure happens when allocating the spinlock array.
Even with GFP_KERNEL its unlikely for such a large allocation
to succeed.
Thomas Graf pointed me at inet_ehash_locks_alloc(), so in addition
to adding NOWARN for atomic allocations this also makes the bucket-array
sizing more conservative.
In commit 095dc8e0c3 ("tcp: fix/cleanup inet_ehash_locks_alloc()"),
Eric Dumazet says: "Budget 2 cache lines per cpu worth of 'spinlocks'".
IOW, consider size needed by a single spinlock when determining
number of locks per cpu. So with 64 byte per cacheline and 4 byte per
spinlock this gives 32 locks per cpu.
Resulting size of the lock-array (sizeof(spinlock) == 4):
cpus: 1 2 4 8 16 32 64
old: 1k 1k 4k 8k 16k 16k 16k
new: 128 256 512 1k 2k 4k 8k
8k allocation should have decent chance of success even
with GFP_ATOMIC, and should not fail with GFP_KERNEL.
With 72-byte spinlock (LOCKDEP):
cpus : 1 2
old: 9k 18k
new: ~2k ~4k
Reported-by: Sander Eikelenboom <linux@eikelenboom.it>
Suggested-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
In certain cases, the 802.11 mesh pathtable code wants to
iterate over all of the entries in the forwarding table from
the receive path, which is inside an RCU read-side critical
section. Enable walks inside atomic sections by allowing
GFP_ATOMIC allocations for the walker state.
Change all existing callsites to pass in GFP_KERNEL.
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: Bob Copeland <me@bobcopeland.com>
[also adjust gfs2/glock.c and rhashtable tests]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
The commit c6ff526829 ("rhashtable:
Fix walker list corruption") causes a suspicious RCU usage warning
because we no longer hold ht->mutex when we dereference ht->tbl.
However, this is a false positive because we now hold ht->lock
which also guarantees that ht->tbl won't disppear from under us.
This patch kills the warning by using rcu_dereference_protected.
Reported-by: kernel test robot <ying.huang@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/geneve.c
Here we had an overlapping change, where in 'net' the extraneous stats
bump was being removed whilst in 'net-next' the final argument to
udp_tunnel6_xmit_skb() was being changed.
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit ba7c95ea38 ("rhashtable:
Fix sleeping inside RCU critical section in walk_stop") introduced
a new spinlock for the walker list. However, it did not convert
all existing users of the list over to the new spin lock. Some
continued to use the old mutext for this purpose. This obviously
led to corruption of the list.
The fix is to use the spin lock everywhere where we touch the list.
This also allows us to do rcu_rad_lock before we take the lock in
rhashtable_walk_start. With the old mutex this would've deadlocked
but it's safe with the new spin lock.
Fixes: ba7c95ea38 ("rhashtable: Fix sleeping inside RCU...")
Reported-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
William Hua <william.hua@canonical.com> wrote:
>
> I wasn't aware there was an enforced minimum size. I simply set the
> nelem_hint in the rhastable_params struct to 1, expecting it to grow as
> needed. This caused a segfault afterwards when trying to insert an
> element.
OK we're doing the size computation before we enforce the limit
on min_size.
---8<---
We need to do the initial hash table size computation after we
have obtained the correct min_size/max_size parameters. Otherwise
we may end up with a hash table whose size is outside the allowed
envelope.
Fixes: a998f712f7 ("rhashtable: Round up/down min/max_size to...")
Reported-by: William Hua <william.hua@canonical.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch 9497df88ab ("rhashtable:
Fix reader/rehash race") added a pair of barriers. In fact the
wmb is superfluous because every subsequent write to the old or
new hash table uses rcu_assign_pointer, which itself carriers a
full barrier prior to the assignment.
Therefore we may remove the explicit wmb.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit d3716f18a7.
vmalloc cannot be used in BH disabled contexts, even
with GFP_ATOMIC. And we certainly want to support
rhashtable users inserting entries with software
interrupts disabled.
Signed-off-by: David S. Miller <davem@davemloft.net>
When an rhashtable user pounds rhashtable hard with back-to-back
insertions we may end up growing the table in GFP_ATOMIC context.
Unfortunately when the table reaches a certain size this often
fails because we don't have enough physically contiguous pages
to hold the new table.
Eric Dumazet suggested (and in fact wrote this patch) using
__vmalloc instead which can be used in GFP_ATOMIC context.
Reported-by: Phil Sutter <phil@nwl.cc>
Suggested-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Thomas and Phil observed that under stress rhashtable insertion
sometimes failed with EBUSY, even though this error should only
ever been seen when we're under attack and our hash chain length
has grown to an unacceptable level, even after a rehash.
It turns out that the logic for detecting whether there is an
existing rehash is faulty. In particular, when two threads both
try to grow the same table at the same time, one of them may see
the newly grown table and thus erroneously conclude that it had
been rehashed. This is what leads to the EBUSY error.
This patch fixes this by remembering the current last table we
used during insertion so that rhashtable_insert_rehash can detect
when another thread has also done a resize/rehash. When this is
detected we will give up our resize/rehash and simply retry the
insertion with the new table.
Reported-by: Thomas Graf <tgraf@suug.ch>
Reported-by: Phil Sutter <phil@nwl.cc>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Tested-by: Phil Sutter <phil@nwl.cc>
Signed-off-by: David S. Miller <davem@davemloft.net>
rhashtable_rehash_one() uses complex logic to update entry->next field,
after INIT_RHT_NULLS_HEAD and NULLS_MARKER expansion:
entry->next = 1 | ((base + off) << 1)
This can be compiled along the lines of:
entry->next = base + off
entry->next <<= 1
entry->next |= 1
Which will break concurrent readers.
NULLS value recomputation is not needed here, so just remove
the complex logic.
The data race was found with KernelThreadSanitizer (KTSAN).
Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
If rhashtable_walk_next detects a resize operation in progress, it jumps
to the new table and continues walking that one. But it misses to drop
the reference to it's current item, leading it to continue traversing
the new table's bucket in which the current item is sorted into, and
after reaching that bucket's end continues traversing the new table's
second bucket instead of the first one, thereby potentially missing
items.
This fixes the rhashtable runtime test for me. Bug probably introduced
by Herbert Xu's patch eddee5ba ("rhashtable: Fix walker behaviour during
rehash") although not explicitly tested.
Fixes: eddee5ba ("rhashtable: Fix walker behaviour during rehash")
Signed-off-by: Phil Sutter <phil@nwl.cc>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
rhashtable uses EXPORT_SYMBOL_GPL() without importing linux/export.h
directly it is only imported indirectly through some other includes.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/cadence/macb.c
drivers/net/phy/phy.c
include/linux/skbuff.h
net/ipv4/tcp.c
net/switchdev/switchdev.c
Switchdev was a case of RTNH_H_{EXTERNAL --> OFFLOAD}
renaming overlapping with net-next changes of various
sorts.
phy.c was a case of two changes, one adding a local
variable to a function whilst the second was removing
one.
tcp.c overlapped a deadlock fix with the addition of new tcp_info
statistic values.
macb.c involved the addition of two zyncq device entries.
skbuff.h involved adding back ipv4_daddr to nf_bridge_info
whilst net-next changes put two other existing members of
that struct into a union.
Signed-off-by: David S. Miller <davem@davemloft.net>
We currently have no limit on the number of elements in a hash table.
This is a problem because some users (tipc) set a ceiling on the
maximum table size and when that is reached the hash table may
degenerate. Others may encounter OOM when growing and if we allow
insertions when that happens the hash table perofrmance may also
suffer.
This patch adds a new paramater insecure_max_entries which becomes
the cap on the table. If unset it defaults to max_size * 2. If
it is also zero it means that there is no cap on the number of
elements in the table. However, the table will grow whenever the
utilisation hits 100% and if that growth fails, you will get ENOMEM
on insertion.
As allowing oversubscription is potentially dangerous, the name
contains the word insecure.
Note that the cap is not a hard limit. This is done for performance
reasons as enforcing a hard limit will result in use of atomic ops
that are heavier than the ones we currently use.
The reasoning is that we're only guarding against a gross over-
subscription of the table, rather than a small breach of the limit.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove useless obj variable and goto logic.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current code currently only stops inserting rehashes into the
chain when no resizes are currently scheduled. As long as resizes
are scheduled and while inserting above the utilization watermark,
more and more rehashes will be scheduled.
This lead to a perfect DoS storm with thousands of rehashes
scheduled which lead to thousands of spinlocks to be taken
sequentially.
Instead, only allow either a series of resizes or a single rehash.
Drop any further rehashes and return -EBUSY.
Fixes: ccd57b1bd3 ("rhashtable: Add immediate rehash during insertion")
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
When rhashtable_insert_rehash() fails with ENOMEM, this indicates that
we can't allocate the necessary memory in the current context but the
limits as set by the user would still allow to grow.
Thus attempt an async resize in the background where we can allocate
using GFP_KERNEL which is more likely to succeed. The insertion itself
will still fail to indicate pressure.
This fixes a bug where the table would never continue growing once the
utilization is above 100%.
Fixes: ccd57b1bd3 ("rhashtable: Add immediate rehash during insertion")
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
nftables sets will be converted to use so called setextensions, moving
the key to a non-fixed position. To hash it, the obj_hashfn must be used,
however it so far doesn't receive the length parameter.
Pass the key length to obj_hashfn() and convert existing users.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
rhashtable_destroy() variant which stops rehashes, iterates over
the table and calls a callback to release resources.
Avoids need for nft_hash to embed rhashtable internals and allows to
get rid of the being_destroyed flag. It also saves a 2nd mutex
lock upon destruction.
Also fixes an RCU lockdep splash on nft set destruction due to
calling rht_for_each_entry_safe() without holding bucket locks.
Open code this loop as we need know that no mutations may occur in
parallel.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce a new bool automatic_shrinking to require the
user to explicitly opt-in to automatic shrinking of tables.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a comment on the choice of the value 16 as the
maximum chain length before we force a rehash.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit 963ecbd41a ("rhashtable:
Fix use-after-free in rhashtable_walk_stop") fixed a real bug
but created another one because we may end up sleeping inside an
RCU critical section.
This patch fixes it properly by replacing the mutex with a spin
lock that specifically protects the walker lists.
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch reintroduces immediate rehash during insertion. If
we find during insertion that the table is full or the chain
length exceeds a set limit (currently 16 but may be disabled
with insecure_elasticity) then we will force an immediate rehash.
The rehash will contain an expansion if the table utilisation
exceeds 75%.
If this rehash fails then the insertion will fail. Otherwise the
insertion will be reattempted in the new hash table.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds the ability to allocate bucket table with GFP_ATOMIC
instead of GFP_KERNEL. This is needed when we perform an immediate
rehash during insertion.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds the missing bits to allow multiple rehashes. The
read-side as well as remove already handle this correctly. So it's
only the rehasher and insertion that need modification to handle
this.
Note that this patch doesn't actually enable it so for now rehashing
is still only performed by the worker thread.
This patch also disables the explicit expand/shrink interface because
the table is meant to expand and shrink automatically, and continuing
to export these interfaces unnecessarily complicates the life of the
rehasher since the rehash process is now composed of two parts.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch changes rhashtable_shrink to shrink to the smallest
size possible rather than halving the table. This is needed
because with multiple rehashing we will defer shrinking until
all other rehashing is done, meaning that when we do shrink
we may be able to shrink a lot.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since every current rhashtable user uses jhash as their hash
function, the fact that jhash is an inline function causes each
user to generate a copy of its code.
This function provides a solution to this problem by allowing
hashfn to be unset. In which case rhashtable will automatically
set it to jhash. Furthermore, if the key length is a multiple
of 4, we will switch over to jhash2.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
The walker is a lockless reader so it too needs an smp_rmb before
reading the future_tbl field in order to see any new tables that
may contain elements that we should have walked over.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that all rhashtable users have been converted over to the
inline interface, this patch removes the unused out-of-line
interface.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch deals with the complaint that we make indirect function
calls on the fast paths unnecessarily in rhashtable. We resolve
it by moving the fast paths into inline functions that take struct
rhashtable_param (which obviously must be the same set of parameters
supplied to rhashtable_init) as an argument.
The only remaining indirect call is to obj_hashfn (or key_hashfn it
obj_hashfn is unset) on the rehash as well as the insert-during-
rehash slow path.
This patch also extends the support of vairable-length keys to
include those where the key is fixed but scattered in the object.
For example, in netlink we want to key off the namespace and the
portid but they're not next to each other.
This patch does this by directly using the object hash function
as the indicator of whether the key is accessible or not. It
also adds a new function obj_cmpfn to compare a key against an
object. This means that the caller no longer needs to supply
explicit compare functions.
All this is done in a backwards compatible manner so no existing
users are affected until they convert to the new interface.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch marks the rhashtable_init params argument const as
there is no reason to modify it since we will always make a copy
of it in the rhashtable.
This patch also fixes a bug where we don't actually round up the
value of min_size unless it is less than HASH_MIN_SIZE.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Round up min_size respectively round down max_size to the next power
of two to make sure we always respect the limit specified by the
user. This is required because we compare the table size against the
limit before we expand or shrink.
Also fixes a minor bug where we modified min_size in the params
provided instead of the copy stored in struct rhashtable.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that nobody uses max_shift and min_shift, we can safely remove
them.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds the parameters max_size and min_size which are
meant to replace max_shift and min_shift.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Keeping both size and shift is silly. We only need one.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Caching the lock pointer avoids having to hash on the object
again to unlock the bucket locks.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes the following sparse warnings:
lib/rhashtable.c:767:5: warning: context imbalance in 'rhashtable_walk_start' - wrong count at exit
lib/rhashtable.c:849:6: warning: context imbalance in 'rhashtable_walk_stop' - unexpected unlock
Fixes: f2dba9c6ff ("rhashtable: Introduce rhashtable_walk_*")
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit 9d901bc051 ("rhashtable:
Free bucket tables asynchronously after rehash") causes gratuitous
failures in rhashtable_remove.
The reason is that it inadvertently introduced multiple rehashing
from the perspective of readers. IOW it is now possible to see
more than two tables during a single RCU critical section.
Fortunately the other reader rhashtable_lookup already deals with
this correctly thanks to c4db8848af
("rhashtable: rhashtable: Move future_tbl into struct bucket_table")
so only rhashtable_remove is broken by this change.
This patch fixes this by looping over every table from the first
one to the last or until we find the element that we were trying
to delete.
Incidentally the simple test for detecting rehashing to prevent
starting another shrinking no longer works. Since it isn't needed
anyway (the work queue and the mutex serves as a natural barrier
to unnecessary rehashes) I've simply killed the test.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit c4db8848af ("rhashtable:
Move future_tbl into struct bucket_table") introduced a use-after-
free bug in rhashtable_walk_stop because it dereferences tbl after
droping the RCU read lock.
This patch fixes it by moving the RCU read unlock down to the bottom
of rhashtable_walk_stop. In fact this was how I had it originally
but it got dropped while rearranging patches because this one
depended on the async freeing of bucket_table.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch moves future_tbl to open up the possibility of having
multiple rehashes on the same table.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a rehash counter to bucket_table to indicate
the last bucket that has been rehashed. This serves two purposes:
1. Any bucket that has been rehashed can never gain a new object.
2. If the rehash counter reaches the size of the table, the table
will forever remain empty.
This patch also downsizes bucket_table->size to an unsigned int
since we do not support sizes greater than 32 bits yet.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is in fact no need to wait for an RCU grace period in the
rehash function, since all insertions are guaranteed to go into
the new table through spin locks.
This patch uses call_rcu to free the old/rehashed table at our
leisure.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
It seems that I have already made every rehash redo the random
seed even though my commit message indicated otherwise :)
Since we have already taken that step, this patch goes one step
further and moves the seed initialisation into bucket_table_alloc.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
We only nest one level deep there is no need to roll our own
subclasses.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously whenever the walker encountered a resize it simply
snaps back to the beginning and starts again. However, this only
works if the rehash started and completed while the walker was
idle.
If the walker attempts to restart while the rehash is still ongoing,
we may miss objects that we shouldn't have.
This patch fixes this by making the walker walk the old table
followed by the new table just like all other readers. If a
rehash is detected we will still signal our caller of the fact
so they can prepare for duplicates but we will simply continue
the walk onto the new table after the old one is finished either
by us or by the rehasher.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>