Followup of commit 634a4b20
Allow tnode_get_child_rcu() to be called either under rcu_read_lock()
protection or with RTNL held.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Allow fib_find_node() to be called either under rcu_read_lock()
protection or with RTNL held.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The FIB algorithim for IPV4 is set at compile time, but kernel goes through
the overhead of function call indirection at runtime. Save some
cycles by turning the indirect calls to direct calls to either
hash or trie code.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Here is rework and cleanup of the resize function.
Some bugs we had. We were using ->parent when we should use
node_parent(). Also we used ->parent which is not assigned by
inflate in inflate loop.
Also a fix to set thresholds to power 2 to fit halve
and double strategy.
max_resize is renamed to max_work which better indicates
it's function.
Reaching max_work is not an error, so warning is removed.
max_work only limits amount of work done per resize.
(limits CPU-usage, outstanding memory etc).
The clean-up makes it relatively easy to add fixed sized
root-nodes if we would like to decrease the memory pressure
on routers with large routing tables and dynamic routing.
If we'll need that...
Its been tested with 280k routes.
Work done together with Robert Olsson.
Signed-off-by: Jens Låås <jens.laas@its.uu.se>
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
String literals are constant, and usually, we can also tag the array
of pointers const too, moving it to the .rodata section.
Signed-off-by: Jan Engelhardt <jengelh@medozas.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
While looking for other fib_trie problems reported by Pawel Staszewski
I noticed there are a few uses of tnode_get_child() and node_parent()
in lookups instead of their rcu versions.
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
During large updates there could be triggered warnings like: "Fix
inflate_threshold_root. Now=25 size=11 bits" if inflate() of the root
node isn't finished in 10 loops. It should be much rarer now, after
changing the threshold from 15 to 25, and a temporary problem, so
this patch tries to handle it automatically using a fix variable to
increase by one inflate threshold for next root resizes (up to the 35
limit, max fix = 10). The fix variable is decreased when root's
inflate() finishes below 7 loops (even if some other, smaller table/
trie is updated -- for simplicity the fix variable is global for now).
Reported-by: Pawel Staszewski <pstaszewski@itcare.pl>
Reported-by: Jorge Boncompte [DTI2] <jorge@dti2.net>
Tested-by: Pawel Staszewski <pstaszewski@itcare.pl>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
During trie_rebalance() we free memory after resizing with call_rcu(),
but large updates, especially with PREEMPT_NONE configs, can cause
memory stresses, so this patch calls synchronize_rcu() in
tnode_free_flush() after each sync_pages to guarantee such freeing
(especially before resizing the root node).
The value of sync_pages = 128 is based on Pawel Staszewski's tests as
the lowest which doesn't hinder updating times. (For testing purposes
there was a sysfs module parameter to change it on demand, but it's
removed until we're sure it could be really useful.)
The patch is based on suggestions by: Paul E. McKenney
<paulmck@linux.vnet.ibm.com>
Reported-by: Pawel Staszewski <pstaszewski@itcare.pl>
Tested-by: Pawel Staszewski <pstaszewski@itcare.pl>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pawel Staszewski wrote:
<blockquote>
Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648
and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back
dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits
...
Fix inflate_threshold_root. Now=15 size=11 bits
cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes.
Main:
Aver depth: 2.28
Max depth: 6
Leaves: 276539
Prefixes: 289922
Internal nodes: 66762
1: 35046 2: 13824 3: 9508 4: 4897 5: 2331 6: 1149 7: 5
9: 1 18: 1
Pointers: 691228
Null ptrs: 347928
Total size: 35709 kB
</blockquote>
It seems, the current threshold for root resizing is too aggressive,
and it causes misleading warnings during big updates, but it might be
also responsible for memory problems, especially with non-preempt
configs, when RCU freeing is delayed long after call_rcu.
It should be also mentioned that because of non-atomic changes during
resizing/rebalancing the current lookup algorithm can miss valid leaves
so it's additional argument to shorten these activities even at a cost
of a minimally longer searching.
This patch restores values before the patch "[IPV4]: fib_trie root
node settings", commit: 965ffea43d from
v2.6.22.
Pawel's report:
<blockquote>
I dont see any big change of (cpu load or faster/slower
routing/propagating routes from bgpd or something else) - in avg there
is from 2% to 3% more of CPU load i dont know why but it is - i change
from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL
1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt"
[...]
cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes.
Main:
Aver depth: 2.44
Max depth: 6
Leaves: 277814
Prefixes: 291306
Internal nodes: 66420
1: 32737 2: 14850 3: 10332 4: 4871 5: 2313 6: 942 7: 371 8: 3 17: 1
Pointers: 599098
Null ptrs: 254865
Total size: 18067 kB
</blockquote>
According to this and other similar reports average depth is slightly
increased (~0.2), and root nodes are shorter (log 17 vs. 18), but
there is no visible performance decrease. So, until memory handling is
improved or added parameters for changing this individually, this
patch resets to safer defaults.
Reported-by: Pawel Staszewski <pstaszewski@itcare.pl>
Reported-by: Jorge Boncompte [DTI2] <jorge@dti2.net>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Tested-by: Pawel Staszewski <pstaszewski@itcare.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Alas current delaying of freeing old tnodes by RCU in trie_rebalance
is still not enough because we can free a top tnode before updating a
t->trie pointer.
Reported-by: Pawel Staszewski <pstaszewski@itcare.pl>
Tested-by: Pawel Staszewski <pstaszewski@itcare.pl>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
My previous patch, which explicitly delays freeing of tnodes by adding
them to the list to flush them after the update is finished, isn't
strict enough. It treats exceptionally tnodes without parent, assuming
they are newly created, so "invisible" for the read side yet.
But the top tnode doesn't have parent as well, so we have to exclude
all exceptions (at least until a better way is found). Additionally we
need to move rcu assignment of this node before flushing, so the
return type of the trie_rebalance() function is changed.
Reported-by: Yan Zheng <zheng.yan@oracle.com>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While doing trie_rebalance(): resize(), inflate(), halve() RCU free
tnodes before updating their parents. It depends on RCU delaying the
real destruction, but if RCU readers start after call_rcu() and before
parent update they could access freed memory.
It is currently prevented with preempt_disable() on the update side,
but it's not safe, except maybe classic RCU, plus it conflicts with
memory allocations with GFP_KERNEL flag used from these functions.
This patch explicitly delays freeing of tnodes by adding them to the
list, which is flushed after the update is finished.
Reported-by: Yan Zheng <zheng.yan@oracle.com>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It seems we can fix this by disabling preemption while we re-balance the
trie. This is with the CONFIG_CLASSIC_RCU. It's been stress-tested at high
loads continuesly taking a full BGP table up/down via iproute -batch.
Note. fib_trie is not updated for CONFIG_PREEMPT_RCU
Reported-by: Andrei Popa
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using NIPQUAD() with NIPQUAD_FMT, %d.%d.%d.%d or %u.%u.%u.%u
can be replaced with %pI4
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are already 7 of them - time to kill some duplicate code.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit a07f5f508a "[IPV4] fib_trie: style
cleanup", the changes to check_leaf() and fn_trie_lookup() were wrong - where
fn_trie_lookup() would previously return a negative error value from
check_leaf(), it now returns 0.
Now fn_trie_lookup() doesn't appear to care about plen, so we can revert
check_leaf() to returning the error value.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Tested-by: William Boughton <bill@boughton.de>
Acked-by: Stephen Heminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes CVS keywords that weren't updated for a long time
from comments.
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Plan C: we can follow the Al Viro's proposal about %n like in this patch.
The same applies to udp, fib (the /proc/net/route file), rt_cache and
sctp debug. This is minus ~150-200 bytes for each.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Avoid unneeded test in the case where object to be freed
has to be a leaf. Don't need to use the generic tnode_free()
function, instead just setup leaf to be freed.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The trie pointer is passed down to flush_list and flush_leaf
but never used.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use vmalloc rather than alloc_pages to avoid wasting memory.
The problem is that tnode structure has a power of 2 sized array,
plus a header. So the current code wastes almost half the memory
allocated because it always needs the next bigger size to hold
that small header.
This is similar to an earlier patch by Eric, but instead of a list
and lock, I used a workqueue to handle the fact that vfree can't
be done in interrupt context.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Without CONFIG_NET_NS, no namespace other than &init_net exists,
no need to store net in seq_net_private.
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Make /proc/net/fib_trie and /proc/net/fib_triestat display
all routing tables, not just local and main.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This gets rid of a warning caused by the test in rcu_assign_pointer.
I tried to fix rcu_assign_pointer, but that devolved into a long set
of discussions about doing it right that came to no real solution.
Since the test in rcu_assign_pointer for constant NULL would never
succeed in fib_trie, just open code instead.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use key/offset caching to change /proc/net/route (use by iputils route)
from O(n^2) to O(n). This improves performance from 30sec with 160,000
routes to 1sec.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes possible problems when trie_firstleaf() returns NULL
to trie_leafindex().
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The line in the /proc/net/fib_trie for route with TOS specified
- has extra \n at the end
- does not have a space after route scope
like below.
|-- 1.1.1.1
/32 universe UNICASTtos =1
Signed-off-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Normally during a dump the key of the last dumped entry is used for
continuation, but since lock is dropped it might be lost. In that case
fallback to the old counter based N^2 behaviour. This means the dump
will end up skipping some routes which matches what FIB_HASH does.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update fib_trie with some fib_hash fixes:
- check for duplicate alternative routes for prefix+tos+priority when
replacing route
- properly insert by matching tos together with priority
- fix alias walking to use list_for_each_entry_continue for insertion
and deletion when fa_head is not NULL
- copy state from fa to new_fa on replace (not a problem for now)
- additionally, avoid replacement without error if new route is same,
as Joonwoo Park suggests.
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since fib_route_seq_show now uses hlist_for_each_entry(), the leaf
info can not be NULL.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This converts dumping (and flushing) of large route tables form O(N^2)
to O(N). If the route dump took multiple pages then the dump routine
gets called again. The old code kept track of location by counter, the
new code instead uses the last key.
This is a really big win ( 0.3 sec vs 12 sec) for big route tables.
One side effect is that if the table changes during the dump, then the
last key will not be found, and we will return -EBUSY.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Get rid of extra search that made route deletion O(n).
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is easier with TRIE to dump the data traversal rather than
interating over every possible prefix. This saves some time and makes
the dump come out in sorted order.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove the complex loop structure of nextleaf() and replace it with a
simpler tree walker. This improves the performance and is much
cleaner.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Match fib_hash, and set NLM_F_MULTI to handle multiple part messages.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The code to dump can use the existing hash chain rather than doing
repeated lookup.
Signed-off-by: Stephen Hemminger <stephen.hemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Compute the number of prefixes when needed, rather than doing bookeeping.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Style cleanups:
* make check_leaf return -1 or plen, rather than by reference
* Get rid of #ifdef that is always set
* split out embedded function calls in if statements.
* checkpatch warnings
Signed-off-by: Stephen Hemminger <stephen.hemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This improves locality for operations that touch all the leaves. Save
space since these entries don't need to be hardware cache aligned.
Signed-off-by: Stephen Hemminger <stephen.hemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
node_parent() and tnode_get_child() currently use rcu_dereference().
These functions are called from both
- readers only paths (where rcu_dereference() is needed), and
- writer path (where rcu_dereference() is not needed)
To make explicit where rcu_dereference() is really needed, I
introduced new node_parent_rcu() and tnode_get_child_rcu() functions
which use rcu_dereference(), while node_parent() and tnode_get_child()
dont use it.
Then I changed calling sites where rcu_dereference() was really needed
to call the _rcu() variants.
This should have no impact but for alpha architecture, and may help
future sparse checks.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Initialization of the slab cache's should be done when IP is
initialized to make sure of available memory, and that code can be
marked __init.
Signed-off-by: Stephen Hemminger <stephen.hemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
