Move replace code out from _cds_lfht_add()

Original patch from Lai Jiangshan, reworked significantly to keep the
iterator semantic, where a "next" node is never re-read after being
looked up within a RCU read-side critical section after it has been
validated.

Although strictly speaking the "replace" code re-checks that the next
node is not logically removed with a cmpxchg, I prefer to ensure that
all lookup paths behave similarly. It will make it easier to track
behavior later on if they all follow the same rules.

Reported-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Rename the parameter name of _cds_lfht_replace()

Although _cds_lfht_replace() is not a plublic API,
but a proper name will help for reading or maintainance.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix dummy node insertion bug

dummy node is the first node of the identical-hash-value chain

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

compare hash value before call compare_fct()

[ Edit by Mathieu Desnoyers:

  This is an optimisation that checks if the reverse hash value is
  equal before calling the comparison function. Given comparing a
  reverse hash is much faster than comparison, this accelarates the
  lookups and add. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: Simplify lookup_bucket()

They are the same, but I don't think the compiler can optimize it.

And it also helps for understanding the following code.

[ Edit by Mathieu Desnoyers:
  - Add a comment that describes the equivalence between get_count_order
    and fls for this lookup of index + 1. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: avoid unneed garbage collect

[ Edit by Mathieu Desnoyers:

If we have a concurrent removal executing while the add is performed,
I just want to ensure we agree that this scenario will be correct:

1) removal succeeds. We have flagged
   a node as "logically removed", but
   it is still in the linked list.

2) garbage collection is initiated.

                                        3) add succeeds. We insert a
                                           node with reverse hash value
                                           higher than the logically
                                           removed node. This means the
                                           add takes care of garbage
                                           collecting the logically
                                           removed node prior to adding
                                           the new node.

4) garbage collection of logically
   removed node fails
   (we ignore the return value of
    (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);)
   retry and iterate until we find the reverse
   hash value higher than the logically
   remove node.

So I think this is OK: given that the add path takes care of garbage
collecting any logically removed node located prior to the insert
position, we don't need to perform gargage collection after the
insertion. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: merge duplicated code for bucket lookup

[ Edit by Mathieu Desnoyers: also remove unnecessary lookups that end up
  returning an already looked up node. Same behavior is guaranteed
  because the hash and size stay invariant across retry. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: make struct rcu_level size power of 2

By adding a small slowpath overhead (added synchronize_rcu call in the
last iteration of the resize), we can reduce the amount of wasted memory
for memory allocators that allocate power of two memory areas. This is
achieved by removing the call_rcu head from struct rcu_level.

[ Edit by Mathieu Desnoyers:
  - add comment about need to manually update the allocation size of
    fields are added to struct rcu_level.
  - create a more explanatory title and changelog. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Reinsert missing test_urcu_*.c files (missing in rename)

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: simplify get_count_order()

[ Edit by Mathieu Desnoyers:

- Updated comments.
- Quick benchmark on Intel(R) Atom(TM) CPU Z520   @ 1.33GHz:

get_count_order_ulong, from 0 to 100000000:

previous:  3.840s
new:       3.187s

- Test comparing bit-exactness for ranges near 0:

/*
 * fls: returns the position of the most significant bit.
 * Returns 0 if no bit is set, else returns the position of the most
 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
 */
static inline
unsigned int fls_u32(uint32_t x)
{
int r;

asm("bsrl %1,%0\n\t"
    "jnz 1f\n\t"
    "movl $-1,%0\n\t"
    "1:\n\t"
    : "=r" (r) : "rm" (x));
return r + 1;
}

static inline
unsigned int fls_u64(uint64_t x)
{
long r;

asm("bsrq %1,%0\n\t"
    "jnz 1f\n\t"
    "movq $-1,%0\n\t"
    "1:\n\t"
    : "=r" (r) : "rm" (x));
return r + 1;
}

static __attribute__((unused))
unsigned int fls_u64(uint64_t x)
{
unsigned int r = 64;

if (!x)
return 0;

if (!(x & 0xFFFFFFFF00000000ULL)) {
x <<= 32;
r -= 32;
}
if (!(x & 0xFFFF000000000000ULL)) {
x <<= 16;
r -= 16;
}
if (!(x & 0xFF00000000000000ULL)) {
x <<= 8;
r -= 8;
}
if (!(x & 0xF000000000000000ULL)) {
x <<= 4;
r -= 4;
}
if (!(x & 0xC000000000000000ULL)) {
x <<= 2;
r -= 2;
}
if (!(x & 0x8000000000000000ULL)) {
x <<= 1;
r -= 1;
}
return r;
}

static __attribute__((unused))
unsigned int fls_u32(uint32_t x)
{
unsigned int r = 32;

if (!x)
return 0;
if (!(x & 0xFFFF0000U)) {
x <<= 16;
r -= 16;
}
if (!(x & 0xFF000000U)) {
x <<= 8;
r -= 8;
}
if (!(x & 0xF0000000U)) {
x <<= 4;
r -= 4;
}
if (!(x & 0xC0000000U)) {
x <<= 2;
r -= 2;
}
if (!(x & 0x80000000U)) {
x <<= 1;
r -= 1;
}
return r;
}

unsigned int fls_ulong(unsigned long x)
{
return fls_u32(x);
return fls_u64(x);
}

int get_count_order_u32(uint32_t x)
{
int order;

order = fls_u32(x) - 1;
if (x & (x - 1))
order++;
return order;
}

int get_count_order_ulong(unsigned long x)
{
int order;

order = fls_ulong(x) - 1;
if (x & (x - 1))
order++;
return order;
}

int test_get_count_order_u32(uint32_t x)
{
if (!x)
return -1;
return fls_u32(x - 1);
}

/* find the minimum order that x <= (1UL << order) */
int test_get_count_order_ulong(unsigned long x)
{
if (!x)
return -1;
return fls_ulong(x - 1);
}

int main(int argc, char **argv)
{
unsigned long i;

for (i = 0UL; i < 10000; i++) {
if (get_count_order_ulong(i) != test_get_count_order_ulong(i))
printf("Error for %lu\n", i);
}

for (i = 4294967293UL; i != 0; i++) {
if (get_count_order_ulong(i) != test_get_count_order_ulong(i))
printf("Error for %lu\n", i);
}

return 0;
}

]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: use dbg_printf() for grow/shrink printout

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: merge thread_id into struct partition_resize_work

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: remove unused rcu_head in partition_resize_work

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: handle write return value

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rcuhash test: fix 32-bit type warning

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

urcu-pointer: fix rcu_set_pointer unset return value

The problem only affected non-_LGPL_SOURCE configs.

Reported-by: Stephen Hemminger <shemminger@vyatta.com>
Fix-proposed-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

Conflicts:
urcu-call-rcu.h

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Enhance API.txt documentation, add to Makefile as EXTRA_DIST

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: do not sample in_progress_destroy in the middle of a level

Caused destroy assert/segfaults.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: break resize loop on destroy

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: decrement resize cnt on destroy

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Check for in progress destroy before calling call_rcu thread

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: Test for initiated destroy before performing resize

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Version 0.6.5

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

call_rcu: Document call_rcu requirements

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

call_rcu: Document call_rcu requirements

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

call_rcu: fix error handling of malloc error

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

urcu call_rcu: Use RCU read-side protection for per-cpu call_rcu data

A concurrent get_cpu_call_rcu_data(), called by get_call_rcu_data(),
could dereference this pointer without holding any mutex. So this
situation would happen if we have a concurrent call_rcu() executing
while we do the create_all_cpu_call_rcu_data().

I think we would need to put a rcu_dereference() around
per_cpu_call_rcu_data read within get_cpu_call_rcu_data() too.
per_cpu_call_rcu_data should be done with rcu_set_pointer.

Also, a rcu read-side critical section would be required around any
usage of per_cpu_call_rcu_data, and the action of tearing down the
per-cpu data would require to wait for a quiescent state. So we would
basically require that the call_rcu users need to be registered as
RCU reader threads.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu,call_rcu: Cleanup call_rcu_data pointers before use in child

[ Edit by Mathieu Desnoyers: create maxcpus_reset to handle cases where
  maxcpus is 0 and -1, depending on the configuration. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu,call_rcu: avoid create call_rcu_data for child when unneed

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu,defer_rcu: Make defer_rcu encoding more compact for marker

When the function changes (and the function is aligned), and only the
data is the marker, we can get away with using only 2 pointers rather
than 3.

[ Edit by Mathieu Desnoyers: patch cleanup, changelog updates ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu_defer: Use cancellation flag instead of pthread_cancel()

- Provides better control over cancellation point location.
- Set the futex to 0 before exiting the defer thread.

This patch combines and enhances patches from Lai Jiangshan:
  urcu,defer_rcu: fix missing respond to a cancellation request
  urcu,defer_rcu: Avoid thread exit unexpected

Reported-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu,call_rcu: protects call_rcu_data_list when remove node

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

Create default call rcu data upon per-cpu call-rcu teardown

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: fix 32-bit hash

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: Use get first/get next to delete all entries

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: fix get first / get next iterator

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix handling of systems without sysconf nr possible cpu support

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash needs local config.h

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: update header documentation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: move init node outside of rcu read-side c.s. (unneeded protection)

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add cds_lfht_first/cds_lfht_next for hash table iteration

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: rename _next into _next_duplicate

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: document use of caa_container_of()

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: cleanup includes

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

powerpc: use __NO_LWSYNC__ check to use appropriate lwsync/sync opcode

We already used it in uatomic code, move it to arch ppc.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

cmm: provide lightweight smp_rmb/smp_wmb on PPC

lwsync orders loads in cacheable memory with respect to other loads,
and stores in cacheable memory with respect to other stores.  Use it
to implement smp_rmb/smp_wmb.

The heavy-weight sync is still used for the "full" rmb/wmb operations,
as well as for smp_mb.

[ Edit by Mathieu Desnoyers: rephrased the comments around the memory
  barriers. ]

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: factor out add_replace and replace

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

atomic: provide seq_cst semantics on powerpc

We provide sequential consistency semantic over all architectures for
cmpxchg and add_return family of primitives, but the powerpc
implementation does not match that.

Change the isync after the atomic primitives to sync, and explain the
scheme.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash tests: make node count RCU aware

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: set next to NULL when node is NULL

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: approximation can be negative

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: min size only needed on shrink, take nr cpus into account

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: type the ht count approx as long

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: handle small and negative table size approximation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: fix node approx counting

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: output approximation of number of nodes in counting

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash cleanup: count percpu deletes in the positive range

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: node alignment is back to 4 bytes

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: remove now unneeded gc flag (combine with removed)

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: implement lock-free replacement

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: support replacement operation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: API rename: remove -> del

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: spawn only number of threads required for resize

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: parallelize resize

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: document flags

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: comment file inclusion order

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: simplify lfht_new API with inline wrapper

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: add TODO for resize worker threads

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: Remove leftover assertions

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: merge table hash and link stages into populate

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'urcu/ht-shrink-help' into urcu/ht-shrink

rculfhash: remove helper scheme

There is a trade-off to consider here:
- The helper scheme would require the helpers to allocate the memory for
  dummy nodes themself for correctness (the current implementation is
  buggy because lookups consider an half-linked dummy node to be ready for
  use), which does not work with a cache-efficient per-level array of
  dummy nodes.
- We want cache-efficiency, so we want to keep the per-level array
  allocation.

Therefore, letting insert/removal/lookups help expand is not the way to
go here. This patch removes the helping scheme.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

avoid leaking crdp for failed path

[ Comment: now that set_cpu_call_rcu_data() is not racy and detects
  overwrites, we can effectively trust its return value and free the
  crdp if already set. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Return -EEXIST when the old cpu call_rcu_data has not been removed

To make it matches the comments.

It is the caller's responsibility to use
set_cpu_call_rcu_data(cpu, NULL) to remove the CPU's
call_rcu_data structure and dispose it.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

protect writing to per_cpu_call_rcu_data[*]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

wake up default call_rcu thread after we move the leftover callbacks

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

avoid memory leak in call_rcu_data_free()

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu call_rcu: fix use after free()

call_rcu_after_fork_child() needs to use cds_list_for_each_entry_safe to
safely iterate on the list as its item is being freed.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

use get_cpu_call_rcu_data() for get_call_rcu_data()

[ Impact: refactor duplicated code ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

init maxcpus before use

[ Edit:
  Covers the per-cpu call_rcu data setup (not all_cpus helper, which is
  why we did not trigger it in our tests. ]

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

call_rcu implementation: add missing static

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

call_rcu implementation: add missing static

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Document QSBR interaction with mutexes

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Document QSBR interaction with mutexes

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: exit upon failure

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: fix add_unique node counting

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: cast rand_r return value directly

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: add verbosity

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: allow different size for lookup, write, init

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

qsbr vs call_rcu : remove exit assertion

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: validate lookups

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: add pool offsets

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: with resize stop and helpers, min size can now be 1

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>