Age | Commit message (Collapse) | Author | Files | Lines |
|
In a case where one pounds on a single kvp in a KVP_AT_BUCKET_LEVEL
table, the code would sporadically return a transitional value (junk)
from a half-deleted kvp. At most, 64-bits worth of the kvp will be
written atomically, so using memset(...) to smear 0xFF's across a kvp
to free it left a lot to be desired.
Performance impact: very mild positive, thanks to FC for doing a
multi-thread host stack perf/scale test.
Added an ASSERT to catch attempts to add a (key,value) pair which
contains the magic "free kvp" value.
Type: fix
Signed-off-by: Dave Barach <dave@barachs.net>
Change-Id: I6a1aa8a2c30bc70bec4b696ce7b17c2839927065
|
|
Use of _vec_len() to set vector length breaks address sanitizer.
Users should use vec_set_len(), vec_inc_len(), vec_dec_len () instead.
Type: improvement
Change-Id: I441ae948771eb21c23a61f3ff9163bdad74a2cb8
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: fix
Signed-off-by: Dave Barach <dave@barachs.net>
Change-Id: Icce7eab4510785e15bdcf97e4d1881b0f46f6899
|
|
Template instances can allocate BIHASH_KVP_PER_PAGE data records
tangent to the bucket, to remove a dependent read / prefetch.
Template instances can ask for immediate memory allocation, to avoid
several branches in the lookup path.
Clean up l2 fib, gpb plugin codes: use clib_bihash_get_bucket(...)
Use hugepages for bihash allocation arenas
Type: improvement
Signed-off-by: Dave Barach <dave@barachs.net>
Signed-off-by: Damjan Marion <damarion@cisco.com>
Change-Id: I92fc11bc58e48d84e2d61f44580916dd1c56361c
|
|
Move the binary api segment above 4gb
Change-Id: I40e8aa7a97722a32397f5a538b5ff8344c50d408
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Applications such as NAT that dynamically create entries require these entries to expire after some time.
Bihash user can now lazily delete expired entries. When inserting and bucket is full, expired entry is overwritten.
Change-Id: I6852305df399b546159407f1729c856afde5a634
Signed-off-by: Matus Fabian <matfabia@cisco.com>
|
|
Add a bucket-level lock bit. Use a spinlock only when actually
allocating, freeing, or splitting a bucket. Should improve
multi-thread add/del performance.
Change-Id: I3e40e2a8371685457f340d6584dea14e3207f2b0
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Directly allocate and carve cache-line-aligned chunks of virtual
memory. To a first approximation, bihash wasn't using
clib_mem_free(...).
We eliminate mheap object header/trailers, which improves space
efficiency. We also eliminate the 4gb bihash table size limit. An 8_8
bihash w/ 100 million random entries uses 3.8 Gbytes.
Change-Id: Icf925fdf99bce7d6ac407ac4edd30560b8f04808
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Reference-count the number of entries in each bucket. If the reference
count goes to zero, free the backing store.
Add long-term churn-testing to test_bihash_template.c, thanks to
Andrew Yourtchenko for the initial implementation.
Change-Id: I4fbd9229cacfaba8027a85cbf87b74afdead6e39
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
According to Maciek, the easiest way to leverage the csit "performance
trend" job is to actually merge the patch once verified. Manual
testing indicates that the patch improves l2 path performance. Other
use-cases are TBD. It's possible that we'll need to back out the patch
depending on what happens.
Change-Id: Ic0a0363de35ef9be953ad7709c57c3936b73fd5a
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Particularly in the DCLIB_VEC64=1 case, using vectors vs. raw
clib_mem_alloc'ed memory causes abysmal memory allocator performance.
Change-Id: I07a4dec0cd69ca357445385e2671cdf23c59b95d
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Change-Id: I55dad7b5cfb3d38c22b1105f7d2d61e7449410ea
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Change-Id: I7b51f88292e057c6443b12224486f2d0c9f8ae23
Signed-off-by: Damjan Marion <damarion@cisco.com>
|