Age | Commit message (Collapse) | Author | Files | Lines |
|
Currently sw_scheduler runs interchangeably over queues of one selected
type either ENCRYPT or DECRYPT, then switches the type for the next run.
This runs perfectly when we have elements in both ENCRYPT and DECRYPT
queues, however, this leads to performance degradation when only one
of the queues have elements i.e either all traffic is to be encrypted or
decrypted.
If all operations are encryption, then 50% of the time, the loop exits
without dequeueing.With this change, that dequeueing happens on every
loop. This increases the performance of single mode operation (ecryption
or decryption) by over 15%.
This change was also added in commit https://github.com/FDio/vpp/commit/61cdc0981084f049067626b0123db700035120df
to fix similar performance issue when the crypto-dispatch node is in interrupt
node, however was removed by https://github.com/FDio/vpp/commit/9a9604b09f15691d7c4ddf29afd99a31e7e31eed
which has its own limitations.
Type: improvement
Change-Id: I15c1375427e06187e9c4faf2461ab79935830802
Signed-off-by: Niyaz Murshed <niyaz.murshed@arm.com>
|
|
Type: refactor
Change-Id: I5235bf3e9aff58af6ba2c14e8c6529c4fc9ec86c
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
As per discussion on the VPP call: since the CSIT tests use these
APIs and the APIs have not changed in quite a while, stabilize the API
from the change process point of view.
Type: improvement
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
Change-Id: Id81999d03cce37764f6ed7d4f77ef5a71fe41ad1
|
|
The checkstyle --fix command remains confused
around the def/foreach/undef usage in convert_async_crypto_id,
but at least the other functions now look correctly indented to me.
Type: style
Change-Id: Ic8f7b580267386b7a6b07d33d9ba7ae9787c0e0a
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
+ Checkstyle demanded indentation edits.
Type: fix
Ticket: VPP-2083
Fixes: 9a9604b09f15691d7c4ddf29afd99a31e7e31eed
Change-Id: Ie2d33d290330247d36435a073675b732bb64ae93
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
This patch can make crypto dispatch node adaptively switching
between pooling and interrupt mode, and improve vpp overall
performance.
Type: improvement
Signed-off-by: Xiaoming Jiang <jiangxiaoming@outlook.com>
Change-Id: I845ed1d29ba9f3c507ea95a337f6dca7f8d6e24e
|
|
Type: fix
Currently sw_scheduler runs interchangeably over queues of one selected
type either ENCRYPT or DECRYPT. Then switches the type for the next run.
This works fine in polling mode as missed frames get processed on the
next run. In interrupt mode if all of the workers miss a frame on the
first run the interrupt flag is lowered so the frame remains pending in
queues waiting for another crypto event to raise the interrupt.
With this fix force sw_scheduler in interrupt mode check the second half
of the queues if the first pass returned no results. This guarantees a
pending frame gets into processing before interrupt is reset.
Change-Id: I7e91d125702336eba72c6a3abaeabcae010d396a
Signed-off-by: Alexander Skorichenko <askorichenko@netgate.com>
|
|
IV requirements vary wildly with the selected mode of operation. For
example, for AES-CBC the IV must be unpredictable whereas for AES
counter mode (CTR or GCM), it can be predictable but reusing an IV with
the same key material is catastrophic.
Because of that, it is hard to generate IV in a generic way, and it is
better left to the crypto user (eg. IPsec).
Type: improvement
Change-Id: I32689c591d8c6572b8d37c4d24f175ea6132d3ec
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
the batch
Type: fix
Signed-off-by: Neale Ranns <neale@graphiant.com>
Change-Id: Icd1e43a5764496784c355c93066273435f16dd35
|
|
When there are several workers, iterator can and will skip
head iterator and it will last until BARRIER_SYNC_TIMEOUT won't
expire and will cause SIGABRT with `worker thread deadlock`
Type: fix
Signed-off-by: Vladimir Ratnikov <vratnikov@netgate.com>
Change-Id: Id4def4d5894e077ae27592367b141ecd822e86af
Signed-off-by: Matthew Smith <mgsmith@netgate.com>
|
|
Type: improvement
Previously multiple sw crypto scheduler queues per core design
caused unaverage frame processing rate for each async op ID –
the lower the op ID is the highly likely they are processed first.
For example, when a RX core is feeding both encryption and
decryption jobs of the same crypto algorithm to the queues at a
high rate, in the mean time the crypto cores have no enough
cycles to process all: the jobs in the decryption queue are less
likely being processed, causing packet drop.
To improve the situation this patch makes every core only owning
a two queues, one for encrypt operations and one for decrypt.
The queue is changed either after checking each core
or after founding a frame to process.
All crypto jobs with different algorithm are pushed to
thoses queues and are treated evenly.
In addition, the crypto async infra now uses unified dequeue handler,
one per engine. Only the active engine will be registered its
dequeue handler in crypto main.
Signed-off-by: DariuszX Kazimierski <dariuszx.kazimierski@intel.com>
Signed-off-by: PiotrX Kleski <piotrx.kleski@intel.com>
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Signed-off-by: Jakub Wysocki <jakubx.wysocki@intel.com>
Change-Id: I517ee8e31633980de5e0dd4b05e1d5db5dea760e
|
|
Type: refactor
Change-Id: Id10cbf52e8f2dd809080a228d8fa282308be84ac
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
When a buffer in the chain comes with a negative current_data offset,
the conversion to sgl will skip it because of resetting offset to 0.
Moreover, crypto_start_offset is relative to the 1st buffer data pointer
so we should not check it against subsequent buffers anyway.
Type: fix
Change-Id: Id177a90bfda242a5372c7e8836cf6668e98c780e
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
Type: fix
Change-Id: I984a3e577a4209e41d046eaf3a8eef8986dc6147
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
The crypto op flags must be reset to frame flags minus invalid values
depending of the operation, instead of forcing them to specific values.
Type: fix
Change-Id: Ib02c2a738bbca6962394b3c03088d516d0da56a0
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
This reverts commit 30ad571cc35e4dc6d4d7e50b81b97f83f8770eea.
Type: fix
Change-Id: If8c6e388e732d2a1b5efd0677d9528a646365f94
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
The sw crypto scheduler converts crypto frames to individual crypto
operations. This is done by reusing per-thread vectors for crypto,
integrity and chained operations.
The crypto op flags must be reset to frame flags minus invalid values
depending of the operation.
The previous tentative also cleared the chained buffer flag, breaking
jumbo support.
Type: fix
Change-Id: Icce6887a9e0dae8c300c56e97b977e203e784713
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
When IPsec async mode is enabled, packets don't pass through the tunnel
if ciphers other than AES GCM are used for child SAs. An error that
arises is "bad-hmac" in the "crypto-dispatch" node.
On the encryption stage, the VNET_CRYPTO_OP_FLAG_HMAC_CHECK flag is set
for the integrity crypto operation when it's not supposed to. It seems
that the flag remains from the previous operation.
With this change, zero flags of crypto operations in the SW scheduler
during operation filling.
Type: fix
Signed-off-by: Alexander Chernavin <achernavin@netgate.com>
Change-Id: Iabac253474e95cb01f9ec0933f3c4860f8a5289c
|
|
Type: fix
To avoid race condition happening in async crypto engines,
async frame state and thread index set should happen before enqueue.
In addition as the enqueue handler already returns the enqueue status,
when an enqueue is failed, the async crypto engine shall not worry
about setting the async frame state but let the submit_open_frame function
to do just that.
Signed-off-by: PiotrX Kleski <piotrx.kleski@intel.com>
Reviewed-by: Fan Zhang <roy.fan.zhang@intel.com>
Change-Id: Ic1b0c94478b3cfd5fab98657218bbd70c46a220a
|
|
Type: fix
Signed-off-by: Neale Ranns <nranns@cisco.com>
Change-Id: I7f98d0c7847ecc40b90b78e5ae83f320575be310
|
|
Type: feature
This adds api calls for the following CLIs:
* set sw_scheuduler worker <N> crypto on|off
* set crypto async dispatch polling|interrupt
* set crypto handler
* set crypto async handler
Change-Id: Ic701d149c440e42ea4575da42b9f69e4c8759602
Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>
|
|
Type: feature
This patch adds new sw_scheduler async crypto engine.
The engine transforms async frames info sync crypto ops and
delegates them to active sync engines. With the patch it
is possible to increase the single worker crypto throughput
by offloading the crypto workload to multiple workers.
By default all workers in the system will attend the crypto
workload processing. However a worker's available cycles
are limited. To avail more cycles to one worker to process
other workload (e.g. the worker core that handles the RX/TX
and IPSec stack processing), a useful cli command is added
to remove itself (or add it back later) from the heavy
crypto workload but only let other workers to process the
crypto. The command is:
- set sw_scheduler worker <idx> crypto <on|off>
It also adds new interrupt mode to async crypto dispatch node.
This mode signals the node when new frames are enqueued
as opposed to polling mode that continuously calls dispatch node.
New cli commands:
- set crypto async dispatch [polling|interrupt]
- show crypto async status (displays mode and nodes' states)
Signed-off-by: PiotrX Kleski <piotrx.kleski@intel.com>
Signed-off-by: DariuszX Kazimierski <dariuszx.kazimierski@intel.com>
Reviewed-by: Fan Zhang <roy.fan.zhang@intel.com>
Change-Id: I332655f347bb9e3bc9c64166e86e393e911bdb39
|