This commit adds support for lock-free (linked list based) stack mempool
handler.
In mempool_perf_autotest the lock-based stack outperforms the
lock-free handler for certain lcore/alloc count/free count
combinations*, however:
- For applications with preemptible pthreads, a standard (lock-based)
stack's worst-case performance (i.e. one thread being preempted while
holding the spinlock) is much worse than the lock-free stack's.
- Using per-thread mempool caches will largely mitigate the performance
difference.
*Test setup: x86_64 build with default config, dual-socket Xeon E5-2699 v4,
running on isolcpus cores with a tickless scheduler. The lock-based stack's
rate_persec was 0.6x-3.5x the lock-free stack's.
Signed-off-by: Gage Eads <gage.eads@intel.com>
Reviewed-by: Olivier Matz <olivier.matz@6wind.com>
This commit adds support for a lock-free (linked list based) stack to the
stack API. This behavior is selected through a new rte_stack_create() flag,
RTE_STACK_F_LF.
The stack consists of a linked list of elements, each containing a data
pointer and a next pointer, and an atomic stack depth counter.
The lock-free push operation enqueues a linked list of pointers by pointing
the tail of the list to the current stack head, and using a CAS to swing
the stack head pointer to the head of the list. The operation retries if it
is unsuccessful (i.e. the list changed between reading the head and
modifying it), else it adjusts the stack length and returns.
The lock-free pop operation first reserves num elements by adjusting the
stack length, to ensure the dequeue operation will succeed without
blocking. It then dequeues pointers by walking the list -- starting from
the head -- then swinging the head pointer (using a CAS as well). While
walking the list, the data pointers are recorded in an object table.
This algorithm stack uses a 128-bit compare-and-swap instruction, which
atomically updates the stack top pointer and a modification counter, to
protect against the ABA problem.
The linked list elements themselves are maintained in a lock-free LIFO
list, and are allocated before stack pushes and freed after stack pops.
Since the stack has a fixed maximum depth, these elements do not need to be
dynamically created.
Signed-off-by: Gage Eads <gage.eads@intel.com>
Reviewed-by: Olivier Matz <olivier.matz@6wind.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
The rte_stack library provides an API for configuration and use of a
bounded stack of pointers. Push and pop operations are MT-safe, allowing
concurrent access, and the interface supports pushing and popping multiple
pointers at a time.
The library's interface is modeled after another DPDK data structure,
rte_ring, and its lock-based implementation is derived from the stack
mempool handler. An upcoming commit will migrate the stack mempool handler
to rte_stack.
Signed-off-by: Gage Eads <gage.eads@intel.com>
Reviewed-by: Olivier Matz <olivier.matz@6wind.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
This patch updates the ipsec library programmer's guide with
the additional algorithms which are now supported.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Due to internal glibc limitations [1], DPDK may exhaust internal
file descriptor limits when using smaller page sizes, which results
in inability to use system calls such as select() by user
applications.
Single file segments option stores lock files per page to ensure
that pages are deleted when there are no more users, however this
is not necessary because the processes will be holding onto the
pages anyway because of mmap(). Thus, removing pages from the
filesystem is safe even though they may be used by some other
secondary process. As a result, single file segments mode no
longer stores inordinate amounts of segment fd's, and the above
issue with fd limits is solved.
However, this will not work for legacy mem mode. For that, simply
document that using bigger page sizes is the only option.
[1] https://mails.dpdk.org/archives/dev/2019-February/124386.html
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
The DPDK APIs expose 3 different modes to work with memory used for DMA:
1. Use the DPDK owned memory (backed by the DPDK provided hugepages).
This memory is allocated by the DPDK libraries, included in the DPDK
memory system (memseg lists) and automatically DMA mapped by the DPDK
layers.
2. Use memory allocated by the user and register to the DPDK memory
systems. Upon registration of memory, the DPDK layers will DMA map it
to all needed devices. After registration, allocation of this memory
will be done with rte_*malloc APIs.
3. Use memory allocated by the user and not registered to the DPDK memory
system. This is for users who wants to have tight control on this
memory (e.g. avoid the rte_malloc header).
The user should create a memory, register it through rte_extmem_register
API, and call DMA map function in order to register such memory to
the different devices.
The scope of the patch focus on #3 above.
Currently the only way to map external memory is through VFIO
(rte_vfio_dma_map). While VFIO is common, there are other vendors
which use different ways to map memory (e.g. Mellanox and NXP).
The work in this patch moves the DMA mapping to vendor agnostic APIs.
Device level DMA map and unmap APIs were added. Implementation of those
APIs was done currently only for PCI devices.
For PCI bus devices, the pci driver can expose its own map and unmap
functions to be used for the mapping. In case the driver doesn't provide
any, the memory will be mapped, if possible, to IOMMU through VFIO APIs.
Application usage with those APIs is quite simple:
* allocate memory
* call rte_extmem_register on the memory chunk.
* take a device, and query its rte_device.
* call the device specific mapping function for this device.
Future work will deprecate the rte_vfio_dma_map and rte_vfio_dma_unmap
APIs, leaving the rte device APIs as the preferred option for the user.
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Gaetan Rivet <gaetan.rivet@6wind.com>
- mbuf_size and mtu are now being calculated according
to the given mb-pool.
- max_mtu is now being set according to the given mtu
the above two changes provide the ability to work with jumbo frames
Signed-off-by: Liron Himi <lironh@marvell.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
EventDev i.e consumer needs to be started before starting the
event producers.
Update documentation of EventDev and EventDev adapters.
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Reviewed-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
Reviewed-by: Abhinandan Gujjar <abhinandan.gujjar@intel.com>
Rather than using linuxapp and bsdapp everywhere, we can change things to
use the, more readable, terms "linux" and "freebsd" in our build configs.
Rather than renaming the configs we can just duplicate the existing ones
with the new names using symlinks, and use the new names exclusively
internally. ["make showconfigs" also only shows the new names to keep the
list short] The result is that backward compatibility is kept fully but any
new builds or development can be done using the newer names, i.e. both
"make config T=x86_64-native-linuxapp-gcc" and "T=x86_64-native-linux-gcc"
work.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
The term "linuxapp" is a legacy one, but just calling the subdirectory
"linux" is just clearer for all concerned.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
The term "bsdapp" is a legacy one, but just calling the subdirectory
"freebsd" is just clearer for all concerned.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Spawning the ctrl threads on anything that is not part of the eal
coremask is not that polite to the rest of the system, especially
when you took good care to pin your processes on cpu resources with
tools like taskset (linux) / cpuset (freebsd).
Rather than introduce yet another eal options to control on which cpu
those ctrl threads are created, let's take the startup cpu affinity
as a reference and remove the eal coremask from it.
If no cpu is left, then we default to the master core.
The cpuset is computed once at init before the original cpu affinity
is lost.
Introduced a RTE_CPU_AND macro to abstract the differences between linux
and freebsd respective macros.
Examples in a 4 cores FreeBSD vm:
$ ./build/app/testpmd -l 2,3 --no-huge --no-pci -m 512 \
-- -i --total-num-mbufs=2048
$ procstat -S 1057
PID TID COMM TDNAME CPU CSID CPU MASK
1057 100131 testpmd - 2 1 2
1057 100140 testpmd eal-intr-thread 1 1 0-1
1057 100141 testpmd rte_mp_handle 1 1 0-1
1057 100142 testpmd lcore-slave-3 3 1 3
$ cpuset -l 1,2,3 ./build/app/testpmd -l 2,3 --no-huge --no-pci -m 512 \
-- -i --total-num-mbufs=2048
$ procstat -S 1061
PID TID COMM TDNAME CPU CSID CPU MASK
1061 100131 testpmd - 2 2 2
1061 100144 testpmd eal-intr-thread 1 2 1
1061 100145 testpmd rte_mp_handle 1 2 1
1061 100147 testpmd lcore-slave-3 3 2 3
$ cpuset -l 2,3 ./build/app/testpmd -l 2,3 --no-huge --no-pci -m 512 \
-- -i --total-num-mbufs=2048
$ procstat -S 1065
PID TID COMM TDNAME CPU CSID CPU MASK
1065 100131 testpmd - 2 2 2
1065 100148 testpmd eal-intr-thread 2 2 2
1065 100149 testpmd rte_mp_handle 2 2 2
1065 100150 testpmd lcore-slave-3 3 2 3
Fixes: d651ee4919cd ("eal: set affinity for control threads")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: Olivier Matz <olivier.matz@6wind.com>
The placeholder for PCI address should be named DBDF
which stands for Domain/Bus/Device/Function.
Fixes: 33af337773ac ("ethdev: add common devargs parser")
Cc: stable@dpdk.org
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Rami Rosen <ramirose@gmail.com>
Highlight that vhost zero copy mbufs should be consumed
as soon as possible.
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Since all other apps have been moved to the "app" folder, the autotest app
remains alone in the test folder. Rather than having an entire top-level
folder for this, we can move it back to where it all started in early
versions of DPDK - the "app/" folder.
This move has a couple of advantages:
* This reduces clutter at the top level of the project, due to one less
folder.
* It eliminates the separate build task necessary for building the
autotests using make "make test-build" which means that developers are
less likely to miss something in their own compilation tests
* It re-aligns the final location of the test binary in the app folder when
building with make with it's location in the source tree.
For meson builds, the autotest app is different from the other apps in that
it needs a series of different test cases defined for it for use by "meson
test". Therefore, it does not get built as part of the main loop in the
app folder, but gets built separately at the end.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Some drivers (mlx, mvpp2, sfc) support the flow isolated mode,
but the feature was not advertised.
A reference to the feature description is added for each driver.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>
A doc page (.rst file) can be referenced with :doc: syntax
instead of :ref: to .. anchor.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>
In the References section in the Power Management overview,
both links pointed to the same l3fwd-power app. Fix the links
so that one points to l3fwd-power, and the other points to
the vm_power_manager sample app.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Marko Kovacevic <marko.kovacevic@intel.com>
No need to keep those file listings, they are very likely to become
outdated.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Marko Kovacevic <marko.kovacevic@intel.com>
This patch adds GRO limitations in the programmer guide.
Fixes: 2c900d09055e ("doc: add GRO guide")
Cc: stable@dpdk.org
Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
This patch fixes a typo in programmer's guide. It should be Frequency,
not Fequence.
Fixes: 450f0791312c ("power: add traffic pattern aware power control")
Cc: stable@dpdk.org
Signed-off-by: Yong Wang <wang.yong19@zte.com.cn>
This patch fixes a typo in SET_MAC_DST action description.
It also adds restriction note for set MAC src/dst actions description.
Fixes: 15dbcdaada77 ("ethdev: add generic MAC address rewrite actions")
Cc: stable@dpdk.org
Signed-off-by: Dekel Peled <dekelp@mellanox.com>
Acked-by: Ori Kam <orika@mellanox.com>
This patch adds a opaque data field to cryptodev symmetric session.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
This patch adds a refcnt field to every session private data in the
cryptodev symmetric session. The counter is used to prevent freeing
symmetric session blindly before it is not cleared by every type of
crypto device in use.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
This patch adds a user_data_sz field to cryptodev symmetric session.
The field is used to check if reading or writing the session's user
data field is eligible.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
This patch updates the rte_cryptodev_sym_session structure for
cryptodev library. The updates include a changed session private
data array and an added nb_drivers field. They are used to
calculate the correct session header size and ensure safe access
of the session private data.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
This patch adds a new API "rte_cryptodev_sym_session_pool_create()" to
cryptodev library. All applications are required to use this API to
create sym session mempool as it adds private data and nb_drivers
information to the mempool private data.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
This patch changes the cryptodev queue pair configure structure
to enable two mempool passed into cryptodev PMD simutaneously.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
Memory mode flags are now shared between primary and secondary
processes, so the in documentation about limitations is no longer
necessary.
Fixes: 64cdfc35aaad ("mem: store memory mode flags in shared config")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Add multiprocess support for externally allocated memory areas that
are not added to DPDK heap (and add relevant doc sections).
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
The general use-case of using external memory is well covered by
existing external memory API's. However, certain use cases require
manual management of externally allocated memory areas, so this
memory should not be added to the heap. It should, however, be
added to DPDK's internal structures, so that API's like
``rte_virt2memseg`` would work on such external memory segments.
This commit adds such an API to DPDK. The new functions will allow
to register and unregister externally allocated memory areas, as
well as documentation for them.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
SPDK uses the rte_mem_event_callback_register API to
create RDMA memory regions (MRs) for newly allocated regions
of memory. This is used in both the SPDK NVMe-oF target
and the NVMe-oF host driver.
DPDK creates internal malloc_elem structures for these
allocated regions. As users malloc and free memory, DPDK
will sometimes merge malloc_elems that originated from
different allocations that were notified through the
registered mem_event callback routine. This results
in subsequent allocations that can span across multiple
RDMA MRs. This requires SPDK to check each DPDK buffer to
see if it crosses an MR boundary, and if so, would have to
add considerable logic and complexity to describe that
buffer before it can be accessed by the RNIC. It is somewhat
analagous to rte_malloc returning a buffer that is not
IOVA-contiguous.
As a malloc_elem gets split and some of these elements
get freed, it can also result in DPDK sending an
RTE_MEM_EVENT_FREE notification for a subset of the
original RTE_MEM_EVENT_ALLOC notification. This is also
problematic for RDMA memory regions, since unregistering
the memory region is all-or-nothing. It is not possible
to unregister part of a memory region.
To support these types of applications, this patch adds
a new --match-allocations EAL init flag. When this
flag is specified, malloc elements from different
hugepage allocations will never be merged. Memory will
also only be freed back to the system (with the requisite
memory event callback) exactly as it was originally
allocated.
Since part of this patch is extending the size of struct
malloc_elem, we also fix up the malloc autotests so they
do not assume its size exactly fits in one cacheline.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
We already changed to use generic IPC in pdump since below commit:
commit 660098d61f57 ("pdump: use generic multi-process channel")
The `rte_pdump_set_socket_dir()`, the `path` parameter of
`rte_pdump_init()` and the `enum rte_pdump_socktype` have been
deprecated since then. This commit removes these deprecated
APIs and also bumps the pdump ABI.
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
Acked-by: Reshma Pattan <reshma.pattan@intel.com>
The DPDK website has a new URL scheme since June 2018.
Cc: stable@dpdk.org
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: John McNamara <john.mcnamara@intel.com>
This commit improves the programmer guide of the hash
library to be more accurate on new features introduced
in 18.11.
Signed-off-by: Yipeng Wang <yipeng1.wang@intel.com>
Signed-off-by: Sameh Gobriel <sameh.gobriel@intel.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
The PCI bus is an independent driver and not part of EAL
as it was in the early days.
EAL must be understood as a generic layer.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: John McNamara <john.mcnamara@intel.com>
The references to the figures and tables in the index
are not maintained.
It is probably better to have no list than an incomplete list.
Anyway the usage of such figures list is not obvious.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: John McNamara <john.mcnamara@intel.com>
Updated the doc on how packets are marked to identify
their timestamp as valid and considered for latency
measurement.
Suggested-by: Bao-Long Tran <longtb5@viettel.com.vn>
Signed-off-by: Reshma Pattan <reshma.pattan@intel.com>
This patch uses EAL option "--iova-mode" to force the IOVA mode to a
particular value. There exists virtual devices that are not directly
attached to the PCI bus, and therefore the auto detection of the IOVA
mode based on probing the PCI bus and IOMMU configuration may not
report the required addressing mode. Using the EAL option permits the
mode to be explicitly configured in this scenario.
Signed-off-by: Eric Zhang <eric.zhang@windriver.com>
Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: Marko Kovacevic <marko.kovacevic@intel.com>
Previous patch introduced the Tx metadata feature, with unnecessary
restrictions on data entry.
This fix updates the documentation, removing the data entry
restrictions on metadata item.
Fixes: 839b20be0e9b ("ethdev: support metadata as flow rule criteria")
Acked-by: Ori Kam <orika@mellanox.com>
Signed-off-by: Dekel Peled <dekelp@mellanox.com>
The hotplug attach/detach features are implemented in EAL layer.
There is a new ethdev iterator to retrieve ports from ethdev layer.
As announced earlier, the (buggy) ethdev functions are now removed.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Andrew Rybchenko <arybchenko@solarflare.com>
Currenlty the encap/decap actions only support encapsulation
of VXLAN and NVGRE L2 packets (L2 encapsulation is where
the inner packet has a valid Ethernet header, while L3 encapsulation
is where the inner packet doesn't have the Ethernet header).
In addtion the parameter to to the encap action is a list of rte items,
this results in 2 extra translation, between the application to the
actioni and from the action to the NIC. This results in negative impact
on the insertion performance.
Looking forward there are going to be a need to support many more tunnel
encapsulations. For example MPLSoGRE, MPLSoUDP.
Adding the new encapsulation will result in duplication of code.
For example the code for handling NVGRE and VXLAN are exactly the same,
and each new tunnel will have the same exact structure.
This patch introduce a raw encapsulation that can support L2 tunnel types
and L3 tunnel types. In addtion the new
encapsulations commands are using raw buffer inorder to save the
converstion time, both for the application and the PMD.
In order to encapsulate L3 tunnel type there is a need to use both
actions in the same rule: The decap to remove the L2 of the original
packet, and then encap command to encapsulate the packet with the
tunnel.
For decap L3 there is also a need to use both commands in the same flow
first the decap command to remove the outer tunnel header and then encap
to add the L2 header.
Signed-off-by: Ori Kam <orika@mellanox.com>
Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
As described in [1], a new rte_flow item is added to support metadata
to use as flow rule match pattern.
The metadata is an opaque item, fully controlled by the application.
The use of metadata is relevant for egress rules only.
It can be set in the flow rule using the RTE_FLOW_ITEM_META.
An additional member 'tx_metadata' is added in union with existing member
'hash' of struct 'rte_mbuf', located to avoid conflicts with existing
fields. This additional member is used to carry the metadata item.
Application should set the packet metadata in the mbuf dedicated field,
and set the PKT_TX_METADATA flag in the mbuf->ol_flags.
The NIC will use the packet metadata as match criteria for relevant
flow rules.
This patch introduces metadata item type for rte_flow RTE_FLOW_ITEM_META,
along with corresponding struct rte_flow_item_meta and ol_flag
PKT_TX_METADATA.
[1] "[RFC,v2] ethdev: support metadata as flow rule criteria"
Signed-off-by: Dekel Peled <dekelp@mellanox.com>
Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
Add module parameter 'carrier='on|off' to set the default carrier state
for linux network interfaces created by the KNI module. The default
carrier state is 'off'.
For KNI interfaces which need to reflect the carrier state of
a physical Ethernet port controlled by the DPDK application, the
default carrier state should be left set to 'off'. The application
can set the carrier state of the KNI interface to reflect the state
of the physical Ethernet port using rte_kni_update_link().
For KNI interfaces which are purely virtual, the default carrier
state can be set to 'on'. This enables the KNI interface to be
used without having to explicity set the carrier state to 'on'
using rte_kni_update_link().
Signed-off-by: Dan Gora <dg@adax.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
Add lock-free read-write concurrency. This is achieved by the
following changes.
1) Add memory ordering to avoid race conditions. The only race
condition that can occur is - using the key store element
before the key write is completed. Hence, while inserting the element
the release memory order is used. Any other race condition is caught
by the key comparison. Memory orderings are added only where needed.
For ex: reads in the writer's context do not need memory ordering
as there is a single writer.
key_idx in the bucket entry and pdata in the key store element are
used for synchronisation. key_idx is used to release an inserted
entry in the bucket to the reader. Use of pdata for synchronisation
is required due to updation of an existing entry where-in only
the pdata is updated without updating key_idx.
2) Reader-writer concurrency issue, caused by moving the keys
to their alternative locations during key insert, is solved
by introducing a global counter(tbl_chng_cnt) indicating a
change in table.
3) Add the flag to enable reader-writer concurrency during
run time.
Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Reviewed-by: Ola Liljedahl <ola.liljedahl@arm.com>
Reviewed-by: Steve Capper <steve.capper@arm.com>
Reviewed-by: Yipeng Wang <yipeng1.wang@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
