When running in no-huge mode, we anonymously allocate our memory.
While this works for regular NICs and vdev's, it's not suitable
for memory sharing scenarios such as virtio with vhost_user
backend.
To fix this, allocate no-huge memory using memfd, and register
it with memalloc just like any other memseg fd. This will enable
using rte_memseg_get_fd() API with --no-huge EAL flag.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Tiwei Bie <tiwei.bie@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Currently, only segment fd's for multi-file segments are supported,
while for memfd-backed no-huge memory we need single-file segments
mode. Add support for single-file segments in the internal API.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Tiwei Bie <tiwei.bie@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
If memfd support was not compiled, or hugepage memfd support
is not available at runtime, the API will now return proper
error code, indicating that this API is unsupported. This
changes the API, so document the changes.
Fixes: 41dbdb68723b ("mem: add external API to retrieve page fd")
Fixes: 3a44687139eb ("mem: allow querying offset into segment fd")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Tiwei Bie <tiwei.bie@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Segment fd API does not support getting segment fd's from
externally allocated memory, so return proper error code
on any attempts to do so. This changes API behavior, so
document the change as well.
Fixes: 5282bb1c3695 ("mem: allow memseg lists to be marked as external")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Tiwei Bie <tiwei.bie@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.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>
Currently, destroying external heap chunk and its memseg list is
part of one process. When we will gain the ability to unregister
external memory from DPDK that doesn't have any heap structures
associated with it, we need to be able to find and destroy
memseg lists as well as heap data separately.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
Currently, creating external malloc heap involves also creating
a memseg list backing that malloc heap. We need to have them as
separate functions, to allow creating memseg lists without
creating a malloc heap.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
The external heaps API already implicitly expects start address
of the external memory area to be page-aligned, but it is not
enforced or documented. Fix this by implementing additional
parameter checks at memory add call, and document the page
alignment requirement explicitly.
Fixes: 7d75c31014f7 ("malloc: allow adding memory to named heaps")
Cc: stable@dpdk.org
Suggested-by: Yongseok Koh <yskoh@mellanox.com>
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
We already trigger a mem event notification inside the walk function,
no need to do it twice.
Fixes: f32c7c9de961 ("malloc: enable event callbacks for external memory")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
When secondary process hotplugs memory, it sends a request
to primary, which then performs the real mmap() and sends
sync requests to all secondary processes. Upon receiving
such sync request, each secondary process will notify the
upper layers of hotplugged memory (and will call all
locally registered event callbacks).
In the end we'll end up with memory event callbacks fired
in all the processes except the primary, which is a bug.
This gets critical if memory is hotplugged while a VFIO
device is attached, as the VFIO memory registration -
which is done from a memory event callback present in the
primary process only - is never called.
After this patch, a primary process fires memory event
callbacks before secondary processes start their
synchronizations - both for hotplug and hotremove.
Fixes: 07dcbfe0101f ("malloc: support multiprocess memory hotplug")
Cc: stable@dpdk.org
Signed-off-by: Seth Howell <seth.howell@intel.com>
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.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>
The RTE_PROC_PRIMARY error handler lost the unlock statement in the
current codes. Now unlock and return in one place to fix it.
Fixes: 49df3db84883 ("memzone: replace memzone array with fbarray")
Cc: stable@dpdk.org
Signed-off-by: Gao Feng <davidfgao@tencent.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Add the check for null peer pointer like the bundle pointer in the mp request
handler. They should follow same style. And add some logs for nomem cases.
Signed-off-by: Gao Feng <davidfgao@tencent.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
When rte_eal_alarm_set failed, need to free the bundle mem in the
error handler of handle_primary_request and handle_secondary_request.
Fixes: 244d5130719c ("eal: enable hotplug on multi-process")
Fixes: ac9e4a17370f ("eal: support attach/detach shared device from secondary")
Cc: stable@dpdk.org
Signed-off-by: Gao Feng <davidfgao@tencent.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Missing brackets around the if means that the loop will end at
its first iteration.
Fixes: 2395332798d0 ("eal: add option register infrastructure")
Cc: stable@dpdk.org
Signed-off-by: Gaetan Rivet <gaetan.rivet@6wind.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
The same issue was fixed on for the ipv4 version of this routine in
commit 8d4d3a4f7337 ("ip_frag: handle MTU sizes not aligned to 8 bytes").
Briefly, the size of an ipv6 header is always 40 bytes. With an MTU of
1500, this will never produce a multiple of 8 bytes for the frag_size
and this routine can never succeed. Since RTE_ASSERTS are disabled by
default, this failure is typically ignored.
To fix this, round down to the nearest 8 bytes and use this when
producing the fragments.
Fixes: 0aa31d7a5929 ("ip_frag: add IPv6 fragmentation support")
Cc: stable@dpdk.org
Signed-off-by: Chas Williams <chas3@att.com>
Acked-by: Luca Boccassi <bluca@debian.org>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
rte_timer_manage() adds expired timers to a "run list", and walks the
list, transitioning each timer from the PENDING to the RUNNING state.
If another lcore resets or stops the timer at precisely this
moment, the timer state would instead be set to CONFIG by that other
lcore, which would cause timer_manage() to skip over it. This is
expected behavior.
However, if a timer expires quickly enough, there exists the
following race condition that causes the timer_manage() routine to
misinterpret a timer in CONFIG state, resulting in lost timers:
- Thread A:
- starts a timer with rte_timer_reset()
- the timer is moved to CONFIG state
- the spinlock associated with the appropriate skiplist is acquired
- timer is inserted into the skiplist
- the spinlock is released
- Thread B:
- executes rte_timer_manage()
- find above timer as expired, add it to run list
- walk run list, see above timer still in CONFIG state, unlink it from
run list and continue on
- Thread A:
- move timer to PENDING state
- return from rte_timer_reset()
- timer is now in PENDING state, but not actually linked into a
pending list or a run list and will never get processed further
by rte_timer_manage()
This commit fixes this race condition by only releasing the spinlock
after the timer state has been transitioned from CONFIG to PENDING,
which prevents rte_timer_manage() from seeing an incorrect state.
Fixes: 9b15ba895b9f ("timer: use a skip list")
Cc: stable@dpdk.org
Signed-off-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Improvements added to throughput test:
- test is run in loop (number of iterations is specified by
TEST_REPETITIONS define) which ensures more accurate results
- length of input data is calculated based on amount of CBs in TB
- maximum number of decoding iterations is gathered from results
- added new functions responsible for printing results
- small fixes for memory management
Signed-off-by: Kamil Chalupnik <kamilx.chalupnik@intel.com>
Acked-by: Amr Mokhtar <amr.mokhtar@intel.com>
Offload cost test was improved in order to collect
more accurate results.
Signed-off-by: Kamil Chalupnik <kamilx.chalupnik@intel.com>
Acked-by: Amr Mokhtar <amr.mokhtar@intel.com>
Fixes incorrect comment on compressdev rte_comp_op structure element.
Comment needed to be updated to be compliant with the use of
chained mbufs.
Fixes: f87bdc1ddcf4 ("compressdev: add compression specific data")
Cc: stable@dpdk.org
Signed-off-by: Lee Daly <lee.daly@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
There's an API to bulk allocate operations,
this adds a corresponding bulk free API.
Signed-off-by: Fiona Trahe <fiona.trahe@intel.com>
Acked-by: Shally Verma <shally.verma@caviumnetworks.com>
Acked-by: Lee Daly <lee.daly@intel.com>
rte_event_eth_tx_adapter_queue_add() - add a check
that returns an error if the ethdev has zero Tx queues
configured.
rte_event_eth_tx_adapter_queue_del() - remove the
checks for ethdev queue count, instead check for
queues added to the adapter which maybe different
from the current ethdev queue count.
Fixes: a3bbf2e09756 ("eventdev: add eth Tx adapter implementation")
Cc: stable@dpdk.org
Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
The eventdev extended stats documentation referred to two non-existent
functions, rte_eventdev_xstats_get and rte_eventdev_get_xstats_by_name.
Fixes: 3ed7fc039a ("eventdev: add extended stats")
Cc: stable@dpdk.org
Signed-off-by: Gage Eads <gage.eads@intel.com>
When the packet length is smaller than the header length,
the calculated payload length will be overflowed and result
in incorrect reassembly behaviors.
Fixes: 1e4cf4d6d4fb ("gro: cleanup")
Fixes: 9e0b9d2ec0f4 ("gro: support VxLAN GRO")
Cc: stable@dpdk.org
Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
When creating process data structures, EAL will create many files
in EAL runtime directory. Because we allow multiple secondary
processes to run, each secondary process gets their own unique
file. With many secondary processes running and exiting on the
system, runtime directory will, over time, create enormous amounts
of sockets, fbarray files and other stuff that just sits there
unused because the process that allocated it has died a long time
ago. This may lead to exhaustion of disk (or RAM) space in the
runtime directory.
Fix this by removing every unlocked file at initialization that
matches either socket or fbarray naming convention. We cannot be
sure of any other files, so we'll leave them alone. Also, remove
similar code from mp socket code.
We do it at the end of init, rather than at the beginning, because
secondary process will use primary process' data structures even
if the primary itself has died, and we don't want to remove those
before we lock them.
Bugzilla ID: 106
Cc: stable@dpdk.org
Reported-by: Vipin Varghese <vipin.varghese@intel.com>
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
When rte_log_register_type_and_pick_level() has been introduced, it has
been correctly added to the EXPERIMENTAL section of the eal map and the
symbol itself has been marked at its definition.
However, the declaration of this symbol in rte_log.h is missing the
__rte_experimental tag.
Because of this, a user can try to call this symbol without being aware
this is an experimental api (neither compilation nor link warning).
Fixes: b22e77c02620 ("eal: register log type and pick level from args")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Prior to this patch, the two affected .c files include <dirent.h>
unnecessarily. This commit removes the include lines.
Signed-off-by: Jeff Shaw <jeffrey.b.shaw@intel.com>
Reviewed-by: Rami Rosen <ramirose@gmail.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@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>
Flags could be updated in a separate process leading to the
inconsistent check.
Additionally, read marked as 'volatile' to highlight the shared
nature of the variable and avoid such issues in the future.
Fixes: d3211c98c456 ("vhost: add helpers for packed virtqueues")
Cc: stable@dpdk.org
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
If mmap() call fails in vhost_user_set_mem_table, dev->mem
is set to NULL. If later, qva_to_vva() is called, a segfault
occurs.
Fixes: 8f972312b8f4 ("vhost: support vhost-user")
Cc: stable@dpdk.org
Reviewed-by: Tiwei Bie <tiwei.bie@intel.com>
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Jens Freimann <jfreimann@redhat.com>
Start version numbering for a new release cycle,
and introduce a template file for release notes.
The release notes comments are updated to mandate
a scope label for API and ABI changes.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
When RTE_EAL_NUMA_AWARE_HUGEPAGES is set to "n", not all memtypes
will be valid, because we skip some due to not supporting other
NUMA nodes, leading to a division by zero error down the line
because the necessary memtype fields weren't populated.
Fix it by limiting number of memtypes to number of memtypes we
have actually created.
Fixes: 1dd342d0fdc4 ("mem: improve segment list preallocation")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: David Hunt <david.hunt@intel.com>
Even if a device failed to plug, it's still a device
object that references the devargs. Those devargs will
be freed automatically together with the device, but
freeing them any earlier - like it's done in the hotplug
error handling path right now - will give us a dangling
pointer and a segfault scenario.
Consider the following case:
* secondary process receives the hotplug request IPC message
* devargs are either created or updated
* the bus is scanned
* a new device object is created with the latest devargs
* the device can't be plugged for whatever reason,
bus->plug returns error
* the devargs are freed, even though they're still referenced
by the device object on the bus
For PCI devices, the generic device name comes from
a buffer within the devargs. Freeing those will make
EAL segfault whenever the device name is checked.
This patch just prevents the hotplug error handling
path from removing the devargs when there's a device
that references them. This is done by simply exiting
early from the hotplug function. As mentioned in the
beginning, those devargs will be freed later, together
with the device itself.
Fixes: 7e8b26650146 ("eal: fix hotplug add / remove")
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Acked-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Device detach triggered through IPC leaked some memory.
It allocated a devargs objects just to use it for
parsing the devargs string in order to retrieve the
device name. Those devargs weren't passed anywhere
and were never freed.
First of all, let's put those devargs on the stack,
so they doesn't need to be freed. Then free the
additional arguments string as soon as it's allocated,
because we won't need it.
Fixes: ac9e4a17370f ("eal: support attach/detach shared device from secondary")
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Acked-by: Qi Zhang <qi.z.zhang@intel.com>
Consider the following scenario:
1) primary process (A) starts, probes the bus
2) a secondary process (B) starts, probes the bus
3) yet another secondary process (C) starts
4) (C) registers the pci driver and hotplugs the device
* an IPC attach req is sent to the primary (A)
* (A) ignores the -EEXIST from process-local probe
* (A) propagates the request to all secondary processes
* (B) responds with -EEXIST
* (A) replies to the original request with the -EEXIST
return code
* the -EEXIST is returned back to the user, although the
device was successfully attached both locally and in
all other processes
This patch makes the primary process reply with rc=0 even if
there was another secondary process with the device already
attached. The primary process already didn't reply with -EEXIST
when the device was attached locally, so now this behavior is
even more consistent. Looking by the code, this seems to be the
originally intended behavior.
Fixes: ac9e4a17370f ("eal: support attach/detach shared device from secondary")
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Acked-by: Qi Zhang <qi.z.zhang@intel.com>
When primary process receives an IPC attach request
of a device that's already locally-attached, it
doesn't setup its variables properly and is prone to
segfaulting on a subsequent rollback.
`ret = local_dev_probe(req->devargs, &dev)`
The above function will set `dev` pointer to the
proper device *unless* it returns with error. One of
those errors is -EEXIST, which the hotplug function
explicitly ignores. For -EEXIST, it proceeds with
attaching the device and expects the dev pointer to
be valid.
This patch makes `local_dev_probe` set the dev pointer
even if it returns -EEXIST.
Fixes: ac9e4a17370f ("eal: support attach/detach shared device from secondary")
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
If a device fails to attach before it's plugged,
the subsequent rollback will still try to detach it,
causing a segfault. Unplugging a device that wasn't
plugged isn't really supported, so this patch adds
an extra error check to prevent that from happening.
While here, fix this also for normal (non-rollback)
detach, which could also theoretically segfault on
non-plugged device.
Fixes: 244d5130719c ("eal: enable hotplug on multi-process")
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
If rte_eal_iopl_init() will be called more than once we'll leak
the file descriptor.
Fixes: b46fe31862ec ("eal/bsd: fix virtio on FreeBSD")
Cc: stable@dpdk.org
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The packed ring defines were declared only if kernel
header does not declare them.
The problem is that they are not applied in upstream kernel,
and some changes in the names have been required.
This patch declares the defines unconditionally, which
fixes potential build issues.
Fixes: 297b1e7350f6 ("vhost: add virtio packed virtqueue defines")
Cc: stable@dpdk.org
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
