Callbacks will be triggered just after allocation and just
before deallocation, to ensure that memory address space
referenced in the callback is always valid by the time
callback is called.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Each process will have its own callbacks. Callbacks will indicate
whether it's allocation and deallocation that's happened, and will
also provide start VA address and length of allocated block.
Since memory hotplug isn't supported on FreeBSD and in legacy mem
mode, it will not be possible to register them in either.
Callbacks are called whenever something happens to the memory map of
current process, therefore at those times memory hotplug subsystem
is write-locked, which leads to deadlocks on attempt to use these
functions. Document the limitation.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This enables multiprocess synchronization for memory hotplug
requests at runtime (as opposed to initialization).
Basic workflow is the following. Primary process always does initial
mapping and unmapping, and secondary processes always follow primary
page map. Only one allocation request can be active at any one time.
When primary allocates memory, it ensures that all other processes
have allocated the same set of hugepages successfully, otherwise
any allocations made are being rolled back, and heap is freed back.
Heap is locked throughout the process, and there is also a global
memory hotplug lock, so no race conditions can happen.
When primary frees memory, it frees the heap, deallocates affected
pages, and notifies other processes of deallocations. Since heap is
freed from that memory chunk, the area basically becomes invisible
to other processes even if they happen to fail to unmap that
specific set of pages, so it's completely safe to ignore results of
sync requests.
When secondary allocates memory, it does not do so by itself.
Instead, it sends a request to primary process to try and allocate
pages of specified size and on specified socket, such that a
specified heap allocation request could complete. Primary process
then sends all secondaries (including the requestor) a separate
notification of allocated pages, and expects all secondary
processes to report success before considering pages as "allocated".
Only after primary process ensures that all memory has been
successfully allocated in all secondary process, it will respond
positively to the initial request, and let secondary proceed with
the allocation. Since the heap now has memory that can satisfy
allocation request, and it was locked all this time (so no other
allocations could take place), secondary process will be able to
allocate memory from the heap.
When secondary frees memory, it hides pages to be deallocated from
the heap. Then, it sends a deallocation request to primary process,
so that it deallocates pages itself, and then sends a separate sync
request to all other processes (including the requestor) to unmap
the same pages. This way, even if secondary fails to notify other
processes of this deallocation, that memory will become invisible
to other processes, and will not be allocated from again.
So, to summarize: address space will only become part of the heap
if primary process can ensure that all other processes have
allocated this memory successfully. If anything goes wrong, the
worst thing that could happen is that a page will "leak" and will
not be available to neither DPDK nor the system, as some process
will still hold onto it. It's not an actual leak, as we can account
for the page - it's just that none of the processes will be able
to use this page for anything useful, until it gets allocated from
by the primary.
Due to underlying DPDK IPC implementation being single-threaded,
some asynchronous magic had to be done, as we need to complete
several requests before we can definitively allow secondary process
to use allocated memory (namely, it has to be present in all other
secondary processes before it can be used). Additionally, only
one allocation request is allowed to be submitted at once.
Memory allocation requests are only allowed when there are no
secondary processes currently initializing. To enforce that,
a shared rwlock is used, that is set to read lock on init (so that
several secondaries could initialize concurrently), and write lock
on making allocation requests (so that either secondary init will
have to wait, or allocation request will have to wait until all
processes have initialized).
Any other function that wishes to iterate over memory or prevent
allocations should be using memory hotplug lock.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This set of changes enables rte_malloc to allocate and free memory
as needed. Currently, it is disabled because legacy mem mode is
enabled unconditionally.
The way it works is, first malloc checks if there is enough memory
already allocated to satisfy user's request. If there isn't, we try
and allocate more memory. The reverse happens with free - we free
an element, check its size (including free element merging due to
adjacency) and see if it's bigger than hugepage size and that its
start and end span a hugepage or more. Then we remove the area from
malloc heap (adjusting element lengths where appropriate), and
deallocate the page.
For legacy mode, runtime alloc/free of pages is disabled.
It is worth noting that memseg lists are being sorted by page size,
and that we try our best to satisfy user's request. That is, if
the user requests an element from a 2MB page memory, we will check
if we can satisfy that request from existing memory, if not we try
and allocate more 2MB pages. If that fails and user also specified
a "size is hint" flag, we then check other page sizes and try to
allocate from there. If that fails too, then, depending on flags,
we may try allocating from other sockets. In other words, we try
our best to give the user what they asked for, but going to other
sockets is last resort - first we try to allocate more memory on
the same socket.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Since we are going to need to map hugepages in both primary and
secondary processes, we need to know where we should look for
hugetlbfs mountpoints. So, share those with secondary processes,
and map them on init.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Add a new (non-legacy) memory init path for EAL. It uses the
new memory hotplug facilities.
If no -m or --socket-mem switches were specified, the new init
will not allocate anything, whereas if those switches were passed,
appropriate amounts of pages would be requested, just like for
legacy init.
Allocated pages will be physically discontiguous (or rather, they're
not guaranteed to be physically contiguous - they may still be so by
accident) unless RTE_IOVA_VA mode is used.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
For non-legacy memory init mode, instead of looking at generic
sysfs path, look at sysfs paths pertaining to each NUMA node
for hugepage counts. Note that per-NUMA node path does not
provide information regarding reserved pages, so we might not
get the best info from these paths, but this saves us from the
whole mapping/remapping business before we're actually able to
tell which page is on which socket, because we no longer require
our memory to be physically contiguous.
Legacy memory init will not use this.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
In preparation for implementing multiprocess support, we are adding
a version number to memseg lists. We will not need any locks, because
memory hotplug will have a global lock (so any time memory map and
thus version number might change, we will already be holding a lock).
There are two ways of implementing multiprocess support for memory
hotplug: either all information about mapped memory is shared
between processes, and secondary processes simply attempt to
map/unmap memory based on requests from the primary, or secondary
processes store their own maps and only check if they are in sync
with the primary process' maps.
This implementation will opt for the latter option: primary process
shared mappings will be authoritative, and each secondary process
will use its own interal view of mapped memory, and will attempt
to synchronize on these mappings using versioning.
Under this model, only primary process will decide which pages get
mapped, and secondary processes will only copy primary's page
maps and get notified of the changes via IPC mechanism (coming
in later commits).
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
For now, memory is always contiguous because legacy mem mode is
enabled unconditionally, but this function will be helpful down
the line when we implement support for allocating physically
non-contiguous memory. We can no longer guarantee physically
contiguous memory unless we're in legacy or IOVA_AS_VA mode, but
we can certainly try and see if we succeed.
In addition, this would be useful for e.g. PMD's who may allocate
chunks that are smaller than the pagesize, but they must not cross
the page boundary, in which case we will be able to accommodate
that request. This function will also support non-hugepage memory.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Currently, DPDK stores all pages as separate files in hugetlbfs.
This option will allow storing all pages in one file (one file
per memseg list).
We do this by using fallocate() calls on FreeBSD, however this is
only supported on fairly recent (4.3+) kernels, so ftruncate()
fallback is provided to grow (but not shrink) hugepage files.
Naming scheme is deterministic, so both primary and secondary
processes will be able to easily map needed files and offsets.
For multi-file segments, we can close fd's right away. For
single-file segments, we can reuse the same fd and reduce the
amount of fd's needed to map/use hugepages. However, we need to
store the fd's somewhere, so we add a tailq.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This isn't used anywhere yet, but the support is now there. Also,
adding cleanup to allocation procedures, so that if we fail to
allocate everything we asked for, we can free all of it back.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Nothing uses this code yet. The bulk of it is copied from old
memory allocation code (linuxapp eal_memory.c). We provide an
EAL-internal API to allocate either one page or multiple pages,
guaranteeing that we'll get contiguous VA for all of the pages
that we requested.
Not supported on FreeBSD.
Locking is done via fcntl() because that way, when it comes to
taking out write locks or unlocking on deallocation, we don't
have to keep original fd's around. Plus, using fcntl() gives us
ability to lock parts of a file, which is useful for single-file
segments, which are coming down the line.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
It's there, so we might as well use it. Some operations will be
sped up by that.
Since we have to allocate an fbarray for memzones, we have to do
it before we initialize memory subsystem, because that, in
secondary processes, will (later) allocate more fbarrays than the
primary process, which will result in inability to attach to
memzone fbarray if we do it after the fact.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Before, we were aggregating multiple pages into one memseg, so the
number of memsegs was small. Now, each page gets its own memseg,
so the list of memsegs is huge. To accommodate the new memseg list
size and to keep the under-the-hood workings sane, the memseg list
is now not just a single list, but multiple lists. To be precise,
each hugepage size available on the system gets one or more memseg
lists, per socket.
In order to support dynamic memory allocation, we reserve all
memory in advance (unless we're in 32-bit legacy mode, in which
case we do not preallocate memory). As in, we do an anonymous
mmap() of the entire maximum size of memory per hugepage size, per
socket (which is limited to either RTE_MAX_MEMSEG_PER_TYPE pages or
RTE_MAX_MEM_MB_PER_TYPE megabytes worth of memory, whichever is the
smaller one), split over multiple lists (which are limited to
either RTE_MAX_MEMSEG_PER_LIST memsegs or RTE_MAX_MEM_MB_PER_LIST
megabytes per list, whichever is the smaller one). There is also
a global limit of CONFIG_RTE_MAX_MEM_MB megabytes, which is mainly
used for 32-bit targets to limit amounts of preallocated memory,
but can be used to place an upper limit on total amount of VA
memory that can be allocated by DPDK application.
So, for each hugepage size, we get (by default) up to 128G worth
of memory, per socket, split into chunks of up to 32G in size.
The address space is claimed at the start, in eal_common_memory.c.
The actual page allocation code is in eal_memalloc.c (Linux-only),
and largely consists of copied EAL memory init code.
Pages in the list are also indexed by address. That is, in order
to figure out where the page belongs, one can simply look at base
address for a memseg list. Similarly, figuring out IOVA address
of a memzone is a matter of finding the right memseg list, getting
offset and dividing by page size to get the appropriate memseg.
This commit also removes rte_eal_dump_physmem_layout() call,
according to deprecation notice [1], and removes that deprecation
notice as well.
On 32-bit targets due to limited VA space, DPDK will no longer
spread memory to different sockets like before. Instead, it will
(by default) allocate all of the memory on socket where master
lcore is. To override this behavior, --socket-mem must be used.
The rest of the changes are really ripple effects from the memseg
change - heap changes, compile fixes, and rewrites to support
fbarray-backed memseg lists. Due to earlier switch to _walk()
functions, most of the changes are simple fixes, however some
of the _walk() calls were switched to memseg list walk, where
it made sense to do so.
Additionally, we are also switching locks from flock() to fcntl().
Down the line, we will be introducing single-file segments option,
and we cannot use flock() locks to lock parts of the file. Therefore,
we will use fcntl() locks for legacy mem as well, in case someone is
unfortunate enough to accidentally start legacy mem primary process
alongside an already working non-legacy mem-based primary process.
[1] http://dpdk.org/dev/patchwork/patch/34002/
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
rte_fbarray is a simple indexed array stored in shared memory
via mapping files into memory. Rationale for its existence is the
following: since we are going to map memory page-by-page, there
could be quite a lot of memory segments to keep track of (for
smaller page sizes, page count can easily reach thousands). We
can't really make page lists truly dynamic and infinitely expandable,
because that involves reallocating memory (which is a big no-no in
multiprocess). What we can do instead is have a maximum capacity as
something really, really large, and decide at allocation time how
big the array is going to be. We map the entire file into memory,
which makes it possible to use fbarray as shared memory, provided
the structure itself is allocated in shared memory. Per-fbarray
locking is also used to avoid index data races (but not contents
data races - that is up to user application to synchronize).
In addition, in understanding that we will frequently need to scan
this array for free space and iterating over array linearly can
become slow, rte_fbarray provides facilities to index array's
usage. The following use cases are covered:
- find next free/used slot (useful either for adding new elements
to fbarray, or walking the list)
- find starting index for next N free/used slots (useful for when
we want to allocate chunk of VA-contiguous memory composed of
several pages)
- find how many contiguous free/used slots there are, starting
from specified index (useful for when we want to figure out
how many pages we have until next hole in allocated memory, to
speed up some bulk operations where we would otherwise have to
walk the array and add pages one by one)
This is accomplished by storing a usage mask in-memory, right
after the data section of the array, and using some bit-level
magic to figure out the info we need.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This adds a "--legacy-mem" command-line switch. It will be used to
go back to the old memory behavior, one where we can't dynamically
allocate/free memory (the downside), but one where the user can
get physically contiguous memory, like before (the upside).
For now, nothing but the legacy behavior exists, non-legacy
memory init sequence will be added later. For FreeBSD, non-legacy
memory init will never be enabled, while for Linux, it is
disabled in this patch to avoid breaking bisect, but will be
enabled once non-legacy mode will be fully operational.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Currently it is not possible to use memory that is not owned by DPDK to
perform DMA. This scenarion might be used in vhost applications (like
SPDK) where guest send its own memory table. To fill this gap provide
API to allow registering arbitrary address in VFIO container.
Signed-off-by: Pawel Wodkowski <pawelx.wodkowski@intel.com>
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Signed-off-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This can be used as a virt2iova function that only looks up
memory that is owned by DPDK (as opposed to doing pagemap walks).
Using this will result in less dependency on internals of mem API.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This is reverse lookup of PA to VA. Using this will make
other code less dependent on internals of mem API.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
This function is meant to walk over first segment of each
VA-contiguous group of memsegs.
For future users of this function, this is done so that
there is less dependency on internals of mem API and less
noise later change sets.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
For code that might need to iterate over list of allocated
segments, using this API will make it more resilient to
internal API changes and will prevent copying the same
iteration code over and over again.
Additionally, down the line there will be locking implemented,
so users of this API will not need to care about locking
either.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
If a user has specified that the zone should have contiguous memory,
add a memzone flag to request contiguous memory. Otherwise, account
for the fact that unless we're in IOVA_AS_VA mode, we cannot
guarantee that the pages would be physically contiguous, so we
calculate the memzone size and alignments as if we were getting
the smallest page size available.
However, for the non-IOVA contiguous case, existing mempool size
calculation function doesn't give us expected results, because it
will return memzone sizes aligned to page size (e.g. a 1MB mempool
may use an entire 1GB page), therefore in cases where we weren't
specifically asked to reserve non-contiguous memory, first try
reserving a memzone as IOVA-contiguous, and if that fails, then
try reserving with page-aligned size/alignment.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Tested-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Tested-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
