[dpdk-dev] [PATCH v2 28/41] eal: add support for multiprocess memory hotplug

Anatoly Burakov anatoly.burakov at intel.com
Wed Mar 7 17:56:56 CET 2018


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, 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).

Signed-off-by: Anatoly Burakov <anatoly.burakov at intel.com>
---

Notes:
    v2: - fixed deadlocking on init problem
        - reverted rte_panic changes (fixed by changes in IPC instead)
    
    This problem is evidently complex to solve without multithreaded
    IPC implementation. An alternative approach would be to process
    each individual message in its own thread (or at least spawn a
    thread per incoming request) - that way, we can send requests
    while responding to another request, and this problem becomes
    trivial to solve (and in fact it was solved that way initially,
    before my aversion to certain other programming languages kicked
    in).
    
    Is the added complexity worth saving a couple of thread spin-ups
    here and there?

 lib/librte_eal/bsdapp/eal/Makefile                |   1 +
 lib/librte_eal/common/eal_common_memory.c         |  16 +-
 lib/librte_eal/common/include/rte_eal_memconfig.h |   3 +
 lib/librte_eal/common/malloc_heap.c               | 255 ++++++--
 lib/librte_eal/common/malloc_mp.c                 | 723 ++++++++++++++++++++++
 lib/librte_eal/common/malloc_mp.h                 |  86 +++
 lib/librte_eal/common/meson.build                 |   1 +
 lib/librte_eal/linuxapp/eal/Makefile              |   1 +
 8 files changed, 1040 insertions(+), 46 deletions(-)
 create mode 100644 lib/librte_eal/common/malloc_mp.c
 create mode 100644 lib/librte_eal/common/malloc_mp.h

diff --git a/lib/librte_eal/bsdapp/eal/Makefile b/lib/librte_eal/bsdapp/eal/Makefile
index 907e30d..250d5c1 100644
--- a/lib/librte_eal/bsdapp/eal/Makefile
+++ b/lib/librte_eal/bsdapp/eal/Makefile
@@ -59,6 +59,7 @@ SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += eal_common_fbarray.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += rte_malloc.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += malloc_elem.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += malloc_heap.c
+SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += malloc_mp.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += rte_keepalive.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += rte_service.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_BSDAPP) += rte_reciprocal.c
diff --git a/lib/librte_eal/common/eal_common_memory.c b/lib/librte_eal/common/eal_common_memory.c
index a571e24..0a0aa88 100644
--- a/lib/librte_eal/common/eal_common_memory.c
+++ b/lib/librte_eal/common/eal_common_memory.c
@@ -543,24 +543,34 @@ rte_mem_lock_page(const void *virt)
 int
 rte_eal_memory_init(void)
 {
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 	int retval;
 	RTE_LOG(DEBUG, EAL, "Setting up physically contiguous memory...\n");
 
+	if (!mcfg)
+		return -1;
+
+	/* lock mem hotplug here, to prevent races while we init */
+	rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
+
 	retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
 			memseg_primary_init() :
 			memseg_secondary_init();
 
 	if (retval < 0)
-		return -1;
+		goto fail;
 
 	retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
 			rte_eal_hugepage_init() :
 			rte_eal_hugepage_attach();
 	if (retval < 0)
-		return -1;
+		goto fail;
 
 	if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
-		return -1;
+		goto fail;
 
 	return 0;
+fail:
+	rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
+	return -1;
 }
diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h
index d653d57..c4b36f6 100644
--- a/lib/librte_eal/common/include/rte_eal_memconfig.h
+++ b/lib/librte_eal/common/include/rte_eal_memconfig.h
@@ -60,6 +60,9 @@ struct rte_mem_config {
 	rte_rwlock_t qlock;   /**< used for tailq operation for thread safe. */
 	rte_rwlock_t mplock;  /**< only used by mempool LIB for thread-safe. */
 
+	rte_rwlock_t memory_hotplug_lock;
+	/**< indicates whether memory hotplug request is in progress. */
+
 	/* memory segments and zones */
 	struct rte_fbarray memzones; /**< Memzone descriptors. */
 
diff --git a/lib/librte_eal/common/malloc_heap.c b/lib/librte_eal/common/malloc_heap.c
index 7a3d0f3..9935238 100644
--- a/lib/librte_eal/common/malloc_heap.c
+++ b/lib/librte_eal/common/malloc_heap.c
@@ -10,6 +10,7 @@
 #include <sys/queue.h>
 
 #include <rte_memory.h>
+#include <rte_errno.h>
 #include <rte_eal.h>
 #include <rte_eal_memconfig.h>
 #include <rte_launch.h>
@@ -26,6 +27,7 @@
 #include "eal_memalloc.h"
 #include "malloc_elem.h"
 #include "malloc_heap.h"
+#include "malloc_mp.h"
 
 static unsigned
 check_hugepage_sz(unsigned flags, uint64_t hugepage_sz)
@@ -81,8 +83,6 @@ malloc_heap_add_memory(struct malloc_heap *heap, struct rte_memseg_list *msl,
 
 	malloc_elem_free_list_insert(elem);
 
-	heap->total_size += len;
-
 	return elem;
 }
 
@@ -146,33 +146,42 @@ heap_alloc(struct malloc_heap *heap, const char *type __rte_unused, size_t size,
 	return elem == NULL ? NULL : (void *)(&elem[1]);
 }
 
-static int
-try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
+/* this function is exposed in malloc_mp.h */
+void
+rollback_expand_heap(struct rte_memseg **ms, int n_pages,
+		struct malloc_elem *elem, void *map_addr, size_t map_len)
+{
+	int i;
+
+	if (elem != NULL) {
+		malloc_elem_free_list_remove(elem);
+		malloc_elem_hide_region(elem, map_addr, map_len);
+	}
+
+	for (i = 0; i < n_pages; i++)
+		eal_memalloc_free_page(ms[i]);
+}
+
+/* this function is exposed in malloc_mp.h */
+struct malloc_elem *
+alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
 		int socket, unsigned int flags, size_t align, size_t bound,
-		bool contig)
+		bool contig, struct rte_memseg **ms, int n_pages)
 {
 	size_t map_len, data_start_offset;
 	struct rte_memseg_list *msl;
-	struct rte_memseg **ms;
-	struct malloc_elem *elem;
-	int i, n_pages, allocd_pages;
+	struct malloc_elem *elem = NULL;
+	int allocd_pages;
 	void *ret, *map_addr, *data_start;
 
-	align = RTE_MAX(align, MALLOC_ELEM_HEADER_LEN);
-	map_len = RTE_ALIGN_CEIL(align + elt_size + MALLOC_ELEM_TRAILER_LEN,
-			pg_sz);
-
-	n_pages = map_len / pg_sz;
+	map_len = n_pages * pg_sz;
 
-	/* we can't know in advance how many pages we'll need, so malloc */
-	ms = malloc(sizeof(*ms) * n_pages);
-
-	allocd_pages = eal_memalloc_alloc_page_bulk(ms, n_pages, pg_sz, socket,
-			true);
+	allocd_pages = eal_memalloc_alloc_page_bulk(ms, n_pages, pg_sz,
+			socket, true);
 
 	/* make sure we've allocated our pages... */
 	if (allocd_pages != n_pages)
-		goto free_ms;
+		return NULL;
 
 	map_addr = ms[0]->addr;
 	msl = rte_mem_virt2memseg_list(map_addr);
@@ -184,7 +193,7 @@ try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
 			n_pages * msl->hugepage_sz)) {
 		RTE_LOG(DEBUG, EAL, "%s(): couldn't allocate physically contiguous space\n",
 				__func__);
-		goto free_pages;
+		goto fail;
 	}
 
 	/* add newly minted memsegs to malloc heap */
@@ -195,7 +204,53 @@ try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
 			contig);
 
 	if (ret == NULL)
+		goto fail;
+
+	return elem;
+
+fail:
+	rollback_expand_heap(ms, n_pages, elem, map_addr, map_len);
+	return NULL;
+}
+
+static int
+try_expand_heap_primary(struct malloc_heap *heap, uint64_t pg_sz,
+		size_t elt_size, int socket, unsigned int flags, size_t align,
+		size_t bound, bool contig)
+{
+	struct malloc_elem *elem;
+	struct rte_memseg **ms;
+	void *map_addr;
+	size_t map_len;
+	int n_pages;
+
+	map_len = RTE_ALIGN_CEIL(align + elt_size +
+			MALLOC_ELEM_TRAILER_LEN, pg_sz);
+	n_pages = map_len / pg_sz;
+
+	/* we can't know in advance how many pages we'll need, so we malloc */
+	ms = malloc(sizeof(*ms) * n_pages);
+
+	if (ms == NULL)
+		return -1;
+
+	elem = alloc_pages_on_heap(heap, pg_sz, elt_size, socket, flags, align,
+			bound, contig, ms, n_pages);
+
+	if (elem == NULL)
+		goto free_ms;
+
+	map_addr = ms[0]->addr;
+
+	/* notify other processes that this has happened */
+	if (request_sync()) {
+		/* we couldn't ensure all processes have mapped memory,
+		 * so free it back and notify everyone that it's been
+		 * freed back.
+		 */
 		goto free_elem;
+	}
+	heap->total_size += map_len;
 
 	RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
 		socket, map_len >> 20ULL);
@@ -205,13 +260,9 @@ try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
 	return 0;
 
 free_elem:
-	malloc_elem_free_list_remove(elem);
-	malloc_elem_hide_region(elem, map_addr, map_len);
-	heap->total_size -= map_len;
+	rollback_expand_heap(ms, n_pages, elem, map_addr, map_len);
 
-free_pages:
-	for (i = 0; i < n_pages; i++)
-		eal_memalloc_free_page(ms[i]);
+	request_sync();
 free_ms:
 	free(ms);
 
@@ -219,6 +270,57 @@ try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
 }
 
 static int
+try_expand_heap_secondary(struct malloc_heap *heap, uint64_t pg_sz,
+		size_t elt_size, int socket, unsigned int flags, size_t align,
+		size_t bound, bool contig)
+{
+	struct malloc_mp_req req;
+	int req_result;
+
+	req.t = REQ_TYPE_ALLOC;
+	req.alloc_req.align = align;
+	req.alloc_req.bound = bound;
+	req.alloc_req.contig = contig;
+	req.alloc_req.flags = flags;
+	req.alloc_req.elt_size = elt_size;
+	req.alloc_req.page_sz = pg_sz;
+	req.alloc_req.socket = socket;
+	req.alloc_req.heap = heap; /* it's in shared memory */
+
+	req_result = request_to_primary(&req);
+
+	if (req_result != 0)
+		return -1;
+
+	if (req.result != REQ_RESULT_SUCCESS)
+		return -1;
+
+	return 0;
+}
+
+static int
+try_expand_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
+		int socket, unsigned int flags, size_t align, size_t bound,
+		bool contig)
+{
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+	int ret;
+
+	rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
+
+	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+		ret = try_expand_heap_primary(heap, pg_sz, elt_size, socket,
+				flags, align, bound, contig);
+	} else {
+		ret = try_expand_heap_secondary(heap, pg_sz, elt_size, socket,
+				flags, align, bound, contig);
+	}
+
+	rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
+	return ret;
+}
+
+static int
 compare_pagesz(const void *a, const void *b)
 {
 	const struct rte_memseg_list * const*mpa = a;
@@ -236,11 +338,10 @@ compare_pagesz(const void *a, const void *b)
 }
 
 static int
-alloc_mem_on_socket(size_t size, int socket, unsigned int flags, size_t align,
-		size_t bound, bool contig)
+alloc_more_mem_on_socket(struct malloc_heap *heap, size_t size, int socket,
+		unsigned int flags, size_t align, size_t bound, bool contig)
 {
 	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
-	struct malloc_heap *heap = &mcfg->malloc_heaps[socket];
 	struct rte_memseg_list *requested_msls[RTE_MAX_MEMSEG_LISTS];
 	struct rte_memseg_list *other_msls[RTE_MAX_MEMSEG_LISTS];
 	uint64_t requested_pg_sz[RTE_MAX_MEMSEG_LISTS];
@@ -355,7 +456,7 @@ heap_alloc_on_socket(const char *type, size_t size, int socket,
 
 	rte_spinlock_lock(&(heap->lock));
 
-	align = align == 0 ? 1 : align;
+	align = RTE_MAX(align == 0 ? 1 : align, MALLOC_ELEM_HEADER_LEN);
 
 	/* for legacy mode, try once and with all flags */
 	if (internal_config.legacy_mem) {
@@ -372,7 +473,8 @@ heap_alloc_on_socket(const char *type, size_t size, int socket,
 	if (ret != NULL)
 		goto alloc_unlock;
 
-	if (!alloc_mem_on_socket(size, socket, flags, align, bound, contig)) {
+	if (!alloc_more_mem_on_socket(heap, size, socket, flags, align, bound,
+			contig)) {
 		ret = heap_alloc(heap, type, size, flags, align, bound, contig);
 
 		/* this should have succeeded */
@@ -424,14 +526,40 @@ malloc_heap_alloc(const char *type, size_t size, int socket_arg,
 	return NULL;
 }
 
+/* this function is exposed in malloc_mp.h */
+int
+malloc_heap_free_pages(void *aligned_start, size_t aligned_len)
+{
+	int n_pages, page_idx, max_page_idx;
+	struct rte_memseg_list *msl;
+
+	msl = rte_mem_virt2memseg_list(aligned_start);
+	if (msl == NULL)
+		return -1;
+
+	n_pages = aligned_len / msl->hugepage_sz;
+	page_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) /
+			msl->hugepage_sz;
+	max_page_idx = page_idx + n_pages;
+
+	for (; page_idx < max_page_idx; page_idx++) {
+		struct rte_memseg *ms;
+
+		ms = rte_fbarray_get(&msl->memseg_arr, page_idx);
+		eal_memalloc_free_page(ms);
+	}
+	return 0;
+}
+
 int
 malloc_heap_free(struct malloc_elem *elem)
 {
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 	struct malloc_heap *heap;
 	void *start, *aligned_start, *end, *aligned_end;
 	size_t len, aligned_len;
 	struct rte_memseg_list *msl;
-	int n_pages, page_idx, max_page_idx, ret;
+	int n_pages, ret;
 
 	if (!malloc_elem_cookies_ok(elem) || elem->state != ELEM_BUSY)
 		return -1;
@@ -463,30 +591,60 @@ malloc_heap_free(struct malloc_elem *elem)
 	aligned_end = RTE_PTR_ALIGN_FLOOR(end, msl->hugepage_sz);
 
 	aligned_len = RTE_PTR_DIFF(aligned_end, aligned_start);
+	n_pages = aligned_len / msl->hugepage_sz;
 
 	/* can't free anything */
-	if (aligned_len < msl->hugepage_sz)
+	if (n_pages == 0)
 		goto free_unlock;
 
+	rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);
+
+	/*
+	 * we allow secondary processes to clear the heap of this allocated
+	 * memory because it is safe to do so, as even if notifications about
+	 * unmapped pages don't make it to other processes, heap is shared
+	 * across all processes, and will become empty of this memory anyway,
+	 * and nothing can allocate it back unless primary process will be able
+	 * to deliver allocation message to every single running process.
+	 */
+
 	malloc_elem_free_list_remove(elem);
 
 	malloc_elem_hide_region(elem, (void *) aligned_start, aligned_len);
 
-	/* we don't really care if we fail to deallocate memory */
-	n_pages = aligned_len / msl->hugepage_sz;
-	page_idx = RTE_PTR_DIFF(aligned_start, msl->base_va) / msl->hugepage_sz;
-	max_page_idx = page_idx + n_pages;
+	heap->total_size -= n_pages * msl->hugepage_sz;
 
-	for (; page_idx < max_page_idx; page_idx++) {
-		struct rte_memseg *ms;
+	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+		/* don't care if any of this fails */
+		malloc_heap_free_pages(aligned_start, aligned_len);
 
-		ms = rte_fbarray_get(&msl->memseg_arr, page_idx);
-		eal_memalloc_free_page(ms);
-		heap->total_size -= msl->hugepage_sz;
+		request_sync();
+	} else {
+		struct malloc_mp_req req;
+
+		req.t = REQ_TYPE_FREE;
+		req.free_req.addr = aligned_start;
+		req.free_req.len = aligned_len;
+
+		/*
+		 * we request primary to deallocate pages, but we don't do it
+		 * in this thread. instead, we notify primary that we would like
+		 * to deallocate pages, and this process will receive another
+		 * request (in parallel) that will do it for us on another
+		 * thread.
+		 *
+		 * we also don't really care if this succeeds - the data is
+		 * already removed from the heap, so it is, for all intents and
+		 * purposes, hidden from the rest of DPDK even if some other
+		 * process (including this one) may have these pages mapped.
+		 */
+		request_to_primary(&req);
 	}
 
 	RTE_LOG(DEBUG, EAL, "Heap on socket %d was shrunk by %zdMB\n",
 		msl->socket_id, aligned_len >> 20ULL);
+
+	rte_rwlock_write_unlock(&mcfg->memory_hotplug_lock);
 free_unlock:
 	rte_spinlock_unlock(&(heap->lock));
 	return ret;
@@ -576,8 +734,16 @@ rte_eal_malloc_heap_init(void)
 	int msl_idx;
 	struct rte_memseg_list *msl;
 
-	if (mcfg == NULL)
+	if (register_mp_requests()) {
+		RTE_LOG(ERR, EAL, "Couldn't register malloc multiprocess actions\n");
 		return -1;
+	}
+
+	/* unlock mem hotplug here. it's safe for primary as no requests can
+	 * even come before primary itself is fully initialized, and secondaries
+	 * do not need to initialize the heap.
+	 */
+	rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
 
 	/* secondary processes don't need to initialize heap */
 	if (rte_eal_process_type() == RTE_PROC_SECONDARY)
@@ -604,6 +770,7 @@ rte_eal_malloc_heap_init(void)
 						rte_fbarray_get(arr, ms_idx);
 				malloc_heap_add_memory(heap, msl,
 						ms->addr, ms->len);
+				heap->total_size += ms->len;
 				ms_idx++;
 				RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
 					msl->socket_id, ms->len >> 20ULL);
@@ -630,6 +797,8 @@ rte_eal_malloc_heap_init(void)
 			 */
 			malloc_heap_add_memory(heap, msl, start_seg->addr, len);
 
+			heap->total_size += len;
+
 			RTE_LOG(DEBUG, EAL, "Heap on socket %d was expanded by %zdMB\n",
 				msl->socket_id, len >> 20ULL);
 
diff --git a/lib/librte_eal/common/malloc_mp.c b/lib/librte_eal/common/malloc_mp.c
new file mode 100644
index 0000000..8052680
--- /dev/null
+++ b/lib/librte_eal/common/malloc_mp.c
@@ -0,0 +1,723 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2018 Intel Corporation
+ */
+
+#include <string.h>
+#include <sys/time.h>
+
+#include <rte_alarm.h>
+#include <rte_errno.h>
+
+#include "eal_memalloc.h"
+
+#include "malloc_elem.h"
+#include "malloc_mp.h"
+
+#define MP_ACTION_SYNC "mp_malloc_sync"
+/**< request sent by primary process to notify of changes in memory map */
+#define MP_ACTION_ROLLBACK "mp_malloc_rollback"
+/**< request sent by primary process to notify of changes in memory map. this is
+ * essentially a regular sync request, but we cannot send sync requests while
+ * another one is in progress, and we might have to - therefore, we do this as
+ * a separate callback.
+ */
+#define MP_ACTION_REQUEST "mp_malloc_request"
+/**< request sent by secondary process to ask for allocation/deallocation */
+#define MP_ACTION_RESPONSE "mp_malloc_response"
+/**< response sent to secondary process to indicate result of request */
+
+#define MP_TIMEOUT_S 5 /**< 5 seconds timeouts */
+
+/* when we're allocating, we need to store some state to ensure that we can
+ * roll back later
+ */
+struct primary_alloc_req_state {
+	struct malloc_heap *heap;
+	struct rte_memseg **ms;
+	int ms_len;
+	struct malloc_elem *elem;
+	void *map_addr;
+	size_t map_len;
+};
+
+enum req_state {
+	REQ_STATE_INACTIVE = 0,
+	REQ_STATE_ACTIVE,
+	REQ_STATE_COMPLETE
+};
+
+struct mp_request {
+	TAILQ_ENTRY(mp_request) next;
+	struct malloc_mp_req user_req; /**< contents of request */
+	pthread_cond_t cond; /**< variable we use to time out on this request */
+	enum req_state state; /**< indicate status of this request */
+	struct primary_alloc_req_state alloc_state;
+};
+
+/*
+ * We could've used just a single request, but it may be possible for
+ * secondaries to timeout earlier than the primary, and send a new request while
+ * primary is still expecting replies to the old one. Therefore, each new
+ * request will get assigned a new ID, which is how we will distinguish between
+ * expected and unexpected messages.
+ */
+TAILQ_HEAD(mp_request_list, mp_request);
+static struct {
+	struct mp_request_list list;
+	pthread_mutex_t lock;
+} mp_request_list = {
+	.list = TAILQ_HEAD_INITIALIZER(mp_request_list.list),
+	.lock = PTHREAD_MUTEX_INITIALIZER
+};
+
+/**
+ * General workflow is the following:
+ *
+ * Allocation:
+ * S: send request to primary
+ * P: attempt to allocate memory
+ *    if failed, sendmsg failure
+ *    if success, send sync request
+ * S: if received msg of failure, quit
+ *    if received sync request, synchronize memory map and reply with result
+ * P: if received sync request result
+ *    if success, sendmsg success
+ *    if failure, roll back allocation and send a rollback request
+ * S: if received msg of success, quit
+ *    if received rollback request, synchronize memory map and reply with result
+ * P: if received sync request result
+ *    sendmsg sync request result
+ * S: if received msg, quit
+ *
+ * Aside from timeouts, there are three points where we can quit:
+ *  - if allocation failed straight away
+ *  - if allocation and sync request succeeded
+ *  - if allocation succeeded, sync request failed, allocation rolled back and
+ *    rollback request received (irrespective of whether it succeeded or failed)
+ *
+ * Deallocation:
+ * S: send request to primary
+ * P: attempt to deallocate memory
+ *    if failed, sendmsg failure
+ *    if success, send sync request
+ * S: if received msg of failure, quit
+ *    if received sync request, synchronize memory map and reply with result
+ * P: if received sync request result
+ *    sendmsg sync request result
+ * S: if received msg, quit
+ *
+ * There is no "rollback" from deallocation, as it's safe to have some memory
+ * mapped in some processes - it's absent from the heap, so it won't get used.
+ */
+
+static struct mp_request *
+find_request_by_id(uint64_t id)
+{
+	struct mp_request *req;
+	TAILQ_FOREACH(req, &mp_request_list.list, next) {
+		if (req->user_req.id == id)
+			break;
+	}
+	return req;
+}
+
+/* this ID is, like, totally guaranteed to be absolutely unique. pinky swear. */
+static uint64_t
+get_unique_id(void)
+{
+	uint64_t id;
+	do {
+		id = rte_rand();
+	} while (find_request_by_id(id) != NULL);
+	return id;
+}
+
+/* secondary will respond to sync requests thusly */
+static int
+handle_sync(const struct rte_mp_msg *msg, const void *peer)
+{
+	struct rte_mp_msg reply = {0};
+	const struct malloc_mp_req *req =
+			(const struct malloc_mp_req *)msg->param;
+	struct malloc_mp_req *resp =
+			(struct malloc_mp_req *)reply.param;
+	int ret;
+
+	if (req->t != REQ_TYPE_SYNC) {
+		RTE_LOG(ERR, EAL, "Unexpected request from primary\n");
+		return -1;
+	}
+
+	reply.num_fds = 0;
+	snprintf(reply.name, sizeof(reply.name), "%s", msg->name);
+	reply.len_param = sizeof(*resp);
+
+	ret = eal_memalloc_sync_with_primary();
+
+	resp->t = REQ_TYPE_SYNC;
+	resp->id = req->id;
+	resp->result = ret == 0 ? REQ_RESULT_SUCCESS : REQ_RESULT_FAIL;
+
+	rte_mp_reply(&reply, peer);
+
+	return 0;
+}
+
+static int
+handle_alloc_request(const struct malloc_mp_req *m,
+		struct mp_request *req)
+{
+	const struct malloc_req_alloc *ar = &m->alloc_req;
+	struct malloc_heap *heap;
+	struct malloc_elem *elem;
+	struct rte_memseg **ms;
+	size_t map_len;
+	int n_pages;
+
+	map_len = RTE_ALIGN_CEIL(ar->align + ar->elt_size +
+			MALLOC_ELEM_TRAILER_LEN, ar->page_sz);
+	n_pages = map_len / ar->page_sz;
+
+	heap = ar->heap;
+
+	/* we can't know in advance how many pages we'll need, so we malloc */
+	ms = malloc(sizeof(*ms) * n_pages);
+
+	if (ms == NULL) {
+		RTE_LOG(ERR, EAL, "Couldn't allocate memory for request state\n");
+		goto fail;
+	}
+
+	elem = alloc_pages_on_heap(heap, ar->page_sz, ar->elt_size, ar->socket,
+			ar->flags, ar->align, ar->bound, ar->contig, ms,
+			n_pages);
+
+	if (elem == NULL)
+		goto fail;
+
+	/* we have succeeded in allocating memory, but we still need to sync
+	 * with other processes. however, since DPDK IPC is single-threaded, we
+	 * send an asynchronous request and exit this callback.
+	 */
+
+	req->alloc_state.ms = ms;
+	req->alloc_state.ms_len = n_pages;
+	req->alloc_state.map_addr = ms[0]->addr;
+	req->alloc_state.map_len = map_len;
+	req->alloc_state.elem = elem;
+	req->alloc_state.heap = heap;
+
+	return 0;
+fail:
+	free(ms);
+	return -1;
+}
+
+/* first stage of primary handling requests from secondary */
+static int
+handle_request(const struct rte_mp_msg *msg, const void *peer __rte_unused)
+{
+	const struct malloc_mp_req *m =
+			(const struct malloc_mp_req *)msg->param;
+	struct mp_request *entry;
+	int ret;
+
+	/* lock access to request */
+	pthread_mutex_lock(&mp_request_list.lock);
+
+	/* make sure it's not a dupe */
+	entry = find_request_by_id(m->id);
+	if (entry != NULL) {
+		RTE_LOG(ERR, EAL, "Duplicate request id\n");
+		goto fail;
+	}
+
+	entry = malloc(sizeof(*entry));
+	if (entry == NULL) {
+		RTE_LOG(ERR, EAL, "Unable to allocate memory for request\n");
+		goto fail;
+	}
+
+	/* erase all data */
+	memset(entry, 0, sizeof(*entry));
+
+	if (m->t == REQ_TYPE_ALLOC) {
+		ret = handle_alloc_request(m, entry);
+	} else if (m->t == REQ_TYPE_FREE) {
+		ret = malloc_heap_free_pages(m->free_req.addr,
+				m->free_req.len);
+	} else {
+		RTE_LOG(ERR, EAL, "Unexpected request from secondary\n");
+		goto fail;
+	}
+
+	if (ret != 0) {
+		struct rte_mp_msg resp_msg;
+		struct malloc_mp_req *resp =
+				(struct malloc_mp_req *)resp_msg.param;
+
+		/* send failure message straight away */
+		resp_msg.num_fds = 0;
+		resp_msg.len_param = sizeof(*resp);
+		snprintf(resp_msg.name, sizeof(resp_msg.name), "%s",
+				MP_ACTION_RESPONSE);
+
+		resp->t = m->t;
+		resp->result = REQ_RESULT_FAIL;
+		resp->id = m->id;
+
+		if (rte_mp_sendmsg(&resp_msg)) {
+			RTE_LOG(ERR, EAL, "Couldn't send response\n");
+			goto fail;
+		}
+		/* we did not modify the request */
+		free(entry);
+	} else {
+		struct rte_mp_msg sr_msg = {0};
+		struct malloc_mp_req *sr =
+				(struct malloc_mp_req *)sr_msg.param;
+		struct timespec ts;
+
+		/* we can do something, so send sync request asynchronously */
+		sr_msg.num_fds = 0;
+		sr_msg.len_param = sizeof(*sr);
+		snprintf(sr_msg.name, sizeof(sr_msg.name), "%s",
+				MP_ACTION_SYNC);
+
+		ts.tv_nsec = 0;
+		ts.tv_sec = MP_TIMEOUT_S;
+
+		/* sync requests carry no data */
+		sr->t = REQ_TYPE_SYNC;
+		sr->id = m->id;
+
+		/* there may be stray timeout still waiting */
+		do {
+			ret = rte_mp_request_async(&sr_msg, &ts);
+		} while (ret != 0 && rte_errno == EEXIST);
+		if (ret != 0) {
+			RTE_LOG(ERR, EAL, "Couldn't send sync request\n");
+			if (m->t == REQ_TYPE_ALLOC)
+				free(entry->alloc_state.ms);
+			goto fail;
+		}
+
+		/* mark request as in progress */
+		memcpy(&entry->user_req, m, sizeof(*m));
+		entry->state = REQ_STATE_ACTIVE;
+
+		TAILQ_INSERT_TAIL(&mp_request_list.list, entry, next);
+	}
+	pthread_mutex_unlock(&mp_request_list.lock);
+	return 0;
+fail:
+	pthread_mutex_unlock(&mp_request_list.lock);
+	free(entry);
+	return -1;
+}
+
+/* callback for asynchronous sync requests for primary. this will either do a
+ * sendmsg with results, or trigger rollback request.
+ */
+static int
+handle_sync_response(const struct rte_mp_msg *request,
+		const struct rte_mp_reply *reply)
+{
+	enum malloc_req_result result;
+	struct mp_request *entry;
+	const struct malloc_mp_req *mpreq =
+			(const struct malloc_mp_req *)request->param;
+	int i;
+
+	/* lock the request */
+	pthread_mutex_lock(&mp_request_list.lock);
+
+	entry = find_request_by_id(mpreq->id);
+	if (entry == NULL) {
+		RTE_LOG(ERR, EAL, "Wrong request ID\n");
+		goto fail;
+	}
+
+	result = REQ_RESULT_SUCCESS;
+
+	if (reply->nb_received != reply->nb_sent)
+		result = REQ_RESULT_FAIL;
+
+	for (i = 0; i < reply->nb_received; i++) {
+		struct malloc_mp_req *resp =
+				(struct malloc_mp_req *)reply->msgs[i].param;
+
+		if (resp->t != REQ_TYPE_SYNC) {
+			RTE_LOG(ERR, EAL, "Unexpected response to sync request\n");
+			result = REQ_RESULT_FAIL;
+			break;
+		}
+		if (resp->id != entry->user_req.id) {
+			RTE_LOG(ERR, EAL, "Response to wrong sync request\n");
+			result = REQ_RESULT_FAIL;
+			break;
+		}
+		if (resp->result == REQ_RESULT_FAIL) {
+			result = REQ_RESULT_FAIL;
+			break;
+		}
+	}
+
+	if (entry->user_req.t == REQ_TYPE_FREE) {
+		struct rte_mp_msg msg = {0};
+		struct malloc_mp_req *resp = (struct malloc_mp_req *)msg.param;
+
+		/* this is a free request, just sendmsg result */
+		resp->t = REQ_TYPE_FREE;
+		resp->result = result;
+		resp->id = entry->user_req.id;
+		msg.num_fds = 0;
+		msg.len_param = sizeof(*resp);
+		snprintf(msg.name, sizeof(msg.name), "%s", MP_ACTION_RESPONSE);
+
+		if (rte_mp_sendmsg(&msg))
+			RTE_LOG(ERR, EAL, "Could not send message to secondary process\n");
+
+		TAILQ_REMOVE(&mp_request_list.list, entry, next);
+		free(entry);
+	} else if (entry->user_req.t == REQ_TYPE_ALLOC &&
+			result == REQ_RESULT_SUCCESS) {
+		struct malloc_heap *heap = entry->alloc_state.heap;
+		struct rte_mp_msg msg = {0};
+		struct malloc_mp_req *resp =
+				(struct malloc_mp_req *)msg.param;
+
+		heap->total_size += entry->alloc_state.map_len;
+
+		/* result is success, so just notify secondary about this */
+		resp->t = REQ_TYPE_ALLOC;
+		resp->result = result;
+		resp->id = entry->user_req.id;
+		msg.num_fds = 0;
+		msg.len_param = sizeof(*resp);
+		snprintf(msg.name, sizeof(msg.name), "%s", MP_ACTION_RESPONSE);
+
+		if (rte_mp_sendmsg(&msg))
+			RTE_LOG(ERR, EAL, "Could not send message to secondary process\n");
+
+		TAILQ_REMOVE(&mp_request_list.list, entry, next);
+		free(entry->alloc_state.ms);
+		free(entry);
+	} else if (entry->user_req.t == REQ_TYPE_ALLOC &&
+			result == REQ_RESULT_FAIL) {
+		struct rte_mp_msg rb_msg = {0};
+		struct malloc_mp_req *rb =
+				(struct malloc_mp_req *)rb_msg.param;
+		struct timespec ts;
+		struct primary_alloc_req_state *state =
+				&entry->alloc_state;
+		int ret;
+
+		/* we've failed to sync, so do a rollback */
+		rollback_expand_heap(state->ms, state->ms_len, state->elem,
+				state->map_addr, state->map_len);
+
+		/* send rollback request */
+		rb_msg.num_fds = 0;
+		rb_msg.len_param = sizeof(*rb);
+		snprintf(rb_msg.name, sizeof(rb_msg.name), "%s",
+				MP_ACTION_ROLLBACK);
+
+		ts.tv_nsec = 0;
+		ts.tv_sec = MP_TIMEOUT_S;
+
+		/* sync requests carry no data */
+		rb->t = REQ_TYPE_SYNC;
+		rb->id = entry->user_req.id;
+
+		/* there may be stray timeout still waiting */
+		do {
+			ret = rte_mp_request_async(&rb_msg, &ts);
+		} while (ret != 0 && rte_errno == EEXIST);
+		if (ret != 0) {
+			RTE_LOG(ERR, EAL, "Could not send rollback request to secondary process\n");
+
+			/* we couldn't send rollback request, but that's OK -
+			 * secondary will time out, and memory has been removed
+			 * from heap anyway.
+			 */
+			TAILQ_REMOVE(&mp_request_list.list, entry, next);
+			free(state->ms);
+			free(entry);
+			goto fail;
+		}
+	} else {
+		RTE_LOG(ERR, EAL, " to sync request of unknown type\n");
+		goto fail;
+	}
+
+	pthread_mutex_unlock(&mp_request_list.lock);
+	return 0;
+fail:
+	pthread_mutex_unlock(&mp_request_list.lock);
+	return -1;
+}
+
+static int
+handle_rollback_response(const struct rte_mp_msg *request,
+		const struct rte_mp_reply *reply __rte_unused)
+{
+	struct rte_mp_msg msg = {0};
+	struct malloc_mp_req *resp = (struct malloc_mp_req *)msg.param;
+	const struct malloc_mp_req *mpreq =
+			(const struct malloc_mp_req *)request->param;
+	struct mp_request *entry;
+
+	/* lock the request */
+	pthread_mutex_lock(&mp_request_list.lock);
+
+	entry = find_request_by_id(mpreq->id);
+	if (entry == NULL) {
+		RTE_LOG(ERR, EAL, "Wrong request ID\n");
+		goto fail;
+	}
+
+	if (entry->user_req.t != REQ_TYPE_ALLOC) {
+		RTE_LOG(ERR, EAL, "Unexpected active request\n");
+		goto fail;
+	}
+
+	/* we don't care if rollback succeeded, request still failed */
+	resp->t = REQ_TYPE_ALLOC;
+	resp->result = REQ_RESULT_FAIL;
+	resp->id = mpreq->id;
+	msg.num_fds = 0;
+	msg.len_param = sizeof(*resp);
+	snprintf(msg.name, sizeof(msg.name), "%s", MP_ACTION_RESPONSE);
+
+	if (rte_mp_sendmsg(&msg))
+		RTE_LOG(ERR, EAL, "Could not send message to secondary process\n");
+
+	/* clean up */
+	TAILQ_REMOVE(&mp_request_list.list, entry, next);
+	free(entry->alloc_state.ms);
+	free(entry);
+
+	pthread_mutex_unlock(&mp_request_list.lock);
+	return 0;
+fail:
+	pthread_mutex_unlock(&mp_request_list.lock);
+	return -1;
+}
+
+/* final stage of the request from secondary */
+static int
+handle_response(const struct rte_mp_msg *msg, const void *peer  __rte_unused)
+{
+	const struct malloc_mp_req *m =
+			(const struct malloc_mp_req *)msg->param;
+	struct mp_request *entry;
+
+	pthread_mutex_lock(&mp_request_list.lock);
+
+	entry = find_request_by_id(m->id);
+	if (entry != NULL) {
+		/* update request status */
+		entry->user_req.result = m->result;
+
+		entry->state = REQ_STATE_COMPLETE;
+
+		/* trigger thread wakeup */
+		pthread_cond_signal(&entry->cond);
+	}
+
+	pthread_mutex_unlock(&mp_request_list.lock);
+
+	return 0;
+}
+
+/* synchronously request memory map sync, this is only called whenever primary
+ * process initiates the allocation.
+ */
+int
+request_sync(void)
+{
+	struct rte_mp_msg msg = {0};
+	struct rte_mp_reply reply = {0};
+	struct malloc_mp_req *req = (struct malloc_mp_req *)msg.param;
+	struct timespec ts;
+	int i, ret;
+
+	/* no need to create tailq entries as this is entirely synchronous */
+
+	msg.num_fds = 0;
+	msg.len_param = sizeof(*req);
+	snprintf(msg.name, sizeof(msg.name), "%s", MP_ACTION_SYNC);
+
+	/* sync request carries no data */
+	req->t = REQ_TYPE_SYNC;
+	req->id = get_unique_id();
+
+	ts.tv_nsec = 0;
+	ts.tv_sec = MP_TIMEOUT_S;
+
+	/* there may be stray timeout still waiting */
+	do {
+		ret = rte_mp_request(&msg, &reply, &ts);
+	} while (ret != 0 && rte_errno == EEXIST);
+	if (ret != 0) {
+		RTE_LOG(ERR, EAL, "Could not send sync request to secondary process\n");
+		ret = -1;
+		goto out;
+	}
+
+	if (reply.nb_received != reply.nb_sent) {
+		RTE_LOG(ERR, EAL, "Not all secondaries have responded\n");
+		ret = -1;
+		goto out;
+	}
+
+	for (i = 0; i < reply.nb_received; i++) {
+		struct malloc_mp_req *resp =
+				(struct malloc_mp_req *)reply.msgs[i].param;
+		if (resp->t != REQ_TYPE_SYNC) {
+			RTE_LOG(ERR, EAL, "Unexpected response from secondary\n");
+			ret = -1;
+			goto out;
+		}
+		if (resp->id != req->id) {
+			RTE_LOG(ERR, EAL, "Wrong request ID\n");
+			ret = -1;
+			goto out;
+		}
+		if (resp->result != REQ_RESULT_SUCCESS) {
+			RTE_LOG(ERR, EAL, "Secondary process failed to synchronize\n");
+			ret = -1;
+			goto out;
+		}
+	}
+
+	ret = 0;
+out:
+	free(reply.msgs);
+	return ret;
+}
+
+/* this is a synchronous wrapper around a bunch of asynchronous requests to
+ * primary process. this will initiate a request and wait until responses come.
+ */
+int
+request_to_primary(struct malloc_mp_req *user_req)
+{
+	struct rte_mp_msg msg = {0};
+	struct malloc_mp_req *msg_req = (struct malloc_mp_req *)msg.param;
+	struct mp_request *entry;
+	struct timespec ts = {0};
+	struct timeval now;
+	int ret;
+
+	pthread_mutex_lock(&mp_request_list.lock);
+
+	entry = malloc(sizeof(*entry));
+	if (entry == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot allocate memory for request\n");
+		goto fail;
+	}
+
+	memset(entry, 0, sizeof(*entry));
+
+	if (gettimeofday(&now, NULL) < 0) {
+		RTE_LOG(ERR, EAL, "Cannot get current time\n");
+		goto fail;
+	}
+
+	ts.tv_nsec = (now.tv_usec * 1000) % 1000000000;
+	ts.tv_sec = now.tv_sec + MP_TIMEOUT_S +
+			(now.tv_usec * 1000) / 1000000000;
+
+	/* initialize the request */
+	pthread_cond_init(&entry->cond, NULL);
+
+	msg.num_fds = 0;
+	msg.len_param = sizeof(*msg_req);
+	snprintf(msg.name, sizeof(msg.name), "%s", MP_ACTION_REQUEST);
+
+	/* (attempt to) get a unique id */
+	user_req->id = get_unique_id();
+
+	/* copy contents of user request into the message */
+	memcpy(msg_req, user_req, sizeof(*msg_req));
+
+	if (rte_mp_sendmsg(&msg)) {
+		RTE_LOG(ERR, EAL, "Cannot send message to primary\n");
+		goto fail;
+	}
+
+	/* copy contents of user request into active request */
+	memcpy(&entry->user_req, user_req, sizeof(*user_req));
+
+	/* mark request as in progress */
+	entry->state = REQ_STATE_ACTIVE;
+
+	TAILQ_INSERT_TAIL(&mp_request_list.list, entry, next);
+
+	/* finally, wait on timeout */
+	do {
+		ret = pthread_cond_timedwait(&entry->cond,
+				&mp_request_list.lock, &ts);
+	} while (ret != 0 && ret != ETIMEDOUT);
+
+	if (entry->state != REQ_STATE_COMPLETE) {
+		RTE_LOG(ERR, EAL, "Request timed out\n");
+		ret = -1;
+	} else {
+		ret = 0;
+		user_req->result = entry->user_req.result;
+	}
+	TAILQ_REMOVE(&mp_request_list.list, entry, next);
+	free(entry);
+
+	pthread_mutex_unlock(&mp_request_list.lock);
+	return ret;
+fail:
+	pthread_mutex_unlock(&mp_request_list.lock);
+	free(entry);
+	return -1;
+}
+
+int
+register_mp_requests(void)
+{
+	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+		if (rte_mp_action_register(MP_ACTION_REQUEST, handle_request)) {
+			RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
+				MP_ACTION_REQUEST);
+			return -1;
+		}
+		if (rte_mp_async_reply_register(MP_ACTION_SYNC,
+				handle_sync_response)) {
+			RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
+				MP_ACTION_SYNC);
+			return -1;
+		}
+		if (rte_mp_async_reply_register(MP_ACTION_ROLLBACK,
+				handle_rollback_response)) {
+			RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
+				MP_ACTION_ROLLBACK);
+			return -1;
+		}
+	} else {
+		if (rte_mp_action_register(MP_ACTION_SYNC, handle_sync)) {
+			RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
+				MP_ACTION_SYNC);
+			return -1;
+		}
+		if (rte_mp_action_register(MP_ACTION_ROLLBACK, handle_sync)) {
+			RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
+				MP_ACTION_SYNC);
+			return -1;
+		}
+		if (rte_mp_action_register(MP_ACTION_RESPONSE,
+				handle_response)) {
+			RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
+				MP_ACTION_RESPONSE);
+			return -1;
+		}
+	}
+	return 0;
+}
diff --git a/lib/librte_eal/common/malloc_mp.h b/lib/librte_eal/common/malloc_mp.h
new file mode 100644
index 0000000..9c79d31
--- /dev/null
+++ b/lib/librte_eal/common/malloc_mp.h
@@ -0,0 +1,86 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2018 Intel Corporation
+ */
+
+#ifndef MALLOC_MP_H
+#define MALLOC_MP_H
+
+#include <stdbool.h>
+#include <stdint.h>
+
+#include <rte_common.h>
+#include <rte_random.h>
+#include <rte_spinlock.h>
+#include <rte_tailq.h>
+
+/* forward declarations */
+struct malloc_heap;
+struct rte_memseg;
+
+/* multiprocess synchronization structures for malloc */
+enum malloc_req_type {
+	REQ_TYPE_ALLOC,     /**< ask primary to allocate */
+	REQ_TYPE_FREE,      /**< ask primary to free */
+	REQ_TYPE_SYNC       /**< ask secondary to synchronize its memory map */
+};
+
+enum malloc_req_result {
+	REQ_RESULT_SUCCESS,
+	REQ_RESULT_FAIL
+};
+
+struct malloc_req_alloc {
+	struct malloc_heap *heap;
+	uint64_t page_sz;
+	size_t elt_size;
+	int socket;
+	unsigned int flags;
+	size_t align;
+	size_t bound;
+	bool contig;
+};
+
+struct malloc_req_free {
+	RTE_STD_C11
+	union {
+		void *addr;
+		uint64_t addr_64;
+	};
+	uint64_t len;
+};
+
+struct malloc_mp_req {
+	enum malloc_req_type t;
+	RTE_STD_C11
+	union {
+		struct malloc_req_alloc alloc_req;
+		struct malloc_req_free free_req;
+	};
+	uint64_t id; /**< not to be populated by caller */
+	enum malloc_req_result result;
+};
+
+int
+register_mp_requests(void);
+
+int
+request_to_primary(struct malloc_mp_req *req);
+
+/* synchronous memory map sync request */
+int
+request_sync(void);
+
+/* functions from malloc_heap exposed here */
+int
+malloc_heap_free_pages(void *aligned_start, size_t aligned_len);
+
+struct malloc_elem *
+alloc_pages_on_heap(struct malloc_heap *heap, uint64_t pg_sz, size_t elt_size,
+		int socket, unsigned int flags, size_t align, size_t bound,
+		bool contig, struct rte_memseg **ms, int n_pages);
+
+void
+rollback_expand_heap(struct rte_memseg **ms, int n_pages,
+		struct malloc_elem *elem, void *map_addr, size_t map_len);
+
+#endif // MALLOC_MP_H
diff --git a/lib/librte_eal/common/meson.build b/lib/librte_eal/common/meson.build
index a1ada24..8a3dcfe 100644
--- a/lib/librte_eal/common/meson.build
+++ b/lib/librte_eal/common/meson.build
@@ -27,6 +27,7 @@ common_sources = files(
 	'eal_common_timer.c',
 	'malloc_elem.c',
 	'malloc_heap.c',
+	'malloc_mp.c',
 	'rte_keepalive.c',
 	'rte_malloc.c',
 	'rte_reciprocal.c',
diff --git a/lib/librte_eal/linuxapp/eal/Makefile b/lib/librte_eal/linuxapp/eal/Makefile
index 5380ba8..542bf7e 100644
--- a/lib/librte_eal/linuxapp/eal/Makefile
+++ b/lib/librte_eal/linuxapp/eal/Makefile
@@ -67,6 +67,7 @@ SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += eal_common_fbarray.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += rte_malloc.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += malloc_elem.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += malloc_heap.c
+SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += malloc_mp.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += rte_keepalive.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += rte_service.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += rte_reciprocal.c
-- 
2.7.4


More information about the dev mailing list