This patch provides infrastructure for deferring buffer frees.
This is a feature ION provided which when used with some form
of a page pool, provides a nice performance boost in an
allocation microbenchmark. The reason it helps is it allows the
page-zeroing to be done out of the normal allocation/free path,
and pushed off to a kthread.
As not all heaps will find this useful, its implemented as
a optional helper library that heaps can utilize.
Cc: Daniel Vetter <[email protected]>
Cc: Sumit Semwal <[email protected]>
Cc: Liam Mark <[email protected]>
Cc: Chris Goldsworthy <[email protected]>
Cc: Laura Abbott <[email protected]>
Cc: Brian Starkey <[email protected]>
Cc: Hridya Valsaraju <[email protected]>
Cc: Suren Baghdasaryan <[email protected]>
Cc: Sandeep Patil <[email protected]>
Cc: Daniel Mentz <[email protected]>
Cc: Ørjan Eide <[email protected]>
Cc: Robin Murphy <[email protected]>
Cc: Ezequiel Garcia <[email protected]>
Cc: Simon Ser <[email protected]>
Cc: James Jones <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: John Stultz <[email protected]>
---
v2:
* Fix sleep in atomic issue from using a mutex, by switching
to a spinlock as Reported-by: kernel test robot <[email protected]>
* Cleanup API to use a reason enum for clarity and add some documentation
comments as suggested by Suren Baghdasaryan.
---
drivers/dma-buf/heaps/Kconfig | 3 +
drivers/dma-buf/heaps/Makefile | 1 +
drivers/dma-buf/heaps/deferred-free-helper.c | 136 +++++++++++++++++++
drivers/dma-buf/heaps/deferred-free-helper.h | 55 ++++++++
4 files changed, 195 insertions(+)
create mode 100644 drivers/dma-buf/heaps/deferred-free-helper.c
create mode 100644 drivers/dma-buf/heaps/deferred-free-helper.h
diff --git a/drivers/dma-buf/heaps/Kconfig b/drivers/dma-buf/heaps/Kconfig
index a5eef06c4226..ecf65204f714 100644
--- a/drivers/dma-buf/heaps/Kconfig
+++ b/drivers/dma-buf/heaps/Kconfig
@@ -1,3 +1,6 @@
+config DMABUF_HEAPS_DEFERRED_FREE
+ bool
+
config DMABUF_HEAPS_SYSTEM
bool "DMA-BUF System Heap"
depends on DMABUF_HEAPS
diff --git a/drivers/dma-buf/heaps/Makefile b/drivers/dma-buf/heaps/Makefile
index 974467791032..4e7839875615 100644
--- a/drivers/dma-buf/heaps/Makefile
+++ b/drivers/dma-buf/heaps/Makefile
@@ -1,3 +1,4 @@
# SPDX-License-Identifier: GPL-2.0
+obj-$(CONFIG_DMABUF_HEAPS_DEFERRED_FREE) += deferred-free-helper.o
obj-$(CONFIG_DMABUF_HEAPS_SYSTEM) += system_heap.o
obj-$(CONFIG_DMABUF_HEAPS_CMA) += cma_heap.o
diff --git a/drivers/dma-buf/heaps/deferred-free-helper.c b/drivers/dma-buf/heaps/deferred-free-helper.c
new file mode 100644
index 000000000000..cf04148167a2
--- /dev/null
+++ b/drivers/dma-buf/heaps/deferred-free-helper.c
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Deferred dmabuf freeing helper
+ *
+ * Copyright (C) 2020 Linaro, Ltd.
+ *
+ * Based on the ION page pool code
+ * Copyright (C) 2011 Google, Inc.
+ */
+
+#include <linux/freezer.h>
+#include <linux/list.h>
+#include <linux/slab.h>
+#include <linux/swap.h>
+#include <linux/sched/signal.h>
+
+#include "deferred-free-helper.h"
+
+static LIST_HEAD(free_list);
+static size_t list_size;
+wait_queue_head_t freelist_waitqueue;
+struct task_struct *freelist_task;
+static DEFINE_SPINLOCK(free_list_lock);
+
+void deferred_free(struct deferred_freelist_item *item,
+ void (*free)(struct deferred_freelist_item*,
+ enum df_reason),
+ size_t size)
+{
+ unsigned long flags;
+
+ INIT_LIST_HEAD(&item->list);
+ item->size = size;
+ item->free = free;
+
+ spin_lock_irqsave(&free_list_lock, flags);
+ list_add(&item->list, &free_list);
+ list_size += size;
+ spin_unlock_irqrestore(&free_list_lock, flags);
+ wake_up(&freelist_waitqueue);
+}
+
+static size_t free_one_item(enum df_reason reason)
+{
+ unsigned long flags;
+ size_t size = 0;
+ struct deferred_freelist_item *item;
+
+ spin_lock_irqsave(&free_list_lock, flags);
+ if (list_empty(&free_list)) {
+ spin_unlock_irqrestore(&free_list_lock, flags);
+ return 0;
+ }
+ item = list_first_entry(&free_list, struct deferred_freelist_item, list);
+ list_del(&item->list);
+ size = item->size;
+ list_size -= size;
+ spin_unlock_irqrestore(&free_list_lock, flags);
+
+ item->free(item, reason);
+ return size;
+}
+
+static unsigned long get_freelist_size(void)
+{
+ unsigned long size;
+ unsigned long flags;
+
+ spin_lock_irqsave(&free_list_lock, flags);
+ size = list_size;
+ spin_unlock_irqrestore(&free_list_lock, flags);
+ return size;
+}
+
+static unsigned long freelist_shrink_count(struct shrinker *shrinker,
+ struct shrink_control *sc)
+{
+ return get_freelist_size();
+}
+
+static unsigned long freelist_shrink_scan(struct shrinker *shrinker,
+ struct shrink_control *sc)
+{
+ int total_freed = 0;
+
+ if (sc->nr_to_scan == 0)
+ return 0;
+
+ while (total_freed < sc->nr_to_scan) {
+ int freed = free_one_item(DF_UNDER_PRESSURE);
+
+ if (!freed)
+ break;
+
+ total_freed += freed;
+ }
+
+ return total_freed;
+}
+
+static struct shrinker freelist_shrinker = {
+ .count_objects = freelist_shrink_count,
+ .scan_objects = freelist_shrink_scan,
+ .seeks = DEFAULT_SEEKS,
+ .batch = 0,
+};
+
+static int deferred_free_thread(void *data)
+{
+ while (true) {
+ wait_event_freezable(freelist_waitqueue,
+ get_freelist_size() > 0);
+
+ free_one_item(DF_NORMAL);
+ }
+
+ return 0;
+}
+
+static int deferred_freelist_init(void)
+{
+ list_size = 0;
+
+ init_waitqueue_head(&freelist_waitqueue);
+ freelist_task = kthread_run(deferred_free_thread, NULL,
+ "%s", "dmabuf-deferred-free-worker");
+ if (IS_ERR(freelist_task)) {
+ pr_err("%s: creating thread for deferred free failed\n",
+ __func__);
+ return -1;
+ }
+ sched_set_normal(freelist_task, 19);
+
+ return register_shrinker(&freelist_shrinker);
+}
+device_initcall(deferred_freelist_init);
diff --git a/drivers/dma-buf/heaps/deferred-free-helper.h b/drivers/dma-buf/heaps/deferred-free-helper.h
new file mode 100644
index 000000000000..2c43dd5a3eda
--- /dev/null
+++ b/drivers/dma-buf/heaps/deferred-free-helper.h
@@ -0,0 +1,55 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef DEFERRED_FREE_HELPER_H
+#define DEFERRED_FREE_HELPER_H
+
+/**
+ * df_reason - enum for reason why item was freed
+ *
+ * This provides a reason for why the free funciton was called
+ * on the item. This is useful when deferred_free is used in
+ * combination with a pagepool, so under pressure the page can
+ * be immediately freed.
+ *
+ * DF_NORMAL: Normal deferred free
+ *
+ * DF_UNDER_PRESSURE: Free was called because the system
+ * is under memory pressure. Usually
+ * from a shrinker. Avoid allocating
+ * memory in the free call, as it may
+ * fail.
+ */
+enum df_reason {
+ DF_NORMAL,
+ DF_UNDER_PRESSURE,
+};
+
+/**
+ * deferred_freelist_item - item structure for deferred freelist
+ *
+ * This is to be added to the structure for whatever you want to
+ * defer freeing on.
+ *
+ * @size: size of the item to be freed
+ * @free: function pointer to be called when freeing the item
+ * @list: list entry for the deferred list
+ */
+struct deferred_freelist_item {
+ size_t size;
+ void (*free)(struct deferred_freelist_item *i,
+ enum df_reason reason);
+ struct list_head list;
+};
+
+/**
+ * deferred_free - call to add item to the deferred free list
+ *
+ * @item: Pointer to deferred_freelist_item field of a structure
+ * @free: Function pointer to the free call
+ * @size: Size of the item to be freed
+ */
+void deferred_free(struct deferred_freelist_item *item,
+ void (*free)(struct deferred_freelist_item *i,
+ enum df_reason reason),
+ size_t size);
+#endif
--
2.17.1
Reuse/abuse the pagepool code from the network code to speed
up allocation performance.
This is similar to the ION pagepool usage, but tries to
utilize generic code instead of a custom implementation.
Cc: Daniel Vetter <[email protected]>
Cc: Sumit Semwal <[email protected]>
Cc: Liam Mark <[email protected]>
Cc: Chris Goldsworthy <[email protected]>
Cc: Laura Abbott <[email protected]>
Cc: Brian Starkey <[email protected]>
Cc: Hridya Valsaraju <[email protected]>
Cc: Suren Baghdasaryan <[email protected]>
Cc: Sandeep Patil <[email protected]>
Cc: Daniel Mentz <[email protected]>
Cc: Ørjan Eide <[email protected]>
Cc: Robin Murphy <[email protected]>
Cc: Ezequiel Garcia <[email protected]>
Cc: Simon Ser <[email protected]>
Cc: James Jones <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: John Stultz <[email protected]>
---
v2:
* Fix build issue caused by selecting PAGE_POOL w/o NET
as Reported-by: kernel test robot <[email protected]>
---
drivers/dma-buf/heaps/Kconfig | 2 +
drivers/dma-buf/heaps/system_heap.c | 68 +++++++++++++++++++++++++++--
2 files changed, 66 insertions(+), 4 deletions(-)
diff --git a/drivers/dma-buf/heaps/Kconfig b/drivers/dma-buf/heaps/Kconfig
index ecf65204f714..748e840e6edd 100644
--- a/drivers/dma-buf/heaps/Kconfig
+++ b/drivers/dma-buf/heaps/Kconfig
@@ -4,6 +4,8 @@ config DMABUF_HEAPS_DEFERRED_FREE
config DMABUF_HEAPS_SYSTEM
bool "DMA-BUF System Heap"
depends on DMABUF_HEAPS
+ select NET
+ select PAGE_POOL
help
Choose this option to enable the system dmabuf heap. The system heap
is backed by pages from the buddy allocator. If in doubt, say Y.
diff --git a/drivers/dma-buf/heaps/system_heap.c b/drivers/dma-buf/heaps/system_heap.c
index 17e0e9a68baf..885e30894b77 100644
--- a/drivers/dma-buf/heaps/system_heap.c
+++ b/drivers/dma-buf/heaps/system_heap.c
@@ -20,6 +20,7 @@
#include <linux/scatterlist.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>
+#include <net/page_pool.h>
static struct dma_heap *sys_heap;
@@ -53,6 +54,7 @@ static gfp_t order_flags[] = {HIGH_ORDER_GFP, LOW_ORDER_GFP, LOW_ORDER_GFP};
*/
static const unsigned int orders[] = {8, 4, 0};
#define NUM_ORDERS ARRAY_SIZE(orders)
+struct page_pool *pools[NUM_ORDERS];
static struct sg_table *dup_sg_table(struct sg_table *table)
{
@@ -281,18 +283,59 @@ static void system_heap_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map)
dma_buf_map_clear(map);
}
+static int system_heap_clear_pages(struct page **pages, int num, pgprot_t pgprot)
+{
+ void *addr = vmap(pages, num, VM_MAP, pgprot);
+
+ if (!addr)
+ return -ENOMEM;
+ memset(addr, 0, PAGE_SIZE * num);
+ vunmap(addr);
+ return 0;
+}
+
+static int system_heap_zero_buffer(struct system_heap_buffer *buffer)
+{
+ struct sg_table *sgt = &buffer->sg_table;
+ struct sg_page_iter piter;
+ struct page *pages[32];
+ int p = 0;
+ int ret = 0;
+
+ for_each_sgtable_page(sgt, &piter, 0) {
+ pages[p++] = sg_page_iter_page(&piter);
+ if (p == ARRAY_SIZE(pages)) {
+ ret = system_heap_clear_pages(pages, p, PAGE_KERNEL);
+ if (ret)
+ return ret;
+ p = 0;
+ }
+ }
+ if (p)
+ ret = system_heap_clear_pages(pages, p, PAGE_KERNEL);
+
+ return ret;
+}
+
static void system_heap_dma_buf_release(struct dma_buf *dmabuf)
{
struct system_heap_buffer *buffer = dmabuf->priv;
struct sg_table *table;
struct scatterlist *sg;
- int i;
+ int i, j;
+
+ /* Zero the buffer pages before adding back to the pool */
+ system_heap_zero_buffer(buffer);
table = &buffer->sg_table;
for_each_sg(table->sgl, sg, table->nents, i) {
struct page *page = sg_page(sg);
- __free_pages(page, compound_order(page));
+ for (j = 0; j < NUM_ORDERS; j++) {
+ if (compound_order(page) == orders[j])
+ break;
+ }
+ page_pool_put_full_page(pools[j], page, false);
}
sg_free_table(table);
kfree(buffer);
@@ -322,8 +365,7 @@ static struct page *alloc_largest_available(unsigned long size,
continue;
if (max_order < orders[i])
continue;
-
- page = alloc_pages(order_flags[i], orders[i]);
+ page = page_pool_alloc_pages(pools[i], order_flags[i]);
if (!page)
continue;
return page;
@@ -428,6 +470,24 @@ static const struct dma_heap_ops system_heap_ops = {
static int system_heap_create(void)
{
struct dma_heap_export_info exp_info;
+ int i;
+
+ for (i = 0; i < NUM_ORDERS; i++) {
+ struct page_pool_params pp;
+
+ memset(&pp, 0, sizeof(pp));
+ pp.order = orders[i];
+ pools[i] = page_pool_create(&pp);
+
+ if (IS_ERR(pools[i])) {
+ int j;
+
+ pr_err("%s: page pool creation failed!\n", __func__);
+ for (j = 0; j < i; j++)
+ page_pool_destroy(pools[j]);
+ return PTR_ERR(pools[i]);
+ }
+ }
exp_info.name = "system";
exp_info.ops = &system_heap_ops;
--
2.17.1
Utilize the deferred free helper library in the system heap.
This provides a nice performance bump and puts the
system heap performance on par with ION.
Cc: Daniel Vetter <[email protected]>
Cc: Sumit Semwal <[email protected]>
Cc: Liam Mark <[email protected]>
Cc: Chris Goldsworthy <[email protected]>
Cc: Laura Abbott <[email protected]>
Cc: Brian Starkey <[email protected]>
Cc: Hridya Valsaraju <[email protected]>
Cc: Suren Baghdasaryan <[email protected]>
Cc: Sandeep Patil <[email protected]>
Cc: Daniel Mentz <[email protected]>
Cc: Ørjan Eide <[email protected]>
Cc: Robin Murphy <[email protected]>
Cc: Ezequiel Garcia <[email protected]>
Cc: Simon Ser <[email protected]>
Cc: James Jones <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: John Stultz <[email protected]>
---
v2:
* Rework deferred-free api to use reason enum as suggested by
Suren Baghdasaryan
---
drivers/dma-buf/heaps/Kconfig | 1 +
drivers/dma-buf/heaps/system_heap.c | 32 ++++++++++++++++++++++-------
2 files changed, 26 insertions(+), 7 deletions(-)
diff --git a/drivers/dma-buf/heaps/Kconfig b/drivers/dma-buf/heaps/Kconfig
index 748e840e6edd..3f4d7b949301 100644
--- a/drivers/dma-buf/heaps/Kconfig
+++ b/drivers/dma-buf/heaps/Kconfig
@@ -6,6 +6,7 @@ config DMABUF_HEAPS_SYSTEM
depends on DMABUF_HEAPS
select NET
select PAGE_POOL
+ select DMABUF_HEAPS_DEFERRED_FREE
help
Choose this option to enable the system dmabuf heap. The system heap
is backed by pages from the buddy allocator. If in doubt, say Y.
diff --git a/drivers/dma-buf/heaps/system_heap.c b/drivers/dma-buf/heaps/system_heap.c
index 885e30894b77..747fa2250e84 100644
--- a/drivers/dma-buf/heaps/system_heap.c
+++ b/drivers/dma-buf/heaps/system_heap.c
@@ -22,6 +22,8 @@
#include <linux/vmalloc.h>
#include <net/page_pool.h>
+#include "deferred-free-helper.h"
+
static struct dma_heap *sys_heap;
struct system_heap_buffer {
@@ -32,6 +34,7 @@ struct system_heap_buffer {
struct sg_table sg_table;
int vmap_cnt;
void *vaddr;
+ struct deferred_freelist_item deferred_free;
};
struct dma_heap_attachment {
@@ -317,30 +320,45 @@ static int system_heap_zero_buffer(struct system_heap_buffer *buffer)
return ret;
}
-static void system_heap_dma_buf_release(struct dma_buf *dmabuf)
+static void system_heap_buf_free(struct deferred_freelist_item *item,
+ enum df_reason reason)
{
- struct system_heap_buffer *buffer = dmabuf->priv;
+ struct system_heap_buffer *buffer;
struct sg_table *table;
struct scatterlist *sg;
int i, j;
+ buffer = container_of(item, struct system_heap_buffer, deferred_free);
/* Zero the buffer pages before adding back to the pool */
- system_heap_zero_buffer(buffer);
+ if (reason == DF_NORMAL)
+ if (system_heap_zero_buffer(buffer))
+ reason = DF_UNDER_PRESSURE; // On failure, just free
table = &buffer->sg_table;
for_each_sg(table->sgl, sg, table->nents, i) {
struct page *page = sg_page(sg);
- for (j = 0; j < NUM_ORDERS; j++) {
- if (compound_order(page) == orders[j])
- break;
+ if (reason == DF_UNDER_PRESSURE) {
+ __free_pages(page, compound_order(page));
+ } else {
+ for (j = 0; j < NUM_ORDERS; j++) {
+ if (compound_order(page) == orders[j])
+ break;
+ }
+ page_pool_put_full_page(pools[j], page, false);
}
- page_pool_put_full_page(pools[j], page, false);
}
sg_free_table(table);
kfree(buffer);
}
+static void system_heap_dma_buf_release(struct dma_buf *dmabuf)
+{
+ struct system_heap_buffer *buffer = dmabuf->priv;
+
+ deferred_free(&buffer->deferred_free, system_heap_buf_free, buffer->len);
+}
+
static const struct dma_buf_ops system_heap_buf_ops = {
.attach = system_heap_attach,
.detach = system_heap_detach,
--
2.17.1
On Fri, Jan 22, 2021 at 7:47 PM John Stultz <[email protected]> wrote:
> +static int system_heap_clear_pages(struct page **pages, int num, pgprot_t pgprot)
> +{
> + void *addr = vmap(pages, num, VM_MAP, pgprot);
> +
> + if (!addr)
> + return -ENOMEM;
> + memset(addr, 0, PAGE_SIZE * num);
> + vunmap(addr);
> + return 0;
> +}
I thought that vmap/vunmap are expensive, and I am wondering if
there's a faster way that avoids vmap.
How about lifting this code from lib/iov_iter.c
static void memzero_page(struct page *page, size_t offset, size_t len)
{
char *addr = kmap_atomic(page);
memset(addr + offset, 0, len);
kunmap_atomic(addr);
}
Or what about lifting that code from the old ion_cma_heap.c
if (PageHighMem(pages)) {
unsigned long nr_clear_pages = nr_pages;
struct page *page = pages;
while (nr_clear_pages > 0) {
void *vaddr = kmap_atomic(page);
memset(vaddr, 0, PAGE_SIZE);
kunmap_atomic(vaddr);
page++;
nr_clear_pages--;
}
} else {
memset(page_address(pages), 0, size);
}
On Wed, Jan 27, 2021 at 12:21 PM Daniel Mentz <[email protected]> wrote:
>
> On Fri, Jan 22, 2021 at 7:47 PM John Stultz <[email protected]> wrote:
> > +static int system_heap_clear_pages(struct page **pages, int num, pgprot_t pgprot)
> > +{
> > + void *addr = vmap(pages, num, VM_MAP, pgprot);
> > +
> > + if (!addr)
> > + return -ENOMEM;
> > + memset(addr, 0, PAGE_SIZE * num);
> > + vunmap(addr);
> > + return 0;
> > +}
>
> I thought that vmap/vunmap are expensive, and I am wondering if
> there's a faster way that avoids vmap.
>
> How about lifting this code from lib/iov_iter.c
> static void memzero_page(struct page *page, size_t offset, size_t len)
> {
> char *addr = kmap_atomic(page);
> memset(addr + offset, 0, len);
> kunmap_atomic(addr);
> }
>
> Or what about lifting that code from the old ion_cma_heap.c
>
> if (PageHighMem(pages)) {
> unsigned long nr_clear_pages = nr_pages;
> struct page *page = pages;
>
> while (nr_clear_pages > 0) {
> void *vaddr = kmap_atomic(page);
>
> memset(vaddr, 0, PAGE_SIZE);
> kunmap_atomic(vaddr);
> page++;
> nr_clear_pages--;
> }
> } else {
> memset(page_address(pages), 0, size);
> }
Though, this last memset only works since CMA is contiguous, so it
probably needs to always do the kmap_atomic for each page, right?
I'm still a little worried if this is right, as the current
implementation with the vmap comes from the old ion_heap_sglist_zero
logic, which similarly tries to batch the vmaps 32 pages at at time,
but I'll give it a try.
thanks
-john
On Wed, Jan 27, 2021 at 9:10 PM John Stultz <[email protected]> wrote:
>
> On Wed, Jan 27, 2021 at 12:21 PM Daniel Mentz <[email protected]> wrote:
> >
> > On Fri, Jan 22, 2021 at 7:47 PM John Stultz <[email protected]> wrote:
> > > +static int system_heap_clear_pages(struct page **pages, int num, pgprot_t pgprot)
> > > +{
> > > + void *addr = vmap(pages, num, VM_MAP, pgprot);
> > > +
> > > + if (!addr)
> > > + return -ENOMEM;
> > > + memset(addr, 0, PAGE_SIZE * num);
> > > + vunmap(addr);
> > > + return 0;
> > > +}
> >
> > I thought that vmap/vunmap are expensive, and I am wondering if
> > there's a faster way that avoids vmap.
> >
> > How about lifting this code from lib/iov_iter.c
> > static void memzero_page(struct page *page, size_t offset, size_t len)
> > {
> > char *addr = kmap_atomic(page);
> > memset(addr + offset, 0, len);
> > kunmap_atomic(addr);
> > }
> >
> > Or what about lifting that code from the old ion_cma_heap.c
> >
> > if (PageHighMem(pages)) {
> > unsigned long nr_clear_pages = nr_pages;
> > struct page *page = pages;
> >
> > while (nr_clear_pages > 0) {
> > void *vaddr = kmap_atomic(page);
> >
> > memset(vaddr, 0, PAGE_SIZE);
> > kunmap_atomic(vaddr);
> > page++;
> > nr_clear_pages--;
> > }
> > } else {
> > memset(page_address(pages), 0, size);
> > }
>
> Though, this last memset only works since CMA is contiguous, so it
> probably needs to always do the kmap_atomic for each page, right?
Yeah, but with the system heap page pool, some of these pages might be
64KB or 1MB large. kmap_atomic(page) just maps to page_address(page)
in most cases. I think iterating over all pages individually in this
manner might still be faster than using vmap.
>
> I'm still a little worried if this is right, as the current
> implementation with the vmap comes from the old ion_heap_sglist_zero
> logic, which similarly tries to batch the vmaps 32 pages at at time,
> but I'll give it a try.