From: Alexander Duyck <alexander.duyck@gmail.com>
To: Shannon Nelson <snelson@pensando.io>
Cc: Netdev <netdev@vger.kernel.org>,
David Miller <davem@davemloft.net>,
Jakub Kicinski <kuba@kernel.org>,
Pensando Drivers <drivers@pensando.io>
Subject: Re: [PATCH net-next 2/6] ionic: implement Rx page reuse
Date: Wed, 10 Mar 2021 18:14:05 -0800 [thread overview]
Message-ID: <CAKgT0UfyjoAN7LTnq0NMZfXRv4v7iTCPyAb9pVr3qWMhop_BVw@mail.gmail.com> (raw)
In-Reply-To: <20210310192631.20022-3-snelson@pensando.io>
On Wed, Mar 10, 2021 at 11:28 AM Shannon Nelson <snelson@pensando.io> wrote:
>
> Rework the Rx buffer allocations to use pages twice when using
> normal MTU in order to cut down on buffer allocation and mapping
> overhead.
>
> Instead of tracking individual pages, in which we may have
> wasted half the space when using standard 1500 MTU, we track
> buffers which use half pages, so we can use the second half
> of the page rather than allocate and map a new page once the
> first buffer has been used.
>
> Signed-off-by: Shannon Nelson <snelson@pensando.io>
So looking at the approach taken here it just seems like you are doing
the linear walk approach and getting 2 uses per 4K page. If you are
taking that route it might make more sense to just split the page and
use both pieces immediately to populate 2 entries instead of waiting
on the next loop through the ring. Then you could just split the page
into multiple buffers and fill your sg list using less total pages
rather than having 2K gaps between your entries. An added advantage
would be that you could simply merge the page fragments in the event
that you have something writing to the full 2K buffers and you cannot
use copybreak.
> ---
> .../net/ethernet/pensando/ionic/ionic_dev.h | 12 +-
> .../net/ethernet/pensando/ionic/ionic_txrx.c | 215 +++++++++++-------
> 2 files changed, 138 insertions(+), 89 deletions(-)
>
> diff --git a/drivers/net/ethernet/pensando/ionic/ionic_dev.h b/drivers/net/ethernet/pensando/ionic/ionic_dev.h
> index 690768ff0143..0f877c86eba6 100644
> --- a/drivers/net/ethernet/pensando/ionic/ionic_dev.h
> +++ b/drivers/net/ethernet/pensando/ionic/ionic_dev.h
> @@ -170,9 +170,15 @@ typedef void (*ionic_desc_cb)(struct ionic_queue *q,
> struct ionic_desc_info *desc_info,
> struct ionic_cq_info *cq_info, void *cb_arg);
>
> -struct ionic_page_info {
> +#define IONIC_PAGE_SIZE PAGE_SIZE
> +#define IONIC_PAGE_SPLIT_SZ (PAGE_SIZE / 2)
This probably doesn't work out too well when the page size gets up to
64K. I don't know of too many networks that support a 32K MTU.. :)
> +#define IONIC_PAGE_GFP_MASK (GFP_ATOMIC | __GFP_NOWARN |\
> + __GFP_COMP | __GFP_MEMALLOC)
> +
> +struct ionic_buf_info {
> struct page *page;
> dma_addr_t dma_addr;
> + u32 page_offset;
> };
I'm not really sure the rename was needed. You are still just working
with a page aren't you? It would actually reduce the complexity of
this patch a bunch as you could drop the renaming changes.
> struct ionic_desc_info {
> @@ -187,8 +193,8 @@ struct ionic_desc_info {
> struct ionic_txq_sg_desc *txq_sg_desc;
> struct ionic_rxq_sg_desc *rxq_sgl_desc;
> };
> - unsigned int npages;
> - struct ionic_page_info pages[IONIC_RX_MAX_SG_ELEMS + 1];
> + unsigned int nbufs;
> + struct ionic_buf_info bufs[IONIC_RX_MAX_SG_ELEMS + 1];
> ionic_desc_cb cb;
> void *cb_arg;
> };
> diff --git a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
> index 70b997f302ac..3e13cfee9ecd 100644
> --- a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
> +++ b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
> @@ -54,7 +54,7 @@ static struct sk_buff *ionic_rx_skb_alloc(struct ionic_queue *q,
> if (frags)
> skb = napi_get_frags(&q_to_qcq(q)->napi);
> else
> - skb = netdev_alloc_skb_ip_align(netdev, len);
> + skb = napi_alloc_skb(&q_to_qcq(q)->napi, len);
>
> if (unlikely(!skb)) {
> net_warn_ratelimited("%s: SKB alloc failed on %s!\n",
> @@ -66,8 +66,15 @@ static struct sk_buff *ionic_rx_skb_alloc(struct ionic_queue *q,
> return skb;
> }
>
> +static void ionic_rx_buf_reset(struct ionic_buf_info *buf_info)
> +{
> + buf_info->page = NULL;
> + buf_info->page_offset = 0;
> + buf_info->dma_addr = 0;
> +}
> +
Technically speaking you probably only need to reset the page value.
You could hold off on resetting the page_offset and dma_addr until you
actually are populating the page.
> static int ionic_rx_page_alloc(struct ionic_queue *q,
> - struct ionic_page_info *page_info)
> + struct ionic_buf_info *buf_info)
> {
> struct ionic_lif *lif = q->lif;
> struct ionic_rx_stats *stats;
> @@ -78,26 +85,26 @@ static int ionic_rx_page_alloc(struct ionic_queue *q,
> dev = lif->ionic->dev;
> stats = q_to_rx_stats(q);
>
> - if (unlikely(!page_info)) {
> - net_err_ratelimited("%s: %s invalid page_info in alloc\n",
> + if (unlikely(!buf_info)) {
> + net_err_ratelimited("%s: %s invalid buf_info in alloc\n",
> netdev->name, q->name);
> return -EINVAL;
> }
>
> - page_info->page = dev_alloc_page();
> - if (unlikely(!page_info->page)) {
> + buf_info->page = alloc_pages(IONIC_PAGE_GFP_MASK, 0);
> + if (unlikely(!buf_info->page)) {
> net_err_ratelimited("%s: %s page alloc failed\n",
> netdev->name, q->name);
> stats->alloc_err++;
> return -ENOMEM;
> }
> + buf_info->page_offset = 0;
>
> - page_info->dma_addr = dma_map_page(dev, page_info->page, 0, PAGE_SIZE,
> - DMA_FROM_DEVICE);
> - if (unlikely(dma_mapping_error(dev, page_info->dma_addr))) {
> - put_page(page_info->page);
> - page_info->dma_addr = 0;
> - page_info->page = NULL;
> + buf_info->dma_addr = dma_map_page(dev, buf_info->page, buf_info->page_offset,
> + IONIC_PAGE_SIZE, DMA_FROM_DEVICE);
> + if (unlikely(dma_mapping_error(dev, buf_info->dma_addr))) {
> + __free_pages(buf_info->page, 0);
> + ionic_rx_buf_reset(buf_info);
> net_err_ratelimited("%s: %s dma map failed\n",
> netdev->name, q->name);
> stats->dma_map_err++;
So one thing I would change about the setup here is that I would not
store anything in buf_info page until after you have stored the
page_offset and dma_addr. That way you know that the other two are
only valid because a page is present. In addition you can avoid having
to do the extra cleanup as they should only be read if info->page is
set.
> @@ -108,32 +115,46 @@ static int ionic_rx_page_alloc(struct ionic_queue *q,
> }
>
> static void ionic_rx_page_free(struct ionic_queue *q,
> - struct ionic_page_info *page_info)
> + struct ionic_buf_info *buf_info)
> {
> - struct ionic_lif *lif = q->lif;
> - struct net_device *netdev;
> - struct device *dev;
> -
> - netdev = lif->netdev;
> - dev = lif->ionic->dev;
> + struct net_device *netdev = q->lif->netdev;
> + struct device *dev = q->lif->ionic->dev;
>
> - if (unlikely(!page_info)) {
> - net_err_ratelimited("%s: %s invalid page_info in free\n",
> + if (unlikely(!buf_info)) {
> + net_err_ratelimited("%s: %s invalid buf_info in free\n",
> netdev->name, q->name);
> return;
> }
>
> - if (unlikely(!page_info->page)) {
> - net_err_ratelimited("%s: %s invalid page in free\n",
> - netdev->name, q->name);
> + if (!buf_info->page)
> return;
> - }
>
> - dma_unmap_page(dev, page_info->dma_addr, PAGE_SIZE, DMA_FROM_DEVICE);
> + dma_unmap_page(dev, buf_info->dma_addr, IONIC_PAGE_SIZE, DMA_FROM_DEVICE);
> + __free_pages(buf_info->page, 0);
> + ionic_rx_buf_reset(buf_info);
> +}
> +
> +static bool ionic_rx_buf_recycle(struct ionic_queue *q,
> + struct ionic_buf_info *buf_info, u32 used)
> +{
> + u32 size;
> +
> + /* don't re-use pages allocated in low-mem condition */
> + if (page_is_pfmemalloc(buf_info->page))
> + return false;
> +
> + /* don't re-use buffers from non-local numa nodes */
> + if (page_to_nid(buf_info->page) != numa_mem_id())
> + return false;
> +
So I am not sure if either of these is really needed if you are just
going to be freeing the page after all 4K is consumed. With the Intel
drivers we were bouncing back and forth between the upper and lower
halves. With this it looks like you do a linear walk and then exit
when you have reached the end of the page. With that being the case
the memory locality check is kind of moot since you will flush the
page after two uses anyway. In addition the pfmemalloc check might
also be moot since it may actually be more efficient to reuse the page
rather than use a full 4K and attempt to allocate yet another page.
> + size = ALIGN(used, IONIC_PAGE_SPLIT_SZ);
> + buf_info->page_offset += size;
> + if (buf_info->page_offset >= IONIC_PAGE_SIZE)
> + return false;
> +
> + get_page(buf_info->page);
The get_page per 2K section will add up in terms of cost as it is an
expensive atomic operation. You might see if you can get away with
batching it to do one atomic add per allocation instead of one per
use.
> - put_page(page_info->page);
> - page_info->dma_addr = 0;
> - page_info->page = NULL;
> + return true;
> }
>
> static struct sk_buff *ionic_rx_frags(struct ionic_queue *q,
> @@ -142,16 +163,16 @@ static struct sk_buff *ionic_rx_frags(struct ionic_queue *q,
> {
> struct ionic_rxq_comp *comp = cq_info->cq_desc;
> struct device *dev = q->lif->ionic->dev;
> - struct ionic_page_info *page_info;
> + struct ionic_buf_info *buf_info;
> struct sk_buff *skb;
> unsigned int i;
> u16 frag_len;
> u16 len;
>
> - page_info = &desc_info->pages[0];
> + buf_info = &desc_info->bufs[0];
> len = le16_to_cpu(comp->len);
>
> - prefetch(page_address(page_info->page) + NET_IP_ALIGN);
> + prefetch(buf_info->page);
Just want to confirm this is what you meant to do. The old code was
prefetching the first line of the data. The new code is just
prefetching the page struct. You may want to change this to a
prefetchw if you are expecting to use this to improve the performance
for the get_page call.
>
> skb = ionic_rx_skb_alloc(q, len, true);
> if (unlikely(!skb))
> @@ -159,7 +180,7 @@ static struct sk_buff *ionic_rx_frags(struct ionic_queue *q,
>
> i = comp->num_sg_elems + 1;
> do {
> - if (unlikely(!page_info->page)) {
> + if (unlikely(!buf_info->page)) {
> struct napi_struct *napi = &q_to_qcq(q)->napi;
>
> napi->skb = NULL;
> @@ -167,15 +188,25 @@ static struct sk_buff *ionic_rx_frags(struct ionic_queue *q,
> return NULL;
> }
>
> - frag_len = min(len, (u16)PAGE_SIZE);
> + frag_len = min_t(u16, len, IONIC_PAGE_SIZE - buf_info->page_offset);
> len -= frag_len;
>
> - dma_unmap_page(dev, dma_unmap_addr(page_info, dma_addr),
> - PAGE_SIZE, DMA_FROM_DEVICE);
> + dma_sync_single_for_cpu(dev,
> + buf_info->dma_addr + buf_info->page_offset,
> + frag_len, DMA_FROM_DEVICE);
> +
> skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
> - page_info->page, 0, frag_len, PAGE_SIZE);
> - page_info->page = NULL;
> - page_info++;
> + buf_info->page, buf_info->page_offset, frag_len,
> + IONIC_PAGE_SIZE);
> +
> + if (!ionic_rx_buf_recycle(q, buf_info, frag_len)) {
> + dma_unmap_page(dev, buf_info->dma_addr,
> + IONIC_PAGE_SIZE, DMA_FROM_DEVICE);
> + ionic_rx_buf_reset(buf_info);
> + }
> +
If you don't unmap it don't you need to sync the remaining portion of
the page for use by the device?
> + buf_info++;
> +
> i--;
> } while (i > 0);
>
> @@ -188,26 +219,26 @@ static struct sk_buff *ionic_rx_copybreak(struct ionic_queue *q,
> {
> struct ionic_rxq_comp *comp = cq_info->cq_desc;
> struct device *dev = q->lif->ionic->dev;
> - struct ionic_page_info *page_info;
> + struct ionic_buf_info *buf_info;
> struct sk_buff *skb;
> u16 len;
>
> - page_info = &desc_info->pages[0];
> + buf_info = &desc_info->bufs[0];
> len = le16_to_cpu(comp->len);
>
> skb = ionic_rx_skb_alloc(q, len, false);
> if (unlikely(!skb))
> return NULL;
>
> - if (unlikely(!page_info->page)) {
> + if (unlikely(!buf_info->page)) {
> dev_kfree_skb(skb);
> return NULL;
> }
>
> - dma_sync_single_for_cpu(dev, dma_unmap_addr(page_info, dma_addr),
> + dma_sync_single_for_cpu(dev, buf_info->dma_addr + buf_info->page_offset,
> len, DMA_FROM_DEVICE);
> - skb_copy_to_linear_data(skb, page_address(page_info->page), len);
> - dma_sync_single_for_device(dev, dma_unmap_addr(page_info, dma_addr),
> + skb_copy_to_linear_data(skb, page_address(buf_info->page) + buf_info->page_offset, len);
> + dma_sync_single_for_device(dev, buf_info->dma_addr + buf_info->page_offset,
> len, DMA_FROM_DEVICE);
>
> skb_put(skb, len);
> @@ -327,64 +358,73 @@ void ionic_rx_fill(struct ionic_queue *q)
> {
> struct net_device *netdev = q->lif->netdev;
> struct ionic_desc_info *desc_info;
> - struct ionic_page_info *page_info;
> struct ionic_rxq_sg_desc *sg_desc;
> struct ionic_rxq_sg_elem *sg_elem;
> + struct ionic_buf_info *buf_info;
> struct ionic_rxq_desc *desc;
> + unsigned int max_sg_elems;
> unsigned int remain_len;
> - unsigned int seg_len;
> + unsigned int frag_len;
> unsigned int nfrags;
> unsigned int i, j;
> unsigned int len;
>
> len = netdev->mtu + ETH_HLEN + VLAN_HLEN;
> - nfrags = round_up(len, PAGE_SIZE) / PAGE_SIZE;
> + max_sg_elems = q->lif->qtype_info[IONIC_QTYPE_RXQ].max_sg_elems;
>
> for (i = ionic_q_space_avail(q); i; i--) {
> + nfrags = 0;
> remain_len = len;
> desc_info = &q->info[q->head_idx];
> desc = desc_info->desc;
> - sg_desc = desc_info->sg_desc;
> - page_info = &desc_info->pages[0];
> + buf_info = &desc_info->bufs[0];
>
> - if (page_info->page) { /* recycle the buffer */
> - ionic_rxq_post(q, false, ionic_rx_clean, NULL);
> - continue;
> - }
> -
> - /* fill main descriptor - pages[0] */
> - desc->opcode = (nfrags > 1) ? IONIC_RXQ_DESC_OPCODE_SG :
> - IONIC_RXQ_DESC_OPCODE_SIMPLE;
> - desc_info->npages = nfrags;
> - if (unlikely(ionic_rx_page_alloc(q, page_info))) {
> - desc->addr = 0;
> - desc->len = 0;
> - return;
> + if (!buf_info->page) { /* alloc a new buffer? */
> + if (unlikely(ionic_rx_page_alloc(q, buf_info))) {
> + desc->addr = 0;
> + desc->len = 0;
> + return;
> + }
> }
> - desc->addr = cpu_to_le64(page_info->dma_addr);
> - seg_len = min_t(unsigned int, PAGE_SIZE, len);
> - desc->len = cpu_to_le16(seg_len);
> - remain_len -= seg_len;
> - page_info++;
>
> - /* fill sg descriptors - pages[1..n] */
> - for (j = 0; j < nfrags - 1; j++) {
> - if (page_info->page) /* recycle the sg buffer */
> - continue;
> + /* fill main descriptor - buf[0] */
> + desc->addr = cpu_to_le64(buf_info->dma_addr + buf_info->page_offset);
> + frag_len = min_t(u16, len, IONIC_PAGE_SIZE - buf_info->page_offset);
> + desc->len = cpu_to_le16(frag_len);
> + remain_len -= frag_len;
> + buf_info++;
> + nfrags++;
>
> + /* fill sg descriptors - buf[1..n] */
> + sg_desc = desc_info->sg_desc;
> + for (j = 0; remain_len > 0 && j < max_sg_elems; j++) {
> sg_elem = &sg_desc->elems[j];
> - if (unlikely(ionic_rx_page_alloc(q, page_info))) {
> - sg_elem->addr = 0;
> - sg_elem->len = 0;
> - return;
> + if (!buf_info->page) { /* alloc a new sg buffer? */
> + if (unlikely(ionic_rx_page_alloc(q, buf_info))) {
> + sg_elem->addr = 0;
> + sg_elem->len = 0;
> + return;
> + }
> }
> - sg_elem->addr = cpu_to_le64(page_info->dma_addr);
> - seg_len = min_t(unsigned int, PAGE_SIZE, remain_len);
> - sg_elem->len = cpu_to_le16(seg_len);
> - remain_len -= seg_len;
> - page_info++;
> +
> + sg_elem->addr = cpu_to_le64(buf_info->dma_addr + buf_info->page_offset);
> + frag_len = min_t(u16, remain_len, IONIC_PAGE_SIZE - buf_info->page_offset);
> + sg_elem->len = cpu_to_le16(frag_len);
> + remain_len -= frag_len;
> + buf_info++;
> + nfrags++;
> }
>
> + /* clear end sg element as a sentinel */
> + if (j < max_sg_elems) {
> + sg_elem = &sg_desc->elems[j];
> + memset(sg_elem, 0, sizeof(*sg_elem));
> + }
> +
> + desc->opcode = (nfrags > 1) ? IONIC_RXQ_DESC_OPCODE_SG :
> + IONIC_RXQ_DESC_OPCODE_SIMPLE;
> + desc_info->nbufs = nfrags;
> +
> ionic_rxq_post(q, false, ionic_rx_clean, NULL);
> }
>
> @@ -395,21 +435,24 @@ void ionic_rx_fill(struct ionic_queue *q)
> void ionic_rx_empty(struct ionic_queue *q)
> {
> struct ionic_desc_info *desc_info;
> - struct ionic_page_info *page_info;
> + struct ionic_buf_info *buf_info;
> unsigned int i, j;
>
> for (i = 0; i < q->num_descs; i++) {
> desc_info = &q->info[i];
> for (j = 0; j < IONIC_RX_MAX_SG_ELEMS + 1; j++) {
> - page_info = &desc_info->pages[j];
> - if (page_info->page)
> - ionic_rx_page_free(q, page_info);
> + buf_info = &desc_info->bufs[j];
> + if (buf_info->page)
> + ionic_rx_page_free(q, buf_info);
> }
>
> - desc_info->npages = 0;
> + desc_info->nbufs = 0;
> desc_info->cb = NULL;
> desc_info->cb_arg = NULL;
> }
> +
> + q->head_idx = 0;
> + q->tail_idx = 0;
> }
>
> static void ionic_dim_update(struct ionic_qcq *qcq)
> --
> 2.17.1
>
next prev parent reply other threads:[~2021-03-11 2:15 UTC|newest]
Thread overview: 10+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-03-10 19:26 [PATCH net-next 0/6] ionic Rx updates Shannon Nelson
2021-03-10 19:26 ` [PATCH net-next 1/6] ionic: move rx_page_alloc and free Shannon Nelson
2021-03-10 19:26 ` [PATCH net-next 2/6] ionic: implement Rx page reuse Shannon Nelson
2021-03-11 2:14 ` Alexander Duyck [this message]
2021-03-12 20:01 ` Shannon Nelson
2021-03-10 19:26 ` [PATCH net-next 3/6] ionic: optimize fastpath struct usage Shannon Nelson
2021-03-10 19:26 ` [PATCH net-next 4/6] ionic: simplify rx skb alloc Shannon Nelson
2021-03-10 19:26 ` [PATCH net-next 5/6] ionic: rebuild debugfs on qcq swap Shannon Nelson
2021-03-10 19:26 ` [PATCH net-next 6/6] ionic: simplify use of completion types Shannon Nelson
2021-03-10 23:40 ` [PATCH net-next 0/6] ionic Rx updates patchwork-bot+netdevbpf
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAKgT0UfyjoAN7LTnq0NMZfXRv4v7iTCPyAb9pVr3qWMhop_BVw@mail.gmail.com \
--to=alexander.duyck@gmail.com \
--cc=davem@davemloft.net \
--cc=drivers@pensando.io \
--cc=kuba@kernel.org \
--cc=netdev@vger.kernel.org \
--cc=snelson@pensando.io \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).