On Fri, 2022-12-23 at 00:57 +0100, lorenzo@xxxxxxxxxx wrote: > > > On Thu, 2022-12-22 at 10:44 +0100, lorenzo@xxxxxxxxxx wrote: > > > > > On Wed, 2022-12-21 at 19:02 +0100, Lorenzo Bianconi wrote: > > > > > > On Dec 21, Felix Fietkau wrote: > > > > > > > Hi Sujuan, > > > > > > > > > > > > > > > Yes, it is so expensive, but if no memcopy, it will > > > > > > > > casue > > > > > > > > memory > > > > > > > > fragmentation (we hit this issue in internal SQC). > > > > > > > > > > > > > > > > as we know, wed needs to exchange rx pkt(belongs to wed > > > > > > > > rx > > > > > > > > buffer > > > > > > > > manager) with wifi driver(dma rx data queue) by > > > > > > > > exchanging > > > > > > > > wfdma > > > > > > > > dmad > > > > > > > > to ensure the free wed rx buffer. > > > > > > > > > > > > > > > > it is possiable that a large number of buffer has been > > > > > > > > exchanged > > > > > > > > to wed > > > > > > > > and can not come back to wlan driver. So, the memory > > > > > > > > from the > > > > > > > > same 32K > > > > > > > > page cache is unable to be released, and it will be > > > > > > > > failed at > > > > > > > > page_frag_alloc in mt76_dma_rx_fill. > > > > > > > > > > > > > > > > Any ideas but memcopy? > > > > > > > > > > > > > > A simple solution would be to simply allocate single > > > > > > > pages, or > > > > > > > half-page fragments. > > > > > > > > > > > > > > - Felix > > > > > > > > > > > > > > > > > > > A simple approach would be allocating a single page for > > > > > > each rx > > > > > > buffer. > > > > > > > > > > > > @Sujuan: can you please double check if it is ok from > > > > > > performance > > > > > > and > > > > > > memory > > > > > > fragmentation point of view? If not I guess we can try > > > > > > to > > > > > > optimize it > > > > > > and allocate multiple buffers in the same page > > > > > > tweeking page > > > > > > refcount. > > > > > > > > > > > > (this patch must be applied on top of Felix's dma fix). > > > > > > > > > > > > > > > > Allocating single page for each rx buffer avoids memory > > > > > fragmentation, > > > > > but it always uses 4K for one rx pkt which only needs 2K, > > > > > right? > > > > > > > > correct, we can optimize it allocating multiple buffers (in > > > > this case > > > > 2, > > > > assuming 4K pages) in a single page and recycling the page. > > > > > > > > > > > > > > I guess performance would be worse without page cache. > > > > > > > > I think it is a trade-off > > > > > > > > > We have tested on the mtk private driver, 7% drop in > > > > > throughput > > > > > when > > > > > setting the 4K page cache compared to the 32K page cache. > > > > > and 10% drop when use slab to allocate buffer. > > > > > > > > IIUC you tested with PAGE_FRAG_CACHE_MAX_SIZE set to 4K (or > > > > with a > > > > private > > > > page_frag_alloc() implementation) and you avoided memory > > > > allocation > > > > failures due to fragmentation but you got ~ 7% drop in > > > > throughput, > > > > correct? > > > > I think this is quite expected since we need to refill ~ 8 > > > > times more > > > > the > > > > page cache. > > > > > > > > Not considering memory fragmentation, have you measured the > > > > impact of > > > > the > > > > memcpy of a full buffer? > > > > > > > > > > well, for pure sw path, it maybe ~300M drop (not sure) when using > > > memcpy(5G/HE80/4*4). > > > but we do memcpy only when wed on, atfer a few unbinding pkts, > > > then all > > > flows are offloaded by hw. it is also a trade-off~ > > > > I guess we have quite a big impact if we enable wed and we do not > > actually > > offload any flow (or we do not enable hw flowtable in netfilter), > > dont' we? > > Hi Sujuan, > > here ([0],[1]) I switched mt76 to page_pool allocator [2]. page_pool > allows us to > allocate pages of a requested order (order 0 for mt76) and split them > between > multiple buffers (2 buffers per page for mt7915). Moreover pp > allocator recycles > the pages in order to avoid the page allocator overhead and keep the > pages DMA mapped > if requested (I implemented the support for mt7915). > The code is stable (with and without WED) but I have not run any > performance > tests. Can you please help with them to see if this code helps > solving the > discussed issue? > Hi Lore, I tried to apply your patches into my codebase(kernel-5.x). However since page pool api coming from kernel-6.x does not exist in kernel- 5.x, I can not build a proper image for testing (even thought we applied 30+ page pool patches). Sorry for not helping you. Reagrds, Sujuan > Regards, > Lorenzo > > [0] > https://github.com/LorenzoBianconi/wireless-drivers-next/commit/61ce8e9be04b40fa8a1c0f2778d266982be32d7a > [1] > https://github.com/LorenzoBianconi/wireless-drivers-next/commit/f54eefce06d389b155574092d246b21631e42d47 > [2] https://docs.kernel.org/networking/page_pool.html > > > > > Regards, > > Lorenzo > > > > > > > > Regards, > > > Sujuan > > > > > > > > > > > > > A single page per rx buffer may cause a throughput drop of > > > > > over 7% > > > > > and > > > > > waste memory, what do you think? > > > > > > > > Implementing the page recycles as it is done in > > > > page_frag_alloc() we > > > > should get > > > > the same results you got with PAGE_FRAG_CACHE_MAX_SIZE set to > > > > 4K. > > > > > > > > Regards, > > > > Lorenzo > > > > > > > > > > > > > > Regards, > > > > > Sujuan > > > > > > > > > > > Regards, > > > > > > Lorenzo > > > > > > > > > > > > diff --git a/drivers/net/wireless/mediatek/mt76/dma.c > > > > > > b/drivers/net/wireless/mediatek/mt76/dma.c > > > > > > index 28a7fe064313..1d9e580977fc 100644 > > > > > > --- a/drivers/net/wireless/mediatek/mt76/dma.c > > > > > > +++ b/drivers/net/wireless/mediatek/mt76/dma.c > > > > > > @@ -580,6 +580,20 @@ mt76_dma_tx_queue_skb(struct mt76_dev > > > > > > *dev, > > > > > > struct mt76_queue *q, > > > > > > return ret; > > > > > > } > > > > > > > > > > > > +static void * > > > > > > +mt76_dma_get_rx_buf(struct mt76_queue *q) > > > > > > +{ > > > > > > + if ((q->flags & MT_QFLAG_WED) && > > > > > > + FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == > > > > > > MT76_WED_Q_RX) { > > > > > > + /* WED RX queue */ > > > > > > + struct page *page = dev_alloc_page(); > > > > > > + > > > > > > + return page ? page_address(page) : NULL; > > > > > > + } > > > > > > + > > > > > > + return page_frag_alloc(&q->rx_page, q->buf_size, > > > > > > GFP_ATOMIC); > > > > > > +} > > > > > > + > > > > > > static int > > > > > > mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue > > > > > > *q) > > > > > > { > > > > > > @@ -596,7 +610,7 @@ mt76_dma_rx_fill(struct mt76_dev *dev, > > > > > > struct > > > > > > mt76_queue *q) > > > > > > struct mt76_queue_buf qbuf; > > > > > > void *buf = NULL; > > > > > > > > > > > > - buf = page_frag_alloc(&q->rx_page, q->buf_size, > > > > > > GFP_ATOMIC); > > > > > > + buf = mt76_dma_get_rx_buf(q); > > > > > > if (!buf) > > > > > > break; > > > > > > > > > > > > diff --git > > > > > > a/drivers/net/wireless/mediatek/mt76/mt7915/mmio.c > > > > > > b/drivers/net/wireless/mediatek/mt76/mt7915/mmio.c > > > > > > index 1a2e4df8d1b5..2924e71e4fbe 100644 > > > > > > --- a/drivers/net/wireless/mediatek/mt76/mt7915/mmio.c > > > > > > +++ b/drivers/net/wireless/mediatek/mt76/mt7915/mmio.c > > > > > > @@ -594,13 +594,9 @@ static void > > > > > > mt7915_mmio_wed_offload_disable(struct mtk_wed_device *wed) > > > > > > static void mt7915_mmio_wed_release_rx_buf(struct > > > > > > mtk_wed_device > > > > > > *wed) > > > > > > { > > > > > > struct mt7915_dev *dev; > > > > > > - u32 length; > > > > > > int i; > > > > > > > > > > > > dev = container_of(wed, struct mt7915_dev, > > > > > > mt76.mmio.wed); > > > > > > - length = SKB_DATA_ALIGN(NET_SKB_PAD + wed->wlan.rx_size > > > > > > + > > > > > > - sizeof(struct > > > > > > skb_shared_info)); > > > > > > - > > > > > > for (i = 0; i < dev->mt76.rx_token_size; i++) { > > > > > > struct mt76_txwi_cache *t; > > > > > > > > > > > > @@ -610,7 +606,7 @@ static void > > > > > > mt7915_mmio_wed_release_rx_buf(struct > > > > > > mtk_wed_device *wed) > > > > > > > > > > > > dma_unmap_single(dev->mt76.dma_dev, t- > > > > > > >dma_addr, > > > > > > wed->wlan.rx_size, > > > > > > DMA_FROM_DEVICE); > > > > > > - __free_pages(virt_to_page(t->ptr), > > > > > > get_order(length)); > > > > > > + free_page(virt_to_page(t->ptr)); > > > > > > t->ptr = NULL; > > > > > > > > > > > > mt76_put_rxwi(&dev->mt76, t); > > > > > > @@ -621,13 +617,9 @@ static u32 > > > > > > mt7915_mmio_wed_init_rx_buf(struct > > > > > > mtk_wed_device *wed, int size) > > > > > > { > > > > > > struct mtk_rxbm_desc *desc = wed->rx_buf_ring.desc; > > > > > > struct mt7915_dev *dev; > > > > > > - u32 length; > > > > > > int i; > > > > > > > > > > > > dev = container_of(wed, struct mt7915_dev, > > > > > > mt76.mmio.wed); > > > > > > - length = SKB_DATA_ALIGN(NET_SKB_PAD + wed->wlan.rx_size > > > > > > + > > > > > > - sizeof(struct > > > > > > skb_shared_info)); > > > > > > - > > > > > > for (i = 0; i < size; i++) { > > > > > > struct mt76_txwi_cache *t = mt76_get_rxwi(&dev- > > > > > > >mt76); > > > > > > dma_addr_t phy_addr; > > > > > > @@ -635,7 +627,7 @@ static u32 > > > > > > mt7915_mmio_wed_init_rx_buf(struct > > > > > > mtk_wed_device *wed, int size) > > > > > > int token; > > > > > > void *ptr; > > > > > > > > > > > > - page = __dev_alloc_pages(GFP_KERNEL, > > > > > > get_order(length)); > > > > > > + page = __dev_alloc_page(GFP_KERNEL); > > > > > > if (!page) > > > > > > goto unmap; > > > > > > > > > > > > @@ -644,7 +636,7 @@ static u32 > > > > > > mt7915_mmio_wed_init_rx_buf(struct > > > > > > mtk_wed_device *wed, int size) > > > > > > wed->wlan.rx_size, > > > > > > DMA_TO_DEVICE); > > > > > > if (unlikely(dma_mapping_error(dev->mt76.dev, > > > > > > phy_addr))) { > > > > > > - __free_pages(page, get_order(length)); > > > > > > + free_page(page); > > > > > > goto unmap; > > > > > > } > > > > > > > > > > > > @@ -653,7 +645,7 @@ static u32 > > > > > > mt7915_mmio_wed_init_rx_buf(struct > > > > > > mtk_wed_device *wed, int size) > > > > > > if (token < 0) { > > > > > > dma_unmap_single(dev->mt76.dma_dev, > > > > > > phy_addr, > > > > > > wed->wlan.rx_size, > > > > > > DMA_TO_DEVICE); > > > > > > - __free_pages(page, get_order(length)); > > > > > > + free_page(page); > > > > > > goto unmap; > > > > > > } > > > > > > > >
