| Message ID | TYCPR01MB8437EE1595CD4AB581C006E898C62@TYCPR01MB8437.jpnprd01.prod.outlook.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v1] mt76: mt7915: wed: find rx token by physical address | expand |
Hi Shih, Can you also give this patch a review? Seems I also need to convert to reverse Xmas tree order for this patch, will do it in next version. Best regards, Shengyu 在 2025/2/22 19:38, Shengyu Qu 写道: > The token id in RxDMAD may be incorrect when it is not the last frame > due to WED HW bug. Lookup correct token id by physical address in sdp0. > > Downstream patch link: https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/737340322ab22b138fd200e020d61ffdbe3e36a9/autobuild/autobuild_5.4_mac80211_release/mt7988_wifi7_mac80211_mlo/package/kernel/mt76/patches/0062-mtk-wifi-mt76-mt7915-wed-find-rx-token-by-physical-a.patch > > Signed-off-by: Peter Chiu <chui-hao.chiu@mediatek.com> > Signed-off-by: Shengyu Qu <wiagn233@outlook.com> > --- > drivers/net/wireless/mediatek/mt76/dma.c | 25 +++++++++++++++++++++++- > 1 file changed, 24 insertions(+), 1 deletion(-) > > diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c > index 844af16ee551..5bf63014263c 100644 > --- a/drivers/net/wireless/mediatek/mt76/dma.c > +++ b/drivers/net/wireless/mediatek/mt76/dma.c > @@ -444,9 +444,32 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx, > mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info); > > if (mt76_queue_is_wed_rx(q)) { > + u32 id, find = 0; > u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1); > - struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token); > + struct mt76_txwi_cache *t; > + > + if (*more) { > + spin_lock_bh(&dev->rx_token_lock); > + > + idr_for_each_entry(&dev->rx_token, t, id) { > + if (t->dma_addr == le32_to_cpu(desc->buf0)) { > + find = 1; > + token = id; > + > + /* Write correct id back to DMA*/ > + u32p_replace_bits(&buf1, id, > + MT_DMA_CTL_TOKEN); > + WRITE_ONCE(desc->buf1, cpu_to_le32(buf1)); > + break; > + } > + } > + > + spin_unlock_bh(&dev->rx_token_lock); > + if (!find) > + return NULL; > + } > > + t = mt76_rx_token_release(dev, token); > if (!t) > return NULL; >
Shengyu Qu <wiagn233@outlook.com> wrote: > > Hi Shih, > > Can you also give this patch a review? Seems I also need to convert to > reverse Xmas tree order for this patch, will do it in next version. Yes. Please do it. > > Best regards, > Shengyu > > 在 2025/2/22 19:38, Shengyu Qu 写道: > > The token id in RxDMAD may be incorrect when it is not the last frame > > due to WED HW bug. Lookup correct token id by physical address in sdp0. > > > > Downstream patch link: > https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/737340322ab22b138fd200e02 > 0d61ffdbe3e36a9/autobuild/autobuild_5.4_mac80211_release/mt7988_wifi7_mac80211_mlo/package/kernel/mt76 > /patches/0062-mtk-wifi-mt76-mt7915-wed-find-rx-token-by-physical-a.patch > > > > Signed-off-by: Peter Chiu <chui-hao.chiu@mediatek.com> > > Signed-off-by: Shengyu Qu <wiagn233@outlook.com> > > --- > > drivers/net/wireless/mediatek/mt76/dma.c | 25 +++++++++++++++++++++++- > > 1 file changed, 24 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c > > index 844af16ee551..5bf63014263c 100644 > > --- a/drivers/net/wireless/mediatek/mt76/dma.c > > +++ b/drivers/net/wireless/mediatek/mt76/dma.c > > @@ -444,9 +444,32 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx, > > mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info); > > > > if (mt76_queue_is_wed_rx(q)) { > > + u32 id, find = 0; > > u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1); > > - struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token); > > + struct mt76_txwi_cache *t; > > + > > + if (*more) { Because of scope, should define 'find' (or 'found'?) as bool here. > > + spin_lock_bh(&dev->rx_token_lock); > > + > > + idr_for_each_entry(&dev->rx_token, t, id) { idr_find() can only get a pointer associated with the ID. Not sure if it can have a reverse way -- get ID from pointer. With sequential search in data path, cost is higher. Maybe you can consider to build association between argument 'idx' and pointer/token by another patch. But logic is okay to me. > > + if (t->dma_addr == le32_to_cpu(desc->buf0)) { > > + find = 1; > > + token = id; > > + > > + /* Write correct id back to DMA*/ > > + u32p_replace_bits(&buf1, id, > > + MT_DMA_CTL_TOKEN); > > + WRITE_ONCE(desc->buf1, cpu_to_le32(buf1)); > > + break; > > + } > > + } > > + > > + spin_unlock_bh(&dev->rx_token_lock); > > + if (!find) > > + return NULL; > > + } > > > > + t = mt76_rx_token_release(dev, token); > > if (!t) > > return NULL; > >
On 12.03.25 02:41, Ping-Ke Shih wrote: > Shengyu Qu <wiagn233@outlook.com> wrote: >> >> Hi Shih, >> >> Can you also give this patch a review? Seems I also need to convert to >> reverse Xmas tree order for this patch, will do it in next version. > > Yes. Please do it. > >> >> Best regards, >> Shengyu >> >> 在 2025/2/22 19:38, Shengyu Qu 写道: >> > The token id in RxDMAD may be incorrect when it is not the last frame >> > due to WED HW bug. Lookup correct token id by physical address in sdp0. >> > >> > Downstream patch link: >> https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/737340322ab22b138fd200e02 >> 0d61ffdbe3e36a9/autobuild/autobuild_5.4_mac80211_release/mt7988_wifi7_mac80211_mlo/package/kernel/mt76 >> /patches/0062-mtk-wifi-mt76-mt7915-wed-find-rx-token-by-physical-a.patch >> > >> > Signed-off-by: Peter Chiu <chui-hao.chiu@mediatek.com> >> > Signed-off-by: Shengyu Qu <wiagn233@outlook.com> >> > --- >> > drivers/net/wireless/mediatek/mt76/dma.c | 25 +++++++++++++++++++++++- >> > 1 file changed, 24 insertions(+), 1 deletion(-) >> > >> > diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c >> > index 844af16ee551..5bf63014263c 100644 >> > --- a/drivers/net/wireless/mediatek/mt76/dma.c >> > +++ b/drivers/net/wireless/mediatek/mt76/dma.c >> > @@ -444,9 +444,32 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx, >> > mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info); >> > >> > if (mt76_queue_is_wed_rx(q)) { >> > + u32 id, find = 0; >> > u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1); >> > - struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token); >> > + struct mt76_txwi_cache *t; >> > + >> > + if (*more) { > > Because of scope, should define 'find' (or 'found'?) as bool here. > >> > + spin_lock_bh(&dev->rx_token_lock); >> > + >> > + idr_for_each_entry(&dev->rx_token, t, id) { > > idr_find() can only get a pointer associated with the ID. Not sure if it can > have a reverse way -- get ID from pointer. With sequential search in data path, > cost is higher. > > Maybe you can consider to build association between argument 'idx' and > pointer/token by another patch. > > But logic is okay to me. Adding to that, it matters how frequently the token ID is incorrect. If the token ID is correct most of the time, then it's worth optimizing for that as the default case. In that case you should avoid the extra idr scan if the pointer for the provided token id already matches. - Felix
Hi Felix, I'm not sure about the token ID incorrect frequency, this is just a port of downstream patch, sorry. Best regards, Shengyu 在 2025/3/12 17:09, Felix Fietkau 写道: > On 12.03.25 02:41, Ping-Ke Shih wrote: >> Shengyu Qu <wiagn233@outlook.com> wrote: >>> >>> Hi Shih, >>> >>> Can you also give this patch a review? Seems I also need to convert to >>> reverse Xmas tree order for this patch, will do it in next version. >> >> Yes. Please do it. >> >>> >>> Best regards, >>> Shengyu >>> >>> 在 2025/2/22 19:38, Shengyu Qu 写道: >>> > The token id in RxDMAD may be incorrect when it is not the last frame >>> > due to WED HW bug. Lookup correct token id by physical address in >>> sdp0. >>> > >>> > Downstream patch link: >>> https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt- >>> feeds/+/737340322ab22b138fd200e02 >>> 0d61ffdbe3e36a9/autobuild/autobuild_5.4_mac80211_release/ >>> mt7988_wifi7_mac80211_mlo/package/kernel/mt76 >>> /patches/0062-mtk-wifi-mt76-mt7915-wed-find-rx-token-by-physical-a.patch >>> > >>> > Signed-off-by: Peter Chiu <chui-hao.chiu@mediatek.com> >>> > Signed-off-by: Shengyu Qu <wiagn233@outlook.com> >>> > --- >>> > drivers/net/wireless/mediatek/mt76/dma.c | 25 +++++++++++++++++++ >>> ++++- >>> > 1 file changed, 24 insertions(+), 1 deletion(-) >>> > >>> > diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/ >>> net/wireless/mediatek/mt76/dma.c >>> > index 844af16ee551..5bf63014263c 100644 >>> > --- a/drivers/net/wireless/mediatek/mt76/dma.c >>> > +++ b/drivers/net/wireless/mediatek/mt76/dma.c >>> > @@ -444,9 +444,32 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct >>> mt76_queue *q, int idx, >>> > mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info); >>> > >>> > if (mt76_queue_is_wed_rx(q)) { >>> > + u32 id, find = 0; >>> > u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1); >>> > - struct mt76_txwi_cache *t = mt76_rx_token_release(dev, >>> token); >>> > + struct mt76_txwi_cache *t; >>> > + >>> > + if (*more) { >> >> Because of scope, should define 'find' (or 'found'?) as bool here. >> >>> > + spin_lock_bh(&dev->rx_token_lock); >>> > + >>> > + idr_for_each_entry(&dev->rx_token, t, id) { >> >> idr_find() can only get a pointer associated with the ID. Not sure if >> it can >> have a reverse way -- get ID from pointer. With sequential search in >> data path, >> cost is higher. >> >> Maybe you can consider to build association between argument 'idx' and >> pointer/token by another patch. >> >> But logic is okay to me. > > Adding to that, it matters how frequently the token ID is incorrect. If > the token ID is correct most of the time, then it's worth optimizing for > that as the default case. In that case you should avoid the extra idr > scan if the pointer for the provided token id already matches. > > - Felix
On 13.03.25 13:43, Shengyu Qu wrote: > Hi Felix, > > I'm not sure about the token ID incorrect frequency, this is just a port > of downstream patch, sorry. Can you reproduce the issue that the downstream patch supposedly fixes? If so, you could add some printk to check how frequently it occurs. - Felix
Can you do the test? I'll be busy for months since tomorrow, so really don't have time to do this test... I'll send the patch v2 first. Best regards, Shengyu 在 2025/3/13 20:47, Felix Fietkau 写道: > On 13.03.25 13:43, Shengyu Qu wrote: >> Hi Felix, >> >> I'm not sure about the token ID incorrect frequency, this is just a port >> of downstream patch, sorry. > > Can you reproduce the issue that the downstream patch supposedly fixes? > If so, you could add some printk to check how frequently it occurs. > > - Felix
diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c index 844af16ee551..5bf63014263c 100644 --- a/drivers/net/wireless/mediatek/mt76/dma.c +++ b/drivers/net/wireless/mediatek/mt76/dma.c @@ -444,9 +444,32 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx, mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info); if (mt76_queue_is_wed_rx(q)) { + u32 id, find = 0; u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1); - struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token); + struct mt76_txwi_cache *t; + + if (*more) { + spin_lock_bh(&dev->rx_token_lock); + + idr_for_each_entry(&dev->rx_token, t, id) { + if (t->dma_addr == le32_to_cpu(desc->buf0)) { + find = 1; + token = id; + + /* Write correct id back to DMA*/ + u32p_replace_bits(&buf1, id, + MT_DMA_CTL_TOKEN); + WRITE_ONCE(desc->buf1, cpu_to_le32(buf1)); + break; + } + } + + spin_unlock_bh(&dev->rx_token_lock); + if (!find) + return NULL; + } + t = mt76_rx_token_release(dev, token); if (!t) return NULL;