| Message ID | 20241021061023.2162701-1-0x1207@gmail.com (mailing list archive) |
|---|---|
| State | New |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net,v1] net: stmmac: TSO: Fix unbalanced DMA map/unmap for non-paged SKB data | expand |
On 2024-10-21 at 11:40:23, Furong Xu (0x1207@gmail.com) wrote: > In case the non-paged data of a SKB carries protocol header and protocol > payload to be transmitted on a certain platform that the DMA AXI address > width is configured to 40-bit/48-bit, or the size of the non-paged data > is bigger than TSO_MAX_BUFF_SIZE on a certain platform that the DMA AXI > address width is configured to 32-bit, then this SKB requires at least > two DMA transmit descriptors to serve it. > > For example, three descriptors are allocated to split one DMA buffer > mapped from one piece of non-paged data: > dma_desc[N + 0], > dma_desc[N + 1], > dma_desc[N + 2]. > Then three elements of tx_q->tx_skbuff_dma[] will be allocated to hold > extra information to be reused in stmmac_tx_clean(): > tx_q->tx_skbuff_dma[N + 0], > tx_q->tx_skbuff_dma[N + 1], > tx_q->tx_skbuff_dma[N + 2]. > Now we focus on tx_q->tx_skbuff_dma[entry].buf, which is the DMA buffer > address returned by DMA mapping call. stmmac_tx_clean() will try to > unmap the DMA buffer _ONLY_IF_ tx_q->tx_skbuff_dma[entry].buf > is a valid buffer address. > > The expected behavior that saves DMA buffer address of this non-paged > data to tx_q->tx_skbuff_dma[entry].buf is: > tx_q->tx_skbuff_dma[N + 0].buf = NULL; > tx_q->tx_skbuff_dma[N + 1].buf = NULL; > tx_q->tx_skbuff_dma[N + 2].buf = dma_map_single(); > Unfortunately, the current code misbehaves like this: > tx_q->tx_skbuff_dma[N + 0].buf = dma_map_single(); > tx_q->tx_skbuff_dma[N + 1].buf = NULL; > tx_q->tx_skbuff_dma[N + 2].buf = NULL; > > On the stmmac_tx_clean() side, when dma_desc[N + 0] is closed by the > DMA engine, tx_q->tx_skbuff_dma[N + 0].buf is a valid buffer address > obviously, then the DMA buffer will be unmapped immediately. > There may be a rare case that the DMA engine does not finish the > pending dma_desc[N + 1], dma_desc[N + 2] yet. Now things will go > horribly wrong, DMA is going to access a unmapped/unreferenced memory > region, corrupted data will be transmited or iommu fault will be > triggered :( > > In contrast, the for-loop that maps SKB fragments behaves perfectly > as expected, and that is how the driver should do for both non-paged > data and paged frags actually. > > This patch corrects DMA map/unmap sequences by fixing the array index > for tx_q->tx_skbuff_dma[entry].buf when assigning DMA buffer address. > > Tested and verified on DWXGMAC CORE 3.20a > > Reported-by: Suraj Jaiswal <quic_jsuraj@quicinc.com> > Fixes: f748be531d70 ("stmmac: support new GMAC4") > Signed-off-by: Furong Xu <0x1207@gmail.com> > --- > .../net/ethernet/stmicro/stmmac/stmmac_main.c | 22 ++++++++++++++----- > 1 file changed, 17 insertions(+), 5 deletions(-) > > diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c > index d3895d7eecfc..208dbc68aaf9 100644 > --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c > +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c > @@ -4304,11 +4304,6 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev) > if (dma_mapping_error(priv->device, des)) > goto dma_map_err; > > - tx_q->tx_skbuff_dma[first_entry].buf = des; > - tx_q->tx_skbuff_dma[first_entry].len = skb_headlen(skb); > - tx_q->tx_skbuff_dma[first_entry].map_as_page = false; > - tx_q->tx_skbuff_dma[first_entry].buf_type = STMMAC_TXBUF_T_SKB; > - > if (priv->dma_cap.addr64 <= 32) { > first->des0 = cpu_to_le32(des); > > @@ -4327,6 +4322,23 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev) > > stmmac_tso_allocator(priv, des, tmp_pay_len, (nfrags == 0), queue); > > + /* In case two or more DMA transmit descriptors are allocated for this > + * non-paged SKB data, the DMA buffer address should be saved to > + * tx_q->tx_skbuff_dma[].buf corresponding to the last descriptor, > + * and leave the other tx_q->tx_skbuff_dma[].buf as NULL to guarantee > + * that stmmac_tx_clean() does not unmap the entire DMA buffer too early > + * since the tail areas of the DMA buffer can be accessed by DMA engine > + * sooner or later. > + * By saving the DMA buffer address to tx_q->tx_skbuff_dma[].buf > + * corresponding to the last descriptor, stmmac_tx_clean() will unmap > + * this DMA buffer right after the DMA engine completely finishes the > + * full buffer transmission. > + */ > + tx_q->tx_skbuff_dma[tx_q->cur_tx].buf = des; > + tx_q->tx_skbuff_dma[tx_q->cur_tx].len = skb_headlen(skb); > + tx_q->tx_skbuff_dma[tx_q->cur_tx].map_as_page = false; > + tx_q->tx_skbuff_dma[tx_q->cur_tx].buf_type = STMMAC_TXBUF_T_SKB; > + > /* Prepare fragments */ > for (i = 0; i < nfrags; i++) { > const skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; > -- Reviewed-by: Hariprasad Kelam <hkelam@marvell.com>
On Mon, Oct 21, 2024 at 02:10:23PM +0800, Furong Xu wrote: > In case the non-paged data of a SKB carries protocol header and protocol > payload to be transmitted on a certain platform that the DMA AXI address > width is configured to 40-bit/48-bit, or the size of the non-paged data > is bigger than TSO_MAX_BUFF_SIZE on a certain platform that the DMA AXI > address width is configured to 32-bit, then this SKB requires at least > two DMA transmit descriptors to serve it. > > For example, three descriptors are allocated to split one DMA buffer > mapped from one piece of non-paged data: > dma_desc[N + 0], > dma_desc[N + 1], > dma_desc[N + 2]. > Then three elements of tx_q->tx_skbuff_dma[] will be allocated to hold > extra information to be reused in stmmac_tx_clean(): > tx_q->tx_skbuff_dma[N + 0], > tx_q->tx_skbuff_dma[N + 1], > tx_q->tx_skbuff_dma[N + 2]. > Now we focus on tx_q->tx_skbuff_dma[entry].buf, which is the DMA buffer > address returned by DMA mapping call. stmmac_tx_clean() will try to > unmap the DMA buffer _ONLY_IF_ tx_q->tx_skbuff_dma[entry].buf > is a valid buffer address. > > The expected behavior that saves DMA buffer address of this non-paged > data to tx_q->tx_skbuff_dma[entry].buf is: > tx_q->tx_skbuff_dma[N + 0].buf = NULL; > tx_q->tx_skbuff_dma[N + 1].buf = NULL; > tx_q->tx_skbuff_dma[N + 2].buf = dma_map_single(); > Unfortunately, the current code misbehaves like this: > tx_q->tx_skbuff_dma[N + 0].buf = dma_map_single(); > tx_q->tx_skbuff_dma[N + 1].buf = NULL; > tx_q->tx_skbuff_dma[N + 2].buf = NULL; > > On the stmmac_tx_clean() side, when dma_desc[N + 0] is closed by the > DMA engine, tx_q->tx_skbuff_dma[N + 0].buf is a valid buffer address > obviously, then the DMA buffer will be unmapped immediately. > There may be a rare case that the DMA engine does not finish the > pending dma_desc[N + 1], dma_desc[N + 2] yet. Now things will go > horribly wrong, DMA is going to access a unmapped/unreferenced memory > region, corrupted data will be transmited or iommu fault will be > triggered :( > > In contrast, the for-loop that maps SKB fragments behaves perfectly > as expected, and that is how the driver should do for both non-paged > data and paged frags actually. > > This patch corrects DMA map/unmap sequences by fixing the array index > for tx_q->tx_skbuff_dma[entry].buf when assigning DMA buffer address. > > Tested and verified on DWXGMAC CORE 3.20a > > Reported-by: Suraj Jaiswal <quic_jsuraj@quicinc.com> > Fixes: f748be531d70 ("stmmac: support new GMAC4") > Signed-off-by: Furong Xu <0x1207@gmail.com> Thanks for the very thorough explanation, much appreciated. Reviewed-by: Simon Horman <horms@kernel.org>
diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index d3895d7eecfc..208dbc68aaf9 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4304,11 +4304,6 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev) if (dma_mapping_error(priv->device, des)) goto dma_map_err; - tx_q->tx_skbuff_dma[first_entry].buf = des; - tx_q->tx_skbuff_dma[first_entry].len = skb_headlen(skb); - tx_q->tx_skbuff_dma[first_entry].map_as_page = false; - tx_q->tx_skbuff_dma[first_entry].buf_type = STMMAC_TXBUF_T_SKB; - if (priv->dma_cap.addr64 <= 32) { first->des0 = cpu_to_le32(des); @@ -4327,6 +4322,23 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev) stmmac_tso_allocator(priv, des, tmp_pay_len, (nfrags == 0), queue); + /* In case two or more DMA transmit descriptors are allocated for this + * non-paged SKB data, the DMA buffer address should be saved to + * tx_q->tx_skbuff_dma[].buf corresponding to the last descriptor, + * and leave the other tx_q->tx_skbuff_dma[].buf as NULL to guarantee + * that stmmac_tx_clean() does not unmap the entire DMA buffer too early + * since the tail areas of the DMA buffer can be accessed by DMA engine + * sooner or later. + * By saving the DMA buffer address to tx_q->tx_skbuff_dma[].buf + * corresponding to the last descriptor, stmmac_tx_clean() will unmap + * this DMA buffer right after the DMA engine completely finishes the + * full buffer transmission. + */ + tx_q->tx_skbuff_dma[tx_q->cur_tx].buf = des; + tx_q->tx_skbuff_dma[tx_q->cur_tx].len = skb_headlen(skb); + tx_q->tx_skbuff_dma[tx_q->cur_tx].map_as_page = false; + tx_q->tx_skbuff_dma[tx_q->cur_tx].buf_type = STMMAC_TXBUF_T_SKB; + /* Prepare fragments */ for (i = 0; i < nfrags; i++) { const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
In case the non-paged data of a SKB carries protocol header and protocol payload to be transmitted on a certain platform that the DMA AXI address width is configured to 40-bit/48-bit, or the size of the non-paged data is bigger than TSO_MAX_BUFF_SIZE on a certain platform that the DMA AXI address width is configured to 32-bit, then this SKB requires at least two DMA transmit descriptors to serve it. For example, three descriptors are allocated to split one DMA buffer mapped from one piece of non-paged data: dma_desc[N + 0], dma_desc[N + 1], dma_desc[N + 2]. Then three elements of tx_q->tx_skbuff_dma[] will be allocated to hold extra information to be reused in stmmac_tx_clean(): tx_q->tx_skbuff_dma[N + 0], tx_q->tx_skbuff_dma[N + 1], tx_q->tx_skbuff_dma[N + 2]. Now we focus on tx_q->tx_skbuff_dma[entry].buf, which is the DMA buffer address returned by DMA mapping call. stmmac_tx_clean() will try to unmap the DMA buffer _ONLY_IF_ tx_q->tx_skbuff_dma[entry].buf is a valid buffer address. The expected behavior that saves DMA buffer address of this non-paged data to tx_q->tx_skbuff_dma[entry].buf is: tx_q->tx_skbuff_dma[N + 0].buf = NULL; tx_q->tx_skbuff_dma[N + 1].buf = NULL; tx_q->tx_skbuff_dma[N + 2].buf = dma_map_single(); Unfortunately, the current code misbehaves like this: tx_q->tx_skbuff_dma[N + 0].buf = dma_map_single(); tx_q->tx_skbuff_dma[N + 1].buf = NULL; tx_q->tx_skbuff_dma[N + 2].buf = NULL; On the stmmac_tx_clean() side, when dma_desc[N + 0] is closed by the DMA engine, tx_q->tx_skbuff_dma[N + 0].buf is a valid buffer address obviously, then the DMA buffer will be unmapped immediately. There may be a rare case that the DMA engine does not finish the pending dma_desc[N + 1], dma_desc[N + 2] yet. Now things will go horribly wrong, DMA is going to access a unmapped/unreferenced memory region, corrupted data will be transmited or iommu fault will be triggered :( In contrast, the for-loop that maps SKB fragments behaves perfectly as expected, and that is how the driver should do for both non-paged data and paged frags actually. This patch corrects DMA map/unmap sequences by fixing the array index for tx_q->tx_skbuff_dma[entry].buf when assigning DMA buffer address. Tested and verified on DWXGMAC CORE 3.20a Reported-by: Suraj Jaiswal <quic_jsuraj@quicinc.com> Fixes: f748be531d70 ("stmmac: support new GMAC4") Signed-off-by: Furong Xu <0x1207@gmail.com> --- .../net/ethernet/stmicro/stmmac/stmmac_main.c | 22 ++++++++++++++----- 1 file changed, 17 insertions(+), 5 deletions(-)