On Wed, Feb 19, 2025 at 01:42:47PM +0800, Wei Fang wrote:
There is an off-by-one issue for the err_chained_bd path, it will free one more tx_swbd than expected. But there is no such issue for the err_map_data path.
It's clear that one of err_chained_bd or err_map_data is wrong, because they operate with a different "count" but same "i". But how did you determine which one is wrong? Is it based on static analysis? Because I think the other one is wrong, more below.
To fix this off-by-one issue and make the two error handling consistent, the loop condition of error handling is modified and the 'count++' operation is moved before enetc_map_tx_tso_data().
Fixes: fb8629e2cbfc ("net: enetc: add support for software TSO") Cc: stable@vger.kernel.org Signed-off-by: Wei Fang wei.fang@nxp.com
drivers/net/ethernet/freescale/enetc/enetc.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c
b/drivers/net/ethernet/freescale/enetc/enetc.c
index 9a24d1176479..fe3967268a19 100644 --- a/drivers/net/ethernet/freescale/enetc/enetc.c +++ b/drivers/net/ethernet/freescale/enetc/enetc.c @@ -832,6 +832,7 @@ static int enetc_map_tx_tso_buffs(struct enetc_bdr
*tx_ring, struct sk_buff *skb
txbd = ENETC_TXBD(*tx_ring, i); tx_swbd = &tx_ring->tx_swbd[i]; prefetchw(txbd);
count++; /* Compute the checksum over this segment of data and * add it to the csum already computed (over the L4@@ -848,7 +849,6 @@ static int enetc_map_tx_tso_buffs(struct enetc_bdr
*tx_ring, struct sk_buff *skb
goto err_map_data; data_len -= size;
count++; bd_data_num++; tso_build_data(skb, &tso, size);@@ -874,13 +874,13 @@ static int enetc_map_tx_tso_buffs(struct
enetc_bdr *tx_ring, struct sk_buff *skb
dev_err(tx_ring->dev, "DMA map error");
err_chained_bd:
- do {
- while (count--) { tx_swbd = &tx_ring->tx_swbd[i]; enetc_free_tx_frame(tx_ring, tx_swbd); if (i == 0) i = tx_ring->bd_count; i--;
- } while (count--);
}
return 0;
}
ah, there you go, here's the 3rd instance of TX DMA buffer unmapping :-/
Forget what I said in reply to patch 1/9 about having common code later. After going through the whole set and now seeing this, I now think it's better that you create the helper now, and consolidate the 2 instances you touch anyway. Later you can make enetc_lso_hw_offload() reuse this helper in net-next.
It should be something like this in the end (sorry, just 1 squashed diff):
I agree with you that we finally need a helper function to replace all the same code blocks, but I'd like to do that for net-next tree, because as I replied in patch 2, we don't need the count variable. Currently I am more focused on solving the problem itself rather than optimizing it. Of course if you think this is necessary, I can add these changes in the next version. :)
diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c b/drivers/net/ethernet/freescale/enetc/enetc.c index 6178157611db..a70e92dcbe2c 100644 --- a/drivers/net/ethernet/freescale/enetc/enetc.c +++ b/drivers/net/ethernet/freescale/enetc/enetc.c @@ -106,6 +106,24 @@ static void enetc_free_tx_frame(struct enetc_bdr *tx_ring, } }
+/**
- enetc_unwind_tx_frame() - Unwind the DMA mappings of a multi-buffer TX
frame
- @tx_ring: Pointer to the TX ring on which the buffer descriptors are located
- @count: Number of TX buffer descriptors which need to be unmapped
- @i: Index of the last successfully mapped TX buffer descriptor
- */
+static void enetc_unwind_tx_frame(struct enetc_bdr *tx_ring, int count, int i) +{
- while (count--) {
struct enetc_tx_swbd *tx_swbd = &tx_ring->tx_swbd[i];enetc_free_tx_frame(tx_ring, tx_swbd);if (i == 0)i = tx_ring->bd_count;i--;- }
+}
/* Let H/W know BD ring has been updated */ static void enetc_update_tx_ring_tail(struct enetc_bdr *tx_ring) { @@ -399,13 +417,7 @@ static int enetc_map_tx_buffs(struct enetc_bdr *tx_ring, struct sk_buff *skb) dma_err: dev_err(tx_ring->dev, "DMA map error");
- while (count--) {
tx_swbd = &tx_ring->tx_swbd[i];enetc_free_tx_frame(tx_ring, tx_swbd);if (i == 0)i = tx_ring->bd_count;i--;- }
enetc_unwind_tx_frame(tx_ring, count, i);
return 0;
} @@ -752,7 +764,6 @@ static int enetc_lso_map_data(struct enetc_bdr *tx_ring, struct sk_buff *skb,
static int enetc_lso_hw_offload(struct enetc_bdr *tx_ring, struct sk_buff *skb) {
- struct enetc_tx_swbd *tx_swbd; struct enetc_lso_t lso = {0}; int err, i, count = 0;
@@ -776,13 +787,7 @@ static int enetc_lso_hw_offload(struct enetc_bdr *tx_ring, struct sk_buff *skb) return count;
dma_err:
- do {
tx_swbd = &tx_ring->tx_swbd[i];enetc_free_tx_frame(tx_ring, tx_swbd);if (i == 0)i = tx_ring->bd_count;i--;- } while (--count);
enetc_unwind_tx_frame(tx_ring, count, i);
return 0;
} @@ -877,13 +882,7 @@ static int enetc_map_tx_tso_buffs(struct enetc_bdr *tx_ring, struct sk_buff *skb dev_err(tx_ring->dev, "DMA map error");
err_chained_bd:
- while (count--) {
tx_swbd = &tx_ring->tx_swbd[i];enetc_free_tx_frame(tx_ring, tx_swbd);if (i == 0)i = tx_ring->bd_count;i--;- }
enetc_unwind_tx_frame(tx_ring, count, i);
return 0;
}
With the definitions laid out explicitly in a kernel-doc, doesn't the rest of the patch look a bit wrong? Why would you increment "count"
Sorry, I don't understand what you mean " With the definitions laid ou explicitly in a kernel-doc", which kernel-doc?
before enetc_map_tx_tso_data() succeeds? Why isn't the problem that "i" needs to be rolled back on enetc_map_tx_tso_data() failure instead?
The root cause of this issue as you said is that "I" and "count" are not synchronized. Either moving count++ before enetc_map_tx_tso_data() or rolling back 'i' after enetc_map_tx_tso_data() fails should solve the issue. There is no problem with the former, it just loops one more time, and actually the loop does nothing.