[PATCH] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[PATCH] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Zhang Tengfei]

Enabling Flow Director reduces RXPBSIZE to reserve SRAM for the
filter table. Clearing the last FDIR rule set the mode back to
none without restoring that buffer, so the next flow create ran
configure again and subtracted the same amount a second time.

Keep the FDIR mode after rules are removed, and subtract RXPBSIZE
only once until the next hardware reset.

Fixes: 301abd26b118 ("net/ixgbe: reset flow state on clear paths")
Cc: stable@dpdk.org
Signed-off-by: Zhang Tengfei <zhtfdev@gmail.com>
---
 drivers/net/intel/ixgbe/ixgbe_ethdev.c |  7 +++++--
 drivers/net/intel/ixgbe/ixgbe_ethdev.h |  1 +
 drivers/net/intel/ixgbe/ixgbe_fdir.c   | 21 ++++++++++++++-------
 drivers/net/intel/ixgbe/ixgbe_flow.c   |  1 -
 4 files changed, 20 insertions(+), 10 deletions(-)

diff --git a/drivers/net/intel/ixgbe/ixgbe_ethdev.c b/drivers/net/intel/ixgbe/ixgbe_ethdev.c
index cbb7b4831c..ce5b4c51f5 100644
--- a/drivers/net/intel/ixgbe/ixgbe_ethdev.c
+++ b/drivers/net/intel/ixgbe/ixgbe_ethdev.c
@@ -1372,6 +1372,7 @@ static int ixgbe_fdir_filter_init(struct rte_eth_dev *eth_dev)
 	}
 	fdir_info->n_flows = 0;
 	fdir_info->mask_added = FALSE;
+	fdir_info->rx_pb_reduced = false;
 
 	/* drop queue is always fixed */
 	IXGBE_DEV_FDIR_CONF(eth_dev)->drop_queue = IXGBE_FDIR_DROP_QUEUE;
@@ -2541,6 +2542,8 @@ ixgbe_dev_start(struct rte_eth_dev *dev)
 		IXGBE_DEV_PRIVATE_TO_TM_CONF(dev->data->dev_private);
 	struct ixgbe_macsec_setting *macsec_setting =
 		IXGBE_DEV_PRIVATE_TO_MACSEC_SETTING(dev->data->dev_private);
+	struct ixgbe_hw_fdir_info *info =
+		IXGBE_DEV_PRIVATE_TO_FDIR_INFO(adapter);
 
 	/*
 	 * This function calls into the base driver, which in turn will use
@@ -2628,9 +2631,9 @@ ixgbe_dev_start(struct rte_eth_dev *dev)
 	/* Configure DCB hw */
 	ixgbe_configure_dcb(dev);
 
+	/* reset_hw restored RXPBSIZE; FDIR must subtract again */
+	info->rx_pb_reduced = false;
 	if (fdir_conf->mode != RTE_FDIR_MODE_NONE) {
-		struct ixgbe_hw_fdir_info *info =
-			IXGBE_DEV_PRIVATE_TO_FDIR_INFO(adapter);
 		err = ixgbe_fdir_configure(adapter, fdir_conf, &info->mask);
 		if (err)
 			goto error;
diff --git a/drivers/net/intel/ixgbe/ixgbe_ethdev.h b/drivers/net/intel/ixgbe/ixgbe_ethdev.h
index ba826b6d8d..495a6ddd22 100644
--- a/drivers/net/intel/ixgbe/ixgbe_ethdev.h
+++ b/drivers/net/intel/ixgbe/ixgbe_ethdev.h
@@ -200,6 +200,7 @@ struct ixgbe_hw_fdir_info {
 	struct rte_hash *hash_handle; /* cuckoo hash handler */
 	uint32_t n_flows;
 	bool mask_added; /* If already got mask from consistent filter */
+	bool rx_pb_reduced; /* RXPBSIZE already reduced for FDIR */
 };
 
 struct ixgbe_rte_flow_rss_conf {
diff --git a/drivers/net/intel/ixgbe/ixgbe_fdir.c b/drivers/net/intel/ixgbe/ixgbe_fdir.c
index b32dc54287..ca96ea85d7 100644
--- a/drivers/net/intel/ixgbe/ixgbe_fdir.c
+++ b/drivers/net/intel/ixgbe/ixgbe_fdir.c
@@ -560,8 +560,10 @@ ixgbe_fdir_configure(struct ixgbe_adapter *adapter,
 		const struct ixgbe_hw_fdir_mask *fdir_mask)
 {
 	struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(adapter);
+	struct ixgbe_hw_fdir_info *info =
+			IXGBE_DEV_PRIVATE_TO_FDIR_INFO(adapter);
 	int err;
-	uint32_t fdirctrl, pbsize;
+	uint32_t fdirctrl;
 	int i;
 	enum rte_fdir_mode mode = fdir_conf->mode;
 
@@ -591,11 +593,18 @@ ixgbe_fdir_configure(struct ixgbe_adapter *adapter,
 	/*
 	 * Before enabling Flow Director, the Rx Packet Buffer size
 	 * must be reduced.  The new value is the current size minus
-	 * flow director memory usage size.
+	 * flow director memory usage size.  Do this only once; a later
+	 * re-configure (e.g. after a failed filter program) must not
+	 * subtract again.
 	 */
-	pbsize = (1 << (PBALLOC_SIZE_SHIFT + (fdirctrl & FDIRCTRL_PBALLOC_MASK)));
-	IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0),
-	    (IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(0)) - pbsize));
+	if (!info->rx_pb_reduced) {
+		uint32_t pbsize = (1 << (PBALLOC_SIZE_SHIFT +
+					(fdirctrl & FDIRCTRL_PBALLOC_MASK)));
+
+		IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0),
+		    (IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(0)) - pbsize));
+		info->rx_pb_reduced = true;
+	}
 
 	/*
 	 * The defaults in the HW for RX PB 1-7 are not zero and so should be
@@ -1359,7 +1368,6 @@ ixgbe_fdir_filter_restore(struct rte_eth_dev *dev)
 int
 ixgbe_clear_all_fdir_filter(struct rte_eth_dev *dev)
 {
-	struct rte_eth_fdir_conf *fdir_conf = IXGBE_DEV_FDIR_CONF(dev);
 	struct ixgbe_hw_fdir_info *fdir_info =
 		IXGBE_DEV_PRIVATE_TO_FDIR_INFO(dev->data->dev_private);
 	struct ixgbe_fdir_filter *fdir_filter;
@@ -1384,7 +1392,6 @@ ixgbe_clear_all_fdir_filter(struct rte_eth_dev *dev)
 	fdir_info->mask = (struct ixgbe_hw_fdir_mask){0};
 	fdir_info->flex_bytes_offset = 0;
 	fdir_info->mask_added = FALSE;
-	fdir_conf->mode = RTE_FDIR_MODE_NONE;
 
 	if (had_flows)
 		ret = ixgbe_fdir_flush(dev);
diff --git a/drivers/net/intel/ixgbe/ixgbe_flow.c b/drivers/net/intel/ixgbe/ixgbe_flow.c
index 6868893d46..1e79ef8b28 100644
--- a/drivers/net/intel/ixgbe/ixgbe_flow.c
+++ b/drivers/net/intel/ixgbe/ixgbe_flow.c
@@ -3164,7 +3164,6 @@ ixgbe_flow_destroy(struct rte_eth_dev *dev,
 				fdir_info->mask_added = false;
 				fdir_info->mask = (struct ixgbe_hw_fdir_mask){0};
 				fdir_info->flex_bytes_offset = 0;
-				fdir_conf->mode = RTE_FDIR_MODE_NONE;
 			}
 		}
 		break;
-- 
2.53.0

Re: [PATCH] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Bruce Richardson]

On Wed, Sep 16, 2026 at 09:58:27PM +0800, Zhang Tengfei wrote:
> Enabling Flow Director reduces RXPBSIZE to reserve SRAM for the
> filter table. Clearing the last FDIR rule set the mode back to
> none without restoring that buffer, so the next flow create ran
> configure again and subtracted the same amount a second time.
> 
> Keep the FDIR mode after rules are removed, and subtract RXPBSIZE
> only once until the next hardware reset.
> 
> Fixes: 301abd26b118 ("net/ixgbe: reset flow state on clear paths")
> Cc: stable@dpdk.org
> Signed-off-by: Zhang Tengfei <zhtfdev@gmail.com>
> ---
>  drivers/net/intel/ixgbe/ixgbe_ethdev.c |  7 +++++--
>  drivers/net/intel/ixgbe/ixgbe_ethdev.h |  1 +
>  drivers/net/intel/ixgbe/ixgbe_fdir.c   | 21 ++++++++++++++-------
>  drivers/net/intel/ixgbe/ixgbe_flow.c   |  1 -
>  4 files changed, 20 insertions(+), 10 deletions(-)
> 
Reviewing this patch, I'm wondering if this approach is actually the best
way to fix this. For example, after this change, if flow director is used
and then all rules removed, the flow director mode is still preserved,
possibly preventing use of other features that conflict with the old fdir
mode.

I wonder if a cleaner fix - which should be shorter too - might be to
instead just restore the RXPBSIZE value back when the fdir mode is set to
NONE. This gives a clean combination of (enable + set RXPBSIZE0) and (disable +
reset RXPBSIZE).

What do you think?

/Bruce

Re: [PATCH] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Burakov, Anatoly]

On 9/16/2026 3:58 PM, Zhang Tengfei wrote:
> Enabling Flow Director reduces RXPBSIZE to reserve SRAM for the
> filter table. Clearing the last FDIR rule set the mode back to
> none without restoring that buffer, so the next flow create ran
> configure again and subtracted the same amount a second time.
> 
> Keep the FDIR mode after rules are removed, and subtract RXPBSIZE
> only once until the next hardware reset.
> 
> Fixes: 301abd26b118 ("net/ixgbe: reset flow state on clear paths")
> Cc: stable@dpdk.org
> Signed-off-by: Zhang Tengfei <zhtfdev@gmail.com>
> ---

A general comment: instead of reducing/bringing things back and storing 
a flag noting whether we did, I would rather do the following:

0) store default rx pb size at init
1) on enabling FDIR, recalculate using that value minus FDIR table size
2) on disabling FDIR[*], restore the default
3) similarly, on enable/disable VMDq, recalculate and/or reset

[*] there is no "disable FDIR" call, only fdir flush which just flushes 
the FDIR tables but does not actually disable FDIR. arguably, we should 
convert it to "disable FDIR" by flushing FDIR *and* writing 0 to 
FDIRCTRL *and* restoring rx pb size to defaults. naturally, after 
running fdir disable function, FDIR will need to be reconfigured for 
next FDIR flow and get rx pb size recalculated again.

So, a bit of a refactor, but I think that would make way more sense.

-- 
Thanks,
Anatoly

Re: [PATCH] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Burakov, Anatoly]

On 9/25/2026 11:47 AM, Burakov, Anatoly wrote:
> A general comment: instead of reducing/bringing things back and storing 
> a flag noting whether we did, I would rather do the following:
> 
> 0) store default rx pb size at init
> 1) on enabling FDIR, recalculate using that value minus FDIR table size
> 2) on disabling FDIR[*], restore the default
> 3) similarly, on enable/disable VMDq, recalculate and/or reset
> 
> [*] there is no "disable FDIR" call, only fdir flush which just flushes 
> the FDIR tables but does not actually disable FDIR. arguably, we should 
> convert it to "disable FDIR" by flushing FDIR *and* writing 0 to 
> FDIRCTRL *and* restoring rx pb size to defaults. naturally, after 
> running fdir disable function, FDIR will need to be reconfigured for 
> next FDIR flow and get rx pb size recalculated again.
> 
> So, a bit of a refactor, but I think that would make way more sense.

I asked an AI to implement a fix based on this, and here's what it came 
up with, it is roughly what I would like to see instead (obviously, 
please review/rework as appropriate e.g. to properly support VMDq as well):

diff --git a/drivers/net/intel/ixgbe/ixgbe_fdir.c 
b/drivers/net/intel/ixgbe/ixgbe_fdir.c
index b32dc542874..9f48a27cb3a 100644
--- a/drivers/net/intel/ixgbe/ixgbe_fdir.c
+++ b/drivers/net/intel/ixgbe/ixgbe_fdir.c
@@ -101,7 +101,6 @@ static int fdir_write_perfect_filter_82599(struct 
ixgbe_hw *hw,
  static int fdir_add_signature_filter_82599(struct ixgbe_hw *hw,
                 union ixgbe_atr_input *input, u8 queue, uint32_t fdircmd,
                 uint32_t fdirhash);
-static int ixgbe_fdir_flush(struct rte_eth_dev *dev);

  /**
   * This function is based on ixgbe_fdir_enable_82599() in 
base/ixgbe_82599.c.
@@ -554,6 +553,20 @@ ixgbe_set_fdir_flex_conf(struct ixgbe_adapter *adapter,
         return 0;
  }

+static void
+ixgbe_fdir_disable(struct ixgbe_hw *hw)
+{
+       uint32_t rx_pb_size;
+       int i;
+
+       IXGBE_WRITE_REG(hw, IXGBE_FDIRCTRL, 0);
+       rx_pb_size = (uint32_t)hw->mac.rx_pb_size << IXGBE_RXPBSIZE_SHIFT;
+       IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0), rx_pb_size);
+       for (i = 1; i < 8; i++)
+               IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), 0);
+       IXGBE_WRITE_FLUSH(hw);
+}
+
  int
  ixgbe_fdir_configure(struct ixgbe_adapter *adapter,
                 const struct rte_eth_fdir_conf *fdir_conf,
@@ -561,7 +574,7 @@ ixgbe_fdir_configure(struct ixgbe_adapter *adapter,
  {
         struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(adapter);
         int err;
-       uint32_t fdirctrl, pbsize;
+       uint32_t fdirctrl, pbsize, rx_pb_size;
         int i;
         enum rte_fdir_mode mode = fdir_conf->mode;

@@ -589,13 +602,14 @@ ixgbe_fdir_configure(struct ixgbe_adapter *adapter,
                 return err;

         /*
-        * Before enabling Flow Director, the Rx Packet Buffer size
-        * must be reduced.  The new value is the current size minus
-        * flow director memory usage size.
+        * Before enabling Flow Director, the Rx Packet Buffer size must be
+        * reduced. The new value is the default size minus flow director
+        * memory usage size.
          */
-       pbsize = (1 << (PBALLOC_SIZE_SHIFT + (fdirctrl & 
FDIRCTRL_PBALLOC_MASK)));
-       IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0),
-           (IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(0)) - pbsize));
+       pbsize = 1 << (PBALLOC_SIZE_SHIFT +
+                       (fdirctrl & FDIRCTRL_PBALLOC_MASK));
+       rx_pb_size = (uint32_t)hw->mac.rx_pb_size << IXGBE_RXPBSIZE_SHIFT;
+       IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0), rx_pb_size - pbsize);

         /*
          * The defaults in the HW for RX PB 1-7 are not zero and so 
should be
@@ -609,21 +623,25 @@ ixgbe_fdir_configure(struct ixgbe_adapter *adapter,
         err = ixgbe_fdir_set_input_mask(adapter, fdir_mask, mode);
         if (err < 0) {
                 PMD_INIT_LOG(ERR, " Error on setting FD mask");
-               return err;
+               goto error;
         }
         err = ixgbe_set_fdir_flex_conf(adapter, &fdir_conf->flex_conf,
                                        &fdirctrl);
         if (err < 0) {
                 PMD_INIT_LOG(ERR, " Error on setting FD flexible 
arguments.");
-               return err;
+               goto error;
         }

         err = fdir_enable_82599(hw, fdirctrl);
         if (err < 0) {
                 PMD_INIT_LOG(ERR, " Error on enabling FD.");
-               return err;
+               goto error;
         }
         return 0;
+
+error:
+       ixgbe_fdir_disable(hw);
+       return err;
  }

  /*
@@ -1180,28 +1198,6 @@ ixgbe_fdir_filter_program(struct ixgbe_adapter 
*adapter,
         return err;
  }

-static int
-ixgbe_fdir_flush(struct rte_eth_dev *dev)
-{
-       struct ixgbe_hw *hw = 
IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
-       struct ixgbe_hw_fdir_info *info =
- 
IXGBE_DEV_PRIVATE_TO_FDIR_INFO(dev->data->dev_private);
-       int ret;
-
-       ret = ixgbe_reinit_fdir_tables_82599(hw);
-       if (ret < 0) {
-               PMD_INIT_LOG(ERR, "Failed to re-initialize FD table.");
-               return ret;
-       }
-
-       info->f_add = 0;
-       info->f_remove = 0;
-       info->add = 0;
-       info->remove = 0;
-
-       return ret;
-}
-
  #define FDIRENTRIES_NUM_SHIFT 10
  void
  ixgbe_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info 
*fdir_info)
@@ -1360,13 +1356,26 @@ int
  ixgbe_clear_all_fdir_filter(struct rte_eth_dev *dev)
  {
         struct rte_eth_fdir_conf *fdir_conf = IXGBE_DEV_FDIR_CONF(dev);
+       struct ixgbe_hw *hw = 
IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
         struct ixgbe_hw_fdir_info *fdir_info =
                 IXGBE_DEV_PRIVATE_TO_FDIR_INFO(dev->data->dev_private);
         struct ixgbe_fdir_filter *fdir_filter;
-       bool had_flows;
-       int ret = 0;
+       int ret;

-       had_flows = (fdir_info->n_flows != 0);
+       if (fdir_conf->mode != RTE_FDIR_MODE_NONE) {
+               ret = ixgbe_reinit_fdir_tables_82599(hw);
+               if (ret < 0) {
+                       PMD_INIT_LOG(ERR, "Failed to re-initialize FD 
table.");
+                       return ret;
+               }
+
+               fdir_info->f_add = 0;
+               fdir_info->f_remove = 0;
+               fdir_info->add = 0;
+               fdir_info->remove = 0;
+
+               ixgbe_fdir_disable(hw);
+       }

         /* flush flow director */
         rte_hash_reset(fdir_info->hash_handle);
@@ -1386,8 +1395,5 @@ ixgbe_clear_all_fdir_filter(struct rte_eth_dev *dev)
         fdir_info->mask_added = FALSE;
         fdir_conf->mode = RTE_FDIR_MODE_NONE;

-       if (had_flows)
-               ret = ixgbe_fdir_flush(dev);
-
-       return ret;
+       return 0;
  }
diff --git a/drivers/net/intel/ixgbe/ixgbe_flow.c 
b/drivers/net/intel/ixgbe/ixgbe_flow.c
index 6868893d46a..da05e61e8b4 100644
--- a/drivers/net/intel/ixgbe/ixgbe_flow.c
+++ b/drivers/net/intel/ixgbe/ixgbe_flow.c
@@ -3157,15 +3157,15 @@ ixgbe_flow_destroy(struct rte_eth_dev *dev,
                 memcpy(&fdir_rule,
                         &fdir_rule_ptr->filter_info,
                         sizeof(struct ixgbe_fdir_rule));
-               ret = ixgbe_fdir_filter_program(adapter, fdir_conf, 
&fdir_rule, TRUE, FALSE);
+               if (fdir_info->n_flows == 1)
+                       ret = ixgbe_clear_all_fdir_filter(dev);
+               else
+                       ret = ixgbe_fdir_filter_program(adapter, fdir_conf,
+                                       &fdir_rule, TRUE, FALSE);
                 if (!ret) {
                         rte_free(fdir_rule_ptr);
-                       if (fdir_info->n_flows > 0 && 
--(fdir_info->n_flows) == 0) {
-                               fdir_info->mask_added = false;
-                               fdir_info->mask = (struct 
ixgbe_hw_fdir_mask){0};
-                               fdir_info->flex_bytes_offset = 0;
-                               fdir_conf->mode = RTE_FDIR_MODE_NONE;
-                       }
+                       if (fdir_info->n_flows > 1)
+                               fdir_info->n_flows--;
                 }
                 break;
         case RTE_ETH_FILTER_L2_TUNNEL:

-- 
Thanks,
Anatoly

Re: [PATCH] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Zhang Tengfei]

On 9/25/26 18:05, Burakov, Anatoly wrote:
> On 9/25/2026 11:47 AM, Burakov, Anatoly wrote:
>> A general comment: instead of reducing/bringing things back and 
> 
> I asked an AI to implement a fix based on this, and here's what it came 
> up with, it is roughly what I would like to see instead (obviously, 
> please review/rework as appropriate e.g. to properly support VMDq as well):
> 

Hi Anatoly, Bruce,

Thanks for the review. Agreed, keeping the FDIR mode after the last rule 
is gone
is wrong (a later rule with another mode fails with "Conflict with
existing fdir mode"). I'll drop the flag and do v2 along the lines of
the draft: RXPBSIZE(0) computed from the default size on enable,
restored on disable, and FDIR disabled on the configure error path and
when the first rule fails after FDIR was just enabled.

Unless you object, I plan to:

1. Use hw->mac.rx_pb_size as the default size (512 on 82599, 384 on
    X540/X550/E610) instead of saving RXPBSIZE at init.

2. Leave VMDq out: ixgbe_fdir_configure() rejects all DCB mq modes, and
    the non-DCB VMDq/SR-IOV paths don't write RXPBSIZE, so PB0 is the
    only buffer whenever FDIR can be enabled.

3. Not address the fact that FDIR enable/disable from the flow API
    changes RXPBSIZE/FDIRCTRL while Rx is enabled if the port is
    started. Enabling on the first rule already works this way today;
    restoring the buffer on the last rule removal adds the disable side.
    dev_start only touches these registers with Rx disabled.

I'll send v2 rebased on next-net-intel in the next few days.

Thanks

[PATCH v2] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Zhang Tengfei]

Enabling Flow Director reduces RXPBSIZE(0) to reserve packet buffer
memory for the filter table. The new size was computed from the current
register value, and nothing restored it when FDIR state was reset. After
the last FDIR rule was removed, or when the first rule failed after FDIR
had been enabled, the next rule created enabled FDIR again and reduced
the buffer a second time.

Compute RXPBSIZE(0) from the default packet buffer size, and disable
FDIR and restore the buffer when the last rule is removed, on flush, and
on the enable error paths.

Fixes: 301abd26b118 ("net/ixgbe: reset flow state on clear paths")
Fixes: 5007ac13189d ("ethdev: remove deprecated Flow Director configuration")
Cc: stable@dpdk.org

Signed-off-by: Zhang Tengfei <zhtfdev@gmail.com>
---
v2:
- Drop the rx_pb_reduced flag and keep resetting the FDIR mode to NONE
  when the last rule is removed (Bruce).
- Compute RXPBSIZE(0) from hw->mac.rx_pb_size, and disable FDIR and
  restore the buffer when the last rule is removed or on flush, based
  on Anatoly's draft.
- Disable FDIR again if the first rule fails after enabling it.
- Do not fail flush if re-initializing the FDIR tables fails, as FDIR
  is disabled right after.
- Add Fixes tag for the first-rule failure path.
- Compile tested only.
- Author name follows the pending .mailmap update: <https://patches.dpdk.org/project/dpdk/patch/20260927035949.36887-1-zhtfdev@gmail.com/>

 drivers/net/intel/ixgbe/ixgbe_ethdev.h |  1 +
 drivers/net/intel/ixgbe/ixgbe_fdir.c   | 70 +++++++++++++-------------
 drivers/net/intel/ixgbe/ixgbe_flow.c   | 34 ++++++++-----
 3 files changed, 58 insertions(+), 47 deletions(-)

diff --git a/drivers/net/intel/ixgbe/ixgbe_ethdev.h b/drivers/net/intel/ixgbe/ixgbe_ethdev.h
index 2e9d166e29..dfc8fce1a8 100644
--- a/drivers/net/intel/ixgbe/ixgbe_ethdev.h
+++ b/drivers/net/intel/ixgbe/ixgbe_ethdev.h
@@ -700,6 +700,7 @@ void ixgbe_filterlist_flush(struct rte_eth_dev *dev);
 int ixgbe_fdir_configure(struct rte_eth_dev *dev,
 			 const struct rte_eth_fdir_conf *fdir_conf,
 			 const struct ixgbe_hw_fdir_mask *fdir_mask);
+void ixgbe_fdir_disable(struct ixgbe_hw *hw);
 int ixgbe_fdir_set_input_mask(struct ixgbe_adapter *adapter,
 			      const struct ixgbe_hw_fdir_mask *mask,
 			      enum rte_fdir_mode mode);
diff --git a/drivers/net/intel/ixgbe/ixgbe_fdir.c b/drivers/net/intel/ixgbe/ixgbe_fdir.c
index 0159c5b3b7..ffa9f5cc4d 100644
--- a/drivers/net/intel/ixgbe/ixgbe_fdir.c
+++ b/drivers/net/intel/ixgbe/ixgbe_fdir.c
@@ -101,7 +101,6 @@ static int fdir_write_perfect_filter_82599(struct ixgbe_hw *hw,
 static int fdir_add_signature_filter_82599(struct ixgbe_hw *hw,
 		union ixgbe_atr_input *input, u8 queue, uint32_t fdircmd,
 		uint32_t fdirhash);
-static int ixgbe_fdir_flush(struct rte_eth_dev *dev);
 
 /**
  * This function is based on ixgbe_fdir_enable_82599() in base/ixgbe_82599.c.
@@ -554,6 +553,19 @@ ixgbe_set_fdir_flex_conf(struct ixgbe_adapter *adapter,
 	return 0;
 }
 
+void
+ixgbe_fdir_disable(struct ixgbe_hw *hw)
+{
+	int i;
+
+	IXGBE_WRITE_REG(hw, IXGBE_FDIRCTRL, 0);
+	IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0),
+			hw->mac.rx_pb_size << IXGBE_RXPBSIZE_SHIFT);
+	for (i = 1; i < 8; i++)
+		IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), 0);
+	IXGBE_WRITE_FLUSH(hw);
+}
+
 int
 ixgbe_fdir_configure(struct rte_eth_dev *dev,
 		const struct rte_eth_fdir_conf *fdir_conf,
@@ -601,12 +613,12 @@ ixgbe_fdir_configure(struct rte_eth_dev *dev,
 
 	/*
 	 * Before enabling Flow Director, the Rx Packet Buffer size
-	 * must be reduced.  The new value is the current size minus
+	 * must be reduced.  The new value is the default size minus
 	 * flow director memory usage size.
 	 */
 	pbsize = (1 << (PBALLOC_SIZE_SHIFT + (fdirctrl & FDIRCTRL_PBALLOC_MASK)));
 	IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(0),
-	    (IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(0)) - pbsize));
+	    (hw->mac.rx_pb_size << IXGBE_RXPBSIZE_SHIFT) - pbsize);
 
 	/*
 	 * The defaults in the HW for RX PB 1-7 are not zero and so should be
@@ -620,21 +632,25 @@ ixgbe_fdir_configure(struct rte_eth_dev *dev,
 	err = ixgbe_fdir_set_input_mask(adapter, fdir_mask, mode);
 	if (err < 0) {
 		PMD_INIT_LOG(ERR, " Error on setting FD mask");
-		return err;
+		goto error;
 	}
 	err = ixgbe_set_fdir_flex_conf(adapter, &fdir_conf->flex_conf,
 				       &fdirctrl);
 	if (err < 0) {
 		PMD_INIT_LOG(ERR, " Error on setting FD flexible arguments.");
-		return err;
+		goto error;
 	}
 
 	err = fdir_enable_82599(hw, fdirctrl);
 	if (err < 0) {
 		PMD_INIT_LOG(ERR, " Error on enabling FD.");
-		return err;
+		goto error;
 	}
 	return 0;
+
+error:
+	ixgbe_fdir_disable(hw);
+	return err;
 }
 
 /*
@@ -1191,28 +1207,6 @@ ixgbe_fdir_filter_program(struct ixgbe_adapter *adapter,
 	return err;
 }
 
-static int
-ixgbe_fdir_flush(struct rte_eth_dev *dev)
-{
-	struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
-	struct ixgbe_hw_fdir_info *info =
-			IXGBE_DEV_PRIVATE_TO_FDIR_INFO(dev->data->dev_private);
-	int ret;
-
-	ret = ixgbe_reinit_fdir_tables_82599(hw);
-	if (ret < 0) {
-		PMD_INIT_LOG(ERR, "Failed to re-initialize FD table.");
-		return ret;
-	}
-
-	info->f_add = 0;
-	info->f_remove = 0;
-	info->add = 0;
-	info->remove = 0;
-
-	return ret;
-}
-
 #define FDIRENTRIES_NUM_SHIFT 10
 void
 ixgbe_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info *fdir_info)
@@ -1371,13 +1365,22 @@ int
 ixgbe_clear_all_fdir_filter(struct rte_eth_dev *dev)
 {
 	struct rte_eth_fdir_conf *fdir_conf = IXGBE_DEV_FDIR_CONF(dev);
+	struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
 	struct ixgbe_hw_fdir_info *fdir_info =
 		IXGBE_DEV_PRIVATE_TO_FDIR_INFO(dev->data->dev_private);
 	struct ixgbe_fdir_filter *fdir_filter;
-	bool had_flows;
-	int ret = 0;
 
-	had_flows = (fdir_info->n_flows != 0);
+	if (fdir_conf->mode != RTE_FDIR_MODE_NONE) {
+		if (ixgbe_reinit_fdir_tables_82599(hw) < 0)
+			PMD_DRV_LOG(WARNING, "Failed to re-initialize FD table");
+
+		fdir_info->f_add = 0;
+		fdir_info->f_remove = 0;
+		fdir_info->add = 0;
+		fdir_info->remove = 0;
+
+		ixgbe_fdir_disable(hw);
+	}
 
 	/* flush flow director */
 	rte_hash_reset(fdir_info->hash_handle);
@@ -1397,8 +1400,5 @@ ixgbe_clear_all_fdir_filter(struct rte_eth_dev *dev)
 	fdir_info->mask_added = FALSE;
 	fdir_conf->mode = RTE_FDIR_MODE_NONE;
 
-	if (had_flows)
-		ret = ixgbe_fdir_flush(dev);
-
-	return ret;
+	return 0;
 }
diff --git a/drivers/net/intel/ixgbe/ixgbe_flow.c b/drivers/net/intel/ixgbe/ixgbe_flow.c
index ff56d7b279..a8d1b269c7 100644
--- a/drivers/net/intel/ixgbe/ixgbe_flow.c
+++ b/drivers/net/intel/ixgbe/ixgbe_flow.c
@@ -2625,6 +2625,7 @@ ixgbe_fdir_flow_program(struct rte_eth_dev *dev,
 	struct rte_eth_fdir_conf local_fdir_conf = *fdir_conf;
 	struct ixgbe_hw_fdir_info *fdir_info =
 			IXGBE_DEV_PRIVATE_TO_FDIR_INFO(adapter);
+	bool fdir_enabled = false;
 	int ret;
 
 	if (fdir_rule->queue >= dev->data->nb_rx_queues) {
@@ -2647,6 +2648,7 @@ ixgbe_fdir_flow_program(struct rte_eth_dev *dev,
 				RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 				NULL, "Failed to configure fdir mode");
 		}
+		fdir_enabled = true;
 	} else if (fdir_conf->mode != fdir_rule->mode) {
 		return rte_flow_error_set(error, EINVAL,
 			RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
@@ -2657,17 +2659,25 @@ ixgbe_fdir_flow_program(struct rte_eth_dev *dev,
 	ret = ixgbe_fdir_process_rule(adapter, fdir_info, fdir_rule,
 		first_mask, error);
 	if (ret)
-		return ret;
+		goto error;
 
 	/* Program the filter */
 	ret = ixgbe_fdir_filter_program(adapter, &local_fdir_conf,
 			fdir_rule, FALSE, FALSE);
-	if (ret)
-		return rte_flow_error_set(error, EINVAL,
+	if (ret) {
+		ret = rte_flow_error_set(error, EINVAL,
 			RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 			NULL, "Failed to add fdir filter");
+		goto error;
+	}
 
 	return 0;
+
+error:
+	/* FDIR mode is only recorded on success, so undo the enable */
+	if (fdir_enabled)
+		ixgbe_fdir_disable(IXGBE_DEV_PRIVATE_TO_HW(adapter));
+	return ret;
 }
 
 /* Flow actions check specific to RSS filter */
@@ -3159,16 +3169,16 @@ ixgbe_flow_destroy(struct rte_eth_dev *dev,
 	case RTE_ETH_FILTER_FDIR:
 		fdir_rule_ptr = (struct ixgbe_fdir_rule_ele *)pmd_flow->rule;
 		fdir_rule = fdir_rule_ptr->filter_info;
-		ret = ixgbe_fdir_filter_program(adapter, fdir_conf, &fdir_rule, TRUE, FALSE);
-		if (!ret) {
-			rte_free(fdir_rule_ptr);
-			if (fdir_info->n_flows > 0 && --(fdir_info->n_flows) == 0) {
-				fdir_info->mask_added = false;
-				fdir_info->mask = (struct ixgbe_hw_fdir_mask){0};
-				fdir_info->flex_bytes_offset = 0;
-				fdir_conf->mode = RTE_FDIR_MODE_NONE;
-			}
+		if (fdir_info->n_flows == 1) {
+			ret = ixgbe_clear_all_fdir_filter(dev);
+		} else {
+			ret = ixgbe_fdir_filter_program(adapter, fdir_conf,
+					&fdir_rule, TRUE, FALSE);
+			if (!ret && fdir_info->n_flows > 0)
+				fdir_info->n_flows--;
 		}
+		if (!ret)
+			rte_free(fdir_rule_ptr);
 		break;
 	case RTE_ETH_FILTER_L2_TUNNEL:
 		l2_tn_filter_ptr = (struct ixgbe_eth_l2_tunnel_conf_ele *)
-- 
2.53.0

Re: [PATCH v2] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Burakov, Anatoly]

On 9/27/2026 6:25 AM, Zhang Tengfei wrote:
> Enabling Flow Director reduces RXPBSIZE(0) to reserve packet buffer
> memory for the filter table. The new size was computed from the current
> register value, and nothing restored it when FDIR state was reset. After
> the last FDIR rule was removed, or when the first rule failed after FDIR
> had been enabled, the next rule created enabled FDIR again and reduced
> the buffer a second time.
> 
> Compute RXPBSIZE(0) from the default packet buffer size, and disable
> FDIR and restore the buffer when the last rule is removed, on flush, and
> on the enable error paths.
> 
> Fixes: 301abd26b118 ("net/ixgbe: reset flow state on clear paths")
> Fixes: 5007ac13189d ("ethdev: remove deprecated Flow Director configuration")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Zhang Tengfei <zhtfdev@gmail.com>
> ---

Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>

-- 
Thanks,
Anatoly

Re: [PATCH v2] net/ixgbe: fix repeated Rx packet buffer shrink for FDIR

[Bruce Richardson]

On Tue, Oct 06, 2026 at 11:23:06AM +0200, Burakov, Anatoly wrote:
> On 9/27/2026 6:25 AM, Zhang Tengfei wrote:
> > Enabling Flow Director reduces RXPBSIZE(0) to reserve packet buffer
> > memory for the filter table. The new size was computed from the current
> > register value, and nothing restored it when FDIR state was reset. After
> > the last FDIR rule was removed, or when the first rule failed after FDIR
> > had been enabled, the next rule created enabled FDIR again and reduced
> > the buffer a second time.
> > 
> > Compute RXPBSIZE(0) from the default packet buffer size, and disable
> > FDIR and restore the buffer when the last rule is removed, on flush, and
> > on the enable error paths.
> > 
> > Fixes: 301abd26b118 ("net/ixgbe: reset flow state on clear paths")
> > Fixes: 5007ac13189d ("ethdev: remove deprecated Flow Director configuration")
> > Cc: stable@dpdk.org
> > 
> > Signed-off-by: Zhang Tengfei <zhtfdev@gmail.com>
> > ---
> 
> Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
> 
Applied to dpdk-next-net-intel

/Bruce