Skip to contentSkip to conversation

igb: cope with large MAX_SKB_FRAGS

PATCH Superseded Corinna Vinschen 3 participants 6 messages Apr 23, 2024 – Jul 18, 2024

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 igb: cope with large MAX_SKB_FRAGS

From: Paolo Abeni <[email protected]>

Sabrina reports that the igb driver does not cope well with large MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload corruption on TX.

An easy reproducer is to run ssh to connect to the machine. With MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails.

The root cause of the issue is that the driver does not take into account properly the (possibly large) shared info size when selecting the ring layout, and will try to fit two packets inside the same 4K page even when the 1st fraglist will trump over the 2nd head.

Address the issue forcing the driver to fit a single packet per page, leaving there enough room to store the (currently) largest possible skb_shared_info.

Fixes 3948b05950fd ("net: introduce a config option to tweak MAX_SKB_FRAG")

---
v2: fix subject, add a simple reproducer

 drivers/net/ethernet/intel/igb/igb_main.c | 1 +
 1 file changed, 1 insertion(+)
drivers/net/ethernet/intel/igb/igb_main.c +1 -0
index a3f100769e39..22fb2c322bca 100644
@@ -4833,6 +4833,7 @@ static void igb_set_rx_buffer_len(struct igb_adapter *adapter,
 
 #if (PAGE_SIZE < 8192)
 	if (adapter->max_frame_size > IGB_MAX_FRAME_BUILD_SKB ||
+	    SKB_HEAD_ALIGN(adapter->max_frame_size) > (PAGE_SIZE / 2) ||
 	    rd32(E1000_RCTL) & E1000_RCTL_SBP)
 		set_ring_uses_large_buffer(rx_ring);
 #endif
Signature · 1 line
2.44.0
#1 lore raw
Quoted 37 lines · Corinna Vinschen

From: Paolo Abeni <[email protected]>

Sabrina reports that the igb driver does not cope well with large MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload corruption on TX.

An easy reproducer is to run ssh to connect to the machine. With MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails.

The root cause of the issue is that the driver does not take into account properly the (possibly large) shared info size when selecting the ring layout, and will try to fit two packets inside the same 4K page even when the 1st fraglist will trump over the 2nd head.

Address the issue forcing the driver to fit a single packet per page, leaving there enough room to store the (currently) largest possible skb_shared_info.

Fixes 3948b05950fd ("net: introduce a config option to tweak MAX_SKB_FRAG")

---
v2: fix subject, add a simple reproducer

 drivers/net/ethernet/intel/igb/igb_main.c | 1 +
 1 file changed, 1 insertion(+)
drivers/net/ethernet/intel/igb/igb_main.c +0 -0
index a3f100769e39..22fb2c322bca 100644

drivers/net/ethernet/intel/igb/igb_main.c static void igb_set_rx_buffer_len(struct igb_adapter *adapter, · line 4836

Quoted 4 lines
drivers/net/ethernet/intel/igb/igb_main.c +1 -0
@@ -4833,6 +4833,7 @@ static void igb_set_rx_buffer_len(struct igb_adapter *adapter,
 
 #if (PAGE_SIZE < 8192)
        if (adapter->max_frame_size > IGB_MAX_FRAME_BUILD_SKB ||
+           SKB_HEAD_ALIGN(adapter->max_frame_size) > (PAGE_SIZE / 2) ||

I wonder if adding IGB_SKB_PAD would be needed ?

adapter->max_frame_size does not seem to include it.

I would try using all mtus between 1200 and 1280 to make sure this works.

Thanks !

Quoted 6 lines
drivers/net/ethernet/intel/igb/igb_main.c +0 -1
            rd32(E1000_RCTL) & E1000_RCTL_SBP)
                set_ring_uses_large_buffer(rx_ring);
 #endif
--

2.44.0

#2 lore raw

Hi Eric,

Quoted 45 lines · Eric Dumazet
Quoted 37 lines · Corinna Vinschen

From: Paolo Abeni <[email protected]>

Sabrina reports that the igb driver does not cope well with large MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload corruption on TX.

An easy reproducer is to run ssh to connect to the machine. With MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails.

The root cause of the issue is that the driver does not take into account properly the (possibly large) shared info size when selecting the ring layout, and will try to fit two packets inside the same 4K page even when the 1st fraglist will trump over the 2nd head.

Address the issue forcing the driver to fit a single packet per page, leaving there enough room to store the (currently) largest possible skb_shared_info.

Fixes 3948b05950fd ("net: introduce a config option to tweak MAX_SKB_FRAG")

---
v2: fix subject, add a simple reproducer

 drivers/net/ethernet/intel/igb/igb_main.c | 1 +
 1 file changed, 1 insertion(+)
drivers/net/ethernet/intel/igb/igb_main.c +0 -0
index a3f100769e39..22fb2c322bca 100644

drivers/net/ethernet/intel/igb/igb_main.c static void igb_set_rx_buffer_len(struct igb_adapter *adapter, · line 4836

Quoted 4 lines
drivers/net/ethernet/intel/igb/igb_main.c +1 -0
@@ -4833,6 +4833,7 @@ static void igb_set_rx_buffer_len(struct igb_adapter *adapter,
 
 #if (PAGE_SIZE < 8192)
        if (adapter->max_frame_size > IGB_MAX_FRAME_BUILD_SKB ||
+           SKB_HEAD_ALIGN(adapter->max_frame_size) > (PAGE_SIZE / 2) ||

I wonder if adding IGB_SKB_PAD would be needed ?

Quoted 3 lines

adapter->max_frame_size does not seem to include it.

I would try using all mtus between 1200 and 1280 to make sure this works.

Erm... did you mean between 1500 and 1580 by any chance? 1200 doesn't really seem to make sense...

I tested this patch now with mtu 1500, 1540 and 1580 successfully.

Either way, I'm just heading into vacation, so I guess I'll pick this up again when I'm back, unless Paolo takes another look during my absence.

Thanks,
Corinna

#3 lore raw
Quoted 49 lines · Corinna Vinschen

Hi Eric,

Quoted 45 lines · Eric Dumazet
Quoted 37 lines · Corinna Vinschen

From: Paolo Abeni <[email protected]>

Sabrina reports that the igb driver does not cope well with large MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload corruption on TX.

An easy reproducer is to run ssh to connect to the machine. With MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails.

The root cause of the issue is that the driver does not take into account properly the (possibly large) shared info size when selecting the ring layout, and will try to fit two packets inside the same 4K page even when the 1st fraglist will trump over the 2nd head.

Address the issue forcing the driver to fit a single packet per page, leaving there enough room to store the (currently) largest possible skb_shared_info.

Fixes 3948b05950fd ("net: introduce a config option to tweak MAX_SKB_FRAG")

---
v2: fix subject, add a simple reproducer

 drivers/net/ethernet/intel/igb/igb_main.c | 1 +
 1 file changed, 1 insertion(+)
drivers/net/ethernet/intel/igb/igb_main.c +0 -0
index a3f100769e39..22fb2c322bca 100644

drivers/net/ethernet/intel/igb/igb_main.c static void igb_set_rx_buffer_len(struct igb_adapter *adapter, · line 4836

Quoted 4 lines
drivers/net/ethernet/intel/igb/igb_main.c +1 -0
@@ -4833,6 +4833,7 @@ static void igb_set_rx_buffer_len(struct igb_adapter *adapter,
 
 #if (PAGE_SIZE < 8192)
        if (adapter->max_frame_size > IGB_MAX_FRAME_BUILD_SKB ||
+           SKB_HEAD_ALIGN(adapter->max_frame_size) > (PAGE_SIZE / 2) ||

I wonder if adding IGB_SKB_PAD would be needed ?

Quoted 6 lines
Quoted 3 lines

adapter->max_frame_size does not seem to include it.

I would try using all mtus between 1200 and 1280 to make sure this works.

Erm... did you mean between 1500 and 1580 by any chance? 1200 doesn't really seem to make sense...

No, I meant 1200 to 1280 . IPv4 should accept these MTU .

1200 + 768 = 1968
1280 + 768 = 2048 (2 KB)

I am worried of some padding that would cross 2048 bytes boundary, while SKB_HEAD_ALIGN(adapter->max_frame_size) could still be < 2048

Quoted 6 lines

I tested this patch now with mtu 1500, 1540 and 1580 successfully.

Either way, I'm just heading into vacation, so I guess I'll pick this up again when I'm back, unless Paolo takes another look during my absence.

I guess your patch is better than nothing, this can be refined if necessary later.

#4 lore raw

Hi Eric,

sorry for the very late reply, I was offline for quite some time and then other things took over.

Quoted 41 lines · Eric Dumazet
Quoted 23 lines · Corinna Vinschen

Hi Eric,

Quoted 19 lines · Eric Dumazet
Quoted 11 lines · Corinna Vinschen

From: Paolo Abeni <[email protected]>

Sabrina reports that the igb driver does not cope well with large MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload corruption on TX.

An easy reproducer is to run ssh to connect to the machine. With MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails. [...]

drivers/net/ethernet/intel/igb/igb_main.c static void igb_set_rx_buffer_len(struct igb_adapter *adapter, · line 4836

Quoted 4 lines
drivers/net/ethernet/intel/igb/igb_main.c +1 -0
@@ -4833,6 +4833,7 @@ static void igb_set_rx_buffer_len(struct igb_adapter *adapter,
 
 #if (PAGE_SIZE < 8192)
        if (adapter->max_frame_size > IGB_MAX_FRAME_BUILD_SKB ||
+           SKB_HEAD_ALIGN(adapter->max_frame_size) > (PAGE_SIZE / 2) ||

I wonder if adding IGB_SKB_PAD would be needed ?

Quoted 6 lines
Quoted 3 lines

adapter->max_frame_size does not seem to include it.

I would try using all mtus between 1200 and 1280 to make sure this works.

Erm... did you mean between 1500 and 1580 by any chance? 1200 doesn't really seem to make sense...

No, I meant 1200 to 1280 . IPv4 should accept these MTU .

1200 + 768 = 1968
1280 + 768 = 2048 (2 KB)

I am worried of some padding that would cross 2048 bytes boundary, while SKB_HEAD_ALIGN(adapter->max_frame_size) could still be < 2048

Quoted 1 line
Quoted 8 lines
Quoted 5 lines

I tested this patch now with mtu 1500, 1540 and 1580 successfully.

Either way, I'm just heading into vacation, so I guess I'll pick this up again when I'm back, unless Paolo takes another look during my absence.

I guess your patch is better than nothing, this can be refined if necessary later.

I now tested the patch with all MTUs between 1200 and 1280 and it works as desired. I just sent a v3.

Thanks,
Corinna

reviewer
#5 lore raw
In reply to Corinna Vinschen
Quoted 14 lines · Corinna Vinschen

From: Paolo Abeni <[email protected]>

Sabrina reports that the igb driver does not cope well with large MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload corruption on TX.

An easy reproducer is to run ssh to connect to the machine. With MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails.

The root cause of the issue is that the driver does not take into account properly the (possibly large) shared info size when selecting the ring layout, and will try to fit two packets inside the same 4K page even when the 1st fraglist will trump over the 2nd head.

Quoted 5 lines

Address the issue forcing the driver to fit a single packet per page, leaving there enough room to store the (currently) largest possible skb_shared_info.

Fixes 3948b05950fd ("net: introduce a config option to tweak MAX_SKB_FRAG")

nit: The trailing "S" in the subject for the fixes tag seems to have been lost.

Fixes 3948b05950fd ("net: introduce a config option to tweak MAX_SKB_FRAGS")
Quoted 6 lines

...