[PATCH net 0/2] bnxt_en: Error recovery bug fixes.

[PATCH net 0/2] bnxt_en: Error recovery bug fixes.

[Michael Chan]

Two error recovery related bug fixes for 2 corner cases.

Please queue patch #2 for -stable.  Thanks.

Edwin Peer (1):
  bnxt_en: reliably allocate IRQ table on reset to avoid crash

Vasundhara Volam (1):
  bnxt_en: Fix race between firmware reset and driver remove.

 drivers/net/ethernet/broadcom/bnxt/bnxt.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)

-- 
2.18.1

[PATCH net 1/2] bnxt_en: Fix race between firmware reset and driver remove.

[Michael Chan]

From: Vasundhara Volam <vasundhara-v.volam@broadcom.com>

The driver's error recovery reset sequence can take many seconds to
complete and only the critical sections are protected by rtnl_lock.
A recent change has introduced a regression in this sequence.

bnxt_remove_one() may be called while the recovery is in progress.
Normally, unregister_netdev() would cause bnxt_close_nic() to be
called and this would cause the error recovery to safely abort
with the BNXT_STATE_ABORT_ERR flag set in bnxt_close_nic().

Recently, we added bnxt_reinit_after_abort() to allow the user to
reopen the device after an aborted recovery.  This causes the
regression in the scenario described above because we would
attempt to re-open even after the netdev has been unregistered.

Fix it by checking the netdev reg_state in
bnxt_reinit_after_abort() and abort if it is unregistered.

Fixes: 6882c36cf82e ("bnxt_en: attempt to reinitialize after aborted reset")
Signed-off-by: Vasundhara Volam <vasundhara-v.volam@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
---
 drivers/net/ethernet/broadcom/bnxt/bnxt.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
index a680fd9c68ea..c55189c7bb36 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
@@ -9890,6 +9890,9 @@ static int bnxt_reinit_after_abort(struct bnxt *bp)
 	if (test_bit(BNXT_STATE_IN_FW_RESET, &bp->state))
 		return -EBUSY;
 
+	if (bp->dev->reg_state == NETREG_UNREGISTERED)
+		return -ENODEV;
+
 	rc = bnxt_fw_init_one(bp);
 	if (!rc) {
 		bnxt_clear_int_mode(bp);
-- 
2.18.1

[PATCH net 2/2] bnxt_en: reliably allocate IRQ table on reset to avoid crash

[Michael Chan]

From: Edwin Peer <edwin.peer@broadcom.com>

The following trace excerpt corresponds with a NULL pointer dereference
of 'bp->irq_tbl' in bnxt_setup_inta() on an Aarch64 system after many
device resets:

    Unable to handle kernel NULL pointer dereference at ... 000000d
    ...
    pc : string+0x3c/0x80
    lr : vsnprintf+0x294/0x7e0
    sp : ffff00000f61ba70 pstate : 20000145
    x29: ffff00000f61ba70 x28: 000000000000000d
    x27: ffff0000009c8b5a x26: ffff00000f61bb80
    x25: ffff0000009c8b5a x24: 0000000000000012
    x23: 00000000ffffffe0 x22: ffff000008990428
    x21: ffff00000f61bb80 x20: 000000000000000d
    x19: 000000000000001f x18: 0000000000000000
    x17: 0000000000000000 x16: ffff800b6d0fb400
    x15: 0000000000000000 x14: ffff800b7fe31ae8
    x13: 00001ed16472c920 x12: ffff000008c6b1c9
    x11: ffff000008cf0580 x10: ffff00000f61bb80
    x9 : 00000000ffffffd8 x8 : 000000000000000c
    x7 : ffff800b684b8000 x6 : 0000000000000000
    x5 : 0000000000000065 x4 : 0000000000000001
    x3 : ffff0a00ffffff04 x2 : 000000000000001f
    x1 : 0000000000000000 x0 : 000000000000000d
    Call trace:
    string+0x3c/0x80
    vsnprintf+0x294/0x7e0
    snprintf+0x44/0x50
    __bnxt_open_nic+0x34c/0x928 [bnxt_en]
    bnxt_open+0xe8/0x238 [bnxt_en]
    __dev_open+0xbc/0x130
    __dev_change_flags+0x12c/0x168
    dev_change_flags+0x20/0x60
    ...

Ordinarily, a call to bnxt_setup_inta() (not in trace due to inlining)
would not be expected on a system supporting MSIX at all. However, if
bnxt_init_int_mode() does not end up being called after the call to
bnxt_clear_int_mode() in bnxt_fw_reset_close(), then the driver will
think that only INTA is supported and bp->irq_tbl will be NULL,
causing the above crash.

In the error recovery scenario, we call bnxt_clear_int_mode() in
bnxt_fw_reset_close() early in the sequence. Ordinarily, we will
call bnxt_init_int_mode() in bnxt_hwrm_if_change() after we
reestablish communication with the firmware after reset.  However,
if the sequence has to abort before we call bnxt_init_int_mode() and
if the user later attempts to re-open the device, then it will cause
the crash above.

We fix it in 2 ways:

1. Check for bp->irq_tbl in bnxt_setup_int_mode(). If it is NULL, call
bnxt_init_init_mode().

2. If we need to abort in bnxt_hwrm_if_change() and cannot complete
the error recovery sequence, set the BNXT_STATE_ABORT_ERR flag.  This
will cause more drastic recovery at the next attempt to re-open the
device, including a call to bnxt_init_int_mode().

Fixes: 3bc7d4a352ef ("bnxt_en: Add BNXT_STATE_IN_FW_RESET state.")
Reviewed-by: Scott Branden <scott.branden@broadcom.com>
Signed-off-by: Edwin Peer <edwin.peer@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
---
 drivers/net/ethernet/broadcom/bnxt/bnxt.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
index c55189c7bb36..b53a0d87371a 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
@@ -8556,10 +8556,18 @@ static void bnxt_setup_inta(struct bnxt *bp)
 	bp->irq_tbl[0].handler = bnxt_inta;
 }
 
+static int bnxt_init_int_mode(struct bnxt *bp);
+
 static int bnxt_setup_int_mode(struct bnxt *bp)
 {
 	int rc;
 
+	if (!bp->irq_tbl) {
+		rc = bnxt_init_int_mode(bp);
+		if (rc || !bp->irq_tbl)
+			return rc ?: -ENODEV;
+	}
+
 	if (bp->flags & BNXT_FLAG_USING_MSIX)
 		bnxt_setup_msix(bp);
 	else
@@ -8744,7 +8752,7 @@ static int bnxt_init_inta(struct bnxt *bp)
 
 static int bnxt_init_int_mode(struct bnxt *bp)
 {
-	int rc = 0;
+	int rc = -ENODEV;
 
 	if (bp->flags & BNXT_FLAG_MSIX_CAP)
 		rc = bnxt_init_msix(bp);
@@ -9514,7 +9522,8 @@ static int bnxt_hwrm_if_change(struct bnxt *bp, bool up)
 {
 	struct hwrm_func_drv_if_change_output *resp = bp->hwrm_cmd_resp_addr;
 	struct hwrm_func_drv_if_change_input req = {0};
-	bool resc_reinit = false, fw_reset = false;
+	bool fw_reset = !bp->irq_tbl;
+	bool resc_reinit = false;
 	int rc, retry = 0;
 	u32 flags = 0;
 
@@ -9557,6 +9566,7 @@ static int bnxt_hwrm_if_change(struct bnxt *bp, bool up)
 
 	if (test_bit(BNXT_STATE_IN_FW_RESET, &bp->state) && !fw_reset) {
 		netdev_err(bp->dev, "RESET_DONE not set during FW reset.\n");
+		set_bit(BNXT_STATE_ABORT_ERR, &bp->state);
 		return -ENODEV;
 	}
 	if (resc_reinit || fw_reset) {
-- 
2.18.1

Re: [PATCH net 0/2] bnxt_en: Error recovery bug fixes.

[patchwork-bot+netdevbpf]

Hello:

This series was applied to netdev/net.git (refs/heads/master):

On Fri, 26 Feb 2021 04:43:08 -0500 you wrote:
> Two error recovery related bug fixes for 2 corner cases.
> 
> Please queue patch #2 for -stable.  Thanks.
> 
> Edwin Peer (1):
>   bnxt_en: reliably allocate IRQ table on reset to avoid crash
> 
> [...]

Here is the summary with links:
  - [net,1/2] bnxt_en: Fix race between firmware reset and driver remove.
    https://git.kernel.org/netdev/net/c/d20cd745218c
  - [net,2/2] bnxt_en: reliably allocate IRQ table on reset to avoid crash
    https://git.kernel.org/netdev/net/c/20d7d1c5c9b1

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html