]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
scsi: ipr: Enable/disable IRQD_NO_BALANCING during reset
authorWen Xiong <wenxiong@linux.ibm.com>
Tue, 28 Oct 2025 14:24:26 +0000 (09:24 -0500)
committerMartin K. Petersen <martin.petersen@oracle.com>
Sat, 29 Nov 2025 20:56:18 +0000 (15:56 -0500)
A dynamic remove/add storage adapter test hits EEH on PowerPC:

  EEH: [c00000000004f75c] __eeh_send_failure_event+0x7c/0x160
  EEH: [c000000000048444] eeh_dev_check_failure.part.0+0x254/0x650
  EEH: [c008000001650678] eeh_readl+0x60/0x90 [ipr]
  EEH: [c00800000166746c] ipr_cancel_op+0x2b8/0x524 [ipr]
  EEH: [c008000001656524] ipr_eh_abort+0x6c/0x130 [ipr]
  EEH: [c000000000ab0d20] scmd_eh_abort_handler+0x140/0x440
  EEH: [c00000000017e558] process_one_work+0x298/0x590
  EEH: [c00000000017eef8] worker_thread+0xa8/0x620
  EEH: [c00000000018be34] kthread+0x124/0x130
  EEH: [c00000000000cd64] ret_from_kernel_thread+0x5c/0x64

A PCIe bus trace reveals that a vector of MSI-X is cleared to 0 by
irqbalance daemon. If we disable irqbalance daemon, we won't see the
issue.

With debug enabled in ipr driver:

  [   44.103071] ipr: Entering __ipr_remove
  [   44.103083] ipr: Entering ipr_initiate_ioa_bringdown
  [   44.103091] ipr: Entering ipr_reset_shutdown_ioa
  [   44.103099] ipr: Leaving ipr_reset_shutdown_ioa
  [   44.103105] ipr: Leaving ipr_initiate_ioa_bringdown
  [   44.149918] ipr: Entering ipr_reset_ucode_download
  [   44.149935] ipr: Entering ipr_reset_alert
  [   44.150032] ipr: Entering ipr_reset_start_timer
  [   44.150038] ipr: Leaving ipr_reset_alert
  [   44.244343] scsi 1:2:3:0: alua: Detached
  [   44.254300] ipr: Entering ipr_reset_start_bist
  [   44.254320] ipr: Entering ipr_reset_start_timer
  [   44.254325] ipr: Leaving ipr_reset_start_bist
  [   44.364329] scsi 1:2:4:0: alua: Detached
  [   45.134341] scsi 1:2:5:0: alua: Detached
  [   45.860949] ipr: Entering ipr_reset_shutdown_ioa
  [   45.860962] ipr: Leaving ipr_reset_shutdown_ioa
  [   45.860966] ipr: Entering ipr_reset_alert
  [   45.861028] ipr: Entering ipr_reset_start_timer
  [   45.861035] ipr: Leaving ipr_reset_alert
  [   45.964302] ipr: Entering ipr_reset_start_bist
  [   45.964309] ipr: Entering ipr_reset_start_timer
  [   45.964313] ipr: Leaving ipr_reset_start_bist
  [   46.264301] ipr: Entering ipr_reset_bist_done
  [   46.264309] ipr: Leaving ipr_reset_bist_done

During adapter reset, ipr device driver blocks config space access but
can't block MMIO access for MSI-X entries.  There is very small window:
irqbalance daemon kicks in during adapter reset before ipr driver calls
pci_restore_state(pdev) to restore MSI-X table.

irqbalance daemon reads back all 0 for that MSI-X vector in
__pci_read_msi_msg().

irqbalance daemon:

  msi_domain_set_affinity()
  ->irq_chip_set_affinity_patent()
  ->xive_irq_set_affinity()
  ->irq_chip_compose_msi_msg()
    ->pseries_msi_compose_msg()
    ->__pci_read_msi_msg(): read all 0 since didn't call pci_restore_state
  ->irq_chip_write_msi_msg()
    -> pci_write_msg_msi(): write 0 to the msix vector entry

When ipr driver calls pci_restore_state(pdev) in
ipr_reset_restore_cfg_space(), the MSI-X vector entry has been cleared
by irqbalance daemon in pci_write_msg_msix().

  pci_restore_state()
  ->__pci_restore_msix_state()

Below is the MSI-X table for ipr adapter after irqbalance daemon kicked
in during adapter reset:

  Dump MSIx table: index=0 address_lo=c800 address_hi=10000000 msg_data=0
  Dump MSIx table: index=1 address_lo=c810 address_hi=10000000 msg_data=0
  Dump MSIx table: index=2 address_lo=c820 address_hi=10000000 msg_data=0
  Dump MSIx table: index=3 address_lo=c830 address_hi=10000000 msg_data=0
  Dump MSIx table: index=4 address_lo=c840 address_hi=10000000 msg_data=0
  Dump MSIx table: index=5 address_lo=c850 address_hi=10000000 msg_data=0
  Dump MSIx table: index=6 address_lo=c860 address_hi=10000000 msg_data=0
  Dump MSIx table: index=7 address_lo=c870 address_hi=10000000 msg_data=0
  Dump MSIx table: index=8 address_lo=0 address_hi=0 msg_data=0
  ---------> Hit EEH since msix vector of index=8 are 0
  Dump MSIx table: index=9 address_lo=c890 address_hi=10000000 msg_data=0
  Dump MSIx table: index=10 address_lo=c8a0 address_hi=10000000 msg_data=0
  Dump MSIx table: index=11 address_lo=c8b0 address_hi=10000000 msg_data=0
  Dump MSIx table: index=12 address_lo=c8c0 address_hi=10000000 msg_data=0
  Dump MSIx table: index=13 address_lo=c8d0 address_hi=10000000 msg_data=0
  Dump MSIx table: index=14 address_lo=c8e0 address_hi=10000000 msg_data=0
  Dump MSIx table: index=15 address_lo=c8f0 address_hi=10000000 msg_data=0

  [   46.264312] ipr: Entering ipr_reset_restore_cfg_space
  [   46.267439] ipr: Entering ipr_fail_all_ops
  [   46.267447] ipr: Leaving ipr_fail_all_ops
  [   46.267451] ipr: Leaving ipr_reset_restore_cfg_space
  [   46.267454] ipr: Entering ipr_ioa_bringdown_done
  [   46.267458] ipr: Leaving ipr_ioa_bringdown_done
  [   46.267467] ipr: Entering ipr_worker_thread
  [   46.267470] ipr: Leaving ipr_worker_thread

IRQ balancing is not required during adapter reset.

Enable "IRQ_NO_BALANCING" flag before starting adapter reset and disable
it after calling pci_restore_state(). The irqbalance daemon is disabled
for this short period of time (~2s).

Co-developed-by: Kyle Mahlkuch <Kyle.Mahlkuch@ibm.com>
Signed-off-by: Kyle Mahlkuch <Kyle.Mahlkuch@ibm.com>
Signed-off-by: Wen Xiong <wenxiong@linux.ibm.com>
Link: https://patch.msgid.link/20251028142427.3969819-2-wenxiong@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
drivers/scsi/ipr.c

index 44214884deaf505feffc00cfa155c7ad205dbdce..d62bb7d0e4164e15a32eec5db052f99bf92210eb 100644 (file)
@@ -61,8 +61,8 @@
 #include <linux/hdreg.h>
 #include <linux/reboot.h>
 #include <linux/stringify.h>
+#include <linux/irq.h>
 #include <asm/io.h>
-#include <asm/irq.h>
 #include <asm/processor.h>
 #include <scsi/scsi.h>
 #include <scsi/scsi_host.h>
@@ -7843,6 +7843,30 @@ static int ipr_dump_mailbox_wait(struct ipr_cmnd *ipr_cmd)
        return IPR_RC_JOB_RETURN;
 }
 
+/**
+ * ipr_set_affinity_nobalance
+ * @ioa_cfg:   ipr_ioa_cfg struct for an ipr device
+ * @flag:      bool
+ *     true: ensable "IRQ_NO_BALANCING" bit for msix interrupt
+ *     false: disable "IRQ_NO_BALANCING" bit for msix interrupt
+ * Description: This function will be called to disable/enable
+ *     "IRQ_NO_BALANCING" to avoid irqbalance daemon
+ *     kicking in during adapter reset.
+ **/
+static void ipr_set_affinity_nobalance(struct ipr_ioa_cfg *ioa_cfg, bool flag)
+{
+       int irq, i;
+
+       for (i = 0; i < ioa_cfg->nvectors; i++) {
+               irq = pci_irq_vector(ioa_cfg->pdev, i);
+
+               if (flag)
+                       irq_set_status_flags(irq, IRQ_NO_BALANCING);
+               else
+                       irq_clear_status_flags(irq, IRQ_NO_BALANCING);
+       }
+}
+
 /**
  * ipr_reset_restore_cfg_space - Restore PCI config space.
  * @ipr_cmd:   ipr command struct
@@ -7867,6 +7891,7 @@ static int ipr_reset_restore_cfg_space(struct ipr_cmnd *ipr_cmd)
                return IPR_RC_JOB_CONTINUE;
        }
 
+       ipr_set_affinity_nobalance(ioa_cfg, false);
        ipr_fail_all_ops(ioa_cfg);
 
        if (ioa_cfg->sis64) {
@@ -7946,6 +7971,7 @@ static int ipr_reset_start_bist(struct ipr_cmnd *ipr_cmd)
                rc = pci_write_config_byte(ioa_cfg->pdev, PCI_BIST, PCI_BIST_START);
 
        if (rc == PCIBIOS_SUCCESSFUL) {
+               ipr_set_affinity_nobalance(ioa_cfg, true);
                ipr_cmd->job_step = ipr_reset_bist_done;
                ipr_reset_start_timer(ipr_cmd, IPR_WAIT_FOR_BIST_TIMEOUT);
                rc = IPR_RC_JOB_RETURN;