Handling an Exhausted Interrupt Supply While Attaching I/O Device Drivers

This following warning on the Oracle Solaris console means that the interrupt supply was exhausted while attaching I/O device drivers:

WARNING: ddi_intr_alloc: cannot fit into interrupt pool

This limitation applies only to the supported SPARC systems prior to the SPARC M7 series servers and SPARC T7 series servers.

The hardware provides a finite number of interrupts, so Oracle Solaris limits how many each device can use. A default limit is designed to match the needs of typical system configurations, however this limit may need adjustment for certain system configurations.

Specifically, the limit may need adjustment if the system is partitioned into multiple logical domains and if too many I/O devices are assigned to any guest domain. Oracle VM Server for SPARC divides the total interrupts into smaller sets given to guest domains. If too many I/O devices are assigned to a guest domain, its supply might be too small to give each device the default limit of interrupts. Thus, it exhausts its supply before it completely attaches all the drivers.

Some drivers provide an optional callback routine which allows Oracle Solaris to automatically adjust their interrupts. The default limit does not apply to these drivers.

To work around this issue, use the ::irmpools and ::irmreqs MDB macros to determine how interrupts are used. The ::irmpools macro shows the overall supply of interrupts divided into pools. The ::irmreqs macro shows which devices are mapped to each pool. For each device, ::irmreqs shows whether the default limit is enforced by an optional callback routine, how many interrupts each driver requested, and how many interrupts the driver is given.

The macros do not show information about drivers that failed to attach. However, the information that is shown helps calculate the extent to which you can adjust the default limit. Any device that uses more than one interrupt without providing a callback routine can be forced to use fewer interrupts by adjusting the default limit. Reducing the default limit below the amount that is used by such a device results in freeing of interrupts for use by other devices.

To adjust the default limit, set the ddi_msix_alloc_limit property to a value from 1 to 8 in the /etc/system file. Then, reboot the system for the change to take effect.

To maximize performance, start by assigning larger values and decrease the values in small increments until the system boots successfully without any warnings. Use the ::irmpools and ::irmreqs macros to measure the adjustment's impact on all attached drivers.

For example, suppose the following warnings are issued while booting the Oracle Solaris OS in a guest domain:

WARNING: emlxs3: interrupt pool too full.
WARNING: ddi_intr_alloc: cannot fit into interrupt pool

The ::irmpools and ::irmreqs macros show the following information:

# echo "::irmpools" | mdb -k
ADDR             OWNER   TYPE   SIZE  REQUESTED  RESERVED
00000400016be970 px#0    MSI/X  36    36         36

# echo "00000400016be970::irmreqs" | mdb -k
ADDR             OWNER   TYPE   CALLBACK NINTRS NREQ NAVAIL
00001000143acaa8 emlxs#0 MSI-X  No       32     8    8
00001000170199f8 emlxs#1 MSI-X  No       32     8    8
000010001400ca28 emlxs#2 MSI-X  No       32     8    8
0000100016151328 igb#3   MSI-X  No       10     3    3
0000100019549d30 igb#2   MSI-X  No       10     3    3
0000040000e0f878 igb#1   MSI-X  No       10     3    3
000010001955a5c8 igb#0   MSI-X  No       10     3    3

The default limit in this example is eight interrupts per device, which is not enough interrupts to accommodate the attachment of the final emlxs3 device to the system. Assuming that all emlxs instances behave in the same way, emlxs3 probably requested 8 interrupts.

By subtracting the 12 interrupts used by all of the igb devices from the total pool size of 36 interrupts, 24 interrupts are available for the emlxs devices. Dividing the 24 interrupts by 4 suggests that 6 interrupts per device would enable all emlxs devices to attach with equal performance. So, the following adjustment is added to the /etc/system file:

set ddi_msix_alloc_limit = 6

When the system successfully boots without warnings, the ::irmpools and ::irmreqs macros show the following updated information:

# echo "::irmpools" | mdb -k
ADDR             OWNER   TYPE   SIZE  REQUESTED  RESERVED
00000400018ca868 px#0    MSI/X  36    36         36
 
# echo "00000400018ca868::irmreqs" | mdb -k
ADDR             OWNER   TYPE   CALLBACK NINTRS NREQ NAVAIL
0000100016143218 emlxs#0 MSI-X  No       32     8    6
0000100014269920 emlxs#1 MSI-X  No       32     8    6
000010001540be30 emlxs#2 MSI-X  No       32     8    6
00001000140cbe10 emlxs#3 MSI-X  No       32     8    6
00001000141210c0 igb#3   MSI-X  No       10     3    3
0000100017549d38 igb#2   MSI-X  No       10     3    3
0000040001ceac40 igb#1   MSI-X  No       10     3    3
000010001acc3480 igb#0   MSI-X  No       10     3    3