Dump Manager Design

Author: Dhruvaraj Subhashchandran dhruvaraj@in.ibm.com

Other contributors:

Created: 12/12/2019

Problem Description

During a crash or a host failure, an event monitor mechanism generates an error log, but the size of the error log is limited to few kilobytes, so all the data from the crash or failure may not fit into an error log. The additional data required for the debugging needs to be collected as a dump. The existing OpenBMC dump interfaces support only the dumps generated on the BMC and dump manager doesn't support download operations.

Glossary

  • System Dump: A dump of the Host's main memory and processor registers. Read More
  • Memory Preserving Reboot(MPR): A method of reboot with preserving the contents of the volatile memory
  • PLDM: An interface and data model to access low-level platform inventory, monitoring, control, event, and data/parameters transfer functions. ReadMore
  • Machine Check Exception: A severe error inside a processor core that causes a processor core to stop all processing activities.
  • BMCWeb: An embedded webserver for OpenBMC. More Info

Background and References

Various types of dumps are created based on the type and source of failure. The dump manager, which is orchestrating the collection and offload, needs to provide methods to create, store the dump details, and offload it. Additionally, some sources allow the dump to be extracted manually without a failure to understand the current state or analyze a suspected problems.

Type of dumps supported.

These are some of the dumps supported by dump manager.

BMC Dump

A dump collected when there is a failure in the BMC with various debug information. This type of dump can be generated by user too to get the current state of the BMC. This dump gets collected on BMC and stored on BMC

System Dump

A system dump is a collection of debugging information from the host, this may include host memory and/or register data. This dump can be initiated by BMC and there can be system reboots while collecting the dump. Dump gets stored in the host memory and offloaded through the BMC or get collected directly to BMC based on the size of dump contents and the available space on the BMC to store the dump.

Resource dump

A special type of host dump is initiated and collected by the host based on the request from a user. No system state change may be necessary during the collection of this kind of a dump. A resource indicator may be used to indicate what data to be collected. The content of the dump can be decided by the host. This dump can be stored in host memory and offloaded through BMC or host can send this dump down to BMC once the collection is completed based on the size of the dump and the availability of space on the BMC.

Hostboot dump

A dump that can be collected during the boot failure of the host. This dump may or may not include the contents of the main memory and/or the processor registers.

Hardware dump

This dump can be collected during a critical failure on the hardware components like the processor while the host is booted and running. The host may stop during this dump and may collect various processor states and/or memory contents to help to debug the failure.

Requirements

Dump use cases - Users are examples, not a mandatory part of implementation

Dump manager needs to provide interfaces for

  • Create a dump: Initiate the creation of the dump, based on an error condition or a user request.
  • List the dumps: List all dumps present in the BMC.
  • Get a dump: Offload the dump to an external entity.
  • Notify: Notify the dump manager that a new dump is created.
  • Delete the dump.
  • Mark a dump as offloaded to an external entity.
  • Set the dump policies like disabling a type of dump or dump overwriting policy.

Proposed Design

There are various types of dumps; interfaces are standard for most of the dumps, but huge dumps which cannot be stored on the BMC needs additional support. This document will explain the design of different types of dumps. The dumps are classified based on where it is collected, stored, and how it is extracted. Two major types are

  • Collected by BMC and stored on BMC.
  • Collected and stored on an attached entity but offloaded through BMC.

This proposal focuses on re-using the existing phosphor-debug-collector, which collects the dumps from BMC.

phosphor-debug-collector

Brief design points of existing phosphor-debug-collector

  • A create interface which assumes the type is BMC dump and returns an ID to the caller for the user-initiated dumps.
  • An external request of dump is considered as a user-initiated BMC dump and initiate BMC dump collection with a tool named dreport with type manual dump
  • The dreport create dump file in the dump path provided by the dump manager code.
  • A watch process is forked by the dump manager to catch the signal for the file close on that path and assume the dump collection is completed.
  • The watch process calls an internal dbus interface to create the dump entry with size and timestamp.
  • The path of dump is based on the predefined base path and the id of the dump.
  • When the request comes for offload, the file is downloaded from the dump base path +id, no update in the entry whether dump is offloaded
  • Deleting a dump by deleting the entry and internal file also will be deleted.
  • There are system generated dumps based on error log or core dump, which works similar to user-initiated dumps with the following difference.
  • No external create D-Bus interface is needed, but a monitor application will monitor for specific conditions like the existence of a core file or a signal for new error log creation of a few selected types.
  • Once the event occurred, the monitor will call an internal D-Bus interface dump manager to create the dump with the type of dump.
  • The dump manager calls dreport with a dump type got from the monitor and write data to a path based on dump id.

Updates proposed to the existing phosphor-debug-collector.

  • External D-Bus interface needs to specify the type of the dump since a user can request multiple types of dumps
  • Create will be returning an id which can be mapped to the actual dump once it is created.
  • A Notify interface is provided for notifying the creation of a dump outside the BMC but offloaded through BMC.
  • The InitiateOffload function will be implemented to download the dump.
  • Status of the dump, whether offloaded or not, will be added to the dump entry.

Dump manager interfaces.

  • Dump Manager DBus object provides interfaces for creating and managing dump

  • Interfaces

    • Create: The interface to create a dump, called by clients to initiate user-initiated dump.

      • AdditionalData: The additional data, if any, for initiating the dump. The key in this case should be an implementation specific enum defined for the specific type of dump either in xyz or in a domain. The values can be either a string or a 64 bit number. The enum-format string is required to come from a parallel class with this specific Enum name. All of the Enum strings should be in the format 'domain.Dump.Create.CreateParameters.ParamName'. e.g.: { "key1": "value1", "key2": "value2" } ends up in AdditionaData like: ["KEY1=value1", "KEY2=value2"]

    • Notify: Notify the dump manager that a new dump is created.

      • ID: ID of the dump, if not 0 this will be the external id of the dump
      • Type: Type of dump that was created.
      • Size: Size of the dump

Dump entry interfaces

-  **InitiateOffload**: Initiate the offload of the dump.
    - OffloadUri: The URI where the dump should be offloaded.

The properties common to all dumps

There will be a base dump entry with properties common to all types of dumps

  • ID: Id of the dump
  • Timestamp: Dump creation timestamp
  • Size: Total size of the dump
  • OffloadComplete: Set to true when offload is completed
  • OffloadURI: The URI for offloading the dump, set while initiating the offload. Specific types need to inherit from this common dump entry class and add specific properties.

Additional propertries based on dump types

BMC Dump
  • No Additional properties
System Dump
  • External Source ID: ID provided by the Host, this id will be used for all communication to the source of the dump, in this case, Host.

Flow of dumps collected and stored in the Host

PLDM is provided as an example dump transport and notification mechanism between Host and BMC.

  • Create: Initiate methods to create the dump in Host.
  • Generating the dump in Host
  • Host notifies the creation of dump through PLDM to BMC.
  • PLDM call Notify to create the dump entry
  • InitiateOffload: Dump manager request Host to start offload
  • The Host sends the dump through PLDM, and PLDM on BMC sends it out.

Alternatives Considered

  • Offloading Host dumps through Host instead of BMC, but considered BMC option due to following reasons - The BMC is considered the "management path" of most servers and often the Host is not connected to the desired network for the offload location. - BMC provides one common point for all dumps generated in the system for external management appliance.

Impacts

  • The existing BMC dump interface needs to be re-used. The current interface is not accepting a dump type, so a new interface to create the dump with type will be provided for BMC dump also without changing the existing interface.
  • Modifying the BMC dump infrastructure to support additional dumps.
  • openpower-proc-control will be updated to call memory preserving chip-ops and to handle memory preserving reboot on POWER platforms.
  • Additional system state to indicate the system is collecting debug data While performing memory preserving reboot.

Testing

  • Unit tests to make sure the dump manager interfaces are working.
  • Following integration tests will be executed to make sure the dump manager is working as expected. - Test creating host dumps and offloading it. - Test deleting host dumps - Create/List/Offload/Delete BMC dumps to make sure existing dump manager functions are not broken.
  • Automated tests for dump Create/List/Offload/Delete to avoid regression.