Skip to content

Latest commit

 

History

History
74 lines (49 loc) · 3.36 KB

trace-format.md

File metadata and controls

74 lines (49 loc) · 3.36 KB
Raw

Described in this file is the version 1.0 of the sqlite block trace format.

Format overview

Several tables in a sqlite file created using rvnmetadata (and the correct format type).

The two tables are described in the following sections.

Blocks

The list of all basic blocks executed in the trace, in order of the first execution of the block.

The first block in the table must be a special block with instruction_data equal to "interrupt", that is referenced by the execution table when non-instructions such as interrupts and faults are executed.

Fields

  • "pc int8 not null," -- The address of the first instruction executed in the block
  • "instruction_data blob not null," -- A blob of the bytes of all the instructions in the block
  • "instruction_count int2 not null," -- The number of instructions in the block
  • "mode int1 not null" -- The execution mode (64, 32 bits or 16 bits, x86 only at the moment). Values can be found in the ExecutionMode enum in block_writer.h.

Execution

The sequential list of all the execution events in the trace. In the context of this library, an execution event corresponds to the execution of an entire basic block, or of part of a basic block.

Fields

  • "transition_id int8 PRIMARY KEY not null," -- The transition id of the first transition that is after the execution of this block.
  • "block_id int4 not null" -- The rowid of the executed block

Implementation detail

This table is "WITHOUT_ROWID" to disable the automatic rowid column added by sqlite. This spares us an explicit index on transition_id, which benefits the size of the database and the speed of access.

It is OK to do so, because two execution events at the same transition is always a bug.

Instruction indices

The list of the offsets (indices) of the instructions in the block.

  • The offset of the first instruction, which is always 0, is not saved.
  • The the size of the last instruction can be computed from the last offset and the size of the entire block, and so is not saved (furthermore, it would be difficult to obtain).

NOTE: due to the fact that blocks are not always executed fully, the last instruction known to this table may not be the last instruction in the block. For the purpose of rvnblock this is not a problem, because the instructions past this one are never executed in the trace. Furthermore, the last instruction known to this table may contain bytes from the unexecuted instructions. This means that downstream clients are expected to use a disassembler to find the actual instruction in this case (like it is the case currently in the reven_backend where we always read 15 bytes for the data of an instruction).

Fields

  • "block_id INTEGER NOT NULL," -- The rowid of the block that contains the instruction
  • "instruction_index INTEGER NOT NULL," -- The offset of the instruction relative to the first instruction of the block

Interrupts

Contains interrupt data

Fields

  • "transition_id int8 PRIMARY KEY NOT NULL,": transition_id of the interrupt
  • "pc int8 NOT NULL,": pc right before the interrupt
  • "mode int1 NOT NULL,": mode right before the interrupt
  • "number INTEGER NOT NULL,": interrupt number. In x86, the index in the interrupt table.
  • "is_hw BOOL NOT NULL,": whether the interrupt is hardware or software
  • "related_instruction_block_id INTEGER NOT NULL": If there is a related instruction, its block id. Otherwise, 0.