Document source: 5a7a75c321ea; code reference: 5a7a75c321ea. zh_CN. Source language: en; requested: en; actual: en; translation: original.

Source documentation does not establish software or physical hardware qualification. See the support and qualification records for their exact scope.

SoC execution budgets

The authored assembly and immutable resolved IR select exact controller modes, capacities and clock profiles. This document explains how to review their execution costs; it is not another capability registry or configuration input. The derived capability matrix owns the maintained selection projection. Product scheduling and overload policy stay external.

Work, time and completion

A finite work count is not a measured time bound. Bootstrap loops use explicit poll limits because the monotonic clock may not yet exist. Runtime absolute deadlines belong to the current boot clock domain. IRQ delay, DMA arbitration, bus stalls, notification latency and service intervals require measurements for the exact firmware and hardware under load.

Each selected mode must document its maximum IRQ work, task service work, critical-section scope, caller storage and terminal predicate. A drain deadline limits the period before quarantine; it never authorizes releasing borrowed memory. The executor must keep calling service/recovery until hardware, IRQ, queues and observers actually withdraw their references.

Current mechanisms

Mode family	Bounded work and progress	Final ownership condition
GPIO	Direct authorized register command; toggle requires serialized writers	External direct writers joined before pin release
UART IRQ bytes/events	At most one TX and one RX hardware byte per IRQ; reads copy one ring item per short critical section	Actual wire TC or reviewed reset abort, TX IRQ withdrawal and captured wake join
UART IRQ blocks	One hardware byte per IRQ; block acquisition/release is explicit	Last prefix published, hardware sources withdrawn and every consumer loan returned
UART finite TX DMA	One selected channel’s terminal facts per IRQ; task service observes DMA and UART independently	DMA engine/source off and actual UART wire completion before SETTLED
UART RX DMA idle/continuous and duplex	Selected bank/IDLE/error facts; caller supplies finite buffers/blocks and replacement capacity	Memory engine stopped and stable prefix/bank proven; RX loans and independent TX drain retained
SPI short polling	Per-call finite length and absolute deadline; no IRQ mode fallback	Both controller and endpoint retain ownership for the complete CS interval
SPI finite full-duplex DMA	Two selected engines; task service checks memory and shifter facts	Both DMA sources withdrawn, wire idle, then CS release and SETTLED
I2C polling	Finite message/byte count; reviewed ACK windows and bounded STOP work	STOP/master/line state satisfies the exact implementation contract
I2C IRQ/DMA messages	Shared V1 IRQ consumes at most two DATA bytes; DMA observes its selected direction	Protocol STOP, wire/master state and DMA/source withdrawal all succeed
ADC polling/trigger blocks	At most the configured scan per task-side rearm; finite block capacity	Trigger and conversion stopped, DMA actually disabled, then publication/reclaim
ADC continuous trigger blocks	Double-bank terminal facts and explicit replacement; no overwrite of borrowed blocks	Closed producer and DMA sources; all outstanding block loans returned
EXTI	Visit only selected line states in deterministic generated order; one queue fact per pending line	This line masked/acknowledged, active shared vector joined, wake withdrawn
PWM/advanced PWM	Fixed authorized channel/period operations and latched break state	Reviewed inactive output before timer ownership release
Timer capture	Finite selected channels and visible capture/overflow facts	Capture IRQ/source withdrawal before event storage release
Classic CAN	Three TX mailboxes; at most three initially queued frames per each of two RX FIFOs	Per-mailbox abort/terminal facts, INAK/RX drain and publisher join
USB FS device PIO	At most eight RX headers and 128 payload words per IRQ; progress scans selected endpoint count	Endpoint request drain, FIFO/source withdrawal, publisher join and clock release
Raw Ethernet/MDIO	Task progress bounded by selected descriptor counts; MDIO has an absolute deadline	Descriptor/DMA ownership withdrawn, IRQ publisher joined and every RX loan returned
Internal Flash	Finite reviewed physical operation and absolute deadline; geometry is chip-specific	Hardware busy cleared or explicit retained recovery responsibility
Watchdog	Finite enable/feed operation; enabling is irreversible	Stop never fabricates release of an active hardware watchdog

The exact implementations and configured limits remain authoritative:

• UART: soc/*/drivers/uart*.c; G474, H743 and U575 are IRQ bytes only. The F407, G474, H743 and U575 wire-drain allowance is 20000000 / baud + 100 microseconds. GD IRQ uses its reviewed reset-abort protocol; it is not assigned the ST wire-drain formula.

• SPI: soc/*/drivers/spi*.c; the finite DMA drain allowance is 200 microseconds.

• I2C: soc/common/i2c_v1_async.c and the SoC wrappers; V1 async drain allowance is 1,000 microseconds. Polling STOP has separate finite work.

• ADC/EXTI/timer: selected SoC driver and private state; scan, queue and channel limits come from resolved inputs or explicit caller storage.

• CAN/USB/Ethernet: io/src/can_controller.c, usb_device_controller.c and eth_mac.c; poll/drain limits and selected capacities are validated at cold admission. IRQ notification remains a hint, never a replacement for predicates.

Cost review and evidence

Keep the exact resolved configuration, compiler identity, flags, linked ELF/map, resource budget and retention outputs together. Compare equivalent CPU/ABI, backend, clock, memory layout and capacities. A source instruction count, object-file size or different-capacity image cannot establish an equivalent whole-image saving. Report Flash, static RAM, reservations and caller storage separately. Disassembly is real tool output, not a cycle measurement.

tools/measurement/soc_costs.py observes actual linked functions and ELF resource costs with complete raw command receipts and input/tool hashes. Its comparison rejects changed routes, bindings, claims, backend or capacity profiles rather than silently attributing their cost to an implementation improvement. The empty versus selected and baremetal versus FreeRTOS dimensions remain useful selection-cost reports, but are not same-profile regression comparisons. The SoC execution record identifies the source scope of completed observations and pending final baselines.

For changed interrupt/concurrency behavior, GoogleTest/GoogleMock calls the real production source with external register/CPU ports modeled. Record actual RED, then GREEN; retain admission, loss ordering, cancel, quarantine, wake and instance isolation regressions. ARM budget/retention verifies actual linking. Software checks do not certify physical IRQ latency or DMA throughput.

No numerical maximum service interval is asserted without the selected producer rate, storage, hardware and interference budget. External products must qualify that interval and the resulting overload policy. A missed service deadline cannot revoke an accepted request or block loan. Current physical measurements remain not executed until an actual station records them.