1. Prepare the operation

Each controller has an Admin SQ/CQ with QID=0, separate from I/O queue identifiers.

AQA gives both lengths; ASQ/ACQ give their physical bases. Admin queues require physical contiguity and are not created with a PC=0 Create Queue PRP list.

PRP

Physical Region Page, a data-page pointer

ASQS/ACQS are zero-based: 64 entries encode as 63. Admin SQEs are 64 bytes and CQEs 16 bytes; the two queues need not have equal lengths.

2. Sequence and completion conditions

Allocate and initialize memory, write the three registers with EN=0 and clear Admin CQ phases. Configure CC, enable, wait for RDY=1, then submit the first Admin command.

Identify can be submitted and matched by CID in Admin CQ. ACQ is associated with interrupt vector 0, while the host can also poll phase tags.

CID

Command Identifier, unique among outstanding commands in one SQ

3. Handle unmet conditions

Size=0 encodes one slot and is not a legal way to disable an Admin queue; enabling in that configuration is undefined. Delete I/O Queue cannot delete Admin queues.

Compare AQA, allocated memory and the first CQE. Correct register readback does not establish device access to the supplied memory.

CQE

Completion Queue Entry

1. Establish the source and scope

Distinguish Admin/I/O and host-memory/CMB configurations. A rule for one is not automatically universal.

CMB

Controller Memory Buffer

Check AQA for Admin, and CAP.MQES/CQR and CC.IOSQES/IOCQES for I/O. CMB exceptions depend on CMBLOC.CQDA/CQPDS and CMBSZ queue support.

CMB

Controller Memory Buffer

2. Fields, units and worked interpretation

Length fields encode entries−1. Ordinary bases and each PRP page of a discontiguous queue align to CC.MPS; PRP1 has offset zero.

PRP

Physical Region Page, a data-page pointer

Choose slots and entry size, calculate bytes, allocate pages and a PRP list if supported, then Create. Do not alter a queue's live PRP list.

Minimum size is two slots; Admin maximum is 4096 and I/O maximum min(65536, MQES+1). One slot remains unavailable to distinguish full from empty.

3. Conditions that change the interpretation

For zero or unsupported I/O QSIZE, the specification recommends Invalid Queue Size (1/02h); a nonzero PRP offset should return PRP Offset Invalid (0/13h). Apply the specific SQ/CQ requirement for PC violating CQR.

A 128-slot NVM SQ needs 8192 bytes, its CQ 2048 bytes. Equal QSIZE=127 does not imply equal byte allocations.

1. Prepare the operation

An SQ selects its CQ at creation; several SQs may use one existing CQ.

Base 2.4 recommends using Set Features Number of Queues during initialization. Also check MQES/CQR, entry sizes and interrupt-vector limits; Create fields remain subject to their applicable requirements. The recommended initialization sequence and individual field requirements have different normative strengths.

Create CQ uses PRP1, QID, QSIZE, PC, IEN and IV. Create SQ additionally uses CQID, QPRIO and NVMSETID; zero NVMSETID means no specific set association.

2. Sequence and completion conditions

CQE

Completion Queue Entry

SCT

Status Code Type

3. Handle unmet conditions

An in-range CQID that has not been created returns Completion Queue Invalid (1/00h); CQID=0 or out of range returns Invalid Queue Identifier (1/01h).

Match CQID against the host's SQ-to-CQ map. Creating a CQ does not automatically associate every SQ with it.

1. What differs

Completion slots and CQ notification are shared; the SQs do not become one CID namespace.

CID

Command Identifier, unique among outstanding commands in one SQ

CQE.SQID+CID identifies the original command; SQHD updates consumption for that SQ.

SQID

Submission Queue Identifier

2. How to choose and verify

Number of Queues returns separate SQ/CQ allocations. Create SQ.CQID selects the mapping; CQ.IEN/IV configures notification.

Create the CQ and multiple SQs referencing it. Dispatch each CQE by SQID, then advance the shared CQ head.

SQID

Submission Queue Identifier

SQ1/CID 7 and SQ2/CID 7 may coexist and must be matched separately, unlike reusing an outstanding CID within one SQ.

3. Conditions that shape the comparison

Valid sharing is not an error. Missing CQIDs follow Create SQ status rules; a full shared CQ does not automatically produce Invalid Queue Size for each SQ.

Compare completion capacity, submission traffic and host consumption rate; SQ count alone does not establish freedom from bottlenecks.

1. Select the target and information

Admin queues require physical contiguity. Discontiguous I/O queues depend on CAP.CQR and memory location.

CQR=1 requires physical contiguity; CQR=0 allows discontiguous I/O queues. CMB placement additionally requires CMBSZ.SQS/CQS and CMBLOC.CQPDS/CQDA checks.

With Create I/O Queue.PC=1, PRP1 is the queue base; with PC=0 it points to a queue-page PRP list, distinct from ordinary data-transfer PRP2 rules.

2. Query sequence and interpretation

Confirm placement/support, build a complete aligned page list and Create. Retain the list's address and contents until successful deletion or reset.

The controller fetches SQEs/writes CQEs through the chosen layout. Successful Create does not authorize rearranging its pages while live.

3. Handle missing or inconsistent evidence

For CQR=1 and PC=0, Create CQ shall return Invalid Field (0/02h), while Create SQ says should. Unsupported CMB usage may return Invalid Use of Controller Memory Buffer (0/12h).

Cross-check CQR, PC, PRP1 interpretation and actual page layout, rather than validating only the completion.

First ask whether PRP1 addresses queue storage or its page list. Confusing them breaks the layout even with valid addresses.

1. Distinguish the failure conditions

This concerns Create I/O Queue Admin commands. Duplicate QIDs are checked within the SQ or CQ identifier space respectively.

Inspect QID, CQID, QSIZE and IV, and initialization of IOSQES/IOCQES. The relevant parameters are in CDW10/11.

Duplicate/zero/out-of-range QID should return 1/01h; zero/unsupported QSIZE should return 1/02h. Create SQ shall return 1/00h for an in-range uncreated CQID, or 1/01h for zero/out-of-range CQID. An IV exceeding the range in Transport 1.4 Figure 7 should return 1/08h. Unless specified otherwise, implementations choose among simultaneously applicable errors.

2. Establish the cause from evidence

CAP.MQES, Number of Queues allocations and PCIe interrupt configuration establish legal ranges.

For a learning test, change one condition at a time from a valid baseline; multiple errors can obscure the expected result.

3. Outcome and follow-up checks

Valid parameters successfully create the queue; after failure, do not treat the requested QID as created.

When available, Error Information Parameter Error Location can identify the field. Without More=1, do not assume every such error must have an entry.

First check that only one field is invalid and the expected SCT is 1, rather than comparing SC alone.

1. Establish the source and scope

FID 07h has controller scope and excludes Admin SQ/CQ.

FID

Feature Identifier

A successful Set Features FID 07h returns allocations in CQE DW0. A successful Get Features with SEL other than 011b also returns them; SEL=011b returns Supported Capabilities, not allocation counts. CAP.MQES separately limits per-queue depth.

2. Fields, units and worked interpretation

Requested NSQR/NCQR and returned NSQA/NCQA are zero-based. Allocations may be smaller or larger than requested because of allocation units.

If the host sets Number of Queues, the specification recommends doing so during initialization before creating any I/O queue. The first successful Set after Controller Level Reset (CLR) allocates the queues, and that allocation remains fixed until the next CLR.

DW0=00030007h allocates four CQs and eight SQs. Each still requires Create, with the respective QID ranges.

3. Checks across reset or power loss

4. Conditions that change the interpretation

Setting after an I/O queue exists shall return Command Sequence Error (0/0Ch). Requested=FFFFh should return Invalid Field (0/02h). Subsequent Set before queue creation should succeed without changing allocation.

Compare allocations, the actual created-queue list and MQES. Allocated counts are not current existence counts.

1. Prepare the operation

A CQ may serve multiple SQs. Delete every associated SQ, not only an SQ with the same numeric QID.

Use the host's SQ-to-CQID map from creation. Number of Queues does not provide this association list.

Delete I/O SQ/CQ selects the target with CDW10.QID; both travel through Admin SQ.

2. Sequence and completion conditions

Stop submissions, appropriately drain work, delete each associated SQ and await success, process needed old completions, then delete the CQ.

Successful Delete CQ allows reclamation of its PRP list and queue resources. Successful Delete SQ separately establishes termination of processing for that SQ's commands.

3. Handle unmet conditions

Delete CQ with associated SQs shall return Invalid Queue Deletion (1/0Ch); zero or invalid QID returns Invalid Queue Identifier (1/01h).

Correlate successful deletion with memory reclamation. Merely submitting Delete does not authorize early reclamation of DMA memory.

DMA

Direct Memory Access, device access to memory

1. Distinguish the failure conditions

SQ deletion affects its outstanding I/O; CQ deletion depends on the existence of every associated SQ.

Delete requires a valid nonzero QID; affected I/O commands retain their own SQID/CID identities.

Invalid QID: 1/01h. CQ still referenced: 1/0Ch. I/O aborted by SQ deletion uses 0/08h, including implicit completion established by successful Delete.

2. Establish the cause from evidence

Consult host creation/deletion and outstanding-command records. These are runtime state, not Identify optional capabilities.

Drain before normal teardown. On successful Delete SQ, previously submitted commands for which the controller had not posted a CQE are implicitly aborted. A CQE already posted but not yet read by the host remains an explicit completion.

3. Outcome and follow-up checks

Before successful Delete SQ, original commands may complete normally or with abort status. After success, no further completions may be posted for those commands.

Distinguish commands whose CQEs were posted before successful Delete from commands implicitly aborted without posted CQEs. A host-read log alone may miss an already posted CQE, so absence of a read is insufficient.

First check whether Delete SQ itself completed successfully. Its outstanding state does not establish that all original work stopped.

1. Mechanism and scope

Deleting an SQ does not delete its shared CQ; already posted completions may remain for host consumption.

Track deleted QID, associated CQID, PRP-list address and memory lifetime. Reused addresses require distinguishing queue lifetimes.

2. Understand it through actions and results

Use successful Delete completion on Admin CQ as the boundary, not submission time or a host timeout.

Stop submissions, successfully delete, then reclaim the corresponding queue/list memory. With SQ-only deletion, handle existing posted entries in the surviving CQ.

After successful Delete SQ, no new completions for its old commands may be posted. Successful Delete CQ permits releasing its list; a deleted queue is no longer a valid DMA target.

DMA

Direct Memory Access, device access to memory

3. Avoid a misleading conclusion

Failure or absence of Delete completion does not authorize memory reclamation. Timeout is not a substitute for success.

Correlate write time, phase, Delete completion and host consumption. Late consumption of an old CQE is not necessarily a prohibited late write.

1. Prepare the operation

Reset the target and its necessary dependencies. Rebuilding a CQ requires deleting its SQs first and recreating them afterward.

Use existing queue associations and Create/Delete rules, not a separate Queue Reset feature.

Use Delete SQ/CQ and Create CQ/SQ. QIDs may be reused, but memory and pointers begin a new lifetime.

2. Sequence and completion conditions

Normally let work finish before deletion. For cancellation, apply explicit/implicit Delete SQ completion rules. Initialize new CQ phases and create CQ before SQ.

New commands execute after reconstruction; old outstanding commands are not automatically migrated into the new SQ.

3. Handle unmet conditions

Deleting a referenced CQ returns 1/0Ch; premature SQ creation returns the applicable CQID error. Queue reset has no additional universal status.

Check teardown/recreation ordering, command-lifetime tracking and continued operation of independent queues.

1. Prepare the operation

Controller Level Reset deletes all I/O SQs/CQs of that controller, unlike deleting only one SQ.

Re-read necessary Identify data and verify queue allocations, entry sizes and interrupt settings rather than assuming previous values remain usable.

Admin registers, CC, Get/Set Features and Create CQ/SQ together establish the recovery path.

2. Sequence and completion conditions

For example, confirm RDY=0, initialize Admin CQ phases and CC, enable and wait for readiness, then restore needed features. Base 2.4 recommends Set Features Number of Queues during initialization; this example follows it and creates the CQ before the SQ. Submit new I/O only after its queue is created successfully.

Each new queue becomes valid after its successful Create completion on Admin CQ, even when reusing an identifier or address.

3. Handle unmet conditions

Updating old queue doorbells is not a valid recovery method. Failures of reconstruction commands follow their specific status rules.

Track queue lifetimes and command mappings across reset in host software; this is not a new NVMe wire field.