Compare commits

...
Sign in to create a new pull request.

13 commits

Author SHA1 Message Date
806ec77a1f voice-proto: long command timeout only for voice blob sends
The blanket 240 s audioreach_send_cmd_sync timeout (cfc2b6a276) broke
boot audio: apm_probe's GET_SPF_STATE ready poll gets no answer when
sent before the DSP audio framework is up (the command is simply lost),
so probe held the shared command mutex for the full 240 s. That stalls
snd_q6apm's module init past udev's worker timeout, q6prm never gets
loaded, the LPI pinctrl never finds its 'core' clock and the sound card
never assembles (every boot since 2026-07-10 had no card). It also
queued every later voice-proto command behind the mutex for minutes,
which sessions 8-11 misread as slow DSP responses.

Restore the mainline 30 s default via audioreach_send_cmd_sync_timeout
and pass 240 s only from the voice-proto blob paths (a cold TX
GRAPH_OPEN through the Fluence FastRPC loads takes >120 s).

Assisted-by: Claude:claude-fable-5
2026-07-11 17:20:11 +02:00
2d3c1ed4cf voice-proto: don't free a reused static-PD heap on spawn failure
If a repeat audiopd spawn fails (e.g. remote invoke timeout), the error
path freed the persistent remote heap that a previous spawn had
SCM-assigned to the LPASS/ADSP VMIDs, without reclaiming it for HLOS.
The protected pages go back to the page allocator and the next 3 MB
dma_alloc dies with a synchronous external abort in dcache_clean_poc
(observed 2026-07-11 after an adsprpcd-audiopd restart). Free the heap
in the error path only when this spawn created it.

Pre-existing mainline bug, candidate for upstream once distilled.

Assisted-by: Claude:claude-fable-5
2026-07-11 16:45:23 +02:00
f3eb810dfe voice-proto: route audiopd to its dedicated context bank (pd-type)
The aDSP refuses to dlopen unsigned dynamic modules (fluence_nn) in an
audio static PD spawned on a generic context bank: it demands a
per-device testsig, which production-fused units cannot satisfy. Stock
(downstream adsprpc) binds audiopd to a dedicated CB — volcano-dsp.dtsi
cb12, SIDs 0x1005/0x1065, pd-type = <2> (AUDIO_STATICPD) — and there the
same unsigned fluence_nn loads fine on identical silicon.

Mirror that: tag compute-cb@5 (same SIDs) with pd-type = <2>, keep
pd-typed banks out of generic round-robin session allocation, and on
createstaticpd of "audiopd" re-route the client to the tagged session
before anything is allocated against it, so every IOVA the DSP sees for
this PD carries the audio bank's SID.

Local milos carry, not upstreamable as-is (upstream binding has no
pd-type model). See journal/calls.md 2026-07-11 session 10.

Assisted-by: Claude:claude-fable-5
2026-07-11 15:58:09 +02:00
cfc2b6a276 voice-proto: AMDB module pre-loading, longer cmd timeout
Three findings from the 2026-07-10 warm-capture session (journal/calls.md):
the stock voice sequence is byte-identical warm vs cold, so our replay
bytes were never the problem; VCPM rejects TX-side cal (EBADPARAM) while
accepting byte-identical RX cal through the same transport; and a voice
GRAPH_OPEN that has to FastRPC-load the Fluence pp modules takes >120 s
cold, so the client-side timeout/verify dance leaves failed opens in
VCPM's session history.

- apm_callback: handle AMDB_CMD_RSP_LOAD_MODULES (AMDB load responds
  with its own opcode, not an ibasic result); log per-module handles
  (handle 0 = not registered - e.g. 0x70010a3, which only appears after
  fluence_nn_module_fvxiii.so.1 is parsed).
- voice-proto: Q6VP_DSTPORT_FLAG record flag - first payload word
  carries the GPR destination port, so blobs can address static services
  other than APM (AMDB port 3 module pre-loading via mkamdbload.py).
- Raise the shared command timeout to 300 s so a cold TX voice
  GRAPH_OPEN (~125 s through the Fluence loads) completes inside ONE
  command with no retry dance.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Jorijn van der Graaf <jorijnvdgraaf@catcrafts.net>
2026-07-10 23:50:20 +02:00
3c381d76fc voice-proto: make cal registration safe; handle REGISTER_CFG responses
- apm_callback: store results for APM_CMD_REGISTER_CFG (0x01001008) so
  registrations complete synchronously instead of always timing out
  (their success/failure was previously invisible and each cost 5 s).
- voice-proto: REGISTER_CFG payloads get permanent bump-allocated
  slices in the shared region instead of reusing the scratch slot -
  the DSP keeps reading registered config memory, and overwriting it
  with the next record's payload while registered wedged the SoC hard
  (two force-reboots on 2026-07-08).
- Grow the region to 8 MiB (the stock sequence registers a 6.4 MiB
  global cal blob) and raise the shared command timeout to 30 s: a
  voice GRAPH_OPEN can block >5 s while the ADSP dynamically loads the
  Fluence modules through FastRPC.

With this the full RX-side stock sequence (open, 139 KB cal register,
VSID/cal-keys, EP configs, prepare, start) completes with status 0 on
a live call; TX GRAPH_START is the one remaining DSP rejection.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Jorijn van der Graaf <jorijnvdgraaf@catcrafts.net>
2026-07-09 00:44:51 +02:00
d36259f54b arm64: dts: qcom: milos: add fastrpc remote heap for the ADSP audio PD
The ADSP audio PD loads its voice post-processing modules (Fluence
NN/EF/pro-vc etc.) as dynamic shared objects fetched from the apps
filesystem at GRAPH_OPEN time through FastRPC (adsprpcd attached with
createstaticpd:audiopd serves the files - /vendor/dsp/adsp on stock).
The static-PD attach allocates a remote heap that must be SCM-assigned
to the DSP: without qcom,vmids the heap stays HLOS-owned and the PD's
first touch kills the SoC instantly (xPU violation, no kernel output).

Add a 12 MiB CMA pool (mirroring downstream's adsp_mem_heap: 32-bit
addressable, 4 MiB aligned) and the LPASS/ADSP_HEAP vmids, following
the sc7280 (kodiak) precedent where upstream audiopd support landed.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Jorijn van der Graaf <jorijnvdgraaf@catcrafts.net>
2026-07-08 01:31:08 +02:00
5f42f9a74a voice-proto: add out-of-band (shared memory) command support
Stock GSL delivers every voice GRAPH_OPEN out-of-band: the whole graph
(all subgraphs plus the cross-subgraph module connections and control
links, 4.3-4.7 KB) must be opened atomically in ONE command, which
exceeds the in-band GPR/GLINK intent limit. Splitting the open
per-subgraph to fit in-band makes the cross-referencing TX DevicePP
subgraph invalid - the root cause of every pp-subgraph open failure
since 2026-07-06 (byte-identical payloads open fine when sent whole;
verified against a gsl_log.bin packet capture of a working call on
stock Android).

Records with bit 31 set in the opcode are copied into a DMA buffer
allocated from the q6apm-dais device (it carries the ADSP SMMU
mapping), mapped once with the DSP via q6apm_map_memory_fixed_region
under a private graph id, and referenced by absolute DSP address in
the apm_cmd_header (mainline maps in absolute mode; GSL uses offset
mode - the DSP rejects offset references against our mapping).

Assisted-by: Claude:claude-fable-5
Signed-off-by: Jorijn van der Graaf <jorijnvdgraaf@catcrafts.net>
2026-07-08 01:31:08 +02:00
57cbd69e36 Merge branch 'voice' into combined 2026-07-07 20:07:09 +02:00
0c63174949 net: qrtr: forward data packets between endpoints (FP6 voice prototype)
The DSP subsystems are star-connected through the application
processor: each has a qrtr link to us but none to each other. Incoming
packets whose dst_node is neither the local node nor broadcast are
currently delivered to a *local* port matching dst_port (usually
nothing) and dropped, so two remote DSPs can never exchange QMI
messages.

On the FP6 (SM7635, AudioReach) the modem's voice stack needs exactly
that: it must reach the ADSP's service-registry notifier / audio
service to attach the vocoder to a VCPM voice session. Without
forwarding, every voice graph starts cleanly but the modem never
streams a single mailbox packet - calls stay silent both ways (see
journal/calls.md 2026-07-07/08).

Forward DATA and RESUME_TX packets destined to another known node onto
that node's endpoint, re-using qrtr_node_enqueue for per-hop flow
control. RESUME_TX additionally releases this hop's flow token on the
arrival link; per-hop counters stay in lockstep with end-to-end ones
because every packet of a forwarded flow transits this node. Control
packets keep flowing to the local ns, which already redistributes
service announcements mesh-wide.

Prototype for the milos carry; needs discussion (loop prevention,
broadcast handling, flow-control semantics) before any upstream
attempt.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Jorijn van der Graaf <jorijnvdgraaf@catcrafts.net>
2026-07-07 20:07:01 +02:00
3cbf6d6fa8 Merge branch 'voice' into combined 2026-07-06 23:17:04 +02:00
2626032a50 voice-proto: continue on DSP rejections, log each record result
Assisted-by: Claude:claude-fable-5
2026-07-06 23:17:04 +02:00
63b5c43f02 Merge branch 'voice' into combined
Voice-call blob-player prototype (q6apm debugfs).
2026-07-06 23:06:35 +02:00
7695d5e6e4 ASoC: qdsp6: q6apm: add voice-call blob-player prototype (FP6 bring-up)
Replay pre-generated APM command sequences (GRAPH_OPEN/SET_CFG/PREPARE/
START, mined from the stock FP6 ACDB by utilities/mkvoiceblobs.py)
through the q6apm GPR channel via
/sys/kernel/debug/q6apm-voice-proto/play.

Bring-up prototype for cellular voice-call audio: the replayed graphs
contain the VCPM voice config; the modem and the ADSP VCPM service
attach the vocoder to them once a call is active. Local carry only,
NOT for upstream (the upstreamable form will be topology-driven).

Assisted-by: Claude:claude-fable-5
2026-07-06 23:06:35 +02:00
9 changed files with 577 additions and 9 deletions

View file

@ -548,6 +548,14 @@
no-map;
};
adsp_rpc_heap_mem: adsp-rpc-heap {
compatible = "shared-dma-pool";
alloc-ranges = <0x0 0x0 0x0 0xffffffff>;
reusable;
alignment = <0x0 0x400000>;
size = <0x0 0xc00000>;
};
pvm_fw_mem: pvm-fw-region@824a0000 {
reg = <0x0 0x824a0000 0x0 0x100000>;
no-map;
@ -1398,6 +1406,9 @@
qcom,glink-channels = "fastrpcglink-apps-dsp";
label = "adsp";
qcom,non-secure-domain;
memory-region = <&adsp_rpc_heap_mem>;
qcom,vmids = <QCOM_SCM_VMID_LPASS
QCOM_SCM_VMID_ADSP_HEAP>;
#address-cells = <1>;
#size-cells = <0>;
@ -1422,6 +1433,8 @@
reg = <5>;
iommus = <&apps_smmu 0x1005 0x0>,
<&apps_smmu 0x1065 0x0>;
/* AUDIO_STATICPD bank, per downstream volcano-dsp.dtsi */
pd-type = <2>;
dma-coherent;
};

View file

@ -103,6 +103,13 @@
#define USER_PD (1)
#define SENSORS_PD (2)
/*
* Downstream DT "pd-type" context-bank designations (volcano-dsp.dtsi).
* A static PD must be spawned on its designated context bank / SMMU
* stream or the DSP treats it as a generic signed PD.
*/
#define FASTRPC_DT_PD_TYPE_AUDIO_STATICPD 2
#define miscdev_to_fdevice(d) container_of(d, struct fastrpc_device, miscdev)
struct fastrpc_phy_page {
@ -255,6 +262,8 @@ struct fastrpc_session_ctx {
int sid;
bool used;
bool valid;
/* downstream DT pd-type this context bank is dedicated to (0 = generic) */
u32 pd_type;
};
struct fastrpc_soc_data {
@ -314,6 +323,11 @@ struct fastrpc_user {
struct kref refcount;
};
static struct fastrpc_session_ctx *fastrpc_session_alloc_pd_type(
struct fastrpc_user *fl, u32 pd_type);
static void fastrpc_session_free(struct fastrpc_channel_ctx *cctx,
struct fastrpc_session_ctx *session);
/* Extract SMMU PA from consolidated IOVA */
static inline dma_addr_t fastrpc_ipa_to_dma_addr(struct fastrpc_channel_ctx *cctx, dma_addr_t iova)
{
@ -1343,7 +1357,7 @@ static int fastrpc_init_create_static_process(struct fastrpc_user *fl,
struct fastrpc_phy_page pages[1];
char *name;
int err;
bool scm_done = false;
bool scm_done = false, heap_created = false;
struct {
int client_id;
u32 namelen;
@ -1371,11 +1385,35 @@ static int fastrpc_init_create_static_process(struct fastrpc_user *fl,
goto err;
}
/*
* The audio static PD must live on its designated context bank so the
* DSP sees its memory on the audio SMMU stream; on a generic bank the
* aDSP enforces module signing in the PD (dlopen of unsigned dynamic
* modules fails demanding a testsig). Re-route before anything is
* allocated against the session. Falls back to the already-held
* generic session if no bank is tagged in DT.
*/
if (init.namelen >= strlen("audiopd") &&
!memcmp(name, "audiopd", strlen("audiopd"))) {
struct fastrpc_session_ctx *audio_sess;
audio_sess = fastrpc_session_alloc_pd_type(fl,
FASTRPC_DT_PD_TYPE_AUDIO_STATICPD);
if (audio_sess) {
fastrpc_session_free(fl->cctx, fl->sctx);
fl->sctx = audio_sess;
}
dev_info(&fl->cctx->rpdev->dev,
"audiopd: using session sid %d (%s)\n", fl->sctx->sid,
audio_sess ? "dedicated audio CB" : "no audio CB in DT");
}
if (!fl->cctx->remote_heap) {
err = fastrpc_remote_heap_alloc(fl, fl->sctx->dev, init.memlen,
&fl->cctx->remote_heap);
if (err)
goto err_name;
heap_created = true;
/* Map if we have any heap VMIDs associated with this ADSP Static Process. */
if (fl->cctx->vmcount) {
@ -1446,8 +1484,16 @@ err_invoke:
&fl->cctx->remote_heap->dma_addr, fl->cctx->remote_heap->size, err);
}
err_map:
/*
* A heap reused from an earlier spawn is still SCM-assigned to the
* remote VMIDs; freeing it here would return protected pages to the
* allocator and a later touch crashes with a synchronous external
* abort. Keep it for the next spawn attempt.
*/
if (heap_created) {
fastrpc_buf_free(fl->cctx->remote_heap);
fl->cctx->remote_heap = NULL;
}
err_name:
kfree(name);
err:
@ -1582,7 +1628,9 @@ static struct fastrpc_session_ctx *fastrpc_session_alloc(
spin_lock_irqsave(&cctx->lock, flags);
for (i = 0; i < cctx->sesscount; i++) {
if (!cctx->session[i].used && cctx->session[i].valid) {
/* dedicated static-PD context banks are assigned by pd_type only */
if (!cctx->session[i].used && cctx->session[i].valid &&
!cctx->session[i].pd_type) {
cctx->session[i].used = true;
session = &cctx->session[i];
/* any non-zero ID will work, session_idx + 1 is the simplest one */
@ -1595,6 +1643,29 @@ static struct fastrpc_session_ctx *fastrpc_session_alloc(
return session;
}
static struct fastrpc_session_ctx *fastrpc_session_alloc_pd_type(
struct fastrpc_user *fl, u32 pd_type)
{
struct fastrpc_channel_ctx *cctx = fl->cctx;
struct fastrpc_session_ctx *session = NULL;
unsigned long flags;
int i;
spin_lock_irqsave(&cctx->lock, flags);
for (i = 0; i < cctx->sesscount; i++) {
if (!cctx->session[i].used && cctx->session[i].valid &&
cctx->session[i].pd_type == pd_type) {
cctx->session[i].used = true;
session = &cctx->session[i];
fl->client_id = i + 1;
break;
}
}
spin_unlock_irqrestore(&cctx->lock, flags);
return session;
}
static void fastrpc_session_free(struct fastrpc_channel_ctx *cctx,
struct fastrpc_session_ctx *session)
{
@ -2251,6 +2322,9 @@ static int fastrpc_cb_probe(struct platform_device *pdev)
if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
dev_info(dev, "FastRPC Session ID not specified in DT\n");
sess->pd_type = 0;
of_property_read_u32(dev->of_node, "pd-type", &sess->pd_type);
if (sessions > 0) {
struct fastrpc_session_ctx *dup_sess;

View file

@ -520,6 +520,54 @@ int qrtr_endpoint_post(struct qrtr_endpoint *ep, const void *data, size_t len)
qrtr_node_assign(node, le32_to_cpu(pkt->server.node));
}
/* The DSPs are star-connected through this node: forward packets
* destined to another node onto that node's endpoint (e.g. the
* modem's voice stack talking to the ADSP's audio service). Only
* DATA and RESUME_TX transit; control packets keep going to the
* local ns, which does its own mesh-wide redistribution. RESUME_TX
* additionally releases this hop's flow-control token on the
* arrival link: per-hop counters advance in lockstep with the
* end-to-end ones since every packet of the flow transits here.
*/
if (cb->dst_node != qrtr_local_nid &&
cb->dst_node != QRTR_NODE_BCAST &&
(cb->type == QRTR_TYPE_DATA || cb->type == QRTR_TYPE_RESUME_TX)) {
struct sockaddr_qrtr from = {AF_QIPCRTR,
cb->src_node, cb->src_port};
struct sockaddr_qrtr to = {AF_QIPCRTR,
cb->dst_node, cb->dst_port};
struct qrtr_node *dst;
dst = qrtr_node_lookup(cb->dst_node);
if (!dst || dst == node) {
if (dst)
qrtr_node_release(dst);
goto err;
}
if (cb->type == QRTR_TYPE_RESUME_TX) {
struct sk_buff *clone;
clone = skb_clone(skb, GFP_ATOMIC);
if (clone)
qrtr_tx_resume(node, clone);
}
pr_debug("qrtr: fwd %u:%u -> %u:%u type %d len %zu\n",
cb->src_node, cb->src_port,
cb->dst_node, cb->dst_port, cb->type, size);
if (skb_cow_head(skb, sizeof(struct qrtr_hdr_v1))) {
qrtr_node_release(dst);
goto err;
}
qrtr_node_enqueue(dst, skb, cb->type, &from, &to);
qrtr_node_release(dst);
return 0;
}
if (cb->type == QRTR_TYPE_RESUME_TX) {
qrtr_tx_resume(node, skb);
} else {

View file

@ -1,6 +1,6 @@
# SPDX-License-Identifier: GPL-2.0-only
snd-q6dsp-common-y := q6dsp-common.o q6dsp-lpass-ports.o q6dsp-lpass-clocks.o
snd-q6apm-y := q6apm.o audioreach.o topology.o
snd-q6apm-y := q6apm.o audioreach.o topology.o q6apm-voice-proto.o
obj-$(CONFIG_SND_SOC_QDSP6_COMMON) += snd-q6dsp-common.o
obj-$(CONFIG_SND_SOC_QDSP6_CORE) += q6core.o

View file

@ -576,10 +576,11 @@ void *audioreach_alloc_graph_pkt(struct q6apm *apm,
}
EXPORT_SYMBOL_GPL(audioreach_alloc_graph_pkt);
int audioreach_send_cmd_sync(struct device *dev, gpr_device_t *gdev,
int audioreach_send_cmd_sync_timeout(struct device *dev, gpr_device_t *gdev,
struct gpr_ibasic_rsp_result_t *result, struct mutex *cmd_lock,
gpr_port_t *port, wait_queue_head_t *cmd_wait,
const struct gpr_pkt *pkt, uint32_t rsp_opcode)
const struct gpr_pkt *pkt, uint32_t rsp_opcode,
unsigned long timeout_s)
{
const struct gpr_hdr *hdr = &pkt->hdr;
@ -601,9 +602,9 @@ int audioreach_send_cmd_sync(struct device *dev, gpr_device_t *gdev,
if (rsp_opcode)
rc = wait_event_timeout(*cmd_wait, (result->opcode == hdr->opcode) ||
(result->opcode == rsp_opcode), 5 * HZ);
(result->opcode == rsp_opcode), timeout_s * HZ);
else
rc = wait_event_timeout(*cmd_wait, (result->opcode == hdr->opcode), 5 * HZ);
rc = wait_event_timeout(*cmd_wait, (result->opcode == hdr->opcode), timeout_s * HZ);
if (!rc) {
dev_err(dev, "CMD timeout for [%x] opcode\n", hdr->opcode);
@ -620,6 +621,16 @@ err:
mutex_unlock(cmd_lock);
return rc;
}
EXPORT_SYMBOL_GPL(audioreach_send_cmd_sync_timeout);
int audioreach_send_cmd_sync(struct device *dev, gpr_device_t *gdev,
struct gpr_ibasic_rsp_result_t *result, struct mutex *cmd_lock,
gpr_port_t *port, wait_queue_head_t *cmd_wait,
const struct gpr_pkt *pkt, uint32_t rsp_opcode)
{
return audioreach_send_cmd_sync_timeout(dev, gdev, result, cmd_lock,
port, cmd_wait, pkt, rsp_opcode, 30);
}
EXPORT_SYMBOL_GPL(audioreach_send_cmd_sync);
int audioreach_graph_send_cmd_sync(struct q6apm_graph *graph, const struct gpr_pkt *pkt,

View file

@ -58,12 +58,15 @@ struct q6apm_graph;
#define APM_CMD_GRAPH_FLUSH 0x01001005
#define APM_CMD_SET_CFG 0x01001006
#define APM_CMD_GET_CFG 0x01001007
#define APM_CMD_REGISTER_CFG 0x01001008
#define APM_CMD_SHARED_MEM_MAP_REGIONS 0x0100100C
#define APM_CMD_SHARED_MEM_UNMAP_REGIONS 0x0100100D
#define APM_CMD_RSP_SHARED_MEM_MAP_REGIONS 0x02001001
#define APM_CMD_RSP_GET_CFG 0x02001000
#define APM_CMD_CLOSE_ALL 0x01001013
#define APM_CMD_REGISTER_SHARED_CFG 0x0100100A
#define AMDB_CMD_LOAD_MODULES 0x01001020
#define AMDB_CMD_RSP_LOAD_MODULES 0x02001008
#define APM_MEMORY_MAP_SHMEM8_4K_POOL 3
@ -842,6 +845,11 @@ void audioreach_graph_free_buf(struct q6apm_graph *graph);
int audioreach_send_cmd_sync(struct device *dev, gpr_device_t *gdev, struct gpr_ibasic_rsp_result_t *result,
struct mutex *cmd_lock, gpr_port_t *port, wait_queue_head_t *cmd_wait,
const struct gpr_pkt *pkt, uint32_t rsp_opcode);
int audioreach_send_cmd_sync_timeout(struct device *dev, gpr_device_t *gdev,
struct gpr_ibasic_rsp_result_t *result,
struct mutex *cmd_lock, gpr_port_t *port, wait_queue_head_t *cmd_wait,
const struct gpr_pkt *pkt, uint32_t rsp_opcode,
unsigned long timeout_s);
int audioreach_graph_send_cmd_sync(struct q6apm_graph *graph, const struct gpr_pkt *pkt,
uint32_t rsp_opcode);
int audioreach_set_media_format(struct q6apm_graph *graph,

View file

@ -0,0 +1,377 @@
// SPDX-License-Identifier: GPL-2.0
/*
* Voice-call bring-up prototype for the Fairphone (Gen. 6): replay
* pre-generated APM command sequences through the q6apm GPR channel.
*
* The blobs reproduce, byte for byte, the GRAPH_OPEN/SET_CFG/PREPARE/START
* sequence the stock Android PAL/AGM/GSL stack sends to set up the
* handset voice-call graphs (mined from the device ACDB; generated by
* utilities/mkvoiceblobs.py in the bring-up repo). The modem's voice
* engine and the ADSP VCPM service wire up the vocoder themselves once
* these graphs are running.
*
* Local prototype only, NOT for upstream. Interface:
* echo <fw-file> > /sys/kernel/debug/q6apm-voice-proto/play
* where <fw-file> is relative to /lib/firmware. Playback aborts on the
* first command the DSP rejects.
*/
#include <linux/debugfs.h>
#include <linux/dma-mapping.h>
#include <linux/firmware.h>
#include <linux/kernel.h>
#include <linux/of_platform.h>
#include <linux/platform_device.h>
#include <linux/slab.h>
#include <linux/soc/qcom/apr.h>
#include <linux/uaccess.h>
#include <sound/soc.h>
#include "audioreach.h"
#include "q6apm.h"
#define Q6VP_MAGIC 0x50563651 /* "Q6VP" little-endian */
#define Q6VP_VERSION 1
/*
* A record opcode with this bit set is sent out-of-band: the payload is
* placed in an ADSP-mapped shared-memory region and the GPR command carries
* only an apm_cmd_header referencing it. Stock GSL sends every voice
* GRAPH_OPEN this way; it is mandatory because a voice graph must be opened
* atomically as one command (all subgraphs + cross-subgraph module
* connections and control links), and the whole payload (~4.3-4.7 KB)
* exceeds the ~4 KiB in-band GPR/GLINK intent limit. Splitting the open
* per-subgraph to fit in-band makes the cross-referencing TX DevicePP
* subgraph invalid (rejected at open) see journal/calls.md 2026-07-08.
*/
#define Q6VP_OOB_FLAG 0x80000000
/*
* A record opcode with this bit set carries the GPR destination port in the
* first payload word (stripped before sending). Lets a blob address static
* services other than APM e.g. AMDB (port 3) module pre-loading, so the
* dynamic Fluence pp modules can be made resident BEFORE the voice
* GRAPH_OPENs. A first open that has to FastRPC-load them can fail or block
* so long that the client-side retry dance corrupts VCPM's per-VSID voice
* session (TX cal EBADPARAM / TX START EFAILED, journal/calls.md
* 2026-07-10). In-band records only.
*/
#define Q6VP_DSTPORT_FLAG 0x40000000
#define Q6VP_GRAPH_ID 0xf000 /* private mem-map handle owner */
#define Q6VP_BUF_SZ 0x800000 /* 8 MiB shared region */
#define Q6VP_SCRATCH_SZ 0x10000 /* [0, 64K): transient commands */
#define Q6VP_SID_MASK 0xf
struct q6vp_hdr {
__le32 magic;
__le32 version;
__le32 num_records;
} __packed;
struct q6vp_rec {
__le32 opcode;
__le32 size;
} __packed;
static DEFINE_MUTEX(q6vp_play_lock);
static struct q6apm *q6vp_apm;
static struct dentry *q6vp_dir;
static struct device *q6vp_dais_dev;
static struct audioreach_graph_info *q6vp_info;
static void *q6vp_buf;
static dma_addr_t q6vp_buf_iova;
static phys_addr_t q6vp_buf_dsp_addr;
static int q6vp_send(uint32_t opcode, const void *payload, uint32_t size)
{
uint32_t dest_port = APM_MODULE_INSTANCE_ID;
uint32_t rsp_opcode = 0;
struct gpr_pkt *pkt;
int ret;
if (opcode & Q6VP_DSTPORT_FLAG) {
if (size < sizeof(__le32))
return -EINVAL;
dest_port = le32_to_cpup(payload);
payload += sizeof(__le32);
size -= sizeof(__le32);
opcode &= ~Q6VP_DSTPORT_FLAG;
}
/* AMDB load responds with its own opcode, not an ibasic result */
if (opcode == AMDB_CMD_LOAD_MODULES)
rsp_opcode = AMDB_CMD_RSP_LOAD_MODULES;
pkt = audioreach_alloc_cmd_pkt(size, opcode, 0, GPR_APM_MODULE_IID,
dest_port);
if (IS_ERR(pkt))
return PTR_ERR(pkt);
memcpy((void *)pkt + GPR_HDR_SIZE + APM_CMD_HDR_SIZE, payload, size);
/*
* Voice blobs need the long wait: a cold TX GRAPH_OPEN that has to
* FastRPC-load the Fluence pp modules takes >120 s. Regular audio
* commands (and especially the probe-time GET_SPF_STATE ready poll)
* must keep the short default or they hold the shared command mutex
* hostage at boot: a 240 s stall in apm_probe delays module loading
* past udev's worker timeout and the sound card never assembles.
*/
ret = q6apm_send_cmd_sync_timeout(q6vp_apm, pkt, rsp_opcode, 240);
kfree(pkt);
return ret;
}
/*
* Lazily set up the shared-memory region used for OOB commands: allocate a
* DMA buffer against the q6apm-dais child device (it carries the iommus
* mapping the ADSP expects, same as the PCM data path), register a private
* graph_info so the APM_CMD_RSP_SHARED_MEM_MAP_REGIONS callback has a slot
* to store the handle in, and map the region with the DSP. The mapping is
* kept until the module goes away.
*/
static int q6vp_oob_init(void)
{
struct device *dev = q6vp_apm->dev;
struct of_phandle_args args;
struct platform_device *pdev;
struct device_node *np;
u64 sid = 0;
u32 graph_id = Q6VP_GRAPH_ID;
int ret;
if (q6vp_info && q6vp_info->mem_map_handle)
return 0;
if (!q6vp_dais_dev) {
np = of_get_compatible_child(dev->of_node, "qcom,q6apm-dais");
if (!np) {
dev_err(dev, "voice-proto: no q6apm-dais node\n");
return -ENODEV;
}
if (!of_parse_phandle_with_fixed_args(np, "iommus", 1, 0,
&args)) {
sid = args.args[0] & Q6VP_SID_MASK;
of_node_put(args.np);
}
pdev = of_find_device_by_node(np);
of_node_put(np);
if (!pdev) {
dev_err(dev, "voice-proto: no q6apm-dais device\n");
return -ENODEV;
}
q6vp_dais_dev = &pdev->dev;
q6vp_buf = dma_alloc_coherent(q6vp_dais_dev, Q6VP_BUF_SZ,
&q6vp_buf_iova, GFP_KERNEL);
if (!q6vp_buf)
return -ENOMEM;
q6vp_buf_dsp_addr = q6vp_buf_iova | (sid << 32);
}
if (!q6vp_info) {
q6vp_info = kzalloc(sizeof(*q6vp_info), GFP_KERNEL);
if (!q6vp_info)
return -ENOMEM;
INIT_LIST_HEAD(&q6vp_info->sg_list);
q6vp_info->id = Q6VP_GRAPH_ID;
ret = idr_alloc_u32(&q6vp_apm->graph_info_idr, q6vp_info,
&graph_id, graph_id, GFP_KERNEL);
if (ret < 0) {
kfree(q6vp_info);
q6vp_info = NULL;
return ret;
}
}
ret = q6apm_map_memory_fixed_region(q6vp_dais_dev, Q6VP_GRAPH_ID,
q6vp_buf_dsp_addr, Q6VP_BUF_SZ);
if (ret) {
dev_err(dev, "voice-proto: mem map failed (%d)\n", ret);
return ret;
}
dev_info(dev, "voice-proto: OOB region mapped, handle 0x%x\n",
q6vp_info->mem_map_handle);
return 0;
}
/*
* Persistent-slice bump offset: APM_CMD_REGISTER_CFG registers the payload
* with the DSP, which keeps reading that memory for as long as the config
* stays registered those payloads must never be overwritten. They get
* bump-allocated slices above Q6VP_SCRATCH_SZ that live until reboot;
* transient commands (opens etc.) reuse the scratch slot at offset 0.
*/
static u32 q6vp_persist_off = Q6VP_SCRATCH_SZ;
static int q6vp_send_oob(uint32_t opcode, const void *payload, uint32_t size)
{
struct apm_cmd_header *cmd_header;
struct gpr_pkt *pkt;
u32 off = 0;
int ret;
ret = q6vp_oob_init();
if (ret)
return ret;
if (opcode == APM_CMD_REGISTER_CFG) {
if (q6vp_persist_off + size > Q6VP_BUF_SZ)
return -EFBIG;
off = q6vp_persist_off;
q6vp_persist_off += ALIGN(size, SZ_4K);
} else if (size > Q6VP_SCRATCH_SZ) {
return -EFBIG;
}
memcpy(q6vp_buf + off, payload, size);
pkt = audioreach_alloc_apm_cmd_pkt(0, opcode, 0);
if (IS_ERR(pkt))
return PTR_ERR(pkt);
/*
* The region is mapped in absolute-address mode (mainline's
* q6apm_map_memory_fixed_region passes property_flag 0, unlike GSL
* which maps with IS_OFFSET_MODE and then references offset 0), so
* the header carries the full DSP-visible address of the buffer.
*/
cmd_header = (void *)pkt + GPR_HDR_SIZE;
cmd_header->payload_address_lsw = lower_32_bits(q6vp_buf_dsp_addr + off);
cmd_header->payload_address_msw = upper_32_bits(q6vp_buf_dsp_addr + off);
cmd_header->mem_map_handle = q6vp_info->mem_map_handle;
cmd_header->payload_size = size;
ret = q6apm_send_cmd_sync_timeout(q6vp_apm, pkt, 0, 240);
kfree(pkt);
return ret;
}
static int q6vp_play(const char *name)
{
const struct firmware *fw;
const struct q6vp_hdr *hdr;
const struct q6vp_rec *rec;
struct device *dev = q6vp_apm->dev;
size_t off;
uint32_t i, num, opcode, size;
int ret, err = 0;
ret = request_firmware(&fw, name, dev);
if (ret) {
dev_err(dev, "voice-proto: cannot load %s (%d)\n", name, ret);
return ret;
}
hdr = (const struct q6vp_hdr *)fw->data;
if (fw->size < sizeof(*hdr) ||
le32_to_cpu(hdr->magic) != Q6VP_MAGIC ||
le32_to_cpu(hdr->version) != Q6VP_VERSION) {
dev_err(dev, "voice-proto: %s: bad header\n", name);
ret = -EINVAL;
goto out;
}
num = le32_to_cpu(hdr->num_records);
off = sizeof(*hdr);
for (i = 0; i < num; i++) {
if (off + sizeof(*rec) > fw->size) {
ret = -EINVAL;
goto out;
}
rec = (const struct q6vp_rec *)(fw->data + off);
opcode = le32_to_cpu(rec->opcode);
size = le32_to_cpu(rec->size);
off += sizeof(*rec);
if (off + size > fw->size) {
ret = -EINVAL;
goto out;
}
if (opcode & Q6VP_OOB_FLAG)
ret = (opcode & Q6VP_DSTPORT_FLAG) ? -EINVAL :
q6vp_send_oob(opcode & ~Q6VP_OOB_FLAG,
fw->data + off, size);
else
ret = q6vp_send(opcode, fw->data + off, size);
dev_info(dev, "voice-proto: %s[%u] opcode 0x%08x size %u -> %d\n",
name, i,
opcode & ~(Q6VP_OOB_FLAG | Q6VP_DSTPORT_FLAG),
size, ret);
/* keep going on DSP rejections; each result is logged */
if (ret)
err = ret;
off += ALIGN(size, 4);
}
out:
release_firmware(fw);
return ret ? ret : err;
}
static ssize_t q6vp_play_write(struct file *file, const char __user *ubuf,
size_t count, loff_t *ppos)
{
char name[128];
int ret;
if (!q6vp_apm)
return -ENODEV;
if (count >= sizeof(name))
return -EINVAL;
if (copy_from_user(name, ubuf, count))
return -EFAULT;
name[count] = '\0';
strim(name);
/*
* Serialize blob playback: concurrent debugfs writers (e.g. a
* dropped ssh session whose write is still stuck in a DSP wait)
* would interleave their records on the GPR channel and corrupt
* both sequences (seen 2026-07-10).
*/
ret = mutex_lock_interruptible(&q6vp_play_lock);
if (ret)
return ret;
ret = q6vp_play(name);
mutex_unlock(&q6vp_play_lock);
return ret ? ret : count;
}
static const struct file_operations q6vp_play_fops = {
.open = simple_open,
.write = q6vp_play_write,
.llseek = default_llseek,
};
void q6apm_voice_proto_init(struct q6apm *apm)
{
q6vp_apm = apm;
q6vp_dir = debugfs_create_dir("q6apm-voice-proto", NULL);
debugfs_create_file("play", 0200, q6vp_dir, NULL, &q6vp_play_fops);
}
void q6apm_voice_proto_exit(void)
{
debugfs_remove_recursive(q6vp_dir);
q6vp_dir = NULL;
if (q6vp_info) {
if (q6vp_info->mem_map_handle)
q6apm_unmap_memory_fixed_region(q6vp_dais_dev,
Q6VP_GRAPH_ID);
idr_remove(&q6vp_apm->graph_info_idr, Q6VP_GRAPH_ID);
kfree(q6vp_info);
q6vp_info = NULL;
}
if (q6vp_buf) {
dma_free_coherent(q6vp_dais_dev, Q6VP_BUF_SZ, q6vp_buf,
q6vp_buf_iova);
q6vp_buf = NULL;
}
q6vp_dais_dev = NULL;
q6vp_apm = NULL;
}

View file

@ -38,6 +38,16 @@ int q6apm_send_cmd_sync(struct q6apm *apm, const struct gpr_pkt *pkt,
NULL, &apm->wait, pkt, rsp_opcode);
}
int q6apm_send_cmd_sync_timeout(struct q6apm *apm, const struct gpr_pkt *pkt,
uint32_t rsp_opcode, unsigned long timeout_s)
{
gpr_device_t *gdev = apm->gdev;
return audioreach_send_cmd_sync_timeout(&gdev->dev, gdev, &apm->result,
&apm->lock, NULL, &apm->wait,
pkt, rsp_opcode, timeout_s);
}
static struct audioreach_graph *q6apm_get_audioreach_graph(struct q6apm *apm, uint32_t graph_id)
{
struct audioreach_graph_info *info;
@ -764,6 +774,8 @@ static int apm_probe(gpr_device_t *gdev)
g_apm = apm;
q6apm_voice_proto_init(apm);
q6apm_get_apm_state(apm);
ret = snd_soc_register_component(dev, &q6apm_audio_component, NULL, 0);
@ -781,6 +793,7 @@ static int apm_probe(gpr_device_t *gdev)
static void apm_remove(gpr_device_t *gdev)
{
q6apm_voice_proto_exit();
of_platform_depopulate(&gdev->dev);
snd_soc_unregister_component(&gdev->dev);
}
@ -817,6 +830,7 @@ static int apm_callback(const struct gpr_resp_pkt *data, void *priv, int op)
case GPR_BASIC_RSP_RESULT:
switch (result->opcode) {
case APM_CMD_SHARED_MEM_MAP_REGIONS:
case APM_CMD_REGISTER_CFG:
case APM_CMD_GRAPH_START:
case APM_CMD_GRAPH_OPEN:
case APM_CMD_GRAPH_PREPARE:
@ -849,6 +863,23 @@ static int apm_callback(const struct gpr_resp_pkt *data, void *priv, int op)
break;
}
break;
case AMDB_CMD_RSP_LOAD_MODULES: {
/* amdb_module_load_unload_t: proc_domain, error_code,
* num_modules, {module_id, handle_lsw, handle_msw}[].
* A zero handle = module not registered with the AMDB.
*/
const u32 *amdb = data->payload;
u32 i, n = amdb[2];
for (i = 0; i < n; i++)
dev_info(dev, "AMDB load: module 0x%08x handle 0x%08x%08x\n",
amdb[3 + 3 * i], amdb[5 + 3 * i],
amdb[4 + 3 * i]);
apm->result.opcode = hdr->opcode;
apm->result.status = amdb[1];
wake_up(&apm->wait);
break;
}
case APM_CMD_RSP_SHARED_MEM_MAP_REGIONS:
apm->result.opcode = hdr->opcode;
apm->result.status = 0;

View file

@ -144,6 +144,8 @@ int q6apm_unmap_memory_fixed_region(struct device *dev, unsigned int graph_id);
/* Helpers */
int q6apm_send_cmd_sync(struct q6apm *apm, const struct gpr_pkt *pkt,
uint32_t rsp_opcode);
int q6apm_send_cmd_sync_timeout(struct q6apm *apm, const struct gpr_pkt *pkt,
uint32_t rsp_opcode, unsigned long timeout_s);
/* Callback for graph specific */
struct audioreach_module *q6apm_find_module_by_mid(struct q6apm_graph *graph,
@ -157,4 +159,8 @@ int q6apm_remove_initial_silence(struct device *dev, struct q6apm_graph *graph,
int q6apm_remove_trailing_silence(struct device *dev, struct q6apm_graph *graph, uint32_t samples);
int q6apm_set_real_module_id(struct device *dev, struct q6apm_graph *graph, uint32_t codec_id);
int q6apm_get_hw_pointer(struct q6apm_graph *graph, int dir);
/* FP6 voice-call bring-up prototype (q6apm-voice-proto.c) */
void q6apm_voice_proto_init(struct q6apm *apm);
void q6apm_voice_proto_exit(void);
#endif /* __APM_GRAPH_ */