* [PATCH v3 0/5] Add missing features to FastRPC driver
@ 2023-11-21 11:24 Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver Ekansh Gupta
` (4 more replies)
0 siblings, 5 replies; 7+ messages in thread
From: Ekansh Gupta @ 2023-11-21 11:24 UTC (permalink / raw
To: srinivas.kandagatla, linux-arm-msm; +Cc: gregkh, linux-kernel
This patch series adds the listed features that have been missing
in upstream fastRPC driver.
- Add early wakeup support to allow DSP user to send early response
to CPU and improve fastrpc performance.
- Add polling mode support with which driver polls on memory to avoid
CPU from going to low power modes.
- Add notifications frameworks to provide users with the DSP PD status
notifications.
- Add a control mechanism to allow users to clean up DSP user PD
- Add wakelock management support
This patch series depends on multimode support series:
[1] https://lore.kernel.org/all/20231121094844.5764-1-quic_ekangupt@quicinc.com/
Ekansh Gupta (5):
misc: fastrpc: Add early wakeup support for fastRPC driver
misc: fastrpc: Add polling mode support for fastRPC driver
misc: fastrpc: Add DSP PD notification support
misc: fastrpc: Add support for users to clean up DSP user PD
misc: fastrpc: Add wakelock management support
drivers/misc/fastrpc.c | 568 +++++++++++++++++++++++++++++++++++-
include/uapi/misc/fastrpc.h | 8 +
2 files changed, 562 insertions(+), 14 deletions(-)
--
2.17.1
^ permalink raw reply [flat|nested] 7+ messages in thread
* [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver
2023-11-21 11:24 [PATCH v3 0/5] Add missing features to FastRPC driver Ekansh Gupta
@ 2023-11-21 11:24 ` Ekansh Gupta
2023-11-24 13:27 ` Srinivas Kandagatla
2023-11-21 11:24 ` [PATCH v3 2/5] misc: fastrpc: Add polling mode " Ekansh Gupta
` (3 subsequent siblings)
4 siblings, 1 reply; 7+ messages in thread
From: Ekansh Gupta @ 2023-11-21 11:24 UTC (permalink / raw
To: srinivas.kandagatla, linux-arm-msm; +Cc: gregkh, linux-kernel
CPU wake up and context switch latency are significant in FastRPC
overhead for remote calls. As part of this change, DSP sends early
signal of completion to CPU and FastRPC driver detects early signal
on the given context and starts polling on a memory for actual
completion. Multiple different response flags are added to support
DSP user early hint of approximate time of completion, early response
from DSP user to wake up CPU and poll on memory for actual completion.
Complete signal is also added which is sent by DSP user in case of
timeout after early response is sent.
Signed-off-by: Ekansh Gupta <quic_ekangupt@quicinc.com>
---
Changes in v3:
- Rebase the patch to latest kernel version
drivers/misc/fastrpc.c | 265 +++++++++++++++++++++++++++++++++++++++--
1 file changed, 252 insertions(+), 13 deletions(-)
diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index 8e77beb3a693..6b6ac3e3328d 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -106,6 +106,19 @@
#define USER_PD (1)
#define SENSORS_PD (2)
+#define FASTRPC_RSP_VERSION2 2
+/* Early wake up poll completion number received from remoteproc */
+#define FASTRPC_EARLY_WAKEUP_POLL (0xabbccdde)
+/* timeout in us for polling until memory barrier */
+#define FASTRPC_POLL_TIME_MEM_UPDATE (500)
+/* timeout in us for busy polling after early response from remoteproc */
+#define FASTRPC_POLL_TIME (4000)
+/* timeout in us for polling completion signal after user early hint */
+#define FASTRPC_USER_EARLY_HINT_TIMEOUT (500)
+/* CPU feature information to DSP */
+#define FASTRPC_CPUINFO_DEFAULT (0)
+#define FASTRPC_CPUINFO_EARLY_WAKEUP (1)
+
#define miscdev_to_fdevice(d) container_of(d, struct fastrpc_device, miscdev)
#define PERF_END ((void)0)
@@ -129,6 +142,15 @@
(uint64_t *)(perf_ptr + offset)\
: (uint64_t *)NULL) : (uint64_t *)NULL)
+enum fastrpc_response_flags {
+ NORMAL_RESPONSE = 0,
+ EARLY_RESPONSE = 1,
+ USER_EARLY_SIGNAL = 2,
+ COMPLETE_SIGNAL = 3,
+ STATUS_RESPONSE = 4,
+ POLL_MODE = 5,
+};
+
static const char *domains[FASTRPC_DEV_MAX] = { "adsp", "mdsp",
"sdsp", "cdsp"};
struct fastrpc_phy_page {
@@ -206,6 +228,14 @@ struct fastrpc_invoke_rsp {
int retval; /* invoke return value */
};
+struct fastrpc_invoke_rspv2 {
+ u64 ctx; /* invoke caller context */
+ int retval; /* invoke return value */
+ u32 flags; /* early response flags */
+ u32 early_wake_time; /* user hint in us */
+ u32 version; /* version number */
+};
+
struct fastrpc_buf_overlap {
u64 start;
u64 end;
@@ -272,11 +302,17 @@ struct fastrpc_invoke_ctx {
int pid;
int tgid;
u32 sc;
+ /* user hint of completion time in us */
+ u32 early_wake_time;
u32 *crc;
u64 *perf_kernel;
u64 *perf_dsp;
u64 ctxid;
u64 msg_sz;
+ /* work done status flag */
+ bool is_work_done;
+ /* response flags from remote processor */
+ enum fastrpc_response_flags rsp_flags;
struct kref refcount;
struct list_head node; /* list of ctxs */
struct completion work;
@@ -321,7 +357,9 @@ struct fastrpc_channel_ctx {
struct list_head invoke_interrupted_mmaps;
bool secure;
bool unsigned_support;
+ bool cpuinfo_status;
u64 dma_mask;
+ u64 cpuinfo_todsp;
};
struct fastrpc_device {
@@ -352,13 +390,21 @@ struct fastrpc_user {
struct mutex mutex;
};
+struct fastrpc_ctrl_latency {
+ u32 enable; /* latency control enable */
+ u32 latency; /* latency request in us */
+};
+
struct fastrpc_ctrl_smmu {
u32 sharedcb; /* Set to SMMU share context bank */
};
struct fastrpc_internal_control {
u32 req;
- struct fastrpc_ctrl_smmu smmu;
+ union {
+ struct fastrpc_ctrl_latency lp;
+ struct fastrpc_ctrl_smmu smmu;
+ };
};
static inline int64_t getnstimediff(struct timespec64 *start)
@@ -692,6 +738,8 @@ static struct fastrpc_invoke_ctx *fastrpc_context_alloc(
ctx->pid = current->pid;
ctx->tgid = user->tgid;
ctx->cctx = cctx;
+ ctx->rsp_flags = NORMAL_RESPONSE;
+ ctx->is_work_done = false;
init_completion(&ctx->work);
INIT_WORK(&ctx->put_work, fastrpc_context_put_wq);
@@ -1302,6 +1350,115 @@ static int fastrpc_invoke_send(struct fastrpc_session_ctx *sctx,
}
+static int poll_for_remote_response(struct fastrpc_invoke_ctx *ctx, u32 timeout)
+{
+ int err = -EIO, ii = 0, jj = 0;
+ u32 sc = ctx->sc;
+ struct fastrpc_invoke_buf *list;
+ struct fastrpc_phy_page *pages;
+ u64 *fdlist = NULL;
+ u32 *crclist = NULL, *poll = NULL;
+ unsigned int inbufs, outbufs, handles;
+
+ /* calculate poll memory location */
+ inbufs = REMOTE_SCALARS_INBUFS(sc);
+ outbufs = REMOTE_SCALARS_OUTBUFS(sc);
+ handles = REMOTE_SCALARS_INHANDLES(sc) + REMOTE_SCALARS_OUTHANDLES(sc);
+ list = fastrpc_invoke_buf_start(ctx->rpra, ctx->nscalars);
+ pages = fastrpc_phy_page_start(list, ctx->nscalars);
+ fdlist = (u64 *)(pages + inbufs + outbufs + handles);
+ crclist = (u32 *)(fdlist + FASTRPC_MAX_FDLIST);
+ poll = (u32 *)(crclist + FASTRPC_MAX_CRCLIST);
+
+ /* poll on memory for DSP response. Return failure on timeout */
+ for (ii = 0, jj = 0; ii < timeout; ii++, jj++) {
+ if (*poll == FASTRPC_EARLY_WAKEUP_POLL) {
+ /* Remote processor sent early response */
+ err = 0;
+ break;
+ }
+ if (jj == FASTRPC_POLL_TIME_MEM_UPDATE) {
+ /* Wait for DSP to finish updating poll memory */
+ rmb();
+ jj = 0;
+ }
+ udelay(1);
+ }
+ return err;
+}
+
+static inline int fastrpc_wait_for_response(struct fastrpc_invoke_ctx *ctx,
+ u32 kernel)
+{
+ int interrupted = 0;
+
+ if (kernel)
+ wait_for_completion(&ctx->work);
+ else
+ interrupted = wait_for_completion_interruptible(&ctx->work);
+
+ return interrupted;
+}
+
+static void fastrpc_wait_for_completion(struct fastrpc_invoke_ctx *ctx,
+ int *ptr_interrupted, u32 kernel)
+{
+ int err = 0, jj = 0;
+ bool wait_resp = false;
+ u32 wTimeout = FASTRPC_USER_EARLY_HINT_TIMEOUT;
+ u32 wakeTime = ctx->early_wake_time;
+
+ do {
+ switch (ctx->rsp_flags) {
+ /* try polling on completion with timeout */
+ case USER_EARLY_SIGNAL:
+ /* try wait if completion time is less than timeout */
+ /* disable preempt to avoid context switch latency */
+ preempt_disable();
+ jj = 0;
+ wait_resp = false;
+ for (; wakeTime < wTimeout && jj < wTimeout; jj++) {
+ wait_resp = try_wait_for_completion(&ctx->work);
+ if (wait_resp)
+ break;
+ udelay(1);
+ }
+ preempt_enable();
+ if (!wait_resp) {
+ *ptr_interrupted = fastrpc_wait_for_response(ctx, kernel);
+ if (*ptr_interrupted || ctx->is_work_done)
+ return;
+ }
+ break;
+ /* busy poll on memory for actual job done */
+ case EARLY_RESPONSE:
+ err = poll_for_remote_response(ctx, FASTRPC_POLL_TIME);
+ /* Mark job done if poll on memory successful */
+ /* Wait for completion if poll on memory timeout */
+ if (!err) {
+ ctx->is_work_done = true;
+ return;
+ }
+ if (!ctx->is_work_done) {
+ *ptr_interrupted = fastrpc_wait_for_response(ctx, kernel);
+ if (*ptr_interrupted || ctx->is_work_done)
+ return;
+ }
+ break;
+ case COMPLETE_SIGNAL:
+ case NORMAL_RESPONSE:
+ *ptr_interrupted = fastrpc_wait_for_response(ctx, kernel);
+ if (*ptr_interrupted || ctx->is_work_done)
+ return;
+ break;
+ default:
+ *ptr_interrupted = -EBADR;
+ dev_err(ctx->fl->sctx->dev, "unsupported response type:0x%x\n", ctx->rsp_flags);
+ break;
+ }
+ } while (!ctx->is_work_done);
+}
+
static void fastrpc_update_invoke_count(u32 handle, u64 *perf_counter,
struct timespec64 *invoket)
{
@@ -1325,7 +1482,7 @@ static int fastrpc_internal_invoke(struct fastrpc_user *fl, u32 kernel,
struct fastrpc_invoke *inv = &invoke->inv;
u32 handle, sc;
u64 *perf_counter = NULL;
- int err = 0, perferr = 0;
+ int err = 0, perferr = 0, interrupted = 0;
struct timespec64 invoket = {0};
if (fl->profile)
@@ -1374,15 +1531,18 @@ static int fastrpc_internal_invoke(struct fastrpc_user *fl, u32 kernel,
PERF_END);
wait:
- if (kernel) {
- if (!wait_for_completion_timeout(&ctx->work, 10 * HZ))
- err = -ETIMEDOUT;
- } else {
- err = wait_for_completion_interruptible(&ctx->work);
+ fastrpc_wait_for_completion(ctx, &interrupted, kernel);
+ if (interrupted != 0) {
+ err = interrupted;
+ goto bail;
}
- if (err)
+ if (!ctx->is_work_done) {
+ err = -ETIMEDOUT;
+ dev_err(fl->sctx->dev, "Error: Invalid workdone state for handle 0x%x, sc 0x%x\n",
+ handle, sc);
goto bail;
+ }
/* make sure that all memory writes by DSP are seen by CPU */
dma_rmb();
@@ -2056,6 +2216,36 @@ static int fastrpc_get_info_from_kernel(struct fastrpc_ioctl_capability *cap,
return 0;
}
+static int fastrpc_send_cpuinfo_to_dsp(struct fastrpc_user *fl)
+{
+ int err = 0;
+ u64 cpuinfo = 0;
+ struct fastrpc_invoke_args args[1];
+ struct fastrpc_enhanced_invoke ioctl;
+
+ if (!fl)
+ return -EBADF;
+
+ cpuinfo = fl->cctx->cpuinfo_todsp;
+ /* return success if already updated to remote processor */
+ if (fl->cctx->cpuinfo_status)
+ return 0;
+
+ args[0].ptr = (u64)(uintptr_t)&cpuinfo;
+ args[0].length = sizeof(cpuinfo);
+ args[0].fd = -1;
+
+ ioctl.inv.handle = FASTRPC_DSP_UTILITIES_HANDLE;
+ ioctl.inv.sc = FASTRPC_SCALARS(1, 1, 0);
+ ioctl.inv.args = (__u64)args;
+
+ err = fastrpc_internal_invoke(fl, true, &ioctl);
+ if (!err)
+ fl->cctx->cpuinfo_status = true;
+
+ return err;
+}
+
static int fastrpc_get_dsp_info(struct fastrpc_user *fl, char __user *argp)
{
struct fastrpc_ioctl_capability cap = {0};
@@ -2407,6 +2597,8 @@ static long fastrpc_device_ioctl(struct file *file, unsigned int cmd,
break;
case FASTRPC_IOCTL_INIT_ATTACH:
err = fastrpc_init_attach(fl, ROOT_PD);
+ if (!err)
+ fastrpc_send_cpuinfo_to_dsp(fl);
break;
case FASTRPC_IOCTL_INIT_ATTACH_SNS:
err = fastrpc_init_attach(fl, SENSORS_PD);
@@ -2627,6 +2819,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
err = fastrpc_device_register(rdev, data, secure_dsp, domains[domain_id]);
if (err)
goto fdev_error;
+ data->cpuinfo_todsp = FASTRPC_CPUINFO_DEFAULT;
break;
case CDSP_DOMAIN_ID:
data->unsigned_support = true;
@@ -2638,6 +2831,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
err = fastrpc_device_register(rdev, data, false, domains[domain_id]);
if (err)
goto fdev_error;
+ data->cpuinfo_todsp = FASTRPC_CPUINFO_EARLY_WAKEUP;
break;
default:
err = -EINVAL;
@@ -2680,10 +2874,12 @@ static void fastrpc_notify_users(struct fastrpc_user *user)
spin_lock(&user->lock);
list_for_each_entry(ctx, &user->pending, node) {
ctx->retval = -EPIPE;
+ ctx->is_work_done = true;
complete(&ctx->work);
}
list_for_each_entry(ctx, &user->interrupted, node) {
ctx->retval = -EPIPE;
+ ctx->is_work_done = true;
complete(&ctx->work);
}
spin_unlock(&user->lock);
@@ -2720,31 +2916,74 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
fastrpc_channel_ctx_put(cctx);
}
+static void fastrpc_notify_user_ctx(struct fastrpc_invoke_ctx *ctx, int retval,
+ u32 rsp_flags, u32 early_wake_time)
+{
+ ctx->retval = retval;
+ ctx->rsp_flags = (enum fastrpc_response_flags)rsp_flags;
+ switch (rsp_flags) {
+ case NORMAL_RESPONSE:
+ case COMPLETE_SIGNAL:
+ /* normal and complete response with return value */
+ ctx->is_work_done = true;
+ complete(&ctx->work);
+ break;
+ case USER_EARLY_SIGNAL:
+ /* user hint of approximate time of completion */
+ ctx->early_wake_time = early_wake_time;
+ break;
+ case EARLY_RESPONSE:
+ /* rpc framework early response with return value */
+ complete(&ctx->work);
+ break;
+ default:
+ break;
+ }
+}
+
static int fastrpc_rpmsg_callback(struct rpmsg_device *rpdev, void *data,
int len, void *priv, u32 addr)
{
struct fastrpc_channel_ctx *cctx = dev_get_drvdata(&rpdev->dev);
struct fastrpc_invoke_rsp *rsp = data;
+ struct fastrpc_invoke_rspv2 *rspv2 = NULL;
struct fastrpc_invoke_ctx *ctx;
unsigned long flags;
unsigned long ctxid;
+ u32 rsp_flags = 0;
+ u32 early_wake_time = 0;
if (len < sizeof(*rsp))
return -EINVAL;
+ if (len >= sizeof(*rspv2)) {
+ rspv2 = data;
+ if (rspv2) {
+ early_wake_time = rspv2->early_wake_time;
+ rsp_flags = rspv2->flags;
+ }
+ }
+
ctxid = ((rsp->ctx & FASTRPC_CTXID_MASK) >> 4);
spin_lock_irqsave(&cctx->lock, flags);
ctx = idr_find(&cctx->ctx_idr, ctxid);
- spin_unlock_irqrestore(&cctx->lock, flags);
if (!ctx) {
- dev_err(&rpdev->dev, "No context ID matches response\n");
- return -ENOENT;
+ dev_info(&cctx->rpdev->dev, "Warning: No context ID matches response\n");
+ spin_unlock_irqrestore(&cctx->lock, flags);
+ return 0;
}
- ctx->retval = rsp->retval;
- complete(&ctx->work);
+ if (rspv2) {
+ if (rspv2->version != FASTRPC_RSP_VERSION2) {
+ dev_err(&cctx->rpdev->dev, "Incorrect response version %d\n", rspv2->version);
+ spin_unlock_irqrestore(&cctx->lock, flags);
+ return -EINVAL;
+ }
+ }
+ fastrpc_notify_user_ctx(ctx, rsp->retval, rsp_flags, early_wake_time);
+ spin_unlock_irqrestore(&cctx->lock, flags);
/*
* The DMA buffer associated with the context cannot be freed in
--
2.17.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH v3 2/5] misc: fastrpc: Add polling mode support for fastRPC driver
2023-11-21 11:24 [PATCH v3 0/5] Add missing features to FastRPC driver Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver Ekansh Gupta
@ 2023-11-21 11:24 ` Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 3/5] misc: fastrpc: Add DSP PD notification support Ekansh Gupta
` (2 subsequent siblings)
4 siblings, 0 replies; 7+ messages in thread
From: Ekansh Gupta @ 2023-11-21 11:24 UTC (permalink / raw
To: srinivas.kandagatla, linux-arm-msm; +Cc: gregkh, linux-kernel
For any remote call to DSP, after sending an invocation message,
fastRPC driver waits for glink response and during this time the
CPU can go into low power modes. Adding a polling mode support
with which fastRPC driver will poll continuously on a memory
after sending a message to remote subsystem which will eliminate
CPU wakeup and scheduling latencies and reduce fastRPC overhead.
With this change, DSP always sends a glink response which will
get ignored if polling mode didn't time out.
Signed-off-by: Ekansh Gupta <quic_ekangupt@quicinc.com>
---
Changes in v3:
- Rebase the patch to latest kernel version
drivers/misc/fastrpc.c | 49 ++++++++++++++++++++++++++++++++++++++++++
1 file changed, 49 insertions(+)
diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index 6b6ac3e3328d..4f8ecfcdf89f 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -109,6 +109,8 @@
#define FASTRPC_RSP_VERSION2 2
/* Early wake up poll completion number received from remoteproc */
#define FASTRPC_EARLY_WAKEUP_POLL (0xabbccdde)
+/* Poll response number from remote processor for call completion */
+#define FASTRPC_POLL_RESPONSE (0xdecaf)
/* timeout in us for polling until memory barrier */
#define FASTRPC_POLL_TIME_MEM_UPDATE (500)
/* timeout in us for busy polling after early response from remoteproc */
@@ -380,10 +382,14 @@ struct fastrpc_user {
struct fastrpc_buf *init_mem;
u32 profile;
+ /* Threads poll for specified timeout and fall back to glink wait */
+ u32 poll_timeout;
int tgid;
int pd;
bool is_secure_dev;
bool sharedcb;
+ /* If set, threads will poll for DSP response instead of glink wait */
+ bool poll_mode;
/* Lock for lists */
spinlock_t lock;
/* lock for allocations */
@@ -1376,6 +1382,11 @@ static int poll_for_remote_response(struct fastrpc_invoke_ctx *ctx, u32 timeout)
/* Remote processor sent early response */
err = 0;
break;
+ } else if (*poll == FASTRPC_POLL_RESPONSE) {
+ err = 0;
+ ctx->is_work_done = true;
+ ctx->retval = 0;
+ break;
}
if (jj == FASTRPC_POLL_TIME_MEM_UPDATE) {
/* Wait for DSP to finish updating poll memory */
@@ -1451,6 +1462,15 @@ static void fastrpc_wait_for_completion(struct fastrpc_invoke_ctx *ctx,
if (*ptr_interrupted || ctx->is_work_done)
return;
break;
+ case POLL_MODE:
+ err = poll_for_remote_response(ctx, ctx->fl->poll_timeout);
+
+ /* If polling timed out, move to normal response state */
+ if (err)
+ ctx->rsp_flags = NORMAL_RESPONSE;
+ else
+ *ptr_interrupted = 0;
+ break;
default:
*ptr_interrupted = -EBADR;
dev_err(ctx->fl->sctx->dev, "unsupported response type:0x%x\n", ctx->rsp_flags);
@@ -2065,6 +2085,32 @@ static int fastrpc_invoke(struct fastrpc_user *fl, char __user *argp)
return err;
}
+static int fastrpc_manage_poll_mode(struct fastrpc_user *fl, u32 enable, u32 timeout)
+{
+ const unsigned int MAX_POLL_TIMEOUT_US = 10000;
+
+ if ((fl->cctx->domain_id != CDSP_DOMAIN_ID) || (fl->pd != USER_PD)) {
+ dev_err(&fl->cctx->rpdev->dev, "poll mode only allowed for dynamic CDSP process\n");
+ return -EPERM;
+ }
+ if (timeout > MAX_POLL_TIMEOUT_US) {
+ dev_err(&fl->cctx->rpdev->dev, "poll timeout %u is greater than max allowed value %u\n",
+ timeout, MAX_POLL_TIMEOUT_US);
+ return -EBADMSG;
+ }
+ spin_lock(&fl->lock);
+ if (enable) {
+ fl->poll_mode = true;
+ fl->poll_timeout = timeout;
+ } else {
+ fl->poll_mode = false;
+ fl->poll_timeout = 0;
+ }
+ spin_unlock(&fl->lock);
+ dev_info(&fl->cctx->rpdev->dev, "updated poll mode to %d, timeout %u\n", enable, timeout);
+ return 0;
+}
+
static int fastrpc_internal_control(struct fastrpc_user *fl,
struct fastrpc_internal_control *cp)
{
@@ -2079,6 +2125,9 @@ static int fastrpc_internal_control(struct fastrpc_user *fl,
case FASTRPC_CONTROL_SMMU:
fl->sharedcb = cp->smmu.sharedcb;
break;
+ case FASTRPC_CONTROL_RPC_POLL:
+ err = fastrpc_manage_poll_mode(fl, cp->lp.enable, cp->lp.latency);
+ break;
default:
err = -EBADRQC;
break;
--
2.17.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH v3 3/5] misc: fastrpc: Add DSP PD notification support
2023-11-21 11:24 [PATCH v3 0/5] Add missing features to FastRPC driver Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 2/5] misc: fastrpc: Add polling mode " Ekansh Gupta
@ 2023-11-21 11:24 ` Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 4/5] misc: fastrpc: Add support for users to clean up DSP user PD Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 5/5] misc: fastrpc: Add wakelock management support Ekansh Gupta
4 siblings, 0 replies; 7+ messages in thread
From: Ekansh Gupta @ 2023-11-21 11:24 UTC (permalink / raw
To: srinivas.kandagatla, linux-arm-msm; +Cc: gregkh, linux-kernel
Current driver design does not provide any notification regarding
the status of used PD on DSP. Only when user makes a FastRPC
invocation, they get to know if the process has been killed on
DSP. Notifying status of user PD can help users to restart the
DSP PD session.
Co-developed-by: Anandu Krishnan E <quic_anane@quicinc.com>
Signed-off-by: Anandu Krishnan E <quic_anane@quicinc.com>
Signed-off-by: Ekansh Gupta <quic_ekangupt@quicinc.com>
---
Changes in v2:
- Added Co-developer tag
Changes in v3:
- Rebase the patch to latest kernel version
drivers/misc/fastrpc.c | 145 +++++++++++++++++++++++++++++++++++-
include/uapi/misc/fastrpc.h | 8 ++
2 files changed, 152 insertions(+), 1 deletion(-)
diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index 4f8ecfcdf89f..72ed14174363 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -120,6 +120,8 @@
/* CPU feature information to DSP */
#define FASTRPC_CPUINFO_DEFAULT (0)
#define FASTRPC_CPUINFO_EARLY_WAKEUP (1)
+/* Process status notifications from DSP will be sent with this unique context */
+#define FASTRPC_NOTIF_CTX_RESERVED 0xABCDABCD
#define miscdev_to_fdevice(d) container_of(d, struct fastrpc_device, miscdev)
@@ -238,6 +240,13 @@ struct fastrpc_invoke_rspv2 {
u32 version; /* version number */
};
+struct dsp_notif_rsp {
+ u64 ctx; /* response context */
+ u32 type; /* Notification type */
+ int pid; /* user process pid */
+ u32 status; /* userpd status notification */
+};
+
struct fastrpc_buf_overlap {
u64 start;
u64 end;
@@ -297,6 +306,27 @@ struct fastrpc_perf {
u64 tid;
};
+struct fastrpc_notif_queue {
+ /* Number of pending status notifications in queue */
+ atomic_t notif_queue_count;
+ /* Wait queue to synchronize notifier thread and response */
+ wait_queue_head_t notif_wait_queue;
+ /* IRQ safe spin lock for protecting notif queue */
+ spinlock_t nqlock;
+};
+
+struct fastrpc_internal_notif_rsp {
+ u32 domain; /* Domain of User PD */
+ u32 session; /* Session ID of User PD */
+ u32 status; /* Status of the process */
+};
+
+struct fastrpc_notif_rsp {
+ struct list_head notifn;
+ u32 domain;
+ enum fastrpc_status_flags status;
+};
+
struct fastrpc_invoke_ctx {
int nscalars;
int nbufs;
@@ -376,10 +406,13 @@ struct fastrpc_user {
struct list_head pending;
struct list_head interrupted;
struct list_head mmaps;
+ struct list_head notif_queue;
struct fastrpc_channel_ctx *cctx;
struct fastrpc_session_ctx *sctx;
struct fastrpc_buf *init_mem;
+ /* Process status notification queue */
+ struct fastrpc_notif_queue proc_state_notif;
u32 profile;
/* Threads poll for specified timeout and fall back to glink wait */
@@ -2085,6 +2118,99 @@ static int fastrpc_invoke(struct fastrpc_user *fl, char __user *argp)
return err;
}
+static void fastrpc_queue_pd_status(struct fastrpc_user *fl, int domain, int status)
+{
+ struct fastrpc_notif_rsp *notif_rsp = NULL;
+ unsigned long flags;
+
+ notif_rsp = kzalloc(sizeof(*notif_rsp), GFP_ATOMIC);
+ if (!notif_rsp)
+ return;
+
+ notif_rsp->status = status;
+ notif_rsp->domain = domain;
+
+ spin_lock_irqsave(&fl->proc_state_notif.nqlock, flags);
+ list_add_tail(¬if_rsp->notifn, &fl->notif_queue);
+ atomic_add(1, &fl->proc_state_notif.notif_queue_count);
+ wake_up_interruptible(&fl->proc_state_notif.notif_wait_queue);
+ spin_unlock_irqrestore(&fl->proc_state_notif.nqlock, flags);
+}
+
+static void fastrpc_notif_find_process(int domain, struct fastrpc_channel_ctx *cctx, struct dsp_notif_rsp *notif)
+{
+ bool is_process_found = false;
+ unsigned long irq_flags = 0;
+ struct fastrpc_user *user;
+
+ spin_lock_irqsave(&cctx->lock, irq_flags);
+ list_for_each_entry(user, &cctx->users, user) {
+ if (user->tgid == notif->pid) {
+ is_process_found = true;
+ break;
+ }
+ }
+ spin_unlock_irqrestore(&cctx->lock, irq_flags);
+
+ if (!is_process_found)
+ return;
+ fastrpc_queue_pd_status(user, domain, notif->status);
+}
+
+static int fastrpc_wait_on_notif_queue(
+ struct fastrpc_internal_notif_rsp *notif_rsp,
+ struct fastrpc_user *fl)
+{
+ int err = 0;
+ unsigned long flags;
+ struct fastrpc_notif_rsp *notif, *inotif, *n;
+
+read_notif_status:
+ err = wait_event_interruptible(fl->proc_state_notif.notif_wait_queue,
+ atomic_read(&fl->proc_state_notif.notif_queue_count));
+ if (err) {
+ kfree(notif);
+ return err;
+ }
+
+ spin_lock_irqsave(&fl->proc_state_notif.nqlock, flags);
+ list_for_each_entry_safe(inotif, n, &fl->notif_queue, notifn) {
+ list_del(&inotif->notifn);
+ atomic_sub(1, &fl->proc_state_notif.notif_queue_count);
+ notif = inotif;
+ break;
+ }
+ spin_unlock_irqrestore(&fl->proc_state_notif.nqlock, flags);
+
+ if (notif) {
+ notif_rsp->status = notif->status;
+ notif_rsp->domain = notif->domain;
+ } else {// Go back to wait if ctx is invalid
+ dev_err(fl->sctx->dev, "Invalid status notification response\n");
+ goto read_notif_status;
+ }
+
+ kfree(notif);
+ return err;
+}
+
+static int fastrpc_get_notif_response(
+ struct fastrpc_internal_notif_rsp *notif,
+ void *param, struct fastrpc_user *fl)
+{
+ int err = 0;
+
+ err = fastrpc_wait_on_notif_queue(notif, fl);
+ if (err)
+ return err;
+
+ if (copy_to_user((void __user *)param, notif,
+ sizeof(struct fastrpc_internal_notif_rsp)))
+ return -EFAULT;
+
+ return 0;
+}
+
static int fastrpc_manage_poll_mode(struct fastrpc_user *fl, u32 enable, u32 timeout)
{
const unsigned int MAX_POLL_TIMEOUT_US = 10000;
@@ -2141,6 +2267,7 @@ static int fastrpc_multimode_invoke(struct fastrpc_user *fl, char __user *argp)
struct fastrpc_invoke_args *args = NULL;
struct fastrpc_ioctl_multimode_invoke invoke;
struct fastrpc_internal_control cp = {0};
+ struct fastrpc_internal_notif_rsp notif;
u32 nscalars;
u64 *perf_kernel;
int err, i;
@@ -2190,6 +2317,10 @@ static int fastrpc_multimode_invoke(struct fastrpc_user *fl, char __user *argp)
err = fastrpc_internal_control(fl, &cp);
break;
+ case FASTRPC_INVOKE_NOTIF:
+ err = fastrpc_get_notif_response(¬if,
+ (void *)invoke.invparam, fl);
+ break;
default:
err = -ENOTTY;
break;
@@ -2944,8 +3075,10 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
/* No invocations past this point */
spin_lock_irqsave(&cctx->lock, flags);
cctx->rpdev = NULL;
- list_for_each_entry(user, &cctx->users, user)
+ list_for_each_entry(user, &cctx->users, user) {
+ fastrpc_queue_pd_status(user, cctx->domain_id, FASTRPC_DSP_SSR);
fastrpc_notify_users(user);
+ }
spin_unlock_irqrestore(&cctx->lock, flags);
if (cctx->fdevice)
@@ -2996,12 +3129,22 @@ static int fastrpc_rpmsg_callback(struct rpmsg_device *rpdev, void *data,
struct fastrpc_channel_ctx *cctx = dev_get_drvdata(&rpdev->dev);
struct fastrpc_invoke_rsp *rsp = data;
struct fastrpc_invoke_rspv2 *rspv2 = NULL;
+ struct dsp_notif_rsp *notif = (struct dsp_notif_rsp *)data;
struct fastrpc_invoke_ctx *ctx;
unsigned long flags;
unsigned long ctxid;
u32 rsp_flags = 0;
u32 early_wake_time = 0;
+ if (notif->ctx == FASTRPC_NOTIF_CTX_RESERVED) {
+ if (notif->type == STATUS_RESPONSE && len >= sizeof(*notif)) {
+ fastrpc_notif_find_process(cctx->domain_id, cctx, notif);
+ return 0;
+ } else {
+ return -ENOENT;
+ }
+ }
+
if (len < sizeof(*rsp))
return -EINVAL;
diff --git a/include/uapi/misc/fastrpc.h b/include/uapi/misc/fastrpc.h
index 3dfd8e95eda8..1f544a35ee4e 100644
--- a/include/uapi/misc/fastrpc.h
+++ b/include/uapi/misc/fastrpc.h
@@ -192,4 +192,12 @@ enum fastrpc_perfkeys {
PERF_KEY_MAX = 10,
};
+enum fastrpc_status_flags {
+ FASTRPC_USERPD_UP = 0,
+ FASTRPC_USERPD_EXIT = 1,
+ FASTRPC_USERPD_FORCE_KILL = 2,
+ FASTRPC_USERPD_EXCEPTION = 3,
+ FASTRPC_DSP_SSR = 4,
+};
+
#endif /* __QCOM_FASTRPC_H__ */
--
2.17.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH v3 4/5] misc: fastrpc: Add support for users to clean up DSP user PD
2023-11-21 11:24 [PATCH v3 0/5] Add missing features to FastRPC driver Ekansh Gupta
` (2 preceding siblings ...)
2023-11-21 11:24 ` [PATCH v3 3/5] misc: fastrpc: Add DSP PD notification support Ekansh Gupta
@ 2023-11-21 11:24 ` Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 5/5] misc: fastrpc: Add wakelock management support Ekansh Gupta
4 siblings, 0 replies; 7+ messages in thread
From: Ekansh Gupta @ 2023-11-21 11:24 UTC (permalink / raw
To: srinivas.kandagatla, linux-arm-msm; +Cc: gregkh, linux-kernel
Add a control mechanism for users to clean up DSP user PD. This
method can be used by users for handling any unexpected hang
scenarios on DSP PD. User can clean up DSP PD and restart the
user PD again.
Signed-off-by: Ekansh Gupta <quic_ekangupt@quicinc.com>
---
Changes in v3:
- Rebase the patch to latest kernel version
drivers/misc/fastrpc.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index 72ed14174363..97895fce1228 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -2254,6 +2254,11 @@ static int fastrpc_internal_control(struct fastrpc_user *fl,
case FASTRPC_CONTROL_RPC_POLL:
err = fastrpc_manage_poll_mode(fl, cp->lp.enable, cp->lp.latency);
break;
+ case FASTRPC_CONTROL_DSPPROCESS_CLEAN:
+ err = fastrpc_release_current_dsp_process(fl);
+ if (!err)
+ fastrpc_queue_pd_status(fl, fl->cctx->domain_id, FASTRPC_USERPD_FORCE_KILL);
+ break;
default:
err = -EBADRQC;
break;
--
2.17.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH v3 5/5] misc: fastrpc: Add wakelock management support
2023-11-21 11:24 [PATCH v3 0/5] Add missing features to FastRPC driver Ekansh Gupta
` (3 preceding siblings ...)
2023-11-21 11:24 ` [PATCH v3 4/5] misc: fastrpc: Add support for users to clean up DSP user PD Ekansh Gupta
@ 2023-11-21 11:24 ` Ekansh Gupta
4 siblings, 0 replies; 7+ messages in thread
From: Ekansh Gupta @ 2023-11-21 11:24 UTC (permalink / raw
To: srinivas.kandagatla, linux-arm-msm; +Cc: gregkh, linux-kernel
CPU can go into suspend mode at anytime. Users might have a
requirement to keep the CPU awake until they get a response for
their remote call to DSP. Wakelock management can be used to
achieve this requirement.
Co-developed-by: Anandu Krishnan E <quic_anane@quicinc.com>
Signed-off-by: Anandu Krishnan E <quic_anane@quicinc.com>
Signed-off-by: Ekansh Gupta <quic_ekangupt@quicinc.com>
---
drivers/misc/fastrpc.c | 104 +++++++++++++++++++++++++++++++++++++++++
1 file changed, 104 insertions(+)
diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index 97895fce1228..dcda287cd899 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -120,6 +120,11 @@
/* CPU feature information to DSP */
#define FASTRPC_CPUINFO_DEFAULT (0)
#define FASTRPC_CPUINFO_EARLY_WAKEUP (1)
+
+/* Maximum PM timeout that can be voted through fastrpc */
+#define FASTRPC_MAX_PM_TIMEOUT_MS 50
+#define FASTRPC_NON_SECURE_WAKE_SOURCE_CLIENT_NAME "fastrpc-non_secure"
+#define FASTRPC_SECURE_WAKE_SOURCE_CLIENT_NAME "fastrpc-secure"
/* Process status notifications from DSP will be sent with this unique context */
#define FASTRPC_NOTIF_CTX_RESERVED 0xABCDABCD
@@ -387,6 +392,10 @@ struct fastrpc_channel_ctx {
struct fastrpc_device *fdevice;
struct fastrpc_buf *remote_heap;
struct list_head invoke_interrupted_mmaps;
+ /* Secure subsystems like ADSP/SLPI will use secure client */
+ struct wakeup_source *wake_source_secure;
+ /* Non-secure subsystem like CDSP will use regular client */
+ struct wakeup_source *wake_source;
bool secure;
bool unsigned_support;
bool cpuinfo_status;
@@ -417,6 +426,8 @@ struct fastrpc_user {
u32 profile;
/* Threads poll for specified timeout and fall back to glink wait */
u32 poll_timeout;
+ u32 ws_timeout;
+ u32 wake_enable;
int tgid;
int pd;
bool is_secure_dev;
@@ -438,11 +449,21 @@ struct fastrpc_ctrl_smmu {
u32 sharedcb; /* Set to SMMU share context bank */
};
+struct fastrpc_ctrl_wakelock {
+ u32 enable; /* wakelock control enable */
+};
+
+struct fastrpc_ctrl_pm {
+ u32 timeout; /* timeout(in ms) for PM to keep system awake */
+};
+
struct fastrpc_internal_control {
u32 req;
union {
struct fastrpc_ctrl_latency lp;
struct fastrpc_ctrl_smmu smmu;
+ struct fastrpc_ctrl_wakelock wp;
+ struct fastrpc_ctrl_pm pm;
};
};
@@ -980,6 +1001,43 @@ static void fastrpc_session_free(struct fastrpc_channel_ctx *cctx,
spin_unlock_irqrestore(&cctx->lock, flags);
}
+static void fastrpc_pm_awake(struct fastrpc_user *fl,
+ u32 is_secure_channel)
+{
+ struct fastrpc_channel_ctx *cctx = fl->cctx;
+ struct wakeup_source *wake_source = NULL;
+
+ /*
+ * Vote with PM to abort any suspend in progress and
+ * keep system awake for specified timeout
+ */
+ if (is_secure_channel)
+ wake_source = cctx->wake_source_secure;
+ else
+ wake_source = cctx->wake_source;
+
+ if (wake_source)
+ pm_wakeup_ws_event(wake_source, fl->ws_timeout, true);
+}
+
+static void fastrpc_pm_relax(struct fastrpc_user *fl,
+ u32 is_secure_channel)
+{
+ struct fastrpc_channel_ctx *cctx = fl->cctx;
+ struct wakeup_source *wake_source = NULL;
+
+ if (!fl->wake_enable)
+ return;
+
+ if (is_secure_channel)
+ wake_source = cctx->wake_source_secure;
+ else
+ wake_source = cctx->wake_source;
+
+ if (wake_source)
+ __pm_relax(wake_source);
+}
+
static int fastrpc_map_create(struct fastrpc_user *fl, int fd,
u64 len, u32 attr, struct fastrpc_map **ppmap)
{
@@ -1971,6 +2029,7 @@ static int fastrpc_device_release(struct inode *inode, struct file *file)
fastrpc_buf_free(buf);
}
+ fastrpc_pm_relax(fl, cctx->secure);
fastrpc_session_free(cctx, fl->sctx);
fastrpc_channel_ctx_put(cctx);
@@ -2251,6 +2310,24 @@ static int fastrpc_internal_control(struct fastrpc_user *fl,
case FASTRPC_CONTROL_SMMU:
fl->sharedcb = cp->smmu.sharedcb;
break;
+ case FASTRPC_CONTROL_WAKELOCK:
+ if (!fl->is_secure_dev) {
+ dev_err(&fl->cctx->rpdev->dev,
+ "PM voting not allowed for non-secure device node");
+ err = -EPERM;
+ return err;
+ }
+ fl->wake_enable = cp->wp.enable;
+ break;
+ case FASTRPC_CONTROL_PM:
+ if (!fl->wake_enable)
+ return -EACCES;
+ if (cp->pm.timeout > FASTRPC_MAX_PM_TIMEOUT_MS)
+ fl->ws_timeout = FASTRPC_MAX_PM_TIMEOUT_MS;
+ else
+ fl->ws_timeout = cp->pm.timeout;
+ fastrpc_pm_awake(fl, fl->cctx->secure);
+ break;
case FASTRPC_CONTROL_RPC_POLL:
err = fastrpc_manage_poll_mode(fl, cp->lp.enable, cp->lp.latency);
break;
@@ -2942,6 +3019,21 @@ static int fastrpc_device_register(struct device *dev, struct fastrpc_channel_ct
return err;
}
+static void fastrpc_register_wakeup_source(struct device *dev,
+ const char *client_name, struct wakeup_source **device_wake_source)
+{
+ struct wakeup_source *wake_source = NULL;
+
+ wake_source = wakeup_source_register(dev, client_name);
+ if (IS_ERR_OR_NULL(wake_source)) {
+ dev_err(dev, "wakeup_source_register failed for dev %s, client %s with err %ld\n",
+ dev_name(dev), client_name, PTR_ERR(wake_source));
+ return;
+ }
+
+ *device_wake_source = wake_source;
+}
+
static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
{
struct device *rdev = &rpdev->dev;
@@ -3023,6 +3115,13 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
goto fdev_error;
}
+ if (data->fdevice)
+ fastrpc_register_wakeup_source(data->fdevice->miscdev.this_device,
+ FASTRPC_NON_SECURE_WAKE_SOURCE_CLIENT_NAME, &data->wake_source);
+ if (data->secure_fdevice)
+ fastrpc_register_wakeup_source(data->secure_fdevice->miscdev.this_device,
+ FASTRPC_SECURE_WAKE_SOURCE_CLIENT_NAME, &data->wake_source_secure);
+
kref_init(&data->refcount);
dev_set_drvdata(&rpdev->dev, data);
@@ -3098,6 +3197,11 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
if (cctx->remote_heap)
fastrpc_buf_free(cctx->remote_heap);
+ if (cctx->wake_source)
+ wakeup_source_unregister(cctx->wake_source);
+ if (cctx->wake_source_secure)
+ wakeup_source_unregister(cctx->wake_source_secure);
+
of_platform_depopulate(&rpdev->dev);
fastrpc_channel_ctx_put(cctx);
--
2.17.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver
2023-11-21 11:24 ` [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver Ekansh Gupta
@ 2023-11-24 13:27 ` Srinivas Kandagatla
0 siblings, 0 replies; 7+ messages in thread
From: Srinivas Kandagatla @ 2023-11-24 13:27 UTC (permalink / raw
To: Ekansh Gupta, linux-arm-msm; +Cc: gregkh, linux-kernel
Thanks Ekansh for the patch,
On 21/11/2023 11:24, Ekansh Gupta wrote:
> CPU wake up and context switch latency are significant in FastRPC
> overhead for remote calls. As part of this change, DSP sends early
> signal of completion to CPU and FastRPC driver detects early signal
> on the given context and starts polling on a memory for actual
> completion. Multiple different response flags are added to support
> DSP user early hint of approximate time of completion, early response
> from DSP user to wake up CPU and poll on memory for actual completion.
> Complete signal is also added which is sent by DSP user in case of
> timeout after early response is sent.
>
> Signed-off-by: Ekansh Gupta <quic_ekangupt@quicinc.com>
> ---
> Changes in v3:
> - Rebase the patch to latest kernel version
>
> drivers/misc/fastrpc.c | 265 +++++++++++++++++++++++++++++++++++++++--
> 1 file changed, 252 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
> index 8e77beb3a693..6b6ac3e3328d 100644
> --- a/drivers/misc/fastrpc.c
> +++ b/drivers/misc/fastrpc.c
> @@ -106,6 +106,19 @@
> #define USER_PD (1)
> #define SENSORS_PD (2)
>
> +#define FASTRPC_RSP_VERSION2 2
> +/* Early wake up poll completion number received from remoteproc */
> +#define FASTRPC_EARLY_WAKEUP_POLL (0xabbccdde)
> +/* timeout in us for polling until memory barrier */
> +#define FASTRPC_POLL_TIME_MEM_UPDATE (500)
> +/* timeout in us for busy polling after early response from remoteproc */
> +#define FASTRPC_POLL_TIME (4000)
> +/* timeout in us for polling completion signal after user early hint */
> +#define FASTRPC_USER_EARLY_HINT_TIMEOUT (500)
> +/* CPU feature information to DSP */
> +#define FASTRPC_CPUINFO_DEFAULT (0)
> +#define FASTRPC_CPUINFO_EARLY_WAKEUP (1)
> +
> #define miscdev_to_fdevice(d) container_of(d, struct fastrpc_device, miscdev)
>
> #define PERF_END ((void)0)
> @@ -129,6 +142,15 @@
> (uint64_t *)(perf_ptr + offset)\
> : (uint64_t *)NULL) : (uint64_t *)NULL)
>
Can you add some documentation to these flags.
> +enum fastrpc_response_flags {
> + NORMAL_RESPONSE = 0,
> + EARLY_RESPONSE = 1,
> + USER_EARLY_SIGNAL = 2,
> + COMPLETE_SIGNAL = 3,
> + STATUS_RESPONSE = 4,
> + POLL_MODE = 5,
> +};
> +
> static const char *domains[FASTRPC_DEV_MAX] = { "adsp", "mdsp",
> "sdsp", "cdsp"};
> struct fastrpc_phy_page {
> @@ -206,6 +228,14 @@ struct fastrpc_invoke_rsp {
> int retval; /* invoke return value */
> };
>
> +struct fastrpc_invoke_rspv2 {
> + u64 ctx; /* invoke caller context */
> + int retval; /* invoke return value */
> + u32 flags; /* early response flags */
> + u32 early_wake_time; /* user hint in us */
> + u32 version; /* version number */
> +};
> +
> struct fastrpc_buf_overlap {
> u64 start;
> u64 end;
> @@ -272,11 +302,17 @@ struct fastrpc_invoke_ctx {
> int pid;
> int tgid;
> u32 sc;
> + /* user hint of completion time in us */
> + u32 early_wake_time;
> u32 *crc;
> u64 *perf_kernel;
> u64 *perf_dsp;
> u64 ctxid;
> u64 msg_sz;
> + /* work done status flag */
> + bool is_work_done;
> + /* response flags from remote processor */
> + enum fastrpc_response_flags rsp_flags;
> struct kref refcount;
> struct list_head node; /* list of ctxs */
> struct completion work;
> @@ -321,7 +357,9 @@ struct fastrpc_channel_ctx {
> struct list_head invoke_interrupted_mmaps;
> bool secure;
> bool unsigned_support;
> + bool cpuinfo_status;
> u64 dma_mask;
> + u64 cpuinfo_todsp;
> };
>
> struct fastrpc_device {
> @@ -352,13 +390,21 @@ struct fastrpc_user {
> struct mutex mutex;
> };
>
> +struct fastrpc_ctrl_latency {
> + u32 enable; /* latency control enable */
> + u32 latency; /* latency request in us */
> +};
> +
What is this struct? am not seeing it used anywhere in this patch.
> struct fastrpc_ctrl_smmu {
> u32 sharedcb; /* Set to SMMU share context bank */
> };
>
> struct fastrpc_internal_control {
> u32 req;
> - struct fastrpc_ctrl_smmu smmu;
> + union {
> + struct fastrpc_ctrl_latency lp;
> + struct fastrpc_ctrl_smmu smmu;
> + };
> };
>
> static inline int64_t getnstimediff(struct timespec64 *start)
> @@ -692,6 +738,8 @@ static struct fastrpc_invoke_ctx *fastrpc_context_alloc(
> ctx->pid = current->pid;
> ctx->tgid = user->tgid;
> ctx->cctx = cctx;
> + ctx->rsp_flags = NORMAL_RESPONSE;
> + ctx->is_work_done = false;
> init_completion(&ctx->work);
> INIT_WORK(&ctx->put_work, fastrpc_context_put_wq);
>
> @@ -1302,6 +1350,115 @@ static int fastrpc_invoke_send(struct fastrpc_session_ctx *sctx,
>
> }
>
> +static int poll_for_remote_response(struct fastrpc_invoke_ctx *ctx, u32 timeout)
> +{
> + int err = -EIO, ii = 0, jj = 0;
variable naming is really odd.
> + u32 sc = ctx->sc;
> + struct fastrpc_invoke_buf *list;
> + struct fastrpc_phy_page *pages;
> + u64 *fdlist = NULL;
> + u32 *crclist = NULL, *poll = NULL;
> + unsigned int inbufs, outbufs, handles;
> +
> + /* calculate poll memory location */
> + inbufs = REMOTE_SCALARS_INBUFS(sc);
> + outbufs = REMOTE_SCALARS_OUTBUFS(sc);
> + handles = REMOTE_SCALARS_INHANDLES(sc) + REMOTE_SCALARS_OUTHANDLES(sc);
> + list = fastrpc_invoke_buf_start(ctx->rpra, ctx->nscalars);
> + pages = fastrpc_phy_page_start(list, ctx->nscalars);
> + fdlist = (u64 *)(pages + inbufs + outbufs + handles);
> + crclist = (u32 *)(fdlist + FASTRPC_MAX_FDLIST);
> + poll = (u32 *)(crclist + FASTRPC_MAX_CRCLIST);
> +
> + /* poll on memory for DSP response. Return failure on timeout */
> + for (ii = 0, jj = 0; ii < timeout; ii++, jj++) {
> + if (*poll == FASTRPC_EARLY_WAKEUP_POLL) {
> + /* Remote processor sent early response */
> + err = 0;
> + break;
> + }
> + if (jj == FASTRPC_POLL_TIME_MEM_UPDATE) {
> + /* Wait for DSP to finish updating poll memory */
> + rmb();
Could you add a comment on why do we need this memory barrier here?
Can we use dma_rmb() ?
> + jj = 0;
> + }
> + udelay(1);
> + }
> + return err;
> +}
> +
> +static inline int fastrpc_wait_for_response(struct fastrpc_invoke_ctx *ctx,
> + u32 kernel)
> +{
> + int interrupted = 0;
> +
> + if (kernel)
> + wait_for_completion(&ctx->work);
> + else
> + interrupted = wait_for_completion_interruptible(&ctx->work);
> +
> + return interrupted;
> +}
> +
> +static void fastrpc_wait_for_completion(struct fastrpc_invoke_ctx *ctx,
> + int *ptr_interrupted, u32 kernel)
> +{
> + int err = 0, jj = 0;
unnecessary intializations.
> + bool wait_resp = false;
> + u32 wTimeout = FASTRPC_USER_EARLY_HINT_TIMEOUT;
> + u32 wakeTime = ctx->early_wake_time;
> +
> + do {
> + switch (ctx->rsp_flags) {
> + /* try polling on completion with timeout */
> + case USER_EARLY_SIGNAL:
> + /* try wait if completion time is less than timeout */
> + /* disable preempt to avoid context switch latency */
pl use proper multi-line commenting style here.
> + preempt_disable();
> + jj = 0;
> + wait_resp = false;
> + for (; wakeTime < wTimeout && jj < wTimeout; jj++) {
> + wait_resp = try_wait_for_completion(&ctx->work);
> + if (wait_resp)
> + break;
> + udelay(1);
> + }
> + preempt_enable();
> + if (!wait_resp) {
> + *ptr_interrupted = fastrpc_wait_for_response(ctx, kernel);
> + if (*ptr_interrupted || ctx->is_work_done)
> + return;
> + }
> + break;
> + /* busy poll on memory for actual job done */
> + case EARLY_RESPONSE:
> + err = poll_for_remote_response(ctx, FASTRPC_POLL_TIME);
> + /* Mark job done if poll on memory successful */
> + /* Wait for completion if poll on memory timeout */
> + if (!err) {
> + ctx->is_work_done = true;
> + return;
> + }
> + if (!ctx->is_work_done) {
> + *ptr_interrupted = fastrpc_wait_for_response(ctx, kernel);
> + if (*ptr_interrupted || ctx->is_work_done)
> + return;
> + }
> + break;
> + case COMPLETE_SIGNAL:
> + case NORMAL_RESPONSE:
> + *ptr_interrupted = fastrpc_wait_for_response(ctx, kernel);
> + if (*ptr_interrupted || ctx->is_work_done)
> + return;
> + break;
> + default:
> + *ptr_interrupted = -EBADR;
> + dev_err(ctx->fl->sctx->dev, "unsupported response type:0x%x\n", ctx->rsp_flags);
> + break;
> + }
> + } while (!ctx->is_work_done);
> +}
> +
> static void fastrpc_update_invoke_count(u32 handle, u64 *perf_counter,
> struct timespec64 *invoket)
> {
> @@ -1325,7 +1482,7 @@ static int fastrpc_internal_invoke(struct fastrpc_user *fl, u32 kernel,
> struct fastrpc_invoke *inv = &invoke->inv;
> u32 handle, sc;
> u64 *perf_counter = NULL;
> - int err = 0, perferr = 0;
> + int err = 0, perferr = 0, interrupted = 0;
> struct timespec64 invoket = {0};
>
> if (fl->profile)
> @@ -1374,15 +1531,18 @@ static int fastrpc_internal_invoke(struct fastrpc_user *fl, u32 kernel,
> PERF_END);
>
> wait:
> - if (kernel) {
> - if (!wait_for_completion_timeout(&ctx->work, 10 * HZ))
> - err = -ETIMEDOUT;
> - } else {
> - err = wait_for_completion_interruptible(&ctx->work);
> + fastrpc_wait_for_completion(ctx, &interrupted, kernel);
> + if (interrupted != 0) {
> + err = interrupted;
> + goto bail;
> }
>
> - if (err)
> + if (!ctx->is_work_done) {
> + err = -ETIMEDOUT;
> + dev_err(fl->sctx->dev, "Error: Invalid workdone state for handle 0x%x, sc 0x%x\n",
> + handle, sc);
> goto bail;
> + }
>
> /* make sure that all memory writes by DSP are seen by CPU */
> dma_rmb();
> @@ -2056,6 +2216,36 @@ static int fastrpc_get_info_from_kernel(struct fastrpc_ioctl_capability *cap,
> return 0;
> }
>
> +static int fastrpc_send_cpuinfo_to_dsp(struct fastrpc_user *fl)
> +{
> + int err = 0;
> + u64 cpuinfo = 0;
> + struct fastrpc_invoke_args args[1];
> + struct fastrpc_enhanced_invoke ioctl;
> +
> + if (!fl)
> + return -EBADF;
> +
> + cpuinfo = fl->cctx->cpuinfo_todsp;
> + /* return success if already updated to remote processor */
> + if (fl->cctx->cpuinfo_status)
> + return 0;
> +
> + args[0].ptr = (u64)(uintptr_t)&cpuinfo;
> + args[0].length = sizeof(cpuinfo);
> + args[0].fd = -1;
> +
> + ioctl.inv.handle = FASTRPC_DSP_UTILITIES_HANDLE;
> + ioctl.inv.sc = FASTRPC_SCALARS(1, 1, 0);
> + ioctl.inv.args = (__u64)args;
> +
> + err = fastrpc_internal_invoke(fl, true, &ioctl);
> + if (!err)
> + fl->cctx->cpuinfo_status = true;
> +
> + return err;
> +}
> +
> static int fastrpc_get_dsp_info(struct fastrpc_user *fl, char __user *argp)
> {
> struct fastrpc_ioctl_capability cap = {0};
> @@ -2407,6 +2597,8 @@ static long fastrpc_device_ioctl(struct file *file, unsigned int cmd,
> break;
> case FASTRPC_IOCTL_INIT_ATTACH:
> err = fastrpc_init_attach(fl, ROOT_PD);
> + if (!err)
> + fastrpc_send_cpuinfo_to_dsp(fl);
> break;
> case FASTRPC_IOCTL_INIT_ATTACH_SNS:
> err = fastrpc_init_attach(fl, SENSORS_PD);
> @@ -2627,6 +2819,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
> err = fastrpc_device_register(rdev, data, secure_dsp, domains[domain_id]);
> if (err)
> goto fdev_error;
> + data->cpuinfo_todsp = FASTRPC_CPUINFO_DEFAULT;
> break;
> case CDSP_DOMAIN_ID:
> data->unsigned_support = true;
> @@ -2638,6 +2831,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
> err = fastrpc_device_register(rdev, data, false, domains[domain_id]);
> if (err)
> goto fdev_error;
> + data->cpuinfo_todsp = FASTRPC_CPUINFO_EARLY_WAKEUP;
> break;
> default:
> err = -EINVAL;
> @@ -2680,10 +2874,12 @@ static void fastrpc_notify_users(struct fastrpc_user *user)
> spin_lock(&user->lock);
> list_for_each_entry(ctx, &user->pending, node) {
> ctx->retval = -EPIPE;
> + ctx->is_work_done = true;
> complete(&ctx->work);
> }
> list_for_each_entry(ctx, &user->interrupted, node) {
> ctx->retval = -EPIPE;
> + ctx->is_work_done = true;
> complete(&ctx->work);
> }
> spin_unlock(&user->lock);
> @@ -2720,31 +2916,74 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
> fastrpc_channel_ctx_put(cctx);
> }
>
> +static void fastrpc_notify_user_ctx(struct fastrpc_invoke_ctx *ctx, int retval,
> + u32 rsp_flags, u32 early_wake_time)
> +{
> + ctx->retval = retval;
> + ctx->rsp_flags = (enum fastrpc_response_flags)rsp_flags;
> + switch (rsp_flags) {
> + case NORMAL_RESPONSE:
> + case COMPLETE_SIGNAL:
> + /* normal and complete response with return value */
> + ctx->is_work_done = true;
> + complete(&ctx->work);
> + break;
> + case USER_EARLY_SIGNAL:
> + /* user hint of approximate time of completion */
> + ctx->early_wake_time = early_wake_time;
> + break;
> + case EARLY_RESPONSE:
> + /* rpc framework early response with return value */
> + complete(&ctx->work);
> + break;
> + default:
> + break;
> + }
> +}
> +
> static int fastrpc_rpmsg_callback(struct rpmsg_device *rpdev, void *data,
> int len, void *priv, u32 addr)
> {
> struct fastrpc_channel_ctx *cctx = dev_get_drvdata(&rpdev->dev);
> struct fastrpc_invoke_rsp *rsp = data;
> + struct fastrpc_invoke_rspv2 *rspv2 = NULL;
> struct fastrpc_invoke_ctx *ctx;
> unsigned long flags;
> unsigned long ctxid;
> + u32 rsp_flags = 0;
> + u32 early_wake_time = 0;
>
> if (len < sizeof(*rsp))
> return -EINVAL;
>
> + if (len >= sizeof(*rspv2)) {
> + rspv2 = data;
> + if (rspv2) {
> + early_wake_time = rspv2->early_wake_time;
> + rsp_flags = rspv2->flags;
> + }
> + }
> +
> ctxid = ((rsp->ctx & FASTRPC_CTXID_MASK) >> 4);
>
> spin_lock_irqsave(&cctx->lock, flags);
> ctx = idr_find(&cctx->ctx_idr, ctxid);
> - spin_unlock_irqrestore(&cctx->lock, flags);
>
> if (!ctx) {
> - dev_err(&rpdev->dev, "No context ID matches response\n");
> - return -ENOENT;
> + dev_info(&cctx->rpdev->dev, "Warning: No context ID matches response\n");
> + spin_unlock_irqrestore(&cctx->lock, flags);
> + return 0;
> }
>
> - ctx->retval = rsp->retval;
> - complete(&ctx->work);
> + if (rspv2) {
> + if (rspv2->version != FASTRPC_RSP_VERSION2) {
> + dev_err(&cctx->rpdev->dev, "Incorrect response version %d\n", rspv2->version);
> + spin_unlock_irqrestore(&cctx->lock, flags);
> + return -EINVAL;
> + }
> + }
> + fastrpc_notify_user_ctx(ctx, rsp->retval, rsp_flags, early_wake_time);
> + spin_unlock_irqrestore(&cctx->lock, flags);
>
> /*
> * The DMA buffer associated with the context cannot be freed in
^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2023-11-24 13:27 UTC | newest]
Thread overview: 7+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2023-11-21 11:24 [PATCH v3 0/5] Add missing features to FastRPC driver Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 1/5] misc: fastrpc: Add early wakeup support for fastRPC driver Ekansh Gupta
2023-11-24 13:27 ` Srinivas Kandagatla
2023-11-21 11:24 ` [PATCH v3 2/5] misc: fastrpc: Add polling mode " Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 3/5] misc: fastrpc: Add DSP PD notification support Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 4/5] misc: fastrpc: Add support for users to clean up DSP user PD Ekansh Gupta
2023-11-21 11:24 ` [PATCH v3 5/5] misc: fastrpc: Add wakelock management support Ekansh Gupta
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.