2020-06-29 20:05:24

by Rishabh Bhatnagar

[permalink] [raw]
Subject: [PATCH v6 0/3] Extend coredump functionality

This patch series moves the coredump functionality to a separate
file and adds "inline" coredump feature. Inline coredump directly
copies segments from device memory during coredump to userspace.
This avoids extra memory usage at the cost of speed. Recovery is
stalled until all data is read by userspace.

Changelog:

v6 -> v5:
- Fix unsigned comaprison with negative bug found on gcc-9.3.0

v5 -> v4:
- Rebase on top of linux-next

v4 -> v3:
- Write a helper function to copy segment memory for every dump format
- Change segment dump fn to add offset and size adn covert mss driver

v3 -> v2:
- Move entire coredump functionality to remoteproc_coredump.c
- Modify rproc_coredump to perform dump according to conf. set by userspace
- Move the userspace configuration to debugfs from sysfs.
- Keep the default coredump implementation as is

v2 -> v1:
- Introduce new file for coredump.
- Add userspace sysfs configuration for dump type.

Rishabh Bhatnagar (3):
remoteproc: Move coredump functionality to a new file
remoteproc: Add inline coredump functionality
remoteproc: Add coredump debugfs entry

drivers/remoteproc/Makefile | 1 +
drivers/remoteproc/qcom_q6v5_mss.c | 9 +-
drivers/remoteproc/remoteproc_core.c | 191 ------------------
drivers/remoteproc/remoteproc_coredump.c | 328 +++++++++++++++++++++++++++++++
drivers/remoteproc/remoteproc_debugfs.c | 86 ++++++++
drivers/remoteproc/remoteproc_internal.h | 4 +
include/linux/remoteproc.h | 21 +-
7 files changed, 443 insertions(+), 197 deletions(-)
create mode 100644 drivers/remoteproc/remoteproc_coredump.c

--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project


2020-06-29 20:06:36

by Rishabh Bhatnagar

[permalink] [raw]
Subject: [PATCH v6 2/3] remoteproc: Add inline coredump functionality

The current coredump implementation uses vmalloc area to copy
all the segments. But this might put strain on low memory targets
as the firmware size sometimes is in tens of MBs. The situation
becomes worse if there are multiple remote processors undergoing
recovery at the same time. This patch adds inline coredump
functionality that avoids extra memory usage. This requires
recovery to be halted until data is read by userspace and free
function is called.

Signed-off-by: Rishabh Bhatnagar <[email protected]>
---
drivers/remoteproc/qcom_q6v5_mss.c | 9 +-
drivers/remoteproc/remoteproc_coredump.c | 160 +++++++++++++++++++++++++++----
include/linux/remoteproc.h | 21 +++-
3 files changed, 165 insertions(+), 25 deletions(-)

diff --git a/drivers/remoteproc/qcom_q6v5_mss.c b/drivers/remoteproc/qcom_q6v5_mss.c
index 903b2bb..d4ff9b8 100644
--- a/drivers/remoteproc/qcom_q6v5_mss.c
+++ b/drivers/remoteproc/qcom_q6v5_mss.c
@@ -1200,12 +1200,13 @@ static int q6v5_mpss_load(struct q6v5 *qproc)

static void qcom_q6v5_dump_segment(struct rproc *rproc,
struct rproc_dump_segment *segment,
- void *dest)
+ void *dest, size_t cp_offset, size_t size)
{
int ret = 0;
struct q6v5 *qproc = rproc->priv;
unsigned long mask = BIT((unsigned long)segment->priv);
int offset = segment->da - qproc->mpss_reloc;
+ size_t cp_size = size ? size : segment->size;
void *ptr = NULL;

/* Unlock mba before copying segments */
@@ -1221,13 +1222,13 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
}

if (!ret)
- ptr = ioremap_wc(qproc->mpss_phys + offset, segment->size);
+ ptr = ioremap_wc(qproc->mpss_phys + offset + cp_offset, cp_size);

if (ptr) {
- memcpy(dest, ptr, segment->size);
+ memcpy(dest, ptr, cp_size);
iounmap(ptr);
} else {
- memset(dest, 0xff, segment->size);
+ memset(dest, 0xff, cp_size);
}

qproc->dump_segment_mask |= mask;
diff --git a/drivers/remoteproc/remoteproc_coredump.c b/drivers/remoteproc/remoteproc_coredump.c
index ded0244..646886f 100644
--- a/drivers/remoteproc/remoteproc_coredump.c
+++ b/drivers/remoteproc/remoteproc_coredump.c
@@ -5,6 +5,7 @@
* Copyright (c) 2020, The Linux Foundation. All rights reserved.
*/

+#include <linux/completion.h>
#include <linux/devcoredump.h>
#include <linux/device.h>
#include <linux/kernel.h>
@@ -12,6 +13,12 @@
#include "remoteproc_internal.h"
#include "remoteproc_elf_helpers.h"

+struct rproc_coredump_state {
+ struct rproc *rproc;
+ void *header;
+ struct completion dump_done;
+};
+
/**
* rproc_coredump_cleanup() - clean up dump_segments list
* @rproc: the remote processor handle
@@ -72,7 +79,8 @@ int rproc_coredump_add_custom_segment(struct rproc *rproc,
dma_addr_t da, size_t size,
void (*dumpfn)(struct rproc *rproc,
struct rproc_dump_segment *segment,
- void *dest),
+ void *dest, size_t offset,
+ size_t size),
void *priv)
{
struct rproc_dump_segment *segment;
@@ -114,12 +122,110 @@ int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16 machine)
}
EXPORT_SYMBOL(rproc_coredump_set_elf_info);

+static void rproc_coredump_free(void *data)
+{
+ struct rproc_coredump_state *dump_state = data;
+
+ complete(&dump_state->dump_done);
+ vfree(dump_state->header);
+}
+
+static void *rproc_coredump_find_segment(loff_t user_offset,
+ struct list_head *segments,
+ size_t *data_left)
+{
+ struct rproc_dump_segment *segment;
+
+ list_for_each_entry(segment, segments, node) {
+ if (user_offset < segment->size) {
+ *data_left = segment->size - user_offset;
+ return segment;
+ }
+ user_offset -= segment->size;
+ }
+
+ *data_left = 0;
+ return NULL;
+}
+
+static void rproc_copy_segment(struct rproc *rproc, void *dest,
+ struct rproc_dump_segment *segment,
+ size_t offset, size_t size)
+{
+ void *ptr;
+
+ if (segment->dump) {
+ segment->dump(rproc, segment, dest, offset, size);
+ } else {
+ ptr = rproc_da_to_va(rproc, segment->da + offset, size);
+ if (!ptr) {
+ dev_err(&rproc->dev,
+ "invalid copy request for segment %pad with offset %zu and size %zu)\n",
+ &segment->da, offset, size);
+ memset(dest, 0xff, size);
+ } else {
+ memcpy(dest, ptr, size);
+ }
+ }
+}
+
+static ssize_t rproc_coredump_read(char *buffer, loff_t offset, size_t count,
+ void *data, size_t header_sz)
+{
+ size_t seg_data, bytes_left = count;
+ ssize_t copy_sz;
+ struct rproc_dump_segment *seg;
+ struct rproc_coredump_state *dump_state = data;
+ struct rproc *rproc = dump_state->rproc;
+ void *elfcore = dump_state->header;
+
+ /* Copy the vmalloc'ed header first. */
+ if (offset < header_sz) {
+ copy_sz = memory_read_from_buffer(buffer, count, &offset,
+ elfcore, header_sz);
+
+ return copy_sz;
+ }
+
+ /*
+ * Find out the segment memory chunk to be copied based on offset.
+ * Keep copying data until count bytes are read.
+ */
+ while (bytes_left) {
+ seg = rproc_coredump_find_segment(offset - header_sz,
+ &rproc->dump_segments,
+ &seg_data);
+ /* EOF check */
+ if (!seg) {
+ dev_info(&rproc->dev, "Ramdump done, %lld bytes read",
+ offset);
+ break;
+ }
+
+ copy_sz = min_t(size_t, bytes_left, seg_data);
+
+ rproc_copy_segment(rproc, buffer, seg, seg->size - seg_data,
+ copy_sz);
+
+ offset += copy_sz;
+ buffer += copy_sz;
+ bytes_left -= copy_sz;
+ }
+
+ return count - bytes_left;
+}
+
/**
* rproc_coredump() - perform coredump
* @rproc: rproc handle
*
* This function will generate an ELF header for the registered segments
- * and create a devcoredump device associated with rproc.
+ * and create a devcoredump device associated with rproc. Based on the
+ * coredump configuration this function will directly copy the segments
+ * from device memory to userspace or copy segments from device memory to
+ * a separate buffer, which can then be read by userspace.
+ * The first approach avoids using extra vmalloc memory. But it will stall
+ * recovery flow until dump is read by userspace.
*/
void rproc_coredump(struct rproc *rproc)
{
@@ -129,11 +235,13 @@ void rproc_coredump(struct rproc *rproc)
size_t data_size;
size_t offset;
void *data;
- void *ptr;
u8 class = rproc->elf_class;
int phnum = 0;
+ struct rproc_coredump_state dump_state;
+ enum rproc_dump_mechanism dump_conf = rproc->dump_conf;

- if (list_empty(&rproc->dump_segments))
+ if (list_empty(&rproc->dump_segments) ||
+ dump_conf == RPROC_COREDUMP_DISABLED)
return;

if (class == ELFCLASSNONE) {
@@ -143,7 +251,14 @@ void rproc_coredump(struct rproc *rproc)

data_size = elf_size_of_hdr(class);
list_for_each_entry(segment, &rproc->dump_segments, node) {
- data_size += elf_size_of_phdr(class) + segment->size;
+ /*
+ * For default configuration buffer includes headers & segments.
+ * For inline dump buffer just includes headers as segments are
+ * directly read from device memory.
+ */
+ data_size += elf_size_of_phdr(class);
+ if (dump_conf == RPROC_COREDUMP_DEFAULT)
+ data_size += segment->size;

phnum++;
}
@@ -182,23 +297,30 @@ void rproc_coredump(struct rproc *rproc)
elf_phdr_set_p_flags(class, phdr, PF_R | PF_W | PF_X);
elf_phdr_set_p_align(class, phdr, 0);

- if (segment->dump) {
- segment->dump(rproc, segment, data + offset);
- } else {
- ptr = rproc_da_to_va(rproc, segment->da, segment->size);
- if (!ptr) {
- dev_err(&rproc->dev,
- "invalid coredump segment (%pad, %zu)\n",
- &segment->da, segment->size);
- memset(data + offset, 0xff, segment->size);
- } else {
- memcpy(data + offset, ptr, segment->size);
- }
- }
+ if (dump_conf == RPROC_COREDUMP_DEFAULT)
+ rproc_copy_segment(rproc, data + offset, segment, 0,
+ segment->size);

offset += elf_phdr_get_p_filesz(class, phdr);
phdr += elf_size_of_phdr(class);
}

- dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
+ if (dump_conf == RPROC_COREDUMP_DEFAULT) {
+ dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
+ return;
+ }
+
+ /* Initialize the dump state struct to be used by rproc_coredump_read */
+ dump_state.rproc = rproc;
+ dump_state.header = data;
+ init_completion(&dump_state.dump_done);
+
+ dev_coredumpm(&rproc->dev, NULL, &dump_state, data_size, GFP_KERNEL,
+ rproc_coredump_read, rproc_coredump_free);
+
+ /*
+ * Wait until the dump is read and free is called. Data is freed
+ * by devcoredump framework automatically after 5 minutes.
+ */
+ wait_for_completion(&dump_state.dump_done);
}
diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
index e7b7bab..43e45a3 100644
--- a/include/linux/remoteproc.h
+++ b/include/linux/remoteproc.h
@@ -435,6 +435,20 @@ enum rproc_crash_type {
};

/**
+ * enum rproc_dump_mechanism - Coredump options for core
+ * @RPROC_COREDUMP_DEFAULT: Copy dump to separate buffer and carry on with
+ recovery
+ * @RPROC_COREDUMP_INLINE: Read segments directly from device memory. Stall
+ recovery until all segments are read
+ * @RPROC_COREDUMP_DISABLED: Don't perform any dump
+ */
+enum rproc_dump_mechanism {
+ RPROC_COREDUMP_DEFAULT,
+ RPROC_COREDUMP_INLINE,
+ RPROC_COREDUMP_DISABLED,
+};
+
+/**
* struct rproc_dump_segment - segment info from ELF header
* @node: list node related to the rproc segment list
* @da: device address of the segment
@@ -451,7 +465,7 @@ struct rproc_dump_segment {

void *priv;
void (*dump)(struct rproc *rproc, struct rproc_dump_segment *segment,
- void *dest);
+ void *dest, size_t offset, size_t size);
loff_t offset;
};

@@ -466,6 +480,7 @@ struct rproc_dump_segment {
* @dev: virtual device for refcounting and common remoteproc behavior
* @power: refcount of users who need this rproc powered up
* @state: state of the device
+ * @dump_conf: Currenlty selected coredump configuration
* @lock: lock which protects concurrent manipulations of the rproc
* @dbg_dir: debugfs directory of this rproc device
* @traces: list of trace buffers
@@ -499,6 +514,7 @@ struct rproc {
struct device dev;
atomic_t power;
unsigned int state;
+ enum rproc_dump_mechanism dump_conf;
struct mutex lock;
struct dentry *dbg_dir;
struct list_head traces;
@@ -630,7 +646,8 @@ int rproc_coredump_add_custom_segment(struct rproc *rproc,
dma_addr_t da, size_t size,
void (*dumpfn)(struct rproc *rproc,
struct rproc_dump_segment *segment,
- void *dest),
+ void *dest, size_t offset,
+ size_t size),
void *priv);
int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16 machine);

--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

2020-07-01 20:21:12

by Sibi Sankar

[permalink] [raw]
Subject: Re: [PATCH v6 2/3] remoteproc: Add inline coredump functionality

On 2020-06-30 01:32, Rishabh Bhatnagar wrote:
> The current coredump implementation uses vmalloc area to copy
> all the segments. But this might put strain on low memory targets
> as the firmware size sometimes is in tens of MBs. The situation
> becomes worse if there are multiple remote processors undergoing
> recovery at the same time. This patch adds inline coredump
> functionality that avoids extra memory usage. This requires
> recovery to be halted until data is read by userspace and free
> function is called.
>
> Signed-off-by: Rishabh Bhatnagar <[email protected]>
> ---
> drivers/remoteproc/qcom_q6v5_mss.c | 9 +-
> drivers/remoteproc/remoteproc_coredump.c | 160
> +++++++++++++++++++++++++++----
> include/linux/remoteproc.h | 21 +++-
> 3 files changed, 165 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/remoteproc/qcom_q6v5_mss.c
> b/drivers/remoteproc/qcom_q6v5_mss.c
> index 903b2bb..d4ff9b8 100644
> --- a/drivers/remoteproc/qcom_q6v5_mss.c
> +++ b/drivers/remoteproc/qcom_q6v5_mss.c
> @@ -1200,12 +1200,13 @@ static int q6v5_mpss_load(struct q6v5 *qproc)
>
> static void qcom_q6v5_dump_segment(struct rproc *rproc,
> struct rproc_dump_segment *segment,
> - void *dest)
> + void *dest, size_t cp_offset, size_t size)
> {
> int ret = 0;
> struct q6v5 *qproc = rproc->priv;
> unsigned long mask = BIT((unsigned long)segment->priv);
> int offset = segment->da - qproc->mpss_reloc;
> + size_t cp_size = size ? size : segment->size;
> void *ptr = NULL;
>
> /* Unlock mba before copying segments */
> @@ -1221,13 +1222,13 @@ static void qcom_q6v5_dump_segment(struct rproc
> *rproc,
> }
>
> if (!ret)
> - ptr = ioremap_wc(qproc->mpss_phys + offset, segment->size);
> + ptr = ioremap_wc(qproc->mpss_phys + offset + cp_offset, cp_size);
>
> if (ptr) {
> - memcpy(dest, ptr, segment->size);
> + memcpy(dest, ptr, cp_size);
> iounmap(ptr);
> } else {
> - memset(dest, 0xff, segment->size);
> + memset(dest, 0xff, cp_size);
> }

https://patchwork.kernel.org/patch/11637157/
To deal with chunk sizes lesser
than the segment size I've sent
a patch. With ^^ you would need
to the following as well on top
of that:

qproc->current_dump_size += cp_size;

Reviewed-by: Sibi Sankar <[email protected]>

>
> qproc->dump_segment_mask |= mask;
> diff --git a/drivers/remoteproc/remoteproc_coredump.c
> b/drivers/remoteproc/remoteproc_coredump.c
> index ded0244..646886f 100644
> --- a/drivers/remoteproc/remoteproc_coredump.c
> +++ b/drivers/remoteproc/remoteproc_coredump.c
> @@ -5,6 +5,7 @@
> * Copyright (c) 2020, The Linux Foundation. All rights reserved.
> */
>
> +#include <linux/completion.h>
> #include <linux/devcoredump.h>
> #include <linux/device.h>
> #include <linux/kernel.h>
> @@ -12,6 +13,12 @@
> #include "remoteproc_internal.h"
> #include "remoteproc_elf_helpers.h"
>
> +struct rproc_coredump_state {
> + struct rproc *rproc;
> + void *header;
> + struct completion dump_done;
> +};
> +
> /**
> * rproc_coredump_cleanup() - clean up dump_segments list
> * @rproc: the remote processor handle
> @@ -72,7 +79,8 @@ int rproc_coredump_add_custom_segment(struct rproc
> *rproc,
> dma_addr_t da, size_t size,
> void (*dumpfn)(struct rproc *rproc,
> struct rproc_dump_segment *segment,
> - void *dest),
> + void *dest, size_t offset,
> + size_t size),
> void *priv)
> {
> struct rproc_dump_segment *segment;
> @@ -114,12 +122,110 @@ int rproc_coredump_set_elf_info(struct rproc
> *rproc, u8 class, u16 machine)
> }
> EXPORT_SYMBOL(rproc_coredump_set_elf_info);
>
> +static void rproc_coredump_free(void *data)
> +{
> + struct rproc_coredump_state *dump_state = data;
> +
> + complete(&dump_state->dump_done);
> + vfree(dump_state->header);
> +}
> +
> +static void *rproc_coredump_find_segment(loff_t user_offset,
> + struct list_head *segments,
> + size_t *data_left)
> +{
> + struct rproc_dump_segment *segment;
> +
> + list_for_each_entry(segment, segments, node) {
> + if (user_offset < segment->size) {
> + *data_left = segment->size - user_offset;
> + return segment;
> + }
> + user_offset -= segment->size;
> + }
> +
> + *data_left = 0;
> + return NULL;
> +}
> +
> +static void rproc_copy_segment(struct rproc *rproc, void *dest,
> + struct rproc_dump_segment *segment,
> + size_t offset, size_t size)
> +{
> + void *ptr;
> +
> + if (segment->dump) {
> + segment->dump(rproc, segment, dest, offset, size);
> + } else {
> + ptr = rproc_da_to_va(rproc, segment->da + offset, size);
> + if (!ptr) {
> + dev_err(&rproc->dev,
> + "invalid copy request for segment %pad with offset %zu and size
> %zu)\n",
> + &segment->da, offset, size);
> + memset(dest, 0xff, size);
> + } else {
> + memcpy(dest, ptr, size);
> + }
> + }
> +}
> +
> +static ssize_t rproc_coredump_read(char *buffer, loff_t offset, size_t
> count,
> + void *data, size_t header_sz)
> +{
> + size_t seg_data, bytes_left = count;
> + ssize_t copy_sz;
> + struct rproc_dump_segment *seg;
> + struct rproc_coredump_state *dump_state = data;
> + struct rproc *rproc = dump_state->rproc;
> + void *elfcore = dump_state->header;
> +
> + /* Copy the vmalloc'ed header first. */
> + if (offset < header_sz) {
> + copy_sz = memory_read_from_buffer(buffer, count, &offset,
> + elfcore, header_sz);
> +
> + return copy_sz;
> + }
> +
> + /*
> + * Find out the segment memory chunk to be copied based on offset.
> + * Keep copying data until count bytes are read.
> + */
> + while (bytes_left) {
> + seg = rproc_coredump_find_segment(offset - header_sz,
> + &rproc->dump_segments,
> + &seg_data);
> + /* EOF check */
> + if (!seg) {
> + dev_info(&rproc->dev, "Ramdump done, %lld bytes read",
> + offset);
> + break;
> + }
> +
> + copy_sz = min_t(size_t, bytes_left, seg_data);
> +
> + rproc_copy_segment(rproc, buffer, seg, seg->size - seg_data,
> + copy_sz);
> +
> + offset += copy_sz;
> + buffer += copy_sz;
> + bytes_left -= copy_sz;
> + }
> +
> + return count - bytes_left;
> +}
> +
> /**
> * rproc_coredump() - perform coredump
> * @rproc: rproc handle
> *
> * This function will generate an ELF header for the registered
> segments
> - * and create a devcoredump device associated with rproc.
> + * and create a devcoredump device associated with rproc. Based on the
> + * coredump configuration this function will directly copy the
> segments
> + * from device memory to userspace or copy segments from device memory
> to
> + * a separate buffer, which can then be read by userspace.
> + * The first approach avoids using extra vmalloc memory. But it will
> stall
> + * recovery flow until dump is read by userspace.
> */
> void rproc_coredump(struct rproc *rproc)
> {
> @@ -129,11 +235,13 @@ void rproc_coredump(struct rproc *rproc)
> size_t data_size;
> size_t offset;
> void *data;
> - void *ptr;
> u8 class = rproc->elf_class;
> int phnum = 0;
> + struct rproc_coredump_state dump_state;
> + enum rproc_dump_mechanism dump_conf = rproc->dump_conf;
>
> - if (list_empty(&rproc->dump_segments))
> + if (list_empty(&rproc->dump_segments) ||
> + dump_conf == RPROC_COREDUMP_DISABLED)
> return;
>
> if (class == ELFCLASSNONE) {
> @@ -143,7 +251,14 @@ void rproc_coredump(struct rproc *rproc)
>
> data_size = elf_size_of_hdr(class);
> list_for_each_entry(segment, &rproc->dump_segments, node) {
> - data_size += elf_size_of_phdr(class) + segment->size;
> + /*
> + * For default configuration buffer includes headers & segments.
> + * For inline dump buffer just includes headers as segments are
> + * directly read from device memory.
> + */
> + data_size += elf_size_of_phdr(class);
> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
> + data_size += segment->size;
>
> phnum++;
> }
> @@ -182,23 +297,30 @@ void rproc_coredump(struct rproc *rproc)
> elf_phdr_set_p_flags(class, phdr, PF_R | PF_W | PF_X);
> elf_phdr_set_p_align(class, phdr, 0);
>
> - if (segment->dump) {
> - segment->dump(rproc, segment, data + offset);
> - } else {
> - ptr = rproc_da_to_va(rproc, segment->da, segment->size);
> - if (!ptr) {
> - dev_err(&rproc->dev,
> - "invalid coredump segment (%pad, %zu)\n",
> - &segment->da, segment->size);
> - memset(data + offset, 0xff, segment->size);
> - } else {
> - memcpy(data + offset, ptr, segment->size);
> - }
> - }
> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
> + rproc_copy_segment(rproc, data + offset, segment, 0,
> + segment->size);
>
> offset += elf_phdr_get_p_filesz(class, phdr);
> phdr += elf_size_of_phdr(class);
> }
>
> - dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
> + if (dump_conf == RPROC_COREDUMP_DEFAULT) {
> + dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
> + return;
> + }
> +
> + /* Initialize the dump state struct to be used by rproc_coredump_read
> */
> + dump_state.rproc = rproc;
> + dump_state.header = data;
> + init_completion(&dump_state.dump_done);
> +
> + dev_coredumpm(&rproc->dev, NULL, &dump_state, data_size, GFP_KERNEL,
> + rproc_coredump_read, rproc_coredump_free);
> +
> + /*
> + * Wait until the dump is read and free is called. Data is freed
> + * by devcoredump framework automatically after 5 minutes.
> + */
> + wait_for_completion(&dump_state.dump_done);
> }
> diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
> index e7b7bab..43e45a3 100644
> --- a/include/linux/remoteproc.h
> +++ b/include/linux/remoteproc.h
> @@ -435,6 +435,20 @@ enum rproc_crash_type {
> };
>
> /**
> + * enum rproc_dump_mechanism - Coredump options for core
> + * @RPROC_COREDUMP_DEFAULT: Copy dump to separate buffer and carry on
> with
> + recovery
> + * @RPROC_COREDUMP_INLINE: Read segments directly from device memory.
> Stall
> + recovery until all segments are read
> + * @RPROC_COREDUMP_DISABLED: Don't perform any dump
> + */
> +enum rproc_dump_mechanism {
> + RPROC_COREDUMP_DEFAULT,
> + RPROC_COREDUMP_INLINE,
> + RPROC_COREDUMP_DISABLED,
> +};
> +
> +/**
> * struct rproc_dump_segment - segment info from ELF header
> * @node: list node related to the rproc segment list
> * @da: device address of the segment
> @@ -451,7 +465,7 @@ struct rproc_dump_segment {
>
> void *priv;
> void (*dump)(struct rproc *rproc, struct rproc_dump_segment *segment,
> - void *dest);
> + void *dest, size_t offset, size_t size);
> loff_t offset;
> };
>
> @@ -466,6 +480,7 @@ struct rproc_dump_segment {
> * @dev: virtual device for refcounting and common remoteproc behavior
> * @power: refcount of users who need this rproc powered up
> * @state: state of the device
> + * @dump_conf: Currenlty selected coredump configuration

s/Currenlty/Currently

> * @lock: lock which protects concurrent manipulations of the rproc
> * @dbg_dir: debugfs directory of this rproc device
> * @traces: list of trace buffers
> @@ -499,6 +514,7 @@ struct rproc {
> struct device dev;
> atomic_t power;
> unsigned int state;
> + enum rproc_dump_mechanism dump_conf;
> struct mutex lock;
> struct dentry *dbg_dir;
> struct list_head traces;
> @@ -630,7 +646,8 @@ int rproc_coredump_add_custom_segment(struct rproc
> *rproc,
> dma_addr_t da, size_t size,
> void (*dumpfn)(struct rproc *rproc,
> struct rproc_dump_segment *segment,
> - void *dest),
> + void *dest, size_t offset,
> + size_t size),
> void *priv);
> int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16
> machine);

--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project.

2020-07-06 17:49:57

by Mathieu Poirier

[permalink] [raw]
Subject: Re: [PATCH v6 2/3] remoteproc: Add inline coredump functionality

On Mon, Jun 29, 2020 at 01:02:12PM -0700, Rishabh Bhatnagar wrote:
> The current coredump implementation uses vmalloc area to copy
> all the segments. But this might put strain on low memory targets
> as the firmware size sometimes is in tens of MBs. The situation
> becomes worse if there are multiple remote processors undergoing
> recovery at the same time. This patch adds inline coredump
> functionality that avoids extra memory usage. This requires
> recovery to be halted until data is read by userspace and free
> function is called.
>
> Signed-off-by: Rishabh Bhatnagar <[email protected]>
> ---
> drivers/remoteproc/qcom_q6v5_mss.c | 9 +-
> drivers/remoteproc/remoteproc_coredump.c | 160 +++++++++++++++++++++++++++----
> include/linux/remoteproc.h | 21 +++-
> 3 files changed, 165 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/remoteproc/qcom_q6v5_mss.c b/drivers/remoteproc/qcom_q6v5_mss.c
> index 903b2bb..d4ff9b8 100644
> --- a/drivers/remoteproc/qcom_q6v5_mss.c
> +++ b/drivers/remoteproc/qcom_q6v5_mss.c
> @@ -1200,12 +1200,13 @@ static int q6v5_mpss_load(struct q6v5 *qproc)
>
> static void qcom_q6v5_dump_segment(struct rproc *rproc,
> struct rproc_dump_segment *segment,
> - void *dest)
> + void *dest, size_t cp_offset, size_t size)
> {
> int ret = 0;
> struct q6v5 *qproc = rproc->priv;
> unsigned long mask = BIT((unsigned long)segment->priv);
> int offset = segment->da - qproc->mpss_reloc;
> + size_t cp_size = size ? size : segment->size;
> void *ptr = NULL;

On the V4 of this set the above line was:

+ void *ptr = rproc_da_to_va(rproc, segment->da + offset, copy_size);

Back then both Bjorn and I had RB'ed this set and all that was required was a
rebase to linux-next. On V5 and V6 our RBs have been removed, the above has
been changed and an iounmap() was been added below. Yet nothing in the cover
letter provides an explanation that justifies the modification.

What is going on here?

>
> /* Unlock mba before copying segments */
> @@ -1221,13 +1222,13 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
> }
>
> if (!ret)
> - ptr = ioremap_wc(qproc->mpss_phys + offset, segment->size);
> + ptr = ioremap_wc(qproc->mpss_phys + offset + cp_offset, cp_size);
>
> if (ptr) {
> - memcpy(dest, ptr, segment->size);
> + memcpy(dest, ptr, cp_size);
> iounmap(ptr);
> } else {
> - memset(dest, 0xff, segment->size);
> + memset(dest, 0xff, cp_size);
> }
>
> qproc->dump_segment_mask |= mask;
> diff --git a/drivers/remoteproc/remoteproc_coredump.c b/drivers/remoteproc/remoteproc_coredump.c
> index ded0244..646886f 100644
> --- a/drivers/remoteproc/remoteproc_coredump.c
> +++ b/drivers/remoteproc/remoteproc_coredump.c
> @@ -5,6 +5,7 @@
> * Copyright (c) 2020, The Linux Foundation. All rights reserved.
> */
>
> +#include <linux/completion.h>
> #include <linux/devcoredump.h>
> #include <linux/device.h>
> #include <linux/kernel.h>
> @@ -12,6 +13,12 @@
> #include "remoteproc_internal.h"
> #include "remoteproc_elf_helpers.h"
>
> +struct rproc_coredump_state {
> + struct rproc *rproc;
> + void *header;
> + struct completion dump_done;
> +};
> +
> /**
> * rproc_coredump_cleanup() - clean up dump_segments list
> * @rproc: the remote processor handle
> @@ -72,7 +79,8 @@ int rproc_coredump_add_custom_segment(struct rproc *rproc,
> dma_addr_t da, size_t size,
> void (*dumpfn)(struct rproc *rproc,
> struct rproc_dump_segment *segment,
> - void *dest),
> + void *dest, size_t offset,
> + size_t size),
> void *priv)
> {
> struct rproc_dump_segment *segment;
> @@ -114,12 +122,110 @@ int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16 machine)
> }
> EXPORT_SYMBOL(rproc_coredump_set_elf_info);
>
> +static void rproc_coredump_free(void *data)
> +{
> + struct rproc_coredump_state *dump_state = data;
> +
> + complete(&dump_state->dump_done);
> + vfree(dump_state->header);
> +}
> +
> +static void *rproc_coredump_find_segment(loff_t user_offset,
> + struct list_head *segments,
> + size_t *data_left)
> +{
> + struct rproc_dump_segment *segment;
> +
> + list_for_each_entry(segment, segments, node) {
> + if (user_offset < segment->size) {
> + *data_left = segment->size - user_offset;
> + return segment;
> + }
> + user_offset -= segment->size;
> + }
> +
> + *data_left = 0;
> + return NULL;
> +}
> +
> +static void rproc_copy_segment(struct rproc *rproc, void *dest,
> + struct rproc_dump_segment *segment,
> + size_t offset, size_t size)
> +{
> + void *ptr;
> +
> + if (segment->dump) {
> + segment->dump(rproc, segment, dest, offset, size);
> + } else {
> + ptr = rproc_da_to_va(rproc, segment->da + offset, size);
> + if (!ptr) {
> + dev_err(&rproc->dev,
> + "invalid copy request for segment %pad with offset %zu and size %zu)\n",
> + &segment->da, offset, size);
> + memset(dest, 0xff, size);
> + } else {
> + memcpy(dest, ptr, size);
> + }
> + }
> +}
> +
> +static ssize_t rproc_coredump_read(char *buffer, loff_t offset, size_t count,
> + void *data, size_t header_sz)
> +{
> + size_t seg_data, bytes_left = count;
> + ssize_t copy_sz;
> + struct rproc_dump_segment *seg;
> + struct rproc_coredump_state *dump_state = data;
> + struct rproc *rproc = dump_state->rproc;
> + void *elfcore = dump_state->header;
> +
> + /* Copy the vmalloc'ed header first. */
> + if (offset < header_sz) {
> + copy_sz = memory_read_from_buffer(buffer, count, &offset,
> + elfcore, header_sz);
> +
> + return copy_sz;
> + }
> +
> + /*
> + * Find out the segment memory chunk to be copied based on offset.
> + * Keep copying data until count bytes are read.
> + */
> + while (bytes_left) {
> + seg = rproc_coredump_find_segment(offset - header_sz,
> + &rproc->dump_segments,
> + &seg_data);
> + /* EOF check */
> + if (!seg) {
> + dev_info(&rproc->dev, "Ramdump done, %lld bytes read",
> + offset);
> + break;
> + }
> +
> + copy_sz = min_t(size_t, bytes_left, seg_data);
> +
> + rproc_copy_segment(rproc, buffer, seg, seg->size - seg_data,
> + copy_sz);
> +
> + offset += copy_sz;
> + buffer += copy_sz;
> + bytes_left -= copy_sz;
> + }
> +
> + return count - bytes_left;
> +}
> +
> /**
> * rproc_coredump() - perform coredump
> * @rproc: rproc handle
> *
> * This function will generate an ELF header for the registered segments
> - * and create a devcoredump device associated with rproc.
> + * and create a devcoredump device associated with rproc. Based on the
> + * coredump configuration this function will directly copy the segments
> + * from device memory to userspace or copy segments from device memory to
> + * a separate buffer, which can then be read by userspace.
> + * The first approach avoids using extra vmalloc memory. But it will stall
> + * recovery flow until dump is read by userspace.
> */
> void rproc_coredump(struct rproc *rproc)
> {
> @@ -129,11 +235,13 @@ void rproc_coredump(struct rproc *rproc)
> size_t data_size;
> size_t offset;
> void *data;
> - void *ptr;
> u8 class = rproc->elf_class;
> int phnum = 0;
> + struct rproc_coredump_state dump_state;
> + enum rproc_dump_mechanism dump_conf = rproc->dump_conf;
>
> - if (list_empty(&rproc->dump_segments))
> + if (list_empty(&rproc->dump_segments) ||
> + dump_conf == RPROC_COREDUMP_DISABLED)
> return;
>
> if (class == ELFCLASSNONE) {
> @@ -143,7 +251,14 @@ void rproc_coredump(struct rproc *rproc)
>
> data_size = elf_size_of_hdr(class);
> list_for_each_entry(segment, &rproc->dump_segments, node) {
> - data_size += elf_size_of_phdr(class) + segment->size;
> + /*
> + * For default configuration buffer includes headers & segments.
> + * For inline dump buffer just includes headers as segments are
> + * directly read from device memory.
> + */
> + data_size += elf_size_of_phdr(class);
> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
> + data_size += segment->size;
>
> phnum++;
> }
> @@ -182,23 +297,30 @@ void rproc_coredump(struct rproc *rproc)
> elf_phdr_set_p_flags(class, phdr, PF_R | PF_W | PF_X);
> elf_phdr_set_p_align(class, phdr, 0);
>
> - if (segment->dump) {
> - segment->dump(rproc, segment, data + offset);
> - } else {
> - ptr = rproc_da_to_va(rproc, segment->da, segment->size);
> - if (!ptr) {
> - dev_err(&rproc->dev,
> - "invalid coredump segment (%pad, %zu)\n",
> - &segment->da, segment->size);
> - memset(data + offset, 0xff, segment->size);
> - } else {
> - memcpy(data + offset, ptr, segment->size);
> - }
> - }
> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
> + rproc_copy_segment(rproc, data + offset, segment, 0,
> + segment->size);
>
> offset += elf_phdr_get_p_filesz(class, phdr);
> phdr += elf_size_of_phdr(class);
> }
>
> - dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
> + if (dump_conf == RPROC_COREDUMP_DEFAULT) {
> + dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
> + return;
> + }
> +
> + /* Initialize the dump state struct to be used by rproc_coredump_read */
> + dump_state.rproc = rproc;
> + dump_state.header = data;
> + init_completion(&dump_state.dump_done);
> +
> + dev_coredumpm(&rproc->dev, NULL, &dump_state, data_size, GFP_KERNEL,
> + rproc_coredump_read, rproc_coredump_free);
> +
> + /*
> + * Wait until the dump is read and free is called. Data is freed
> + * by devcoredump framework automatically after 5 minutes.
> + */
> + wait_for_completion(&dump_state.dump_done);
> }
> diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
> index e7b7bab..43e45a3 100644
> --- a/include/linux/remoteproc.h
> +++ b/include/linux/remoteproc.h
> @@ -435,6 +435,20 @@ enum rproc_crash_type {
> };
>
> /**
> + * enum rproc_dump_mechanism - Coredump options for core
> + * @RPROC_COREDUMP_DEFAULT: Copy dump to separate buffer and carry on with
> + recovery
> + * @RPROC_COREDUMP_INLINE: Read segments directly from device memory. Stall
> + recovery until all segments are read
> + * @RPROC_COREDUMP_DISABLED: Don't perform any dump
> + */
> +enum rproc_dump_mechanism {
> + RPROC_COREDUMP_DEFAULT,
> + RPROC_COREDUMP_INLINE,
> + RPROC_COREDUMP_DISABLED,
> +};
> +
> +/**
> * struct rproc_dump_segment - segment info from ELF header
> * @node: list node related to the rproc segment list
> * @da: device address of the segment
> @@ -451,7 +465,7 @@ struct rproc_dump_segment {
>
> void *priv;
> void (*dump)(struct rproc *rproc, struct rproc_dump_segment *segment,
> - void *dest);
> + void *dest, size_t offset, size_t size);
> loff_t offset;
> };
>
> @@ -466,6 +480,7 @@ struct rproc_dump_segment {
> * @dev: virtual device for refcounting and common remoteproc behavior
> * @power: refcount of users who need this rproc powered up
> * @state: state of the device
> + * @dump_conf: Currenlty selected coredump configuration
> * @lock: lock which protects concurrent manipulations of the rproc
> * @dbg_dir: debugfs directory of this rproc device
> * @traces: list of trace buffers
> @@ -499,6 +514,7 @@ struct rproc {
> struct device dev;
> atomic_t power;
> unsigned int state;
> + enum rproc_dump_mechanism dump_conf;
> struct mutex lock;
> struct dentry *dbg_dir;
> struct list_head traces;
> @@ -630,7 +646,8 @@ int rproc_coredump_add_custom_segment(struct rproc *rproc,
> dma_addr_t da, size_t size,
> void (*dumpfn)(struct rproc *rproc,
> struct rproc_dump_segment *segment,
> - void *dest),
> + void *dest, size_t offset,
> + size_t size),
> void *priv);
> int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16 machine);
>
> --
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> a Linux Foundation Collaborative Project
>

2020-07-07 17:15:10

by Rishabh Bhatnagar

[permalink] [raw]
Subject: Re: [PATCH v6 2/3] remoteproc: Add inline coredump functionality

On 2020-07-06 10:47, Mathieu Poirier wrote:
> On Mon, Jun 29, 2020 at 01:02:12PM -0700, Rishabh Bhatnagar wrote:
>> The current coredump implementation uses vmalloc area to copy
>> all the segments. But this might put strain on low memory targets
>> as the firmware size sometimes is in tens of MBs. The situation
>> becomes worse if there are multiple remote processors undergoing
>> recovery at the same time. This patch adds inline coredump
>> functionality that avoids extra memory usage. This requires
>> recovery to be halted until data is read by userspace and free
>> function is called.
>>
>> Signed-off-by: Rishabh Bhatnagar <[email protected]>
>> ---
>> drivers/remoteproc/qcom_q6v5_mss.c | 9 +-
>> drivers/remoteproc/remoteproc_coredump.c | 160
>> +++++++++++++++++++++++++++----
>> include/linux/remoteproc.h | 21 +++-
>> 3 files changed, 165 insertions(+), 25 deletions(-)
>>
>> diff --git a/drivers/remoteproc/qcom_q6v5_mss.c
>> b/drivers/remoteproc/qcom_q6v5_mss.c
>> index 903b2bb..d4ff9b8 100644
>> --- a/drivers/remoteproc/qcom_q6v5_mss.c
>> +++ b/drivers/remoteproc/qcom_q6v5_mss.c
>> @@ -1200,12 +1200,13 @@ static int q6v5_mpss_load(struct q6v5 *qproc)
>>
>> static void qcom_q6v5_dump_segment(struct rproc *rproc,
>> struct rproc_dump_segment *segment,
>> - void *dest)
>> + void *dest, size_t cp_offset, size_t size)
>> {
>> int ret = 0;
>> struct q6v5 *qproc = rproc->priv;
>> unsigned long mask = BIT((unsigned long)segment->priv);
>> int offset = segment->da - qproc->mpss_reloc;
>> + size_t cp_size = size ? size : segment->size;
>> void *ptr = NULL;
>
> On the V4 of this set the above line was:
>
> + void *ptr = rproc_da_to_va(rproc, segment->da + offset,
> copy_size);
>
> Back then both Bjorn and I had RB'ed this set and all that was required
> was a
> rebase to linux-next. On V5 and V6 our RBs have been removed, the
> above has
> been changed and an iounmap() was been added below. Yet nothing in the
> cover
> letter provides an explanation that justifies the modification.
>
> What is going on here?
>
Hi Mathieu,
The qcom_q6v5_dump_segment code has changed in the latest tip. That's
why I had to
make modifications to my patch. Sibi reviewed this patch as he made the
above
changes to mss driver and he has reviewed and tested from his side. He
has
also created a minor fix in the driver to make inline dumps work for
mss.
https://patchwork.kernel.org/patch/11637159/

I removed the reviewed-by because I made some modifications to this
patch
and was not sure if the reviewed-by still applied. Let me know if i
should add it
again.
>>
>> /* Unlock mba before copying segments */
>> @@ -1221,13 +1222,13 @@ static void qcom_q6v5_dump_segment(struct
>> rproc *rproc,
>> }
>>
>> if (!ret)
>> - ptr = ioremap_wc(qproc->mpss_phys + offset, segment->size);
>> + ptr = ioremap_wc(qproc->mpss_phys + offset + cp_offset, cp_size);
>>
>> if (ptr) {
>> - memcpy(dest, ptr, segment->size);
>> + memcpy(dest, ptr, cp_size);
>> iounmap(ptr);
>> } else {
>> - memset(dest, 0xff, segment->size);
>> + memset(dest, 0xff, cp_size);
>> }
>>
>> qproc->dump_segment_mask |= mask;
>> diff --git a/drivers/remoteproc/remoteproc_coredump.c
>> b/drivers/remoteproc/remoteproc_coredump.c
>> index ded0244..646886f 100644
>> --- a/drivers/remoteproc/remoteproc_coredump.c
>> +++ b/drivers/remoteproc/remoteproc_coredump.c
>> @@ -5,6 +5,7 @@
>> * Copyright (c) 2020, The Linux Foundation. All rights reserved.
>> */
>>
>> +#include <linux/completion.h>
>> #include <linux/devcoredump.h>
>> #include <linux/device.h>
>> #include <linux/kernel.h>
>> @@ -12,6 +13,12 @@
>> #include "remoteproc_internal.h"
>> #include "remoteproc_elf_helpers.h"
>>
>> +struct rproc_coredump_state {
>> + struct rproc *rproc;
>> + void *header;
>> + struct completion dump_done;
>> +};
>> +
>> /**
>> * rproc_coredump_cleanup() - clean up dump_segments list
>> * @rproc: the remote processor handle
>> @@ -72,7 +79,8 @@ int rproc_coredump_add_custom_segment(struct rproc
>> *rproc,
>> dma_addr_t da, size_t size,
>> void (*dumpfn)(struct rproc *rproc,
>> struct rproc_dump_segment *segment,
>> - void *dest),
>> + void *dest, size_t offset,
>> + size_t size),
>> void *priv)
>> {
>> struct rproc_dump_segment *segment;
>> @@ -114,12 +122,110 @@ int rproc_coredump_set_elf_info(struct rproc
>> *rproc, u8 class, u16 machine)
>> }
>> EXPORT_SYMBOL(rproc_coredump_set_elf_info);
>>
>> +static void rproc_coredump_free(void *data)
>> +{
>> + struct rproc_coredump_state *dump_state = data;
>> +
>> + complete(&dump_state->dump_done);
>> + vfree(dump_state->header);
>> +}
>> +
>> +static void *rproc_coredump_find_segment(loff_t user_offset,
>> + struct list_head *segments,
>> + size_t *data_left)
>> +{
>> + struct rproc_dump_segment *segment;
>> +
>> + list_for_each_entry(segment, segments, node) {
>> + if (user_offset < segment->size) {
>> + *data_left = segment->size - user_offset;
>> + return segment;
>> + }
>> + user_offset -= segment->size;
>> + }
>> +
>> + *data_left = 0;
>> + return NULL;
>> +}
>> +
>> +static void rproc_copy_segment(struct rproc *rproc, void *dest,
>> + struct rproc_dump_segment *segment,
>> + size_t offset, size_t size)
>> +{
>> + void *ptr;
>> +
>> + if (segment->dump) {
>> + segment->dump(rproc, segment, dest, offset, size);
>> + } else {
>> + ptr = rproc_da_to_va(rproc, segment->da + offset, size);
>> + if (!ptr) {
>> + dev_err(&rproc->dev,
>> + "invalid copy request for segment %pad with offset %zu and size
>> %zu)\n",
>> + &segment->da, offset, size);
>> + memset(dest, 0xff, size);
>> + } else {
>> + memcpy(dest, ptr, size);
>> + }
>> + }
>> +}
>> +
>> +static ssize_t rproc_coredump_read(char *buffer, loff_t offset,
>> size_t count,
>> + void *data, size_t header_sz)
>> +{
>> + size_t seg_data, bytes_left = count;
>> + ssize_t copy_sz;
>> + struct rproc_dump_segment *seg;
>> + struct rproc_coredump_state *dump_state = data;
>> + struct rproc *rproc = dump_state->rproc;
>> + void *elfcore = dump_state->header;
>> +
>> + /* Copy the vmalloc'ed header first. */
>> + if (offset < header_sz) {
>> + copy_sz = memory_read_from_buffer(buffer, count, &offset,
>> + elfcore, header_sz);
>> +
>> + return copy_sz;
>> + }
>> +
>> + /*
>> + * Find out the segment memory chunk to be copied based on offset.
>> + * Keep copying data until count bytes are read.
>> + */
>> + while (bytes_left) {
>> + seg = rproc_coredump_find_segment(offset - header_sz,
>> + &rproc->dump_segments,
>> + &seg_data);
>> + /* EOF check */
>> + if (!seg) {
>> + dev_info(&rproc->dev, "Ramdump done, %lld bytes read",
>> + offset);
>> + break;
>> + }
>> +
>> + copy_sz = min_t(size_t, bytes_left, seg_data);
>> +
>> + rproc_copy_segment(rproc, buffer, seg, seg->size - seg_data,
>> + copy_sz);
>> +
>> + offset += copy_sz;
>> + buffer += copy_sz;
>> + bytes_left -= copy_sz;
>> + }
>> +
>> + return count - bytes_left;
>> +}
>> +
>> /**
>> * rproc_coredump() - perform coredump
>> * @rproc: rproc handle
>> *
>> * This function will generate an ELF header for the registered
>> segments
>> - * and create a devcoredump device associated with rproc.
>> + * and create a devcoredump device associated with rproc. Based on
>> the
>> + * coredump configuration this function will directly copy the
>> segments
>> + * from device memory to userspace or copy segments from device
>> memory to
>> + * a separate buffer, which can then be read by userspace.
>> + * The first approach avoids using extra vmalloc memory. But it will
>> stall
>> + * recovery flow until dump is read by userspace.
>> */
>> void rproc_coredump(struct rproc *rproc)
>> {
>> @@ -129,11 +235,13 @@ void rproc_coredump(struct rproc *rproc)
>> size_t data_size;
>> size_t offset;
>> void *data;
>> - void *ptr;
>> u8 class = rproc->elf_class;
>> int phnum = 0;
>> + struct rproc_coredump_state dump_state;
>> + enum rproc_dump_mechanism dump_conf = rproc->dump_conf;
>>
>> - if (list_empty(&rproc->dump_segments))
>> + if (list_empty(&rproc->dump_segments) ||
>> + dump_conf == RPROC_COREDUMP_DISABLED)
>> return;
>>
>> if (class == ELFCLASSNONE) {
>> @@ -143,7 +251,14 @@ void rproc_coredump(struct rproc *rproc)
>>
>> data_size = elf_size_of_hdr(class);
>> list_for_each_entry(segment, &rproc->dump_segments, node) {
>> - data_size += elf_size_of_phdr(class) + segment->size;
>> + /*
>> + * For default configuration buffer includes headers & segments.
>> + * For inline dump buffer just includes headers as segments are
>> + * directly read from device memory.
>> + */
>> + data_size += elf_size_of_phdr(class);
>> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
>> + data_size += segment->size;
>>
>> phnum++;
>> }
>> @@ -182,23 +297,30 @@ void rproc_coredump(struct rproc *rproc)
>> elf_phdr_set_p_flags(class, phdr, PF_R | PF_W | PF_X);
>> elf_phdr_set_p_align(class, phdr, 0);
>>
>> - if (segment->dump) {
>> - segment->dump(rproc, segment, data + offset);
>> - } else {
>> - ptr = rproc_da_to_va(rproc, segment->da, segment->size);
>> - if (!ptr) {
>> - dev_err(&rproc->dev,
>> - "invalid coredump segment (%pad, %zu)\n",
>> - &segment->da, segment->size);
>> - memset(data + offset, 0xff, segment->size);
>> - } else {
>> - memcpy(data + offset, ptr, segment->size);
>> - }
>> - }
>> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
>> + rproc_copy_segment(rproc, data + offset, segment, 0,
>> + segment->size);
>>
>> offset += elf_phdr_get_p_filesz(class, phdr);
>> phdr += elf_size_of_phdr(class);
>> }
>>
>> - dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
>> + if (dump_conf == RPROC_COREDUMP_DEFAULT) {
>> + dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
>> + return;
>> + }
>> +
>> + /* Initialize the dump state struct to be used by
>> rproc_coredump_read */
>> + dump_state.rproc = rproc;
>> + dump_state.header = data;
>> + init_completion(&dump_state.dump_done);
>> +
>> + dev_coredumpm(&rproc->dev, NULL, &dump_state, data_size, GFP_KERNEL,
>> + rproc_coredump_read, rproc_coredump_free);
>> +
>> + /*
>> + * Wait until the dump is read and free is called. Data is freed
>> + * by devcoredump framework automatically after 5 minutes.
>> + */
>> + wait_for_completion(&dump_state.dump_done);
>> }
>> diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
>> index e7b7bab..43e45a3 100644
>> --- a/include/linux/remoteproc.h
>> +++ b/include/linux/remoteproc.h
>> @@ -435,6 +435,20 @@ enum rproc_crash_type {
>> };
>>
>> /**
>> + * enum rproc_dump_mechanism - Coredump options for core
>> + * @RPROC_COREDUMP_DEFAULT: Copy dump to separate buffer and carry on
>> with
>> + recovery
>> + * @RPROC_COREDUMP_INLINE: Read segments directly from device memory.
>> Stall
>> + recovery until all segments are read
>> + * @RPROC_COREDUMP_DISABLED: Don't perform any dump
>> + */
>> +enum rproc_dump_mechanism {
>> + RPROC_COREDUMP_DEFAULT,
>> + RPROC_COREDUMP_INLINE,
>> + RPROC_COREDUMP_DISABLED,
>> +};
>> +
>> +/**
>> * struct rproc_dump_segment - segment info from ELF header
>> * @node: list node related to the rproc segment list
>> * @da: device address of the segment
>> @@ -451,7 +465,7 @@ struct rproc_dump_segment {
>>
>> void *priv;
>> void (*dump)(struct rproc *rproc, struct rproc_dump_segment
>> *segment,
>> - void *dest);
>> + void *dest, size_t offset, size_t size);
>> loff_t offset;
>> };
>>
>> @@ -466,6 +480,7 @@ struct rproc_dump_segment {
>> * @dev: virtual device for refcounting and common remoteproc
>> behavior
>> * @power: refcount of users who need this rproc powered up
>> * @state: state of the device
>> + * @dump_conf: Currenlty selected coredump configuration
>> * @lock: lock which protects concurrent manipulations of the rproc
>> * @dbg_dir: debugfs directory of this rproc device
>> * @traces: list of trace buffers
>> @@ -499,6 +514,7 @@ struct rproc {
>> struct device dev;
>> atomic_t power;
>> unsigned int state;
>> + enum rproc_dump_mechanism dump_conf;
>> struct mutex lock;
>> struct dentry *dbg_dir;
>> struct list_head traces;
>> @@ -630,7 +646,8 @@ int rproc_coredump_add_custom_segment(struct rproc
>> *rproc,
>> dma_addr_t da, size_t size,
>> void (*dumpfn)(struct rproc *rproc,
>> struct rproc_dump_segment *segment,
>> - void *dest),
>> + void *dest, size_t offset,
>> + size_t size),
>> void *priv);
>> int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16
>> machine);
>>
>> --
>> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
>> Forum,
>> a Linux Foundation Collaborative Project
>>

2020-07-07 21:58:52

by Mathieu Poirier

[permalink] [raw]
Subject: Re: [PATCH v6 2/3] remoteproc: Add inline coredump functionality

On Tue, 7 Jul 2020 at 11:12, <[email protected]> wrote:
>
> On 2020-07-06 10:47, Mathieu Poirier wrote:
> > On Mon, Jun 29, 2020 at 01:02:12PM -0700, Rishabh Bhatnagar wrote:
> >> The current coredump implementation uses vmalloc area to copy
> >> all the segments. But this might put strain on low memory targets
> >> as the firmware size sometimes is in tens of MBs. The situation
> >> becomes worse if there are multiple remote processors undergoing
> >> recovery at the same time. This patch adds inline coredump
> >> functionality that avoids extra memory usage. This requires
> >> recovery to be halted until data is read by userspace and free
> >> function is called.
> >>
> >> Signed-off-by: Rishabh Bhatnagar <[email protected]>
> >> ---
> >> drivers/remoteproc/qcom_q6v5_mss.c | 9 +-
> >> drivers/remoteproc/remoteproc_coredump.c | 160
> >> +++++++++++++++++++++++++++----
> >> include/linux/remoteproc.h | 21 +++-
> >> 3 files changed, 165 insertions(+), 25 deletions(-)
> >>
> >> diff --git a/drivers/remoteproc/qcom_q6v5_mss.c
> >> b/drivers/remoteproc/qcom_q6v5_mss.c
> >> index 903b2bb..d4ff9b8 100644
> >> --- a/drivers/remoteproc/qcom_q6v5_mss.c
> >> +++ b/drivers/remoteproc/qcom_q6v5_mss.c
> >> @@ -1200,12 +1200,13 @@ static int q6v5_mpss_load(struct q6v5 *qproc)
> >>
> >> static void qcom_q6v5_dump_segment(struct rproc *rproc,
> >> struct rproc_dump_segment *segment,
> >> - void *dest)
> >> + void *dest, size_t cp_offset, size_t size)
> >> {
> >> int ret = 0;
> >> struct q6v5 *qproc = rproc->priv;
> >> unsigned long mask = BIT((unsigned long)segment->priv);
> >> int offset = segment->da - qproc->mpss_reloc;
> >> + size_t cp_size = size ? size : segment->size;
> >> void *ptr = NULL;
> >
> > On the V4 of this set the above line was:
> >
> > + void *ptr = rproc_da_to_va(rproc, segment->da + offset,
> > copy_size);
> >
> > Back then both Bjorn and I had RB'ed this set and all that was required
> > was a
> > rebase to linux-next. On V5 and V6 our RBs have been removed, the
> > above has
> > been changed and an iounmap() was been added below. Yet nothing in the
> > cover
> > letter provides an explanation that justifies the modification.
> >
> > What is going on here?
> >
> Hi Mathieu,
> The qcom_q6v5_dump_segment code has changed in the latest tip. That's
> why I had to
> make modifications to my patch. Sibi reviewed this patch as he made the
> above
> changes to mss driver and he has reviewed and tested from his side. He
> has
> also created a minor fix in the driver to make inline dumps work for
> mss.
> https://patchwork.kernel.org/patch/11637159/
>
> I removed the reviewed-by because I made some modifications to this
> patch
> and was not sure if the reviewed-by still applied. Let me know if i
> should add it
> again.

Removing the reviewed-by was the right thing to do but not mentioning
that something had changed in the cover letter was perplexing. I had
to fish out previous revisions and compare each line to figure out
what has happened, an exercise that was time consuming.

I will take another look at your patchset but it may not be until next week.

Thanks,
Mathieu

> >>
> >> /* Unlock mba before copying segments */
> >> @@ -1221,13 +1222,13 @@ static void qcom_q6v5_dump_segment(struct
> >> rproc *rproc,
> >> }
> >>
> >> if (!ret)
> >> - ptr = ioremap_wc(qproc->mpss_phys + offset, segment->size);
> >> + ptr = ioremap_wc(qproc->mpss_phys + offset + cp_offset, cp_size);
> >>
> >> if (ptr) {
> >> - memcpy(dest, ptr, segment->size);
> >> + memcpy(dest, ptr, cp_size);
> >> iounmap(ptr);
> >> } else {
> >> - memset(dest, 0xff, segment->size);
> >> + memset(dest, 0xff, cp_size);
> >> }
> >>
> >> qproc->dump_segment_mask |= mask;
> >> diff --git a/drivers/remoteproc/remoteproc_coredump.c
> >> b/drivers/remoteproc/remoteproc_coredump.c
> >> index ded0244..646886f 100644
> >> --- a/drivers/remoteproc/remoteproc_coredump.c
> >> +++ b/drivers/remoteproc/remoteproc_coredump.c
> >> @@ -5,6 +5,7 @@
> >> * Copyright (c) 2020, The Linux Foundation. All rights reserved.
> >> */
> >>
> >> +#include <linux/completion.h>
> >> #include <linux/devcoredump.h>
> >> #include <linux/device.h>
> >> #include <linux/kernel.h>
> >> @@ -12,6 +13,12 @@
> >> #include "remoteproc_internal.h"
> >> #include "remoteproc_elf_helpers.h"
> >>
> >> +struct rproc_coredump_state {
> >> + struct rproc *rproc;
> >> + void *header;
> >> + struct completion dump_done;
> >> +};
> >> +
> >> /**
> >> * rproc_coredump_cleanup() - clean up dump_segments list
> >> * @rproc: the remote processor handle
> >> @@ -72,7 +79,8 @@ int rproc_coredump_add_custom_segment(struct rproc
> >> *rproc,
> >> dma_addr_t da, size_t size,
> >> void (*dumpfn)(struct rproc *rproc,
> >> struct rproc_dump_segment *segment,
> >> - void *dest),
> >> + void *dest, size_t offset,
> >> + size_t size),
> >> void *priv)
> >> {
> >> struct rproc_dump_segment *segment;
> >> @@ -114,12 +122,110 @@ int rproc_coredump_set_elf_info(struct rproc
> >> *rproc, u8 class, u16 machine)
> >> }
> >> EXPORT_SYMBOL(rproc_coredump_set_elf_info);
> >>
> >> +static void rproc_coredump_free(void *data)
> >> +{
> >> + struct rproc_coredump_state *dump_state = data;
> >> +
> >> + complete(&dump_state->dump_done);
> >> + vfree(dump_state->header);
> >> +}
> >> +
> >> +static void *rproc_coredump_find_segment(loff_t user_offset,
> >> + struct list_head *segments,
> >> + size_t *data_left)
> >> +{
> >> + struct rproc_dump_segment *segment;
> >> +
> >> + list_for_each_entry(segment, segments, node) {
> >> + if (user_offset < segment->size) {
> >> + *data_left = segment->size - user_offset;
> >> + return segment;
> >> + }
> >> + user_offset -= segment->size;
> >> + }
> >> +
> >> + *data_left = 0;
> >> + return NULL;
> >> +}
> >> +
> >> +static void rproc_copy_segment(struct rproc *rproc, void *dest,
> >> + struct rproc_dump_segment *segment,
> >> + size_t offset, size_t size)
> >> +{
> >> + void *ptr;
> >> +
> >> + if (segment->dump) {
> >> + segment->dump(rproc, segment, dest, offset, size);
> >> + } else {
> >> + ptr = rproc_da_to_va(rproc, segment->da + offset, size);
> >> + if (!ptr) {
> >> + dev_err(&rproc->dev,
> >> + "invalid copy request for segment %pad with offset %zu and size
> >> %zu)\n",
> >> + &segment->da, offset, size);
> >> + memset(dest, 0xff, size);
> >> + } else {
> >> + memcpy(dest, ptr, size);
> >> + }
> >> + }
> >> +}
> >> +
> >> +static ssize_t rproc_coredump_read(char *buffer, loff_t offset,
> >> size_t count,
> >> + void *data, size_t header_sz)
> >> +{
> >> + size_t seg_data, bytes_left = count;
> >> + ssize_t copy_sz;
> >> + struct rproc_dump_segment *seg;
> >> + struct rproc_coredump_state *dump_state = data;
> >> + struct rproc *rproc = dump_state->rproc;
> >> + void *elfcore = dump_state->header;
> >> +
> >> + /* Copy the vmalloc'ed header first. */
> >> + if (offset < header_sz) {
> >> + copy_sz = memory_read_from_buffer(buffer, count, &offset,
> >> + elfcore, header_sz);
> >> +
> >> + return copy_sz;
> >> + }
> >> +
> >> + /*
> >> + * Find out the segment memory chunk to be copied based on offset.
> >> + * Keep copying data until count bytes are read.
> >> + */
> >> + while (bytes_left) {
> >> + seg = rproc_coredump_find_segment(offset - header_sz,
> >> + &rproc->dump_segments,
> >> + &seg_data);
> >> + /* EOF check */
> >> + if (!seg) {
> >> + dev_info(&rproc->dev, "Ramdump done, %lld bytes read",
> >> + offset);
> >> + break;
> >> + }
> >> +
> >> + copy_sz = min_t(size_t, bytes_left, seg_data);
> >> +
> >> + rproc_copy_segment(rproc, buffer, seg, seg->size - seg_data,
> >> + copy_sz);
> >> +
> >> + offset += copy_sz;
> >> + buffer += copy_sz;
> >> + bytes_left -= copy_sz;
> >> + }
> >> +
> >> + return count - bytes_left;
> >> +}
> >> +
> >> /**
> >> * rproc_coredump() - perform coredump
> >> * @rproc: rproc handle
> >> *
> >> * This function will generate an ELF header for the registered
> >> segments
> >> - * and create a devcoredump device associated with rproc.
> >> + * and create a devcoredump device associated with rproc. Based on
> >> the
> >> + * coredump configuration this function will directly copy the
> >> segments
> >> + * from device memory to userspace or copy segments from device
> >> memory to
> >> + * a separate buffer, which can then be read by userspace.
> >> + * The first approach avoids using extra vmalloc memory. But it will
> >> stall
> >> + * recovery flow until dump is read by userspace.
> >> */
> >> void rproc_coredump(struct rproc *rproc)
> >> {
> >> @@ -129,11 +235,13 @@ void rproc_coredump(struct rproc *rproc)
> >> size_t data_size;
> >> size_t offset;
> >> void *data;
> >> - void *ptr;
> >> u8 class = rproc->elf_class;
> >> int phnum = 0;
> >> + struct rproc_coredump_state dump_state;
> >> + enum rproc_dump_mechanism dump_conf = rproc->dump_conf;
> >>
> >> - if (list_empty(&rproc->dump_segments))
> >> + if (list_empty(&rproc->dump_segments) ||
> >> + dump_conf == RPROC_COREDUMP_DISABLED)
> >> return;
> >>
> >> if (class == ELFCLASSNONE) {
> >> @@ -143,7 +251,14 @@ void rproc_coredump(struct rproc *rproc)
> >>
> >> data_size = elf_size_of_hdr(class);
> >> list_for_each_entry(segment, &rproc->dump_segments, node) {
> >> - data_size += elf_size_of_phdr(class) + segment->size;
> >> + /*
> >> + * For default configuration buffer includes headers & segments.
> >> + * For inline dump buffer just includes headers as segments are
> >> + * directly read from device memory.
> >> + */
> >> + data_size += elf_size_of_phdr(class);
> >> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
> >> + data_size += segment->size;
> >>
> >> phnum++;
> >> }
> >> @@ -182,23 +297,30 @@ void rproc_coredump(struct rproc *rproc)
> >> elf_phdr_set_p_flags(class, phdr, PF_R | PF_W | PF_X);
> >> elf_phdr_set_p_align(class, phdr, 0);
> >>
> >> - if (segment->dump) {
> >> - segment->dump(rproc, segment, data + offset);
> >> - } else {
> >> - ptr = rproc_da_to_va(rproc, segment->da, segment->size);
> >> - if (!ptr) {
> >> - dev_err(&rproc->dev,
> >> - "invalid coredump segment (%pad, %zu)\n",
> >> - &segment->da, segment->size);
> >> - memset(data + offset, 0xff, segment->size);
> >> - } else {
> >> - memcpy(data + offset, ptr, segment->size);
> >> - }
> >> - }
> >> + if (dump_conf == RPROC_COREDUMP_DEFAULT)
> >> + rproc_copy_segment(rproc, data + offset, segment, 0,
> >> + segment->size);
> >>
> >> offset += elf_phdr_get_p_filesz(class, phdr);
> >> phdr += elf_size_of_phdr(class);
> >> }
> >>
> >> - dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
> >> + if (dump_conf == RPROC_COREDUMP_DEFAULT) {
> >> + dev_coredumpv(&rproc->dev, data, data_size, GFP_KERNEL);
> >> + return;
> >> + }
> >> +
> >> + /* Initialize the dump state struct to be used by
> >> rproc_coredump_read */
> >> + dump_state.rproc = rproc;
> >> + dump_state.header = data;
> >> + init_completion(&dump_state.dump_done);
> >> +
> >> + dev_coredumpm(&rproc->dev, NULL, &dump_state, data_size, GFP_KERNEL,
> >> + rproc_coredump_read, rproc_coredump_free);
> >> +
> >> + /*
> >> + * Wait until the dump is read and free is called. Data is freed
> >> + * by devcoredump framework automatically after 5 minutes.
> >> + */
> >> + wait_for_completion(&dump_state.dump_done);
> >> }
> >> diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
> >> index e7b7bab..43e45a3 100644
> >> --- a/include/linux/remoteproc.h
> >> +++ b/include/linux/remoteproc.h
> >> @@ -435,6 +435,20 @@ enum rproc_crash_type {
> >> };
> >>
> >> /**
> >> + * enum rproc_dump_mechanism - Coredump options for core
> >> + * @RPROC_COREDUMP_DEFAULT: Copy dump to separate buffer and carry on
> >> with
> >> + recovery
> >> + * @RPROC_COREDUMP_INLINE: Read segments directly from device memory.
> >> Stall
> >> + recovery until all segments are read
> >> + * @RPROC_COREDUMP_DISABLED: Don't perform any dump
> >> + */
> >> +enum rproc_dump_mechanism {
> >> + RPROC_COREDUMP_DEFAULT,
> >> + RPROC_COREDUMP_INLINE,
> >> + RPROC_COREDUMP_DISABLED,
> >> +};
> >> +
> >> +/**
> >> * struct rproc_dump_segment - segment info from ELF header
> >> * @node: list node related to the rproc segment list
> >> * @da: device address of the segment
> >> @@ -451,7 +465,7 @@ struct rproc_dump_segment {
> >>
> >> void *priv;
> >> void (*dump)(struct rproc *rproc, struct rproc_dump_segment
> >> *segment,
> >> - void *dest);
> >> + void *dest, size_t offset, size_t size);
> >> loff_t offset;
> >> };
> >>
> >> @@ -466,6 +480,7 @@ struct rproc_dump_segment {
> >> * @dev: virtual device for refcounting and common remoteproc
> >> behavior
> >> * @power: refcount of users who need this rproc powered up
> >> * @state: state of the device
> >> + * @dump_conf: Currenlty selected coredump configuration
> >> * @lock: lock which protects concurrent manipulations of the rproc
> >> * @dbg_dir: debugfs directory of this rproc device
> >> * @traces: list of trace buffers
> >> @@ -499,6 +514,7 @@ struct rproc {
> >> struct device dev;
> >> atomic_t power;
> >> unsigned int state;
> >> + enum rproc_dump_mechanism dump_conf;
> >> struct mutex lock;
> >> struct dentry *dbg_dir;
> >> struct list_head traces;
> >> @@ -630,7 +646,8 @@ int rproc_coredump_add_custom_segment(struct rproc
> >> *rproc,
> >> dma_addr_t da, size_t size,
> >> void (*dumpfn)(struct rproc *rproc,
> >> struct rproc_dump_segment *segment,
> >> - void *dest),
> >> + void *dest, size_t offset,
> >> + size_t size),
> >> void *priv);
> >> int rproc_coredump_set_elf_info(struct rproc *rproc, u8 class, u16
> >> machine);
> >>
> >> --
> >> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
> >> Forum,
> >> a Linux Foundation Collaborative Project
> >>