Skip to contentSkip to conversation

Extend coredump functionality

PATCH Superseded Rishabh Bhatnagar 3 participants 15 messages Jul 9, 2020 – Jul 20, 2020

Subsystems (2)

v7 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

v7 cover 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.
This patchset also includes Sibi Sankar's patch to deal with chunk sizes lesser than segment size to make inline coredump work for modem.
https://patchwork.kernel.org/patch/11637157/

Changelog:

v7 -> v6:
- Include Sibi's patch as part of this patchset
- Add a linefeed when displaying coredump conf in debugfs
- Fix a typo in remoteproc.h

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

v5 -> v4:
- Rebase on top of linux-next
- Modify qcom_q6v5_mss driver as a result of rebasing on latest tip.

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

Sibi Sankar (1):
remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

 drivers/remoteproc/Makefile              |   1 +
 drivers/remoteproc/qcom_q6v5_mss.c       |  22 +--
 drivers/remoteproc/remoteproc_core.c     | 191 ------------------
 drivers/remoteproc/remoteproc_coredump.c | 326 +++++++++++++++++++++++++++++++
 drivers/remoteproc/remoteproc_debugfs.c  |  90 +++++++++
 drivers/remoteproc/remoteproc_internal.h |   4 +
 include/linux/remoteproc.h               |  21 +-
 7 files changed, 451 insertions(+), 204 deletions(-)
 create mode 100644 drivers/remoteproc/remoteproc_coredump.c
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#1 lore raw

PATCH 2/4 remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

From: Sibi Sankar <[email protected]>

In order to land inline coredump support for mss, the dump_segment function would need to support granularities less than the segment size. This is achieved by replacing mask based tracking with size.

---
 drivers/remoteproc/qcom_q6v5_mss.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +7 -8
index feb70283b..c6ce032 100644
@@ -181,8 +181,8 @@ struct q6v5 {
 	bool running;
 
 	bool dump_mba_loaded;
-	unsigned long dump_segment_mask;
-	unsigned long dump_complete_mask;
+	size_t current_dump_size;
+	size_t total_dump_size;
 
 	phys_addr_t mba_phys;
 	void *mba_region;
@@ -1203,7 +1203,6 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
-	unsigned long mask = BIT((unsigned long)segment->priv);
 	int offset = segment->da - qproc->mpss_reloc;
 	void *ptr = NULL;
 
@@ -1229,10 +1228,10 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
 		memset(dest, 0xff, segment->size);
 	}
 
-	qproc->dump_segment_mask |= mask;
+	qproc->current_dump_size += segment->size;
 
 	/* Reclaim mba after copying segments */
-	if (qproc->dump_segment_mask == qproc->dump_complete_mask) {
+	if (qproc->current_dump_size == qproc->total_dump_size) {
 		if (qproc->dump_mba_loaded) {
 			/* Try to reset ownership back to Q6 */
 			q6v5_xfer_mem_ownership(qproc, &qproc->mpss_perm,
@@ -1274,7 +1273,7 @@ static int q6v5_start(struct rproc *rproc)
 			"Failed to reclaim mba buffer system may become unstable\n");
 
 	/* Reset Dump Segment Mask */
-	qproc->dump_segment_mask = 0;
+	qproc->current_dump_size = 0;
 	qproc->running = true;
 
 	return 0;
@@ -1323,7 +1322,7 @@ static int qcom_q6v5_register_dump_segments(struct rproc *rproc,
 
 	ehdr = (struct elf32_hdr *)fw->data;
 	phdrs = (struct elf32_phdr *)(ehdr + 1);
-	qproc->dump_complete_mask = 0;
+	qproc->total_dump_size = 0;
 
 	for (i = 0; i < ehdr->e_phnum; i++) {
 		phdr = &phdrs[i];
@@ -1338,7 +1337,7 @@ static int qcom_q6v5_register_dump_segments(struct rproc *rproc,
 		if (ret)
 			break;
 
-		qproc->dump_complete_mask |= BIT(i);
+		qproc->total_dump_size += phdr->p_memsz;
 	}
 
 	release_firmware(fw);
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#8 lore raw

PATCH v7 3/4 remoteproc: Add inline coredump functionality

In reply to Rishabh Bhatnagar

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. Also modify the qcom_q6v5_mss driver to include size and offset in the segment dump function.

---
 drivers/remoteproc/qcom_q6v5_mss.c       |  11 ++-
 drivers/remoteproc/remoteproc_coredump.c | 160 +++++++++++++++++++++++++++----
 include/linux/remoteproc.h               |  21 +++-
 3 files changed, 166 insertions(+), 26 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +6 -5
index c6ce032..79df354 100644
@@ -1199,11 +1199,12 @@ 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;
 	int offset = segment->da - qproc->mpss_reloc;
+	size_t cp_size = size ? size : segment->size;
 	void *ptr = NULL;
 
 	/* Unlock mba before copying segments */
@@ -1219,16 +1220,16 @@ 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->current_dump_size += segment->size;
+	qproc->current_dump_size += cp_size;
 
 	/* Reclaim mba after copying segments */
 	if (qproc->current_dump_size == qproc->total_dump_size) {
drivers/remoteproc/remoteproc_coredump.c +141 -19

213 lines not on this page · raw message

include/linux/remoteproc.h +19 -2
index e7b7bab..38d037d 100644
@@ -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: Currently 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);
 
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#14 lore raw

PATCH v7 4/4 remoteproc: Add coredump debugfs entry

In reply to Rishabh Bhatnagar

Add coredump debugfs entry to configure the type of dump that will be collected during recovery. User can select between default or inline coredump functionality. Also coredump collection can be disabled through this interface.
This functionality can be configured differently for different remote processors.

---
 drivers/remoteproc/remoteproc_debugfs.c | 90 +++++++++++++++++++++++++++++++++
 1 file changed, 90 insertions(+)
drivers/remoteproc/remoteproc_debugfs.c +90 -0
index 732770e..16a4196 100644
@@ -28,6 +28,94 @@
 static struct dentry *rproc_dbg;
 
 /*
+ * A coredump-configuration-to-string lookup table, for exposing a
+ * human readable configuration via debugfs. Always keep in sync with
+ * enum rproc_coredump_mechanism
+ */
+static const char * const rproc_coredump_str[] = {
+	[RPROC_COREDUMP_DEFAULT]	= "default",
+	[RPROC_COREDUMP_INLINE]		= "inline",
+	[RPROC_COREDUMP_DISABLED]	= "disabled",
+};
+
+/* Expose the current coredump configuration via debugfs */
+static ssize_t rproc_coredump_read(struct file *filp, char __user *userbuf,
+				   size_t count, loff_t *ppos)
+{
+	struct rproc *rproc = filp->private_data;
+	char buf[20];
+	int len;
+
+	len = scnprintf(buf, sizeof(buf), "%s\n",
+			rproc_coredump_str[rproc->dump_conf]);
+
+	return simple_read_from_buffer(userbuf, count, ppos, buf, len);
+}
+
+/*
+ * By writing to the 'coredump' debugfs entry, we control the behavior of the
+ * coredump mechanism dynamically. The default value of this entry is "default".
+ *
+ * The 'coredump' debugfs entry supports these commands:
+ *
+ * default:	This is the default coredump mechanism. When the remoteproc
+ *		crashes the entire coredump will be copied to a separate buffer
+ *		and exposed to userspace.
+ *
+ * inline:	The coredump will not be copied to a separate buffer and the
+ *		recovery process will have to wait until data is read by
+ *		userspace. But this avoid usage of extra memory.
+ *
+ * disabled:	This will disable coredump. Recovery will proceed without
+ *		collecting any dump.
+ */
+static ssize_t rproc_coredump_write(struct file *filp,
+				    const char __user *user_buf, size_t count,
+				    loff_t *ppos)
+{
+	struct rproc *rproc = filp->private_data;
+	int ret, err = 0;
+	char buf[20];
+
+	if (count > sizeof(buf))
+		return -EINVAL;
+
+	ret = copy_from_user(buf, user_buf, count);
+	if (ret)
+		return -EFAULT;
+
+	/* remove end of line */
+	if (buf[count - 1] == '\n')
+		buf[count - 1] = '\0';
+
+	if (rproc->state == RPROC_CRASHED) {
+		dev_err(&rproc->dev, "can't change coredump configuration\n");
+		err = -EBUSY;
+		goto out;
+	}
+
+	if (!strncmp(buf, "disable", count)) {
+		rproc->dump_conf = RPROC_COREDUMP_DISABLED;
+	} else if (!strncmp(buf, "inline", count)) {
+		rproc->dump_conf = RPROC_COREDUMP_INLINE;
+	} else if (!strncmp(buf, "default", count)) {
+		rproc->dump_conf = RPROC_COREDUMP_DEFAULT;
+	} else {
+		dev_err(&rproc->dev, "Invalid coredump configuration\n");
+		err = -EINVAL;
+	}
+out:
+	return err ? err : count;
+}
+
+static const struct file_operations rproc_coredump_fops = {
+	.read = rproc_coredump_read,
+	.write = rproc_coredump_write,
+	.open = simple_open,
+	.llseek = generic_file_llseek,
+};
+
+/*
  * Some remote processors may support dumping trace logs into a shared
  * memory buffer. We expose this trace buffer using debugfs, so users
  * can easily tell what's going on remotely.
@@ -337,6 +425,8 @@ void rproc_create_debug_dir(struct rproc *rproc)
 			    rproc, &rproc_rsc_table_fops);
 	debugfs_create_file("carveout_memories", 0400, rproc->dbg_dir,
 			    rproc, &rproc_carveouts_fops);
+	debugfs_create_file("coredump", 0600, rproc->dbg_dir,
+			    rproc, &rproc_coredump_fops);
 }
 
 void __init rproc_init_debugfs(void)
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#7 lore raw

PATCH v7 1/4 remoteproc: Move coredump functionality to a new file

In reply to Rishabh Bhatnagar

Move all coredump functionality to an individual file. This is being done so that the current functionality can be extended in future patchsets.

---
 drivers/remoteproc/Makefile              |   1 +
 drivers/remoteproc/remoteproc_core.c     | 191 -----------------------------
 drivers/remoteproc/remoteproc_coredump.c | 204 +++++++++++++++++++++++++++++++
 drivers/remoteproc/remoteproc_internal.h |   4 +
 4 files changed, 209 insertions(+), 191 deletions(-)
 create mode 100644 drivers/remoteproc/remoteproc_coredump.c
drivers/remoteproc/Makefile +1 -0
index e8b886e..8702a4e 100644
@@ -5,6 +5,7 @@
 
 obj-$(CONFIG_REMOTEPROC)		+= remoteproc.o
 remoteproc-y				:= remoteproc_core.o
+remoteproc-y				+= remoteproc_coredump.o
 remoteproc-y				+= remoteproc_debugfs.o
 remoteproc-y				+= remoteproc_sysfs.o
 remoteproc-y				+= remoteproc_virtio.o
drivers/remoteproc/remoteproc_core.c +0 -191

215 lines not on this page · raw message

drivers/remoteproc/remoteproc_coredump.c new file +204 -0

204 lines not on this page · raw message

drivers/remoteproc/remoteproc_internal.h +4 -0
index 4ba7cb5..97d441b 100644
@@ -47,6 +47,10 @@ extern struct class rproc_class;
 int rproc_init_sysfs(void);
 void rproc_exit_sysfs(void);
 
+/* from remoteproc_coredump.c */
+void rproc_coredump_cleanup(struct rproc *rproc);
+void rproc_coredump(struct rproc *rproc);
+
 void rproc_free_vring(struct rproc_vring *rvring);
 int rproc_alloc_vring(struct rproc_vdev *rvdev, int i);
 
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
maintainer
#2 lore raw

Re: [PATCH v1 2/4] remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

In reply to Rishabh Bhatnagar
Quoted 67 lines · Rishabh Bhatnagar

From: Sibi Sankar <[email protected]>

In order to land inline coredump support for mss, the dump_segment function would need to support granularities less than the segment size. This is achieved by replacing mask based tracking with size.

---
 drivers/remoteproc/qcom_q6v5_mss.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +6 -7
index feb70283b..c6ce032 100644
@@ -181,8 +181,8 @@ struct q6v5 {
 	bool running;
 
 	bool dump_mba_loaded;
-	unsigned long dump_segment_mask;
-	unsigned long dump_complete_mask;
+	size_t current_dump_size;
+	size_t total_dump_size;
 
 	phys_addr_t mba_phys;
 	void *mba_region;
@@ -1203,7 +1203,6 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
-	unsigned long mask = BIT((unsigned long)segment->priv);
 	int offset = segment->da - qproc->mpss_reloc;
 	void *ptr = NULL;
 
@@ -1229,10 +1228,10 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
 		memset(dest, 0xff, segment->size);
 	}
 
-	qproc->dump_segment_mask |= mask;
+	qproc->current_dump_size += segment->size;
 
 	/* Reclaim mba after copying segments */
-	if (qproc->dump_segment_mask == qproc->dump_complete_mask) {
+	if (qproc->current_dump_size == qproc->total_dump_size) {
 		if (qproc->dump_mba_loaded) {
 			/* Try to reset ownership back to Q6 */
 			q6v5_xfer_mem_ownership(qproc, &qproc->mpss_perm,
@@ -1274,7 +1273,7 @@ static int q6v5_start(struct rproc *rproc)
 			"Failed to reclaim mba buffer system may become unstable\n");
 
 	/* Reset Dump Segment Mask */
-	qproc->dump_segment_mask = 0;
+	qproc->current_dump_size = 0;
 	qproc->running = true;
 
 	return 0;
@@ -1323,7 +1322,7 @@ static int qcom_q6v5_register_dump_segments(struct rproc *rproc,
 
 	ehdr = (struct elf32_hdr *)fw->data;
 	phdrs = (struct elf32_phdr *)(ehdr + 1);
-	qproc->dump_complete_mask = 0;
+	qproc->total_dump_size = 0;
 
 	for (i = 0; i < ehdr->e_phnum; i++) {
 		phdr = &phdrs[i];

drivers/remoteproc/qcom_q6v5_mss.c static int qcom_q6v5_register_dump_segments(struct rproc *rproc, · line 1338

Quoted 2 lines
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
@@ -1338,7 +1337,7 @@ static int qcom_q6v5_register_dump_segments(struct rproc *rproc,
 		if (ret)
 			break;

There is also no longer a need to carry the 'i' in:

                ret = rproc_coredump_add_custom_segment(rproc, phdr->p_paddr,
                                                        phdr->p_memsz,
                                                        qcom_q6v5_dump_segment,
                                                        (void *)i);
Quoted 10 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -2
-		qproc->dump_complete_mask |= BIT(i);
+		qproc->total_dump_size += phdr->p_memsz;
 	}
 
 	release_firmware(fw);
-- 

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

maintainer
#9 lore raw

Re: [PATCH v7 3/4] remoteproc: Add inline coredump functionality

In reply to Rishabh Bhatnagar
Quoted 23 lines · Rishabh Bhatnagar

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. Also modify the qcom_q6v5_mss driver to include size and offset in the segment dump function.

---
 drivers/remoteproc/qcom_q6v5_mss.c       |  11 ++-
 drivers/remoteproc/remoteproc_coredump.c | 160 +++++++++++++++++++++++++++----
 include/linux/remoteproc.h               |  21 +++-
 3 files changed, 166 insertions(+), 26 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
index c6ce032..79df354 100644

drivers/remoteproc/qcom_q6v5_mss.c static int q6v5_mpss_load(struct q6v5 *qproc) · line 1202

Quoted 5 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -1
@@ -1199,11 +1199,12 @@ 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)

Now that I'm taking another look at this, refactoring qcom_q6v5_dump_segment() should be in another patch - please split them.

drivers/remoteproc/qcom_q6v5_mss.c static int q6v5_mpss_load(struct q6v5 *qproc) · line 1207

Quoted 5 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -0
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
 	int offset = segment->da - qproc->mpss_reloc;
+	size_t cp_size = size ? size : segment->size;

I can't find a scenario where size is 0 - can you explain when such condition occurs?

Thanks,
Mathieu

Quoted 312 lines
drivers/remoteproc/qcom_q6v5_mss.c +4 -4
 	void *ptr = NULL;
 
 	/* Unlock mba before copying segments */
@@ -1219,16 +1220,16 @@ 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->current_dump_size += segment->size;
+	qproc->current_dump_size += cp_size;
 
 	/* Reclaim mba after copying segments */
 	if (qproc->current_dump_size == qproc->total_dump_size) {
drivers/remoteproc/remoteproc_coredump.c +141 -19
index ded0244..646886f 100644
@@ -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);
 }
include/linux/remoteproc.h +19 -2
index e7b7bab..38d037d 100644
@@ -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: Currently 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);
 
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#3 lore raw

Re: [PATCH v1 2/4] remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

In reply to Mathieu Poirier
Quoted 77 lines · Mathieu Poirier
Quoted 70 lines · Rishabh Bhatnagar

From: Sibi Sankar <[email protected]>

In order to land inline coredump support for mss, the dump_segment function would need to support granularities less than the segment size. This is achieved by replacing mask based tracking with size.

---
 drivers/remoteproc/qcom_q6v5_mss.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +4 -5
index feb70283b..c6ce032 100644
@@ -181,8 +181,8 @@ struct q6v5 {
 	bool running;
 
 	bool dump_mba_loaded;
-	unsigned long dump_segment_mask;
-	unsigned long dump_complete_mask;
+	size_t current_dump_size;
+	size_t total_dump_size;
 
 	phys_addr_t mba_phys;
 	void *mba_region;
@@ -1203,7 +1203,6 @@ static void qcom_q6v5_dump_segment(struct rproc 
 *rproc,
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
-	unsigned long mask = BIT((unsigned long)segment->priv);
 	int offset = segment->da - qproc->mpss_reloc;
 	void *ptr = NULL;
@@ -1229,10 +1228,10 @@ static void qcom_q6v5_dump_segment(struct 
 rproc *rproc,
 		memset(dest, 0xff, segment->size);
 	}
 
-	qproc->dump_segment_mask |= mask;
+	qproc->current_dump_size += segment->size;
 
 	/* Reclaim mba after copying segments */
-	if (qproc->dump_segment_mask == qproc->dump_complete_mask) {
+	if (qproc->current_dump_size == qproc->total_dump_size) {
 		if (qproc->dump_mba_loaded) {
 			/* Try to reset ownership back to Q6 */
drivers/remoteproc/qcom_q6v5_mss.c +2 -2
 			q6v5_xfer_mem_ownership(qproc, &qproc->mpss_perm,
@@ -1274,7 +1273,7 @@ static int q6v5_start(struct rproc *rproc)
 			"Failed to reclaim mba buffer system may become unstable\n");
 
 	/* Reset Dump Segment Mask */
-	qproc->dump_segment_mask = 0;
+	qproc->current_dump_size = 0;
 	qproc->running = true;
 
 	return 0;
@@ -1323,7 +1322,7 @@ static int 
 qcom_q6v5_register_dump_segments(struct rproc *rproc,
 
 	ehdr = (struct elf32_hdr *)fw->data;
 	phdrs = (struct elf32_phdr *)(ehdr + 1);
-	qproc->dump_complete_mask = 0;
+	qproc->total_dump_size = 0;
 
 	for (i = 0; i < ehdr->e_phnum; i++) {
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
 		phdr = &phdrs[i];

drivers/remoteproc/qcom_q6v5_mss.c static int · line 1339

Quoted 3 lines
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
@@ -1338,7 +1337,7 @@ static int 
 qcom_q6v5_register_dump_segments(struct rproc *rproc,
 		if (ret)
 			break;

There is also no longer a need to carry the 'i' in:

Quoted 6 lines

ret = rproc_coredump_add_custom_segment(rproc,
phdr->p_paddr,

                                                        phdr->p_memsz,

qcom_q6v5_dump_segment,

                                                        (void *)i);

Mathieu,
Thanks for the review. Will pass in
NULL instead in the next re-spin.

Quoted 10 lines
Quoted 10 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -2
-		qproc->dump_complete_mask |= BIT(i);
+		qproc->total_dump_size += phdr->p_memsz;
 	}
 
 	release_firmware(fw);
--

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

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project.
#10 lore raw

Re: [PATCH v7 3/4] remoteproc: Add inline coredump functionality

In reply to Mathieu Poirier
Quoted 26 lines · Mathieu Poirier
Quoted 25 lines · Rishabh Bhatnagar

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. Also modify the qcom_q6v5_mss driver to include size and offset in the segment dump function.

---
 drivers/remoteproc/qcom_q6v5_mss.c       |  11 ++-
 drivers/remoteproc/remoteproc_coredump.c | 160 
+++++++++++++++++++++++++++----
 include/linux/remoteproc.h               |  21 +++-
 3 files changed, 166 insertions(+), 26 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
index c6ce032..79df354 100644
@@ -1199,11 +1199,12 @@ static int q6v5_mpss_load(struct q6v5 *qproc)
 
Quoted 8 lines

drivers/remoteproc/qcom_q6v5_mss.c static int q6v5_mpss_load(struct q6v5 *qproc) · line 1202

Quoted 4 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -1
 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)

Now that I'm taking another look at this, refactoring qcom_q6v5_dump_segment() should be in another patch - please split them.

Hi Mathieu,
Thanks for reviewing.

Since I'm changing the API signature if we just take this patch without q6v5_mss changes (in a separate patch) it would cause compilation errors. Is that ok?

Quoted 9 lines

drivers/remoteproc/qcom_q6v5_mss.c static int q6v5_mpss_load(struct q6v5 *qproc) · line 1207

Quoted 5 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -0
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
 	int offset = segment->da - qproc->mpss_reloc;
+	size_t cp_size = size ? size : segment->size;

I can't find a scenario where size is 0 - can you explain when such condition occurs?

There is no such case right now. I can get rid of the check.

Quoted 333 lines

Thanks,
Mathieu

Quoted 329 lines
drivers/remoteproc/qcom_q6v5_mss.c +4 -4
 	void *ptr = NULL;
 
 	/* Unlock mba before copying segments */
@@ -1219,16 +1220,16 @@ 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->current_dump_size += segment->size;
+	qproc->current_dump_size += cp_size;
 
 	/* Reclaim mba after copying segments */
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
 	if (qproc->current_dump_size == qproc->total_dump_size) {
drivers/remoteproc/remoteproc_coredump.c +9 -1
index ded0244..646886f 100644
@@ -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)
 {
drivers/remoteproc/remoteproc_coredump.c +93 -0
 	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)
{

drivers/remoteproc/remoteproc_coredump.c +21 -17
@@ -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 */

drivers/remoteproc/remoteproc_coredump.c +12 -0
+	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);
 }
include/linux/remoteproc.h +2 -0
index e7b7bab..38d037d 100644
@@ -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

include/linux/remoteproc.h +2 -0
+				recovery
+ * @RPROC_COREDUMP_INLINE:	Read segments directly from device memory. 

Stall

include/linux/remoteproc.h +11 -1
+				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: Currently 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

include/linux/remoteproc.h +3 -1
@@ -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

maintainer
#11 lore raw

Re: [PATCH v7 3/4] remoteproc: Add inline coredump functionality

Quoted 38 lines · On Wed, 15 Jul 2020 at 13:22, <[email protected]> wrote:
Quoted 27 lines · Mathieu Poirier
Quoted 26 lines · Rishabh Bhatnagar

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. Also modify the qcom_q6v5_mss driver to include size and offset in the segment dump function.

---
 drivers/remoteproc/qcom_q6v5_mss.c       |  11 ++-
 drivers/remoteproc/remoteproc_coredump.c | 160
+++++++++++++++++++++++++++----
 include/linux/remoteproc.h               |  21 +++-
 3 files changed, 166 insertions(+), 26 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
index c6ce032..79df354 100644
@@ -1199,11 +1199,12 @@ static int q6v5_mpss_load(struct q6v5 *qproc)
 
Quoted 9 lines

drivers/remoteproc/qcom_q6v5_mss.c static int q6v5_mpss_load(struct q6v5 *qproc) · line 1202

Quoted 4 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -1
 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)

Now that I'm taking another look at this, refactoring qcom_q6v5_dump_segment()
should be in another patch - please split them.

Quoted 7 lines

Hi Mathieu,
Thanks for reviewing.

Since I'm changing the API signature if we just take this patch without q6v5_mss
changes (in a separate patch) it would cause compilation errors. Is that ok?

Just fix the arguments in q6v5_mss to match the signature and then do another patch to implement the inline coredump feature.

Quoted 345 lines
Quoted 9 lines

drivers/remoteproc/qcom_q6v5_mss.c static int q6v5_mpss_load(struct q6v5 *qproc) · line 1207

Quoted 5 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -0
 {
     int ret = 0;
     struct q6v5 *qproc = rproc->priv;
     int offset = segment->da - qproc->mpss_reloc;
+    size_t cp_size = size ? size : segment->size;

I can't find a scenario where size is 0 - can you explain when such condition occurs?

There is no such case right now. I can get rid of the check.

Quoted 335 lines

Thanks,
Mathieu

Quoted 331 lines
drivers/remoteproc/qcom_q6v5_mss.c +4 -4
     void *ptr = NULL;
 
     /* Unlock mba before copying segments */
@@ -1219,16 +1220,16 @@ 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->current_dump_size += segment->size;
+    qproc->current_dump_size += cp_size;
 
     /* Reclaim mba after copying segments */
     if (qproc->current_dump_size == qproc->total_dump_size) {
drivers/remoteproc/remoteproc_coredump.c +9 -1
index ded0244..646886f 100644
@@ -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;
drivers/remoteproc/remoteproc_coredump.c +93 -0
@@ -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)
{

drivers/remoteproc/remoteproc_coredump.c +21 -17
@@ -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 */

drivers/remoteproc/remoteproc_coredump.c +12 -0
+    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);
 }
include/linux/remoteproc.h +2 -0
index e7b7bab..38d037d 100644
@@ -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

include/linux/remoteproc.h +11 -1
+                            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: Currently 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

include/linux/remoteproc.h +3 -1
@@ -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

#4 lore raw

Re: [PATCH v1 2/4] remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

In reply to Mathieu Poirier
Quoted 77 lines · Mathieu Poirier
Quoted 70 lines · Rishabh Bhatnagar

From: Sibi Sankar <[email protected]>

In order to land inline coredump support for mss, the dump_segment function would need to support granularities less than the segment size. This is achieved by replacing mask based tracking with size.

---
 drivers/remoteproc/qcom_q6v5_mss.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +4 -5
index feb70283b..c6ce032 100644
@@ -181,8 +181,8 @@ struct q6v5 {
 	bool running;
 
 	bool dump_mba_loaded;
-	unsigned long dump_segment_mask;
-	unsigned long dump_complete_mask;
+	size_t current_dump_size;
+	size_t total_dump_size;
 
 	phys_addr_t mba_phys;
 	void *mba_region;
@@ -1203,7 +1203,6 @@ static void qcom_q6v5_dump_segment(struct rproc 
 *rproc,
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
-	unsigned long mask = BIT((unsigned long)segment->priv);
 	int offset = segment->da - qproc->mpss_reloc;
 	void *ptr = NULL;
@@ -1229,10 +1228,10 @@ static void qcom_q6v5_dump_segment(struct 
 rproc *rproc,
 		memset(dest, 0xff, segment->size);
 	}
 
-	qproc->dump_segment_mask |= mask;
+	qproc->current_dump_size += segment->size;
 
 	/* Reclaim mba after copying segments */
-	if (qproc->dump_segment_mask == qproc->dump_complete_mask) {
+	if (qproc->current_dump_size == qproc->total_dump_size) {
 		if (qproc->dump_mba_loaded) {
 			/* Try to reset ownership back to Q6 */
drivers/remoteproc/qcom_q6v5_mss.c +2 -2
 			q6v5_xfer_mem_ownership(qproc, &qproc->mpss_perm,
@@ -1274,7 +1273,7 @@ static int q6v5_start(struct rproc *rproc)
 			"Failed to reclaim mba buffer system may become unstable\n");
 
 	/* Reset Dump Segment Mask */
-	qproc->dump_segment_mask = 0;
+	qproc->current_dump_size = 0;
 	qproc->running = true;
 
 	return 0;
@@ -1323,7 +1322,7 @@ static int 
 qcom_q6v5_register_dump_segments(struct rproc *rproc,
 
 	ehdr = (struct elf32_hdr *)fw->data;
 	phdrs = (struct elf32_phdr *)(ehdr + 1);
-	qproc->dump_complete_mask = 0;
+	qproc->total_dump_size = 0;
 
 	for (i = 0; i < ehdr->e_phnum; i++) {
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
 		phdr = &phdrs[i];

drivers/remoteproc/qcom_q6v5_mss.c static int · line 1339

Quoted 3 lines
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
@@ -1338,7 +1337,7 @@ static int 
 qcom_q6v5_register_dump_segments(struct rproc *rproc,
 		if (ret)
 			break;

There is also no longer a need to carry the 'i' in:

Quoted 6 lines

ret = rproc_coredump_add_custom_segment(rproc,
phdr->p_paddr,

                                                        phdr->p_memsz,

qcom_q6v5_dump_segment,

                                                        (void *)i);

I assume Rishabh will re-spin the
series today and this will be taken
care as well.

Quoted 10 lines
Quoted 10 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -2
-		qproc->dump_complete_mask |= BIT(i);
+		qproc->total_dump_size += phdr->p_memsz;
 	}
 
 	release_firmware(fw);
--

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

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project.
#12 lore raw

Re: [PATCH v7 3/4] remoteproc: Add inline coredump functionality

In reply to Rishabh Bhatnagar
Quoted 98 lines · Rishabh Bhatnagar

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. Also modify the qcom_q6v5_mss driver to include size and offset in the segment dump function.

---
 drivers/remoteproc/qcom_q6v5_mss.c       |  11 ++-
 drivers/remoteproc/remoteproc_coredump.c | 160 
+++++++++++++++++++++++++++----
 include/linux/remoteproc.h               |  21 +++-
 3 files changed, 166 insertions(+), 26 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +6 -5
index c6ce032..79df354 100644
@@ -1199,11 +1199,12 @@ 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;
 	int offset = segment->da - qproc->mpss_reloc;
+	size_t cp_size = size ? size : segment->size;
 	void *ptr = NULL;
 
 	/* Unlock mba before copying segments */
@@ -1219,16 +1220,16 @@ 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->current_dump_size += segment->size;
+	qproc->current_dump_size += cp_size;
 
 	/* Reclaim mba after copying segments */
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
 	if (qproc->current_dump_size == qproc->total_dump_size) {
drivers/remoteproc/remoteproc_coredump.c +9 -1
index ded0244..646886f 100644
@@ -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)
 {
drivers/remoteproc/remoteproc_coredump.c +0 -0
 	struct rproc_dump_segment *segment;

drivers/remoteproc/remoteproc_coredump.c int rproc_coredump_set_elf_info(struct rproc · line 131

Quoted 10 lines
drivers/remoteproc/remoteproc_coredump.c +6 -0
@@ -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);

Rishabh,

I observed a system hang when I trigger
recovery of a number of remoteproc
simultaneously (after a number of iterations
consistently). This goes away if I free the
header and then issue the dump completion
and that does seem to be the right thing to
do.

Quoted 249 lines
drivers/remoteproc/remoteproc_coredump.c +87 -0
+}
+
+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)
{

drivers/remoteproc/remoteproc_coredump.c +33 -17
@@ -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);

}

include/linux/remoteproc.h +2 -0
index e7b7bab..38d037d 100644
@@ -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

include/linux/remoteproc.h +2 -0
+				recovery
+ * @RPROC_COREDUMP_INLINE:	Read segments directly from device memory. 

Stall

include/linux/remoteproc.h +15 -2
+				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: Currently 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);

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project.
#5 lore raw

Re: [PATCH v1 2/4] remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

In reply to Sibi Sankar
Quoted 82 lines · Sibi Sankar
Quoted 81 lines · Mathieu Poirier
Quoted 70 lines · Rishabh Bhatnagar

From: Sibi Sankar <[email protected]>

In order to land inline coredump support for mss, the dump_segment function would need to support granularities less than the segment size. This is achieved by replacing mask based tracking with size.

---
 drivers/remoteproc/qcom_q6v5_mss.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +4 -5
index feb70283b..c6ce032 100644
@@ -181,8 +181,8 @@ struct q6v5 {
 	bool running;
 
 	bool dump_mba_loaded;
-	unsigned long dump_segment_mask;
-	unsigned long dump_complete_mask;
+	size_t current_dump_size;
+	size_t total_dump_size;
 
 	phys_addr_t mba_phys;
 	void *mba_region;
@@ -1203,7 +1203,6 @@ static void qcom_q6v5_dump_segment(struct rproc 
 *rproc,
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
-	unsigned long mask = BIT((unsigned long)segment->priv);
 	int offset = segment->da - qproc->mpss_reloc;
 	void *ptr = NULL;
@@ -1229,10 +1228,10 @@ static void qcom_q6v5_dump_segment(struct 
 rproc *rproc,
 		memset(dest, 0xff, segment->size);
 	}
 
-	qproc->dump_segment_mask |= mask;
+	qproc->current_dump_size += segment->size;
 
 	/* Reclaim mba after copying segments */
-	if (qproc->dump_segment_mask == qproc->dump_complete_mask) {
+	if (qproc->current_dump_size == qproc->total_dump_size) {
 		if (qproc->dump_mba_loaded) {
 			/* Try to reset ownership back to Q6 */
drivers/remoteproc/qcom_q6v5_mss.c +2 -2
 			q6v5_xfer_mem_ownership(qproc, &qproc->mpss_perm,
@@ -1274,7 +1273,7 @@ static int q6v5_start(struct rproc *rproc)
 			"Failed to reclaim mba buffer system may become unstable\n");
 
 	/* Reset Dump Segment Mask */
-	qproc->dump_segment_mask = 0;
+	qproc->current_dump_size = 0;
 	qproc->running = true;
 
 	return 0;
@@ -1323,7 +1322,7 @@ static int 
 qcom_q6v5_register_dump_segments(struct rproc *rproc,
 
 	ehdr = (struct elf32_hdr *)fw->data;
 	phdrs = (struct elf32_phdr *)(ehdr + 1);
-	qproc->dump_complete_mask = 0;
+	qproc->total_dump_size = 0;
 
 	for (i = 0; i < ehdr->e_phnum; i++) {
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
 		phdr = &phdrs[i];

drivers/remoteproc/qcom_q6v5_mss.c static int · line 1339

Quoted 3 lines
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
@@ -1338,7 +1337,7 @@ static int 
 qcom_q6v5_register_dump_segments(struct rproc *rproc,
 		if (ret)
 			break;

There is also no longer a need to carry the 'i' in:

ret = rproc_coredump_add_custom_segment(rproc,
phdr->p_paddr,

                                                        phdr->p_memsz,
Quoted 7 lines
Quoted 2 lines

qcom_q6v5_dump_segment,

                                                        (void *)i);

I assume Rishabh will re-spin the
series today and this will be taken
care as well.

Hi Sibi,
I'll respin and add the fixes.
Thanks

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
drivers/remoteproc/qcom_q6v5_mss.c +1 -2
-		qproc->dump_complete_mask |= BIT(i);
+		qproc->total_dump_size += phdr->p_memsz;
 	}
 
 	release_firmware(fw);
--

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

#13 lore raw

Re: [PATCH v7 3/4] remoteproc: Add inline coredump functionality

In reply to Sibi Sankar
Quoted 112 lines · Sibi Sankar
Quoted 98 lines · Rishabh Bhatnagar

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. Also modify the qcom_q6v5_mss driver to include size and offset in the segment dump function.

---
 drivers/remoteproc/qcom_q6v5_mss.c       |  11 ++-
 drivers/remoteproc/remoteproc_coredump.c | 160 
+++++++++++++++++++++++++++----
 include/linux/remoteproc.h               |  21 +++-
 3 files changed, 166 insertions(+), 26 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +6 -5
index c6ce032..79df354 100644
@@ -1199,11 +1199,12 @@ 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;
 	int offset = segment->da - qproc->mpss_reloc;
+	size_t cp_size = size ? size : segment->size;
 	void *ptr = NULL;
 
 	/* Unlock mba before copying segments */
@@ -1219,16 +1220,16 @@ 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->current_dump_size += segment->size;
+	qproc->current_dump_size += cp_size;
 
 	/* Reclaim mba after copying segments */
drivers/remoteproc/qcom_q6v5_mss.c +0 -0
 	if (qproc->current_dump_size == qproc->total_dump_size) {
drivers/remoteproc/remoteproc_coredump.c +9 -1
index ded0244..646886f 100644
@@ -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)
 {
drivers/remoteproc/remoteproc_coredump.c +0 -0
 	struct rproc_dump_segment *segment;

drivers/remoteproc/remoteproc_coredump.c int rproc_coredump_set_elf_info(struct rproc · line 131

Quoted 10 lines
drivers/remoteproc/remoteproc_coredump.c +6 -0
@@ -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);

Rishabh,

Quoted 8 lines

I observed a system hang when I trigger recovery of a number of remoteproc simultaneously (after a number of iterations consistently). This goes away if I free the header and then issue the dump completion and that does seem to be the right thing to do.

Makes sense i'll switch it in the next patchset.

Quoted 252 lines
Quoted 252 lines
drivers/remoteproc/remoteproc_coredump.c +87 -0
+}
+
+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)
{

drivers/remoteproc/remoteproc_coredump.c +21 -17
@@ -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 */

drivers/remoteproc/remoteproc_coredump.c +12 -0
+	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);
 }
include/linux/remoteproc.h +2 -0
index e7b7bab..38d037d 100644
@@ -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

include/linux/remoteproc.h +2 -0
+				recovery
+ * @RPROC_COREDUMP_INLINE:	Read segments directly from device memory. 

Stall

include/linux/remoteproc.h +11 -1
+				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: Currently 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

include/linux/remoteproc.h +3 -1
@@ -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);

maintainer Acked-by on v1 2/4
#6 lore raw

Re: [PATCH v1 2/4] remoteproc: qcom_q6v5_mss: Replace mask based tracking with size

In reply to Rishabh Bhatnagar
Quoted 8 lines · Rishabh Bhatnagar

From: Sibi Sankar <[email protected]>

In order to land inline coredump support for mss, the dump_segment function would need to support granularities less than the segment size. This is achieved by replacing mask based tracking with size.

Quoted 71 lines
---
 drivers/remoteproc/qcom_q6v5_mss.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/remoteproc/qcom_q6v5_mss.c +7 -8
index feb70283b..c6ce032 100644
@@ -181,8 +181,8 @@ struct q6v5 {
 	bool running;
 
 	bool dump_mba_loaded;
-	unsigned long dump_segment_mask;
-	unsigned long dump_complete_mask;
+	size_t current_dump_size;
+	size_t total_dump_size;
 
 	phys_addr_t mba_phys;
 	void *mba_region;
@@ -1203,7 +1203,6 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
 {
 	int ret = 0;
 	struct q6v5 *qproc = rproc->priv;
-	unsigned long mask = BIT((unsigned long)segment->priv);
 	int offset = segment->da - qproc->mpss_reloc;
 	void *ptr = NULL;
 
@@ -1229,10 +1228,10 @@ static void qcom_q6v5_dump_segment(struct rproc *rproc,
 		memset(dest, 0xff, segment->size);
 	}
 
-	qproc->dump_segment_mask |= mask;
+	qproc->current_dump_size += segment->size;
 
 	/* Reclaim mba after copying segments */
-	if (qproc->dump_segment_mask == qproc->dump_complete_mask) {
+	if (qproc->current_dump_size == qproc->total_dump_size) {
 		if (qproc->dump_mba_loaded) {
 			/* Try to reset ownership back to Q6 */
 			q6v5_xfer_mem_ownership(qproc, &qproc->mpss_perm,
@@ -1274,7 +1273,7 @@ static int q6v5_start(struct rproc *rproc)
 			"Failed to reclaim mba buffer system may become unstable\n");
 
 	/* Reset Dump Segment Mask */
-	qproc->dump_segment_mask = 0;
+	qproc->current_dump_size = 0;
 	qproc->running = true;
 
 	return 0;
@@ -1323,7 +1322,7 @@ static int qcom_q6v5_register_dump_segments(struct rproc *rproc,
 
 	ehdr = (struct elf32_hdr *)fw->data;
 	phdrs = (struct elf32_phdr *)(ehdr + 1);
-	qproc->dump_complete_mask = 0;
+	qproc->total_dump_size = 0;
 
 	for (i = 0; i < ehdr->e_phnum; i++) {
 		phdr = &phdrs[i];
@@ -1338,7 +1337,7 @@ static int qcom_q6v5_register_dump_segments(struct rproc *rproc,
 		if (ret)
 			break;
 
-		qproc->dump_complete_mask |= BIT(i);
+		qproc->total_dump_size += phdr->p_memsz;
 	}
 
 	release_firmware(fw);
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project