Re: [PATCH 2/3] remoteproc: Add inline coredump functionality

[Bjorn Andersson]

On Mon 11 May 17:11 PDT 2020, rishabhb@xxxxxxxxxxxxxx wrote:

> On 2020-05-07 13:21, Bjorn Andersson wrote:
> > On Thu 16 Apr 11:38 PDT 2020, Rishabh Bhatnagar wrote:
> > 
> > > This patch adds the inline coredump functionality. The current
> > > coredump implementation uses vmalloc area to copy all the segments.
> > > But this might put a lot of strain on low memory targets as the
> > > firmware size sometimes is in ten's of MBs. The situation becomes
> > > worse if there are multiple remote processors  undergoing recovery
> > > at the same time. This patch directly copies the device memory to
> > > userspace buffer and 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 <rishabhb@xxxxxxxxxxxxxx>
> > > ---
> > >  drivers/remoteproc/remoteproc_coredump.c | 130
> > > +++++++++++++++++++++++++++++++
> > >  drivers/remoteproc/remoteproc_internal.h |  23 +++++-
> > >  include/linux/remoteproc.h               |   2 +
> > >  3 files changed, 153 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/drivers/remoteproc/remoteproc_coredump.c
> > > b/drivers/remoteproc/remoteproc_coredump.c
> > > index 9de0467..888b7dec91 100644
> > > --- a/drivers/remoteproc/remoteproc_coredump.c
> > > +++ b/drivers/remoteproc/remoteproc_coredump.c
> > > @@ -12,6 +12,84 @@
> > >  #include <linux/remoteproc.h>
> > >  #include "remoteproc_internal.h"
> > > 
> > > +static void rproc_free_dump(void *data)
> > 
> > rproc_coredump_free()
> > 
> > > +{
> > > +	struct rproc_coredump_state *dump_state = data;
> > > +
> > > +	complete(&dump_state->dump_done);
> > 
> > vfree(dump_state->header);
> > 
> > > +}
> > > +
> > > +static unsigned long resolve_addr(loff_t user_offset,
> > 
> > rproc_coredump_find_segment()
> > 
> > > +				   struct list_head *segments,
> > > +				   unsigned long *data_left)
> > > +{
> > > +	struct rproc_dump_segment *segment;
> > > +
> > > +	list_for_each_entry(segment, segments, node) {
> > > +		if (user_offset >= segment->size)
> > > +			user_offset -= segment->size;
> > > +		else
> > > +			break;
> > 
> > 		if (user_offset < segment->size) {
> > 			*data_left = segment->size - user_offset;
> > 			return segment->da + user_offset;
> > 		}
> > 
> > 		user_offset -= segment->size;
> > > +	}
> > 
> > 	*data_left = 0;
> > 	return 0;
> > 
> > > +
> > > +	if (&segment->node == segments) {
> > > +		*data_left = 0;
> > > +		return 0;
> > > +	}
> > > +
> > > +	*data_left = segment->size - user_offset;
> > > +
> > > +	return segment->da + user_offset;
> > > +}
> > > +
> > > +static ssize_t rproc_read_dump(char *buffer, loff_t offset, size_t
> > > count,
> > > +				void *data, size_t header_size)
> > > +{
> > > +	void *device_mem;
> > > +	size_t data_left, copy_size, bytes_left = count;
> > > +	unsigned long addr;
> > > +	struct rproc_coredump_state *dump_state = data;
> > > +	struct rproc *rproc = dump_state->rproc;
> > > +	void *elfcore = dump_state->header;
> > > +
> > > +	/* Copy the header first */
> > > +	if (offset < header_size) {
> > > +		copy_size = header_size - offset;
> > > +		copy_size = min(copy_size, bytes_left);
> > > +
> > > +		memcpy(buffer, elfcore + offset, copy_size);
> > > +		offset += copy_size;
> > > +		bytes_left -= copy_size;
> > > +		buffer += copy_size;
> > > +	}
> > 
> > Perhaps you can take inspiration from devcd_readv() here?
> > 
> > > +
> > > +	while (bytes_left) {
> > > +		addr = resolve_addr(offset - header_size,
> > > +				    &rproc->dump_segments, &data_left);
> > > +		/* EOF check */
> > > +		if (data_left == 0) {
> > 
> > Afaict data_left denotes the amount of data left in this particular
> > segment, rather than in the entire core.
> > 
> Yes, but it only returns 0 when the final segment has been copied
> completely.  Otherwise it gives data left to copy for every segment
> and moves to next segment once the current one is copied.

You're right.

> > I think you should start by making bytes_left the minimum of the core
> > size and @count and then have this loop as long as bytes_left, copying
> > data to the buffer either from header or an appropriate segment based on
> > the current offset.
> > 
> That would require an extra function that calculates entire core size,
> as its not available right now. Do you see any missed corner cases with this
> approach?

You're looping over all the segments as you're building the header
anyways, so you could simply store this in the dump_state. I think this
depend more on the ability to reuse the read function between inline and
default coredump.

Regards,
Bjorn

> > > +			pr_info("Ramdump complete %lld bytes read", offset);
> > 
> > dev_dbg(&rproc->dev, ...)
> > 
> > > +			break;
> > > +		}
> > > +
> > > +		copy_size = min_t(size_t, bytes_left, data_left);
> > > +
> > > +		device_mem = rproc->ops->da_to_va(rproc, addr, copy_size);
> > 
> > rproc_da_to_va()
> > 
> > > +		if (!device_mem) {
> > > +			pr_err("Address:%lx with size %zd out of remoteproc carveout\n",
> > 
> > dev_err(&rproc->dev, "coredump: %#lx size %#zx outside of carveouts\n",
> > ..);
> > 
> > > +				addr, copy_size);
> > > +			return -ENOMEM;
> > > +		}
> > > +		memcpy(buffer, device_mem, copy_size);
> > > +
> > > +		offset += copy_size;
> > > +		buffer += copy_size;
> > > +		bytes_left -= copy_size;
> > > +	}
> > > +
> > > +	return count - bytes_left;
> > > +}
> > > +
> > >  static void create_elf_header(void *data, int phnum, struct rproc
> > > *rproc)
> > >  {
> > >  	struct elf32_phdr *phdr;
> > > @@ -55,6 +133,58 @@ static void create_elf_header(void *data, int
> > > phnum, struct rproc *rproc)
> > >  }
> > > 
> > >  /**
> > > + * rproc_inline_coredump() - perform synchronized coredump
> > > + * @rproc:	rproc handle
> > > + *
> > > + * This function will generate an ELF header for the registered
> > > segments
> > > + * and create a devcoredump device associated with rproc. This
> > > function
> > > + * directly copies the segments from device memory to userspace. The
> > > + * recovery is stalled until the enitire coredump is read. This
> > > approach
> > > + * avoids using extra vmalloc memory(which can be really large).
> > > + */
> > > +void rproc_inline_coredump(struct rproc *rproc)
> > > +{
> > > +	struct rproc_dump_segment *segment;
> > > +	struct elf32_phdr *phdr;
> > > +	struct elf32_hdr *ehdr;
> > > +	struct rproc_coredump_state *dump_state;
> > 
> > This can live on the stack, unless you follow my suggestion below...
> > 
> > > +	size_t header_size;
> > > +	void *data;
> > > +	int phnum = 0;
> > > +
> > > +	if (list_empty(&rproc->dump_segments))
> > > +		return;
> > > +
> > > +	header_size = sizeof(*ehdr);
> > > +	list_for_each_entry(segment, &rproc->dump_segments, node) {
> > > +		header_size += sizeof(*phdr);
> > > +
> > > +		phnum++;
> > > +	}
> > > +
> > > +	data = vmalloc(header_size);
> > > +	if (!data)
> > > +		return;
> > > +
> > > +	ehdr = data;
> > 
> > ehdr is unused.
> > 
> > > +	create_elf_header(data, phnum, rproc);
> > > +
> > > +	dump_state = kzalloc(sizeof(*dump_state), GFP_KERNEL);
> > > +	dump_state->rproc = rproc;
> > > +	dump_state->header = data;
> > > +	init_completion(&dump_state->dump_done);
> > > +
> > > +	dev_coredumpm(&rproc->dev, NULL, dump_state, header_size,
> > > GFP_KERNEL,
> > > +		      rproc_read_dump, rproc_free_dump);
> > 
> > I can help feeling that if you vmalloc() either the header or the entire
> > thing depending on DEFAULT vs INLINE and populate it with either all
> > segments or just the header, then you should be able to use the same
> > (custom) read function to serve both cases.
> > 
> > You should by doing this be able to avoid some duplication, your two
> > code paths would not diverge and the main difference would be if you
> > wait or not below (the kfree would have to go in the rproc_free_dump).
> > 
> > > +
> > > +	/* Wait until the dump is read and free is called */
> > > +	wait_for_completion(&dump_state->dump_done);
> > > +
> > > +	kfree(dump_state);
> > > +}
> > > +EXPORT_SYMBOL(rproc_inline_coredump);
> > > +
> > > +/**
> > >   * rproc_default_coredump() - perform coredump
> > >   * @rproc:	rproc handle
> > >   *
> > > diff --git a/drivers/remoteproc/remoteproc_internal.h
> > > b/drivers/remoteproc/remoteproc_internal.h
> > > index 28b6af2..ea6146e 100644
> > > --- a/drivers/remoteproc/remoteproc_internal.h
> > > +++ b/drivers/remoteproc/remoteproc_internal.h
> > > @@ -24,6 +24,18 @@ struct rproc_debug_trace {
> > >  	struct rproc_mem_entry trace_mem;
> > >  };
> > > 
> > > +struct rproc_coredump_state {
> > 
> > This is only used within remoteproc_coredump.c, so please move it there.
> > 
> > > +	struct rproc *rproc;
> > > +	void *header;
> > > +	struct completion dump_done;
> > > +};
> > > +
> > > +enum rproc_coredump_conf {
> > 
> > How about rproc_coredump_mechanism?
> > 
> > > +	COREDUMP_DEFAULT,
> > > +	COREDUMP_INLINE,
> > > +	COREDUMP_DISABLED,
> > > +};
> > > +
> > >  /* from remoteproc_core.c */
> > >  void rproc_release(struct kref *kref);
> > >  irqreturn_t rproc_vq_interrupt(struct rproc *rproc, int vq_id);
> > > @@ -49,6 +61,7 @@ struct dentry *rproc_create_trace_file(const char
> > > *name, struct rproc *rproc,
> > > 
> > >  /* from remoteproc_coredump.c */
> > >  void rproc_default_coredump(struct rproc *rproc);
> > > +void rproc_inline_coredump(struct rproc *rproc);
> > > 
> > >  void rproc_free_vring(struct rproc_vring *rvring);
> > >  int rproc_alloc_vring(struct rproc_vdev *rvdev, int i);
> > > @@ -125,8 +138,14 @@ struct resource_table
> > > *rproc_find_loaded_rsc_table(struct rproc *rproc,
> > >  static inline
> > >  void rproc_coredump(struct rproc *rproc)
> > >  {
> > > -	return rproc_default_coredump(rproc);
> > > -
> > > +	switch (rproc->coredump_conf) {
> > > +	case COREDUMP_DEFAULT:
> > > +		return rproc_default_coredump(rproc);
> > > +	case COREDUMP_INLINE:
> > > +		return rproc_inline_coredump(rproc);
> > > +	default:
> > > +		break;
> > > +	}
> > 
> > I think this better belong inside remoteproc_coredump.c
> > 
> > Regards,
> > Bjorn
> > 
> > >  }
> > > 
> > >  #endif /* REMOTEPROC_INTERNAL_H */
> > > diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
> > > index 16ad666..23298ce 100644
> > > --- a/include/linux/remoteproc.h
> > > +++ b/include/linux/remoteproc.h
> > > @@ -459,6 +459,7 @@ 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
> > > + * @coredump_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
> > > @@ -492,6 +493,7 @@ struct rproc {
> > >  	struct device dev;
> > >  	atomic_t power;
> > >  	unsigned int state;
> > > +	unsigned int coredump_conf;
> > >  	struct mutex lock;
> > >  	struct dentry *dbg_dir;
> > >  	struct list_head traces;
> > > --
> > > The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
> > > Forum,
> > > a Linux Foundation Collaborative Project