Re: [PATCH next v3 12/15] printk: introduce a kmsg_dump iterator

On 2021-03-01, Petr Mladek [off-list ref] wrote:

quoted

diff --git a/arch/powerpc/kernel/nvram_64.c b/arch/powerpc/kernel/nvram_64.c
index 532f22637783..5a64b24a91c2 100644
--- a/arch/powerpc/kernel/nvram_64.c
+++ b/arch/powerpc/kernel/nvram_64.c

@@ -72,8 +72,7 @@ static const char *nvram_os_partitions[] = {
 	NULL
 };
 
-static void oops_to_nvram(struct kmsg_dumper *dumper,
-			  enum kmsg_dump_reason reason);
+static void oops_to_nvram(enum kmsg_dump_reason reason);
 
 static struct kmsg_dumper nvram_kmsg_dumper = {
 	.dump = oops_to_nvram

@@ -642,11 +641,11 @@ void __init nvram_init_oops_partition(int rtas_partition_exists)
  * that we think will compress sufficiently to fit in the lnx,oops-log
  * partition.  If that's too much, go back and capture uncompressed text.
  */
-static void oops_to_nvram(struct kmsg_dumper *dumper,
-			  enum kmsg_dump_reason reason)
+static void oops_to_nvram(enum kmsg_dump_reason reason)
 {
 	struct oops_log_info *oops_hdr = (struct oops_log_info *)oops_buf;
 	static unsigned int oops_count = 0;
+	static struct kmsg_dump_iter iter;
 	static bool panicking = false;
 	static DEFINE_SPINLOCK(lock);
 	unsigned long flags;

@@ -681,13 +680,14 @@ static void oops_to_nvram(struct kmsg_dumper *dumper,
 		return;
 
 	if (big_oops_buf) {
-		kmsg_dump_get_buffer(dumper, false,
+		kmsg_dump_rewind(&iter);

It would be nice to get rid of the kmsg_dump_rewind(&iter) calls
in all callers.

A solution might be to create the following in include/linux/kmsg_dump.h

#define KMSG_DUMP_ITER_INIT(iter) {	\
	.cur_seq = 0,			\
	.next_seq = U64_MAX,		\
	}

#define DEFINE_KMSG_DUMP_ITER(iter)	\
	struct kmsg_dump_iter iter = KMSG_DUMP_ITER_INIT(iter)

For this caller (arch/powerpc/kernel/nvram_64.c) and for
(kernel/debug/kdb/kdb_main.c), kmsg_dump_rewind() is called twice within
the dumper. So rewind will still be used there.

Then we could do the following at the beginning of both
kmsg_dump_get_buffer() and kmsg_dump_get_line():

	u64 clear_seq = latched_seq_read_nolock(&clear_seq);

	if (iter->cur_seq < clear_seq)
		cur_seq = clear_seq;

I suppose we need to add this part anyway, if we want to enforce that
records before @clear_seq are not to be available for dumpers.

I am not completely sure about next_seq:

   + kmsg_dump_get_buffer() will set it for the next call anyway.
     It reads the blocks of messages from the newest.

   + kmsg_dump_get_line() wants to read the entire buffer anyway.
     But there is a small risk of an infinite loop when new messages
     are printed when dumping each line.

It might be better to avoid the infinite loop. We could do the following:

static void check_and_set_iter(struct kmsg_dump_iter)
{
	if (iter->cur_seq == 0 && iter->next_seq == U64_MAX) {
		kmsg_dump_rewind(iter);
}

and call this at the beginning of both kmsg_dump_get_buffer()
and kmsg_dump_get_line()

What do you think?

On a technical level, it does not make any difference. It is pure
cosmetic.

Personally, I prefer the rewind directly before the kmsg_dump_get calls
because it puts the initializer directly next to the user.

As an example to illustrate my view, I prefer:

    for (i = 0; i < n; i++)
        ...;

instead of:

    int i = 0;

    ...

    for (; i < n; i++)
        ...;

Also, I do not really like the special use of 0/U64_MAX to identify
special actions of the kmsg_dump_get functions.

Note that I do not resist on it. But it might make the API easier to
use from my POV.

Since you do not resist, I will keep the API the same for v4. But I will
add the @clear_seq check to the kmsg_dump_get functions.

John Ogness