Skip to contentSkip to conversation

bio-integrity: Copy bip_buf and bip_size in bio_integrity_clone()

RFC PATCH In review Alberto Bertogli 2 participants 5 messages May 24, 2009 – May 26, 2009 VFS

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 bio-integrity: Copy bip_buf and bip_size in bio_integrity_clone()

Hi!

I'm trying to add bio-integrity support to a device-mapper target I posted a couple of days ago (http://lkml.org/lkml/2009/5/21/235).

While at it, I found that bio_integrity_clone() does not clone neither bip_buf nor bip_size, which already copies the bvec, which should have the same data because it's allocated in bio_integrity_prep().

Quoted 1 line

From a quick grep, it seems they're only used inside fs/bio-integrity.c, and

through there they get passed to the functions each block device registered for integrity verification and generation, and used in bio_integrity_tag().

Also, bio_integrity_free() already takes into account that and does not free bip_buf if it's freeing a cloned bio.

Is there any reason I'm missing why they shouldn't be copied in bio_integrity_clone(), as illustrated in the following patch?

Thanks a lot,
Alberto

------- 8< ------- 8< ------- 8< ------- 8< ------- 8< ------- 8< -------

From: Alberto Bertogli <[email protected]>
Date: Sun, 24 May 2009 01:02:08 -0300
Subject: [PATCH] bio-integrity: Copy bip_buf and bip_size in bio_integrity_clone()

---
 fs/bio-integrity.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)
fs/bio-integrity.c +3 -0
index 31c46a2..fa9457d 100644
@@ -682,6 +682,9 @@ int bio_integrity_clone(struct bio *bio, struct bio *bio_src, gfp_t gfp_mask)
 	if (bip == NULL)
 		return -EIO;
 
+	bip->bip_buf = bip_src->bip_buf;
+	bip->bip_size = bip_src->bip_size;
+
 	memcpy(bip->bip_vec, bip_src->bip_vec,
 	       bip_src->bip_vcnt * sizeof(struct bio_vec));
 
Signature · 1 line
1.6.2.2.646.gb214
#1 lore raw
Quoted 3 lines · Alberto Bertogli

While at it, I found that bio_integrity_clone() does not clone neither bip_buf nor bip_size, which already copies the bvec, which should have the same data because it's allocated in bio_integrity_prep().

In case that makes sense, updating bip_buf and bip_size (if bip_buf is not NULL) in bio_integrity_trim() does too, doesn't it?

Thanks a lot,
Alberto

#2 lore raw
In reply to Alberto Bertogli
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Alberto" == Alberto Bertogli <[email protected]> writes:

Alberto> While at it, I found that bio_integrity_clone() does not clone Alberto> neither bip_buf nor bip_size, which already copies the bvec, Alberto> which should have the same data because it's allocated in Alberto> bio_integrity_prep().

Alberto> Is there any reason I'm missing why they shouldn't be copied in Alberto> bio_integrity_clone(), as illustrated in the following patch?

Yes. The bip_buf is strictly an internal housekeeping construct. It contains a pointer to the kernel buffer that contains the protection information if the protection was added by the block layer itself (via bio_integrity_prep()). However, that's not always the case. The protection information may be passed down from a filesystem or (eventually) a userland application. So the bip_buf is for use by the top-level of the block layer exclusively. The bip_vec is what you want to use for accessing the actual protection information.

Also, the cloned bio may be truncated or split. In that case the bip_buf isn't going to match the data bvec. So in any case blindly cloning bip_buf isn't going to work.

Right now the integrity vector itself is cloned together with the bio because of the way the block layer works (advancing bvec offset for partial completion). However, I'm brewing on a patch that gets rid of that so we can avoid cloning the vector and thus cut down on memory allocations as the I/O goes through each remapping layer. With my patch in place the bip_vec becomes immutable and bip_buf will go away.

I took a quick look at your DM patch last week and I didn't see any reason why it couldn't hook into the block integrity infrastructure. Take a look at drivers/scsi/sd_dif.c for clues on how to do that.

Let me know if you have questions...

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering
#3 lore raw
Quoted 23 lines · Martin K. Petersen
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Alberto" == Alberto Bertogli <[email protected]> writes:

Alberto> While at it, I found that bio_integrity_clone() does not clone Alberto> neither bip_buf nor bip_size, which already copies the bvec, Alberto> which should have the same data because it's allocated in Alberto> bio_integrity_prep().

Alberto> Is there any reason I'm missing why they shouldn't be copied in Alberto> bio_integrity_clone(), as illustrated in the following patch?

Yes. The bip_buf is strictly an internal housekeeping construct. It contains a pointer to the kernel buffer that contains the protection information if the protection was added by the block layer itself (via bio_integrity_prep()). However, that's not always the case. The protection information may be passed down from a filesystem or (eventually) a userland application. So the bip_buf is for use by the top-level of the block layer exclusively. The bip_vec is what you want to use for accessing the actual protection information.

Also, the cloned bio may be truncated or split. In that case the bip_buf isn't going to match the data bvec. So in any case blindly cloning bip_buf isn't going to work.

Quoted 6 lines

Right now the integrity vector itself is cloned together with the bio because of the way the block layer works (advancing bvec offset for partial completion). However, I'm brewing on a patch that gets rid of that so we can avoid cloning the vector and thus cut down on memory allocations as the I/O goes through each remapping layer. With my patch in place the bip_vec becomes immutable and bip_buf will go away.

That makes sense, thanks for the explanation.

The case I was thinking about was something like a filesystem calling bio_integrity_get_tag() on a cloned bio, since it depends on having a bip_buf available. But if you're going to remove it altogether then it's a moot question.

Quoted 3 lines

I took a quick look at your DM patch last week and I didn't see any reason why it couldn't hook into the block integrity infrastructure. Take a look at drivers/scsi/sd_dif.c for clues on how to do that.

Thanks, I've already implemented it, and will post an updated patch soon, after I clean it up a little.

Quoted 1 line

Let me know if you have questions...

Actually, I have two minor questions, if you don't mind:

- What would be a decent name to use in struct blk_integrity for a module such as mine? Is LINUX-DMCSUM-V0-CCITT reasonable?
- How can I test the tag functions? From a quick grep I found no in-tree users of bio_integrity_get/set_tag().

Thanks a lot,
Alberto

#4 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Alberto" == Alberto Bertogli <[email protected]> writes:

Alberto> The case I was thinking about was something like a filesystem Alberto> calling bio_integrity_get_tag() on a cloned bio,

Filesystems never see cloned bios.

Alberto> - What would be a decent name to use in struct blk_integrity
Alberto> for a module such as mine? Is LINUX-DMCSUM-V0-CCITT
Alberto> reasonable?

Yeah. Doesn't matter much, really. The original intent was simply to capture the variations in the (hardware) protocols .

Alberto> - How can I test the tag functions? From a quick grep I found
Alberto> no in-tree users of bio_integrity_get/set_tag().

There are no users yet.

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering