iomap: enforce DIO alignment check in iomap]
Subsystems (2)
PATCH 1/1 iomap: enforce DIO alignment check in iomap]
Sorry, I screwed up two addresses, bouncing them to the right addresses to avoid spamming the lists again.
----- Forwarded message from [email protected] -----
Date: Wed, 10 Jun 2026 16:52:11 +0200
From: [email protected]
To: [email protected]
Cc: [email protected], [email protected],
[email protected], [email protected], Carlos Maiolino
<[email protected]>, Keith Busch <[email protected]>, Hannes Reinecke <[email protected]>, "Martin K. Petersen" <[email protected]>, Christoph Hellwig <[email protected]>, Jens Axboe <[email protected]>
Subject: [PATCH] iomap: enforce DIO alignment check in iomap
Message-ID: <[email protected]>
X-Mailer: git-send-email 2.54.0
Tags: sent
From: Carlos Maiolino <[email protected]>
The DIO alignment check has been lifted from iomap layer to rely on the block layer to enforce proper alignment when issuing direct IO operations. This though, depending on the IO size and buffer address passed to the IO operation may lead to user-visible behavior change.
This has been caught initially by LTP test diotest4 running on PPC architecture, where the test fails because a read() operation with a supposedly misaligned buffer succeeds instead of an expected -EINVAL.
This has no direct relationship with PPC, but seems to do with the IO size crossing page borders or not.
The test allocates a 4k buffer, and then increments the buffer pointer by a single byte to enforce a misaligned address. It then issues a 4k read() using such buffer. The read is supposed to return an -EINVAL but it ends up succeeding.
The allocated buffer is at least a single page, so the read() size being smaller will end up most of the time within the very same page initially allocated which seems to suffice the block layer to accept the IO.
On x86 though, the same 4k read will end up crossing page boundaries causing a bio_split which ends up properly checking the address and rejecting it due to misalignment.
The test itself is buggy (which seems by design) because it ends up attempting to read 4096 bytes into a 4095, but I believe the test expected the address to be rejected prior to any write attempt.
The problematic behavior is reproducible on x86 by reducing the IO size to something < PAGE_SIZE, so the misaligned read()s will also be accepted by the block layer.
Fixing this is just a matter of enforcing daddr and memory alignment back into iomap.
This behavior is reproducible in ext4 and xfs due to both relying on iomap layer, btrfs does not present this behavior change as it does its own DIO alignment checking.
---
While I didn't spot any memory/disk corruption looking into this, it changes the user behavior that dictates buffer addresses must be properly aligned when issuing direct IO operations so I thought making iomap check again for the buffer address alignment is reasonable.
fs/iomap/direct-io.c | 3 +++ 1 file changed, 3 insertions(+)
fs/iomap/direct-io.c +3 -0
@@ -400,6 +400,9 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)if ((pos | length) & (alignment -1 ))return -EINVAL ; +if (iov_iter_alignment (dio->submit.iter) & (alignment -1 )) +return -EINVAL ; +if (dio->flags &IOMAP_DIO_WRITE ) {bool need_completion_work = true;
Signature · 5 lines
2.54.0 ----- End forwarded message -----
Quoted 9 lines · Carlos Maiolino
The DIO alignment check has been lifted from iomap layer to rely on the block layer to enforce proper alignment when issuing direct IO operations. This though, depending on the IO size and buffer address passed to the IO operation may lead to user-visible behavior change.
This has been caught initially by LTP test diotest4 running on PPC architecture, where the test fails because a
read()operation with a supposedly misaligned buffer succeeds instead of an expected-EINVAL.
It's not supposed to matter where in the stack we determined this be an invalid request: it should still fail if it's misaligned. Could you clarify how this is succeeding?
Quoted 6 lines · Keith Busch
Quoted 5 lines · Carlos Maiolino
The DIO alignment check has been lifted from iomap layer to rely on the block layer to enforce proper alignment when issuing direct IO operations. This though, depending on the IO size and buffer address passed to the IO operation may lead to user-visible behavior change.
Quoted 8 lines
Quoted 4 lines
This has been caught initially by LTP test diotest4 running on PPC architecture, where the test fails because a
read()operation with a supposedly misaligned buffer succeeds instead of an expected-EINVAL.It's not supposed to matter where in the stack we determined this be an invalid request: it should still fail if it's misaligned. Could you clarify how this is succeeding?
Fair enough, can you point me to where the alignment is supposed to be checked? I've been seeing kind of different behaviors with different machines so I kind of am not sure where the alignment is supposed to be validated within the block layer (I should probably have tagged this patch as RFC as my understanding of the block layer is superficial).
A few runs when the test failed (because the read() call succeded) I could see this for example:
openat(dirfd=AT_FDCWD, pathname="testdata-4.2063", flags=O_RDWR|O_DIRECT) = 3 _llseek(fd=3, offset=4096, result=[4096], whence=SEEK_SET) = 0 read(arg1=0x3, arg2=0x25960001, arg3=0x1000) = 0x1000
FWIW, on my laptop, the test fails because read() returns an -EIO (while the test expects -EINVAL), I'm pointing the fingers to a faulty hardware now but I didn't discard the possibility of an EIO being returned EINVAL should have been returned instead.
I'm trying to find another machine to test and see whatever differences appear...
Cheers
Quoted 12 lines · Carlos Maiolino
Quoted 11 lines · Keith Busch
Quoted 9 lines · Carlos Maiolino
The DIO alignment check has been lifted from iomap layer to rely on the block layer to enforce proper alignment when issuing direct IO operations. This though, depending on the IO size and buffer address passed to the IO operation may lead to user-visible behavior change.
This has been caught initially by LTP test diotest4 running on PPC architecture, where the test fails because a
read()operation with a supposedly misaligned buffer succeeds instead of an expected-EINVAL.
Quoted 6 lines
Quoted 3 lines
It's not supposed to matter where in the stack we determined this be an invalid request: it should still fail if it's misaligned. Could you clarify how this is succeeding?
Fair enough, can you point me to where the alignment is supposed to be checked?
https://elixir.bootlin.com/linux/v7.1-rc7/source/block/blk-merge.c#L352
It does require that someone calls the bio split-to-limits routine, which I had taken for granted as a given, but I realize that some drivers don't do that. What block device are you using for your test?
Quoted 17 lines · Keith Busch
Quoted 16 lines · Carlos Maiolino
Quoted 14 lines · Keith Busch
Quoted 9 lines · Carlos Maiolino
The DIO alignment check has been lifted from iomap layer to rely on the block layer to enforce proper alignment when issuing direct IO operations. This though, depending on the IO size and buffer address passed to the IO operation may lead to user-visible behavior change.
This has been caught initially by LTP test diotest4 running on PPC architecture, where the test fails because a
read()operation with a supposedly misaligned buffer succeeds instead of an expected-EINVAL.It's not supposed to matter where in the stack we determined this be an invalid request: it should still fail if it's misaligned. Could you clarify how this is succeeding?
Quoted 8 lines
Quoted 2 lines
Fair enough, can you point me to where the alignment is supposed to be checked?
https://elixir.bootlin.com/linux/v7.1-rc7/source/block/blk-merge.c#L352
It does require that someone calls the bio split-to-limits routine, which I had taken for granted as a given, but I realize that some drivers don't do that. What block device are you using for your test?
In the PPC machine, it's a virtual scsi vdasd device from one of the virtual nodes
NAME HCTL TYPE VENDOR MODEL REV SERIAL TRAN sda 0:0:1:0 disk AIX VDASD 0001 000a508a00007a0000000175dcba35ac.5
ibmvfc 262144 0 ibmvscsi 196608 2
For my x86 machine (remind I reduce the buffer size to 512 on x86), it's a commodity sata samsung SSD:
sdb 6:0:0:0 disk ATA Samsung SSD 860 EVO 1TB RVT02B6Q S3Z9NB0KB82416J sata
Both with an ext4 straight on top of a partition (no device-mapper or any other volume layer in between), although I also tried with a linear device-mapper (lvm) device and the results seemed the same.
Those are two machines I have had more reliable results. My laptop with a samsung NVME makes read() return an -EIO but I can't tell if it's a device failure or just a wrong error being returned by now
Quoted 9 lines · Carlos Maiolino
Quoted 4 lines · Keith Busch
It does require that someone calls the bio split-to-limits routine, which I had taken for granted as a given, but I realize that some drivers don't do that. What block device are you using for your test?
In the PPC machine, it's a virtual scsi vdasd device from one of the virtual nodes
Quoted 8 lines
NAME HCTL TYPE VENDOR MODEL REV SERIAL TRAN sda 0:0:1:0 disk AIX VDASD 0001 000a508a00007a0000000175dcba35ac.5ibmvfc 262144 0 ibmvscsi 196608 2For my x86 machine (remind I reduce the buffer size to 512 on x86), it's a commodity sata samsung SSD:
Okay, these are under blk-mq so always call __bio_split_to_limits. However, I see there's an optimization to skip the checks we're depending on if bio_may_need_split doesn't think it needs to be split, which is a problem for your observation. I don't think the current expecations can allow us to take this optimization anymore when page offsets are used.
This should fix it:
---
block/blk.h +1 -1
@@ -404,7 +404,7 @@ static inline bool bio_may_need_split(struct bio *bio, bv =__bvec_iter_bvec (bio->bi_io_vec, bio->bi_iter);if (bio->bi_iter.bi_size > bv->bv_len - bio->bi_iter.bi_bvec_done)return true; -return bv->bv_len + bv->bv_offset > lim->max_fast_segment_size; +return bv.bv_offset || bv->bv_len > lim->max_fast_segment_size; }/**
--
Quoted 34 lines · Keith Busch
Quoted 9 lines · Carlos Maiolino
Quoted 4 lines · Keith Busch
It does require that someone calls the bio split-to-limits routine, which I had taken for granted as a given, but I realize that some drivers don't do that. What block device are you using for your test?
In the PPC machine, it's a virtual scsi vdasd device from one of the virtual nodes
Quoted 8 lines
NAME HCTL TYPE VENDOR MODEL REV SERIAL TRAN sda 0:0:1:0 disk AIX VDASD 0001 000a508a00007a0000000175dcba35ac.5ibmvfc 262144 0 ibmvscsi 196608 2For my x86 machine (remind I reduce the buffer size to 512 on x86), it's a commodity sata samsung SSD:
Okay, these are under blk-mq so always call
__bio_split_to_limits. However, I see there's an optimization to skip the checks we're depending on ifbio_may_need_splitdoesn't think it needs to be split, which is a problem for your observation. I don't think the current expecations can allow us to take this optimization anymore when page offsets are used.This should fix it:
---
block/blk.h +0
-0
block/blk.h static inline bool bio_may_need_split(struct bio *bio, · line 408
Quoted 6 lines
block/blk.h +1
-1@@ -404,7 +404,7 @@ static inline bool bio_may_need_split(struct bio *bio, bv =__bvec_iter_bvec (bio->bi_io_vec, bio->bi_iter);if (bio->bi_iter.bi_size > bv->bv_len - bio->bi_iter.bi_bvec_done)return true; -return bv->bv_len + bv->bv_offset > lim->max_fast_segment_size; +return bv.bv_offset || bv->bv_len > lim->max_fast_segment_size; }
This should work for the un-aligned DMA buffer, but might hurt perf for any sub-page IO.
Given you have switched to validate dio buffer alignment to bio splitting, it should be fine to check ->dma_alignment here by putting the three limits fields into same cache line.
Thanks,
Ming