libceph: fix last_piece calculation in ceph_msg_data_pages_advance
PATCH 1/1 libceph: fix last_piece calculation in ceph_msg_data_pages_advance
It's possible we'll have less than a page's worth of residual data, that is stradding the last two pages in the array. That will make it incorrectly set the last_piece boolean when it shouldn't.
Account for a non-zero cursor->page_offset when advancing.
--- net/ceph/messenger.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
net/ceph/messenger.c +1 -2
@@ -894,10 +894,9 @@ static bool ceph_msg_data_pages_advance(struct ceph_msg_data_cursor *cursor,return false;/* no more data */ /* Move on to the next page; offset is already at 0 */ -BUG_ON (cursor->page_index >= cursor->page_count); cursor->page_index++; - cursor->last_piece = cursor->resid <=PAGE_SIZE ; + cursor->last_piece = cursor->page_offset + cursor->resid <=PAGE_SIZE ;return true; }
Signature · 1 line
2.35.1
Quoted 5 lines · Jeff Layton
It's possible we'll have less than a page's worth of residual data, that is stradding the last two pages in the array. That will make it incorrectly set the
last_pieceboolean when it shouldn't.Account for a non-zero
cursor->page_offsetwhen advancing.
It's been quite a while I looked at this code, but isn't cursor->resid supposed to be the number of bytes remaining, irrespective of the offset?
Have you found this to cause a failure?
-Alex
Quoted 21 lines
Signed-off-by Jeff Layton <[email protected]>--- net/ceph/messenger.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)net/ceph/messenger.c +1
-2@@ -894,10 +894,9 @@ static bool ceph_msg_data_pages_advance(struct ceph_msg_data_cursor *cursor,return false;/* no more data */ /* Move on to the next page; offset is already at 0 */ -BUG_ON (cursor->page_index >= cursor->page_count); cursor->page_index++; - cursor->last_piece = cursor->resid <=PAGE_SIZE ; + cursor->last_piece = cursor->page_offset + cursor->resid <=PAGE_SIZE ;return true; }
Quoted 11 lines · Alex Elder
Quoted 5 lines · Jeff Layton
It's possible we'll have less than a page's worth of residual data, that is stradding the last two pages in the array. That will make it incorrectly set the
last_pieceboolean when it shouldn't.Account for a non-zero
cursor->page_offsetwhen advancing.It's been quite a while I looked at this code, but isn't
cursor->residsupposed to be the number of bytes remaining, irrespective of the offset?
Correct. The "residual" bytes in the cursor, AFAICT.
Quoted 2 lines
Have you found this to cause a failure?
Not with the existing code in libceph, as it only ever seems to advance to the end of the page. I'm working on some patches to allow for sparse reads though, and with those in place I need to sometimes advance to arbitrary positions in the array, and this reliably causes a BUG_ON() to trip.
Quoted 24 lines
-Alex
Quoted 21 lines
Signed-off-by Jeff Layton <[email protected]>--- net/ceph/messenger.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)net/ceph/messenger.c +1
-2@@ -894,10 +894,9 @@ static bool ceph_msg_data_pages_advance(struct ceph_msg_data_cursor *cursor,return false;/* no more data */ /* Move on to the next page; offset is already at 0 */ -BUG_ON (cursor->page_index >= cursor->page_count); cursor->page_index++; - cursor->last_piece = cursor->resid <=PAGE_SIZE ; + cursor->last_piece = cursor->page_offset + cursor->resid <=PAGE_SIZE ;return true; }
Signature · 1 line
Jeff Layton <[email protected]>
Ilya Dryomov
maintainerQuoted 16 lines · Jeff Layton
Quoted 11 lines · Alex Elder
Quoted 5 lines · Jeff Layton
It's possible we'll have less than a page's worth of residual data, that is stradding the last two pages in the array. That will make it incorrectly set the
last_pieceboolean when it shouldn't.Account for a non-zero
cursor->page_offsetwhen advancing.It's been quite a while I looked at this code, but isn't
cursor->residsupposed to be the number of bytes remaining, irrespective of the offset?Correct. The "residual" bytes in the cursor, AFAICT.
Quoted 8 lines
Quoted 2 lines
Have you found this to cause a failure?
Not with the existing code in libceph, as it only ever seems to advance to the end of the page. I'm working on some patches to allow for sparse reads though, and with those in place I need to sometimes advance to arbitrary positions in the array, and this reliably causes a
BUG_ON()to trip.
Hi Jeff,
Which BUG_ON? Can you explain what "advance to arbitrary positions in the array" means?
I think you may be misusing ceph_msg_data_pages_advance() because cursor->page_offset _has_ to be zero at that point: we have just determined that we are done with the current page and incremented cursor->page_index (i.e. moved to the next page). The way we determine whether we are done with the current page is by testing cursor->page_offset == 0, so either your change is a no-op or something else is broken.
Thanks,
Ilya
Quoted 31 lines · Ilya Dryomov
Quoted 16 lines · Jeff Layton
Quoted 11 lines · Alex Elder
Quoted 5 lines · Jeff Layton
It's possible we'll have less than a page's worth of residual data, that is stradding the last two pages in the array. That will make it incorrectly set the
last_pieceboolean when it shouldn't.Account for a non-zero
cursor->page_offsetwhen advancing.It's been quite a while I looked at this code, but isn't
cursor->residsupposed to be the number of bytes remaining, irrespective of the offset?Correct. The "residual" bytes in the cursor, AFAICT.
Quoted 8 lines
Quoted 2 lines
Have you found this to cause a failure?
Not with the existing code in libceph, as it only ever seems to advance to the end of the page. I'm working on some patches to allow for sparse reads though, and with those in place I need to sometimes advance to arbitrary positions in the array, and this reliably causes a
BUG_ON()to trip.Hi Jeff,
Which
BUG_ON? Can you explain what "advance to arbitrary positions in the array" means?
Quoted 8 lines
I think you may be misusing
ceph_msg_data_pages_advance()becausecursor->page_offset_has_ to be zero at that point: we have just determined that we are done with the current page and incrementedcursor->page_index(i.e. moved to the next page). The way we determine whether we are done with the current page is by testingcursor->page_offset== 0, so either your change is a no-op or something else is broken.
My apologies, I've got a stack of other patches sitting on top of this, and this patch should have been folded into one of them instead of being sent separately.
In the sparse_read code I have now, I'm creating some temporary ceph_msg_data_cursors on the stack and revamped the "advance" code to allow advancing an arbirtrary amount instead of only allowing it to go to the end of the current page. This fix is necessary with those changes.
I'm not sure I'm going to need that going forward though, as I may rework the patches (again) to use the cursor in the con, and in that case I'll only need to walk through it a page at a time like we do now.
We can drop this patch.
--
Jeff Layton <[email protected]>