Server crash during nbackup 4.0.5

[gerdus]

Server crash during nbackup resulting in delta file corruption.
When access db:

database file appears corrupt (/databases/filename.FDB4) wrong page type page 0 is of wrong type (expected database header, found index root)

after remove delta and fixup database is working.

in log there is nothing helpful

servername	Sun Mar  8 21:01:31 2026
	/opt/firebird/bin/fbguard: /opt/firebird/bin/firebird terminated abnormally (-1)

server version: 4.0.5
OS: linux

Crash dump available at https://infireal.com/github-issues/firebird-crash-2026-03-09-1105/

could be #8237 but sounds different because error talks about database header page but idk.

Thank you

[EPluribusUnum]

@gerdus , "wrong page type page" happend for us, when in gbak the database access was not via TCP/IP.

See #7301 and #7151

[gerdus]

@EPluribusUnum access was via tcp/ip (via firebird c# driver) so should not be same problem.

[AlexPeshkoff]

Was this crash during access DB permanent (i.e. was it reproduced after server restart)?

[gerdus]

It did not crash again after restart if that is what you are asking.
Database remained non accessible due to the wrong page type page 0, only became useful again after delta file delete and fixup.

Don't know if it is useful but i am attaching the delta file, can't really share the main db because it is customer data.

RWRWSZC2PDBU6VLU7UP94JP4D3.FDB4.delta.zip

[mrotteveel]

@gerdus The delta file can also contain customer data, as it's where database pages are written while the database is locked when nbackup is in progress.

[gerdus]

@mrotteveel Indeed but it's corrupted and tiny so it's a calculated risk.

[AlexPeshkoff]

On 3/9/26 21:53, Gerdus van Zyl wrote: *gerdus* left a comment (FirebirdSQL/firebird#8935) <#8935 (comment)> It did not crash again after restart if that is what you are asking. Database remained non accessible due to the wrong page type page 0, only became useful again after delta file delete and fixup. Don't know if it is useful but i am attaching the delta file, can't really share the main db because it is customer data. RWRWSZC2PDBU6VLU7UP94JP4D3.FDB4.delta.zip <https://github.com/user-attachments/files/25851065/RWRWSZC2PDBU6VLU7UP94JP4D3.FDB4.delta.zip>

Yes, file is useful - at least now I can say that this is definitely not #8237 What is worse - no idea how it could happen. Afraid asking about reproduceable test case is unreal here. @hvlad Any idea how could IRT be written into delta file instead header?

[hvlad]

Any idea how could IRT be written into delta file instead header?

Wrong page index at allocation table, stored in delta file, for example.
BTW, what is value of FW setting on database ?

[AlexPeshkoff]

On 3/10/26 13:03, Vlad Khorsun wrote: *hvlad* left a comment (FirebirdSQL/firebird#8935) <#8935 (comment)> Any idea how could IRT be written into delta file instead header? Wrong page index at allocation table, stored in delta file, for example.

Yes, that's the first thing coming to the mind - but I see no way for it t get there, except passed from cch. In     alloc_buffer[++alloc_buffer[0]] = db_page; rather hard to have problems. Writing first few bytes of page buffer and not writing next - also hardly possible.

[gerdus]

force write was enabled on database

[hvlad]

Allocation table contains 3 entry: 0x0000E735, 0x0000E751, 0x00000000. I.e. delta file contains 4 pages. As delta file size is 32KB, we get database page size of 8KB.

At offset 0x2000 there is page header that describes SCN's inventory page (type 0xA) with SCN = 0x00018DF7 and pageno = 0x0000E735 (same as alloc table)

At offset 0x4000 there is page header that describes Transaction inventory page (type 0x3) with SCN = 0x00018DF7 and pageno = 0x0000E751 (same as alloc table)

At offset 0x6000 there is page header that describes Index root page (type 0x6) with SCN = 0xFEB0F640 and pageno = 0x00007FB0 (not as alloc table !)

Note, when new page is allocated in delta file, engine writes zero's to it first. Looks like this page was never written to the delta file and contains random garbage (too big SCN and zeros in next 16 bytes after header confirms it).

Note, also, that 3rd page in delta very likely should be the header page - it is expected to write header page immediately after TIP page (when any transaction ends).

[hvlad]

On 3/10/26 13:03, Vlad Khorsun wrote:
hvlad left a comment (FirebirdSQL/firebird#8935)
<#8935 (comment)>

 Any idea how could IRT be written into delta file instead header?

Wrong page index at allocation table, stored in delta file, for example.
Yes, that's the first thing coming to the mind - but I see no way for it
t get there, except passed from cch. In

 alloc_buffer[++alloc_buffer[0]] = db_page;

rather hard to have problems. Writing first few bytes of page buffer and
not writing next - also hardly possible.

It just confirms that 3rd page in delta were header page. And its content were never written into delta.