GitLab

Update EnsureCanonicalObject so that it can optionally skip reading the
file/object length, based on the new argument R7. Then trace back
through the callers to have them pass in an appropriate value based on
whether they care about the length or not.

TopPath_DoBusinessToPath has also gained a new flag bit to suppress
reading the length, but so far I've only updated DoThePath to use it.

Result: If you have a file in PipeFS which is open for writing, and then
attempt to open it for reading using OS_Find, it'll now open immediately
instead of PipeFS trying to block/sleep until it's no longer open for
writing.

Why: Since at least as far back as RISC OS 3.6, PipeFS tries to sleep
inside fsfile_ReadInfo if the file is open for writing. But a
combination of buggy sleep logic and TaskWindow's lazy approach to
UpCall 6 meant that (prior to TaskWindow 0.65) it wouldn't sleep for
long and would instead allow the open to proceed while the file is still
open for write. The ab...

Instead of manually resetting the SVC stack pointer when starting a new
task, prefer to use the stack reset routine that's exposed by
OS_TaskControl 0 in RISC OS Select / 5. If that routine isn't available,
fall back to the old method of doing it manually.

Requires Kernel-6_61 to build.

Each of the 3 wildcard accepting OS_FSControl options takes its own copies of
the paths in Util_CommonStart then frees the PassedFilename (and co). This is
generally OK because the src name is kept in a 'copy block' struct relative
to R7 (and, for Copy, the dst in a second struct relative to R8).

However, when needing to get info on objects or other low level operations,
if the FS is marked in the 'DuffFS' table it would then do a safety check for
multiple $'s in the PassedFilename and in doing so dereferences a NULL pointer.

Ref https://www.riscosopen.org/forum/forums/4/topics/17259

Since the filenames constructed by resolving wildcards for Count/Copy/Wipe are
all sanitised by Process_WildPathname we can skip the safety check; there will
only be 1 $ by then.

FSUtils3.s: Correct misleading comment
LowLevel.s: Avoid NULL pointer dereference, drop noisy debug on every FS lookup

Version 2.89. Tagged as 'FileSwitch-2_89'

Detail:
* Makefile: get directory separator and header extension from shared makefiles.
* Build variable ImageSize (inherited by amu/make from environment) needs
  passing to objasm on command line.
* Consolidated .gitignore to root.
* Enable GitLab CI.


Version 2.88. Not tagged

This reverts commit ec553fef.

The above commit causes various BASIC programs to fail as they rely on the
location of quotes in order to parse their command line.


Version 2.88. Not tagged

Detail:
  Quotes in an alias mean
    If "<Alias$@RunType_FD1>" = "" Then ...
  doesn't work. The quotes aren't necessary, because filenames can't have spaces in anyway.
Admin:
  Tested on Linux and RPCEmu. Boots and fixes above problem.

Version 2.88. Not tagged

Programmer's Reference Manual Volume 4 Table C
defines a filetype for non-tokenised BASIC programs.

Add support for this to FileSwitch with a Filetype name,
load and run action.

Version 2.88. Tagged as 'FileSwitch-2_88'

This makes editing the source easier on other operating systems.


Version 2.87. Not tagged

The code path that filing systems the register with dontuseload omitted the line printed when *OPT 1 is non-zero.
In particular this meant FileCore (which has the fsinfo word flag set) didn't report any monitor name.
For dontusesave, the initial call to do the equivalent of *Create covers the output even though it's a bit early. However, the flag check appeared to be inverted, so an FS with it set would call OS_File, and one clear would call OS_GBPB.

FileSwHdr.s: Remove use_fsfile_Save and use_fsfile_Load switches, since clients declare at runtime what their preference is (since RISC OS 3).
LowLevel.s: Extract monitor name reporting code, replace with function call to common code.
FSShared.s: Common function ReportMonitor.
OSFile.s: After doing a load via OS_GBPB call the monitor name reporting code. Tune up the StrongARM code flush code a little to not need R5 as a temporary. Fix inverted sense of dontusesave.

Version 2.87. Tagged as 'FileSwitch-2_87'

Detail:
  s/FSControl - Change the post-Service_UKCompression call to OS_SynchroniseCodeAreas to be a ranged up rather than a global one.
Admin:
  Tested on iMx6
  Improves performance with SMP ROMs, where global D-cache clean isn't really possible


Version 2.86. Tagged as 'FileSwitch-2_86'

Detail:
  PRM2-97 describes filing system name matching as case insensitive. However, the Internat_LowerCase macro is a no-op when there's no case conversion table, such as prior to the territory loading.
  Change FindFSName to use an case insensitive ASCII match until the territory starts.
  Fixes a problem with EtherUSB failing to start because it was using Devices: (rather than devices:). A global search of the sources shows SerMouse and the Kernel would also be affected.
Admin:
  Submission from Colin Granville.

Version 2.85. Tagged as 'FileSwitch-2_85'

Add fsfunc_Opt subreasons, fsfunc_DirIs subreasons.
Expand tabs on fsargs_IOCtl.
No code change, retagged as FileSwitch-2_84.

Detail:
  s/CtrlUtils, s/OSGBPB - Avoid unnecessary remainder calculations in DivRem macro
Admin:
  Tested on Cortex-A15


Version 2.84. Tagged as 'FileSwitch-2_84'

Detail:
  s/CtrlUtils, s/DirStore, s/FSCtrl2, s/FSPath, s/FileSwBody, s/SysVars - Escape some dollars contained in strings to avoid warnings from objasm
Admin:
  Resulting binary unchanged


Version 2.83. Retagged as 'FileSwitch-2_83'

Fix buffer overflow in *cat when dealing with long paths. Fix spurious errors when examining files with certain load addresses

Detail:
  s/CtrlUtils - Increase MaxFilenameSpace (used to reserve stack space for *cat directory path) from 400 bytes to StaticName_length (i.e. 1K)
  s/FileTypes - Avoid IsFileTyped erroneously returning with V set if r2 is within a certain range (e.g. &7ff00000). Now all exits go via code that ensures V is clear.
Admin:
  Tested on Raspberry Pi


Version 2.83. Tagged as 'FileSwitch-2_83'

LowFSI: Max_BuffSize define changed.
FileSwBody: Module workspace claim changed to claim Max_BuffSize in addition to the workspace required, so that all the variables in the workspace remain within ADR range.
OSFind: Vetting of acceptable buffer sizes now produced with a WHILE/WEND loop.
StreamBits: Changed to use 'free' flag rather then comparing with 'Nowt'.
MkClean/MkRom: Throwback enabled.

Tested in IOMD ROM, workspace manually inspected as OK.

Version 2.82. Tagged as 'FileSwitch-2_82'

Detail:
  s/FSCommon - Adjust ValidateR2R5_WriteToCoreCodeLoad to try and increase the wimpslot if it isn't already big enough to contain the region being validated (i.e. the program being loaded)
  s/FSControl - FSControl 4 now checks AIF memory limts and tries to extend wimpslot as necessary both before and after decompression. An extra 8K is added to the calculated memory requirements to take into account memory which isn't declared in the header (root stack chunk?) which is required by SCL initialisation.
Admin:
  Tested in Iyonix ROM softload
  This should eliminate most/all of the "No writeable memory at this address", abort on data transfer, and "Not enough memory for C library" errors users receive due to the wimpslot being too small when all they want to do is run a program


Version 2.81. Tagged as 'FileSwitch-2_81'

Checked with a ROM build for conflicts, no clashes seen.

Version 2.80. Not tagged

Version 2.80. Not tagged

In OS_Find there is a call to TopPath_DoBusinessToPath in DoTheOpen, any V set exits from this are propagated to the user correctly.
Non V set exits (ie. the object type is returned) are passed to DoTheOpen_Common which in turn calls TryToOpenFile. TryToOpenFile (if fscb is not Nowt) dereferences it to get at the info word.
A problem exists where TopPath_DoBusinessToPath fails the policing check - when multipart paths are permitted (such as OS_Find) it swallows the error, steps the path, and eventually exits with an object type 0. However, this code path never sets up the fscb.
So when TryToOpenFile dereferences it (in testing it seemed to have some address in the kernel lingering) depending on your luck the fsinfo_alwaysopen bit might be set. This in turn triggers a forced call to the low level FS entry point, and the entry point is some garbage address leading to an abort.
The minimalist test cases were
 SYS"OS_Find",&41,"MyName","MyPath$"
 SYS"OS_Find",&49,"MyName","MyPath$"
both of which fail the policing step because the incorrect use of the path type bits mean the attempted name becomes MyPath$MyName and you can't have '$' in the middle of a filename. Other more complex examples exist, including having a multipart path where one of the path elements in invalid.
Solved here by setting fscb up to a default (valid) value.
Fixes ticket #287.

Version 2.80. Tagged as 'FileSwitch-2_80'

The corresponding external interface, OS_Args 6, takes a byte size in R2 which is passed on verbatim. However, when FileSwitch wants to do the same operation it rounds up to the nearest buffer size and complains the operation can't complete ("File too big") when the same operation through OS_Args would succeed.
Internally FileSwitch doesn't actually use the allocated size for anything (other than to avoid unnecesary ensures, and to cache the value for read back via OS_Args 4), so the full byte range of R2 is now allowed internally as well as externally.
FSEntry_Open may also pass back 0xFFFFFFFF as the allocated size on opening similarly.
FSEntry_Args 7 and FSEntry_Args 4 are the corresponding low level API, these are unchanged in meaning.

The FSEntry_Args 8, write 0s to file (usually triggered by extending with EXT#handle=number), is defined as always passing the file offset as a multiple of the buffer size, and the number of 0s to write as a multiple of the buffer size. To allow this to write up to 4G of 0s while not spilling over R3, FileSwitch now splits such a call into one buffer, followed by the remainder.

Tested with LanManFS 2.38, checking operations are split appropriately.

Version 2.79. Tagged as 'FileSwitch-2_79'

Not tagged.

Detail:
  With *Copy A, and the related SWI OS_FSControl 26 with R3 bit 9 clear, if
  a directory was copied over an image file, the contents of the directory
  would be copied (assuming recursive) but then when the attributes were
  copied over, the directory's filetype (usually FFD, Data) and user access
  bits (both 0) were copied to the destination image file, rendering it
  unusable.
Admin:
  Tested on a Raspberry Pi with the DOS partition in the standard disc image.

Version 2.78. Tagged as 'FileSwitch-2_78'

Detail:
  s/FSControl - faff_boot_option_startup (OS_FSControl 47/48) was acting as if it was given a path variable in r3, but neither of those calls use path variables, leading to aborts in some situations due to random bits of memory being peeked at.
  s/FSUtils - When CopyWildObject performs a single object copy, it wasn't taking into account the fact that the src/dest leafname pointers could be null, potentially causing a crash if high processor vectors are in use.
Since I'm not 100% sure what the correct course of action is (in the case I saw it crash, both src and dest leafnames were null), and there aren't any bugs I know of with file copies using the wrong names, the code now uses the dest leafname if the dest leafname is null.
This sounds wrong, but matches the behaviour of the old code, since byte &00000000 is highly unlikely to contain a '*' character
Admin:
  Tested on BB-xM with high processor vectors


Version 2.77. Tagged as 'FileSwitch-2_77'

Detail:
  s/TopPath - Code which attempted to free the cleaned up user's path incorrectly assumed that it could find the string via the stacked R0. However stacked regs R0-R6 get updated with the found objects details, so we need to store a copy of the users R0 elsewhere and free the string via that instead.
Admin:
  Fixes OS_FSControl 37 (and other operations?) filling the system heap if given paths which will generate errors (e.g. if drive empty/missing)


Version 2.76. Tagged as 'FileSwitch-2_76'

Retagged as FileSwitch-2_75.

Was returning internal value 'Nowt' instead of zero.
This bug seems to have appeared in FileSwitch 1.70 in RISC OS 3.00 and as such for backwards compatibility the value 0 and the value 0x40000000 should be taken to mean 'none'.

Version 2.75. Tagged as 'FileSwitch-2_75'

Version 2.74. Tagged as 'FileSwitch-2_74'

Delete dead macro.
Ammend OS_FSControl 28 API to aloow report of 64 bit results.
On failing to get any memory at all for *COPY an attempt is made to grab an 'excessive' amount of RMA to get an error message, but this was only requesting 32MB which is now quite feasible - now asks for 2GB.

Version 2.73. Tagged as 'FileSwitch-2_73'

Collapse old switches.
Remove 'Version' GBLA.
*COUNT summary now says "total 1234 bytes" not "1234 bytes" to match *COPY
Empty *COUNT syntax message corrected.
Binary identical to last version (messages different clearly).

Version 2.72. Tagged as 'FileSwitch-2_72'

Delete unused 'KernelFSW'.
HighFSI
 Retire some crusty definitions
 Reserve OS_FSControl reason codes used by ROL

Version 2.71. Tagged as 'FileSwitch-2_71'

Same drill, V was being set or not preserved when it shouldn't or should be. In light of this all locations that do CMP/SUBS/ADDS involving fileptr/extent/anything ending in size or length have been reviewed and ammended where needed.
Trim some dead code.
Discard 'Version' file.

Version 2.70. Tagged as 'FileSwitch-2_70'

Retagged as FileSwitch-2_69.

In OS_Args 1 a compare was followed by a conditional store, but with files > 2G some circumstances end up with VS so the store wasn't taken.
In OS_GBPB 3/4 the assessment of how much could be done was based on adding the current fileptr to the request but not respecting the carry out. Reordered the equation to get this calculation right.
Changed a signed compare of a byte variable to use unsigned so auditing of signed condition codes is clearer.
Corrected spelling of 'openning'.
Fix copy_srchandle and copy_dsthandle to not assume handles fit in 1 byte.
Tested on a variety of contrived file sizes at the boundaries of 2G and 4G.

Version 2.69. Tagged as 'FileSwitch-2_69'

Detail:
  s/FileSwBody, s/FileSwHdr - Try getting DomainId pointer via OS_ReadSysInfo 6 before falling back on legacy address. Store result in workspace for speedy lookup.
  s/FSCommon - Use workspace DomainId pointer
  s/CtrlUtils - Update int_ConstructFullPathWithoutFSAndSpecial to not attempt to append the contents of null string pointers.
  s/FSControl - Update AppendStringIfNotNull to ignore null pointers
  s/LowLevel - Update CallFSFile_Given to not try copying from null special field strings
Admin:
  Tested on rev A2 BB-xM.
  AFAIK the null pointer bugs are all FileSwitch bugs and not bugs in whatever called FileSwitch, although I may be wrong. Not entirely sure how some of the code managed to read strings from null pointers without creating garbage filenames as a result!


Version 2.68. Tagged as 'FileSwitch-2_68'

Detail:
  s.LowLevel: STRNE should have been STRNEB. Looks like this meant that
  OS_File 5 was broken if the object name contained a filing system special
  field.
  Added ENTRY directive to permit building of GPA debug listing.
Admin:
  Built but not tested (should be a safe change).

Version 2.67. Tagged as 'FileSwitch-2_67'

Detail:
  Uses suffixed file extensions for compatiblity with both objasm and asasm.
Admin:
  Supplied by Peter Naulls, tested at ROOL

Version 2.66. Not tagged

Version 2.66. Tagged as 'FileSwitch-2_66'