From Newsgroup: comp.arch.embedded

The request is very common: when some interested events occur in the
system, they, with the related timestamp, must be saved in a log. The
log must be saved in an external SPI Flash connected to a MCU. The log
has a maximum number of events. After the log is completely filled, the
new event overwrite the oldest event.

I tried to implement such library, but I eventually found it's not a
simple task, mostly if you want a reliable library that works even when
errors occur (for example, when one writing fails).

I started by assigning 5 sectors of SPI Flash to the log. There are 256
events in a sector (the event is a fixed struct).
In this scenario, the maximum log size is 1024 events, because the 5th
sector can be erased when the pointer reaches the end of a sector.

The first challenge is how the system can understand what is the first (newest) event in the log at startup. I solved saving a 16-bits counter
ID to each event. The initialization routine starts reading all the IDs
and taks the greatest as the last event.
However initially the log is empty, so all the IDs are 0xFFFF, the
maximum. One solution is to stop reading events when 0xFFFF is read and wrap-around ID at 0xFFFE and not 0xFFFF.

However there's another problem. What happens after writing 65535 events
in the log? The ID restarts from 0, so the latest event hasn't the
greatest ID anymore.

These are the saved IDs after 65536 events:

1^ SECT 2^ SECT 3^ SECT 4^ SECT 5^SECT---------->
0xFB00 ... 0xFC00 ... 0xFD00 ... 0xFE00 ... 0xFF00 ... 0xFFFF

The rule "newest event has greatest ID" is correct yet. Now a new event
is written:

1^ SECT-------> 2^ SECT 3^ SECT 4^ SECT 5^SECT--------->
0x0000 0xFB01.. 0xFC00 .. 0xFD00 .. 0xFE00 .. 0xFF00 .. 0xFFFF

Now the rule doesn't work anymore. The solution I found is to detect discontinuity. The IDs are consecutive, so the initialization routine continues reading while the ID(n+1)=ID(n)+1. When there's a gap, the
init function stops and found the ID and position of the newest event.

But he problems don't stop here. What happens if an event write failed?
Should I verify the real written values, by reading back them and
comparing with original data? In a reliable system, yes, I should.

I was thinking to protect an event with a CRC, so to understand at any
time if an event is corrupted (bad written). However the possibility to
have isolated corrupted events increase the complexity of the task a lot.

Suppose to write the 4th event with ID=3 at position 4 in the sector.
Now suppose the write failed. I can try to re-write the same event at poisition 5. Should I use ID=4 or ID=5? At first, I think it's better to
use ID=5. The CRC should detect that at position 4 is stored a corrupted event.
After that two events are written as well, ID=6 and 7.

Now the higher application wants to read the 4th event of the log
(starting from the newest). We know that the newest is ID=7, so the 4th
event is ID=7-4+1=4. However the event with ID=4 is corrupted, so we
should check the previous ID=3... and so on.
Now we can't have a mathematical function that returns the position of
an event starting from it's ordinal number from the newest event.


Eventually I think it's better to use a public domain library, if it exists.

--- Synchronet 3.20a-Linux NewsLink 1.114

From Newsgroup: comp.arch.embedded

On 18/04/2024 16:38, pozz wrote:

The request is very common: when some interested events occur in the
system, they, with the related timestamp, must be saved in a log. The
log must be saved in an external SPI Flash connected to a MCU. The log
has a maximum number of events. After the log is completely filled, the
new event overwrite the oldest event.

I tried to implement such library, but I eventually found it's not a
simple task, mostly if you want a reliable library that works even when errors occur (for example, when one writing fails).

I started by assigning 5 sectors of SPI Flash to the log. There are 256 events in a sector (the event is a fixed struct).
In this scenario, the maximum log size is 1024 events, because the 5th sector can be erased when the pointer reaches the end of a sector.

The first challenge is how the system can understand what is the first (newest) event in the log at startup. I solved saving a 16-bits counter
ID to each event. The initialization routine starts reading all the IDs
and taks the greatest as the last event.
However initially the log is empty, so all the IDs are 0xFFFF, the
maximum. One solution is to stop reading events when 0xFFFF is read and wrap-around ID at 0xFFFE and not 0xFFFF.

Start at 0xfffe, and count down. Or xor with 0xffff for storage. Or
wrap at 0xfffe, as you suggest. Or use 32-bit values. Or have another
way to indicate that the log entry is valid. Or, since you have a
timestamp, there's no need to track an ID - the timestamp will be
increasing monotonically.

However there's another problem. What happens after writing 65535 events
in the log? The ID restarts from 0, so the latest event hasn't the
greatest ID anymore.

These are the saved IDs after 65536 events:

    1^ SECT    2^ SECT    3^ SECT    4^ SECT    5^SECT---------->
    0xFB00 ... 0xFC00 ... 0xFD00 ... 0xFE00 ... 0xFF00 ... 0xFFFF

The rule "newest event has greatest ID" is correct yet. Now a new event
is written:

    1^ SECT-------> 2^ SECT   3^ SECT   4^ SECT   5^SECT--------->
    0x0000 0xFB01.. 0xFC00 .. 0xFD00 .. 0xFE00 .. 0xFF00 .. 0xFFFF

Now the rule doesn't work anymore. The solution I found is to detect discontinuity. The IDs are consecutive, so the initialization routine continues reading while the ID(n+1)=ID(n)+1. When there's a gap, the
init function stops and found the ID and position of the newest event.

Make your counts from 0 to 256*5 - 1, then wrap. Log entry "n" will be
at address n * sizeof(log entry), with up to 256 log entries blank.
Then you don't need to store a number at all.

But he problems don't stop here. What happens if an event write failed? Should I verify the real written values, by reading back them and
comparing with original data? In a reliable system, yes, I should.

I was thinking to protect an event with a CRC, so to understand at any
time if an event is corrupted (bad written). However the possibility to
have isolated corrupted events increase the complexity of the task a lot.

An 8-bit or 16-bit CRC is peanuts to calculate and check. Write the
whole log entry except for a byte or two (whatever is the minimum write
size for the flashes), which must be something other than 0xff's. Check
the entry after writing if you really want, then write the final byte or
two. Then if there's a power-fail in the middle, it's obvious that the
log entry is bad.

For SPI NOR flash, the risk of a bad write - other than through
unexpected resets or power fails - is negligible for most purposes.
Don't over-complicate things worrying about something that will never
happen.

Suppose to write the 4th event with ID=3 at position 4 in the sector.
Now suppose the write failed. I can try to re-write the same event at poisition 5. Should I use ID=4 or ID=5? At first, I think it's better to
use ID=5. The CRC should detect that at position 4 is stored a corrupted event.
After that two events are written as well, ID=6 and 7.

Now the higher application wants to read the 4th event of the log
(starting from the newest). We know that the newest is ID=7, so the 4th event is ID=7-4+1=4. However the event with ID=4 is corrupted, so we
should check the previous ID=3... and so on.
Now we can't have a mathematical function that returns the position of
an event starting from it's ordinal number from the newest event.

Eventually I think it's better to use a public domain library, if it
exists.

--- Synchronet 3.20a-Linux NewsLink 1.114

From Newsgroup: comp.arch.embedded

Il 18/04/2024 21:30, David Brown ha scritto:

On 18/04/2024 16:38, pozz wrote:

The request is very common: when some interested events occur in the
system, they, with the related timestamp, must be saved in a log. The
log must be saved in an external SPI Flash connected to a MCU. The log
has a maximum number of events. After the log is completely filled,
the new event overwrite the oldest event.

I tried to implement such library, but I eventually found it's not a
simple task, mostly if you want a reliable library that works even
when errors occur (for example, when one writing fails).

I started by assigning 5 sectors of SPI Flash to the log. There are
256 events in a sector (the event is a fixed struct).
In this scenario, the maximum log size is 1024 events, because the 5th
sector can be erased when the pointer reaches the end of a sector.

The first challenge is how the system can understand what is the first
(newest) event in the log at startup. I solved saving a 16-bits
counter ID to each event. The initialization routine starts reading
all the IDs and taks the greatest as the last event.
However initially the log is empty, so all the IDs are 0xFFFF, the
maximum. One solution is to stop reading events when 0xFFFF is read
and wrap-around ID at 0xFFFE and not 0xFFFF.

Start at 0xfffe, and count down.

And what to do when the counter reaches zero? It can wrap-around up to
0xfffe (that is very similar to an increasing counter from 0x0000-0xFFFE).

Or xor with 0xffff for storage.  Or
wrap at 0xfffe, as you suggest.  Or use 32-bit values.  Or have another way to indicate that the log entry is valid.

I will add a CRC for each entry and that can be used to validate the
event. An empty/erased slot filled with 0xFF will not pass CRC validation.

Or, since you have a
timestamp, there's no need to track an ID - the timestamp will be
increasing monotonically.

I don't want to use timestamps for two reasons:

- the system wall clock can be changed (the system is isolated)
- the library I'm writing doesn't know the content of "events", for
it the event is an opaque sequence of bytes.

However there's another problem. What happens after writing 65535
events in the log? The ID restarts from 0, so the latest event hasn't
the greatest ID anymore.

These are the saved IDs after 65536 events:

     1^ SECT    2^ SECT    3^ SECT    4^ SECT    5^SECT---------->
     0xFB00 ... 0xFC00 ... 0xFD00 ... 0xFE00 ... 0xFF00 ... 0xFFFF

The rule "newest event has greatest ID" is correct yet. Now a new
event is written:

     1^ SECT-------> 2^ SECT   3^ SECT   4^ SECT   5^SECT--------->
     0x0000 0xFB01.. 0xFC00 .. 0xFD00 .. 0xFE00 .. 0xFF00 .. 0xFFFF

Now the rule doesn't work anymore. The solution I found is to detect
discontinuity. The IDs are consecutive, so the initialization routine
continues reading while the ID(n+1)=ID(n)+1. When there's a gap, the
init function stops and found the ID and position of the newest event.

Make your counts from 0 to 256*5 - 1, then wrap.  Log entry "n" will be
at address n * sizeof(log entry), with up to 256 log entries blank. Then
you don't need to store a number at all.

What do you mean with log entry "0"? Is it the oldest or the newest? I
think the oldest, because that formula is imho correct in this case.

However it doesn't appear correct when the log has rotated, that happens
after writing 5x256+1 events. In this case the newest entry ("n"=1024)
is at address 0, not n*sizeof(entry).

But he problems don't stop here. What happens if an event write
failed? Should I verify the real written values, by reading back them
and comparing with original data? In a reliable system, yes, I should.

I was thinking to protect an event with a CRC, so to understand at any
time if an event is corrupted (bad written). However the possibility
to have isolated corrupted events increase the complexity of the task
a lot.

An 8-bit or 16-bit CRC is peanuts to calculate and check.

I know. Here the increased complexity wasn't related to the CRC
calculation, but to the possibility to have isolated corrupted slots in
the buffer. Taking into account these corrupted slots isn't so simple
for me.

Write the
whole log entry except for a byte or two (whatever is the minimum write
size for the flashes), which must be something other than 0xff's.  Check the entry after writing if you really want, then write the final byte or two.  Then if there's a power-fail in the middle, it's obvious that the
log entry is bad.

I don't understand why to write in two steps. I think I can write all
the bytes of an entry, including CRC. After writing, I can read back all
the bytes and check integrity.

For SPI NOR flash, the risk of a bad write - other than through
unexpected resets or power fails - is negligible for most purposes.
Don't over-complicate things worrying about something that will never happen.

Indeed unexpected resets, but mainly power failes, are possible. Rare,
but possible.

Suppose to write the 4th event with ID=3 at position 4 in the sector.
Now suppose the write failed. I can try to re-write the same event at
poisition 5. Should I use ID=4 or ID=5? At first, I think it's better
to use ID=5. The CRC should detect that at position 4 is stored a
corrupted event.
After that two events are written as well, ID=6 and 7.

Now the higher application wants to read the 4th event of the log
(starting from the newest). We know that the newest is ID=7, so the
4th event is ID=7-4+1=4. However the event with ID=4 is corrupted, so
we should check the previous ID=3... and so on.
Now we can't have a mathematical function that returns the position of
an event starting from it's ordinal number from the newest event.


Eventually I think it's better to use a public domain library, if it
exists.

--- Synchronet 3.20a-Linux NewsLink 1.114

From Newsgroup: comp.arch.embedded

On 4/18/2024 7:38 AM, pozz wrote:

The request is very common: when some interested events occur in the system, they, with the related timestamp, must be saved in a log. The log must be saved
in an external SPI Flash connected to a MCU. The log has a maximum number of events. After the log is completely filled, the new event overwrite the oldest
event.

A circular buffer.

I tried to implement such library, but I eventually found it's not a simple task, mostly if you want a reliable library that works even when errors occur
(for example, when one writing fails).

That depends on the sorts of failures that can occur and that are likely
to occur. If, for example, power fails during a write... (do you know
how your hardware will handle such an event? even if you THINK you
have some guarantee that it can't happen -- because you have an early
warning indication from the power supply?)

Can a write APPEAR successful and yet the data degrade later (so your
notion of where the next message should start changes based on whether
or not you currently KNOW where it should start vs. having to DISCOVER
where it should start?

I started by assigning 5 sectors of SPI Flash to the log. There are 256 events
in a sector (the event is a fixed struct).
In this scenario, the maximum log size is 1024 events, because the 5th sector
can be erased when the pointer reaches the end of a sector.

You're being overly conservative; expecting new events to occur WHILE you
are erasing the "extra" sector. If you can cache the errors elsewhere
while the erase is in progress, then you can use the extra sector, too
(and defer erasing until ALL sectors are full AND another event occurs)

The first challenge is how the system can understand what is the first (newest)
event in the log at startup. I solved saving a 16-bits counter ID to each event. The initialization routine starts reading all the IDs and taks the greatest as the last event.
However initially the log is empty, so all the IDs are 0xFFFF, the maximum. One
solution is to stop reading events when 0xFFFF is read and wrap-around ID at 0xFFFE and not 0xFFFF.

In your scheme, one sector will be erased or in the process of being erased. So, you can (almost) rely on its contents (unless the erase can be interrupted by a power event)

However there's another problem. What happens after writing 65535 events in the
log? The ID restarts from 0, so the latest event hasn't the greatest ID anymore.

This is the same as handling a circular buffer; how address N<M can actually represent something that occurs AFTER the data at M. Why do you see it as anything different?

These are the saved IDs after 65536 events:

    1^ SECT    2^ SECT    3^ SECT    4^ SECT    5^SECT---------->
    0xFB00 ... 0xFC00 ... 0xFD00 ... 0xFE00 ... 0xFF00 ... 0xFFFF

The rule "newest event has greatest ID" is correct yet. Now a new event is written:

    1^ SECT-------> 2^ SECT   3^ SECT   4^ SECT   5^SECT--------->
    0x0000 0xFB01.. 0xFC00 .. 0xFD00 .. 0xFE00 .. 0xFF00 .. 0xFFFF

Now the rule doesn't work anymore. The solution I found is to detect discontinuity. The IDs are consecutive, so the initialization routine continues
reading while the ID(n+1)=ID(n)+1. When there's a gap, the init function stops
and found the ID and position of the newest event.

Unless, of course, the entry at n or n+1 are corrupt...

But he problems don't stop here. What happens if an event write failed? Should
I verify the real written values, by reading back them and comparing with original data? In a reliable system, yes, I should.

So, you've answered your own question. If you want the log to have
value, then you have to put effort into ensure its integrity.

Note my earlier comment "Can a write APPEAR successful and yet..."

I was thinking to protect an event with a CRC, so to understand at any time if
an event is corrupted (bad written). However the possibility to have isolated
corrupted events increase the complexity of the task a lot.

You can't assume ANYTHING about a corrupted event.

Suppose to write the 4th event with ID=3 at position 4 in the sector. Now suppose the write failed. I can try to re-write the same event at poisition 5.
Should I use ID=4 or ID=5? At first, I think it's better to use ID=5. The CRC
should detect that at position 4 is stored a corrupted event.
After that two events are written as well, ID=6 and 7.

Now the higher application wants to read the 4th event of the log (starting from the newest). We know that the newest is ID=7, so the 4th event is ID=7-4+1=4. However the event with ID=4 is corrupted, so we should check the previous ID=3... and so on.
Now we can't have a mathematical function that returns the position of an event
starting from it's ordinal number from the newest event.

Eventually I think it's better to use a public domain library, if it exists.

You want an existing library to be able to address arbitrary log entry formats on arbitrary media with arbitrary failure modes (hardware and software)?

The only time you really care about recovering data from the log is when you actually *want* (need) to recover data. (??) Can the log automatically restart with each device reset? (and, if you want to preserve the log,
have a mechanism that inhibits logging after a "diagnostic reset")

Assuming there is no external control data for the log (i.e., that you have
to infer the structure of the log from the log's contents), a generic algorithm is to start at the first memory location that *can* hold the start of a message. Then, look at that piece of memory (taking into account any "wrap" caused by the circular nature of the buffer) and determine if it represents
a valid log message (e.g., if you store a hash of the message in the message, then the hash must be correct in addition to the format/content of the message).

If a valid message is encountered, return SUCCESS along with the length of
the message (so you will know where to start looking for the NEXT message). [You already know where THIS message starts]

If an invalid message is encountered, return FAIL along with a number of
bytes that can safely be skipped over to begin searching for the next
message. In the degenerate case, this is '1'.

Based on YOUR expected failure modes (and your write strategy), you can then walk through the log in search of "good" messages and extract their ID's (sequence numbers).

[Note that this scheme allows messages to be of variable length]

If you rely on sequence numbers or "short" timestamps that can wrap in the context of the log buffer's length, then you need to use a numbering scheme that is relatively prime wrt the buffer length otherwise you could encounter sets of messages that can appear to start "anywhere".

Note that your IDs are just abbreviated timestamps; with arbitrary time intervals between them. Using a "system time" that is always going to
increase monotonically often offers more USEFUL information as it lets
you decide how events are temporarily related (did event #5 happen
immediately after event #4? Or, two weeks later??)

Having a system time that you always reinitialize with the last noted
value (assuming you can't track time when powered off) means that chronology/sequence is always evident. (you can add events like "User
set Wall Clock Time to XX:YY:ZZ" so you can map the logged events to
"real" time for convenience -- is something happening at some particular
time of day that may be pertinent??)

--- Synchronet 3.20a-Linux NewsLink 1.114

From Newsgroup: comp.arch.embedded

On 4/19/2024 12:58 AM, Don Y wrote:

Note that your IDs are just abbreviated timestamps; with arbitrary time intervals between them.  Using a "system time" that is always going to increase monotonically often offers more USEFUL information as it lets
you decide how events are temporarily related (did event #5 happen immediately after event #4?  Or, two weeks later??)

s/temporarily/temporally/

--- Synchronet 3.20a-Linux NewsLink 1.114

From Newsgroup: comp.arch.embedded

On 18/04/2024 22:36, pozz wrote:

Il 18/04/2024 21:30, David Brown ha scritto:

On 18/04/2024 16:38, pozz wrote:

The request is very common: when some interested events occur in the
system, they, with the related timestamp, must be saved in a log. The
log must be saved in an external SPI Flash connected to a MCU. The
log has a maximum number of events. After the log is completely
filled, the new event overwrite the oldest event.

I tried to implement such library, but I eventually found it's not a
simple task, mostly if you want a reliable library that works even
when errors occur (for example, when one writing fails).

I started by assigning 5 sectors of SPI Flash to the log. There are
256 events in a sector (the event is a fixed struct).
In this scenario, the maximum log size is 1024 events, because the
5th sector can be erased when the pointer reaches the end of a sector.

The first challenge is how the system can understand what is the
first (newest) event in the log at startup. I solved saving a 16-bits
counter ID to each event. The initialization routine starts reading
all the IDs and taks the greatest as the last event.
However initially the log is empty, so all the IDs are 0xFFFF, the
maximum. One solution is to stop reading events when 0xFFFF is read
and wrap-around ID at 0xFFFE and not 0xFFFF.

Start at 0xfffe, and count down.

And what to do when the counter reaches zero? It can wrap-around up to 0xfffe (that is very similar to an increasing counter from 0x0000-0xFFFE).

How big are your log entries? How many entries are realistic in the
lifetime of the system?

Or xor with 0xffff for storage.  Or
wrap at 0xfffe, as you suggest.  Or use 32-bit values.  Or have
another way to indicate that the log entry is valid.

I will add a CRC for each entry and that can be used to validate the
event. An empty/erased slot filled with 0xFF will not pass CRC validation.

That's usually fine.

Or, since you have a timestamp, there's no need to track an ID - the
timestamp will be increasing monotonically.

I don't want to use timestamps for two reasons:

- the system wall clock can be changed (the system is isolated)
- the library I'm writing doesn't know the content of "events", for
  it the event is an opaque sequence of bytes.

OK.

However there's another problem. What happens after writing 65535
events in the log? The ID restarts from 0, so the latest event hasn't
the greatest ID anymore.

These are the saved IDs after 65536 events:

     1^ SECT    2^ SECT    3^ SECT    4^ SECT    5^SECT---------->
     0xFB00 ... 0xFC00 ... 0xFD00 ... 0xFE00 ... 0xFF00 ... 0xFFFF

The rule "newest event has greatest ID" is correct yet. Now a new
event is written:

     1^ SECT-------> 2^ SECT   3^ SECT   4^ SECT   5^SECT--------->
     0x0000 0xFB01.. 0xFC00 .. 0xFD00 .. 0xFE00 .. 0xFF00 .. 0xFFFF

Now the rule doesn't work anymore. The solution I found is to detect
discontinuity. The IDs are consecutive, so the initialization routine
continues reading while the ID(n+1)=ID(n)+1. When there's a gap, the
init function stops and found the ID and position of the newest event.

Make your counts from 0 to 256*5 - 1, then wrap.  Log entry "n" will
be at address n * sizeof(log entry), with up to 256 log entries blank.
Then you don't need to store a number at all.

What do you mean with log entry "0"? Is it the oldest or the newest? I
think the oldest, because that formula is imho correct in this case.

However it doesn't appear correct when the log has rotated, that happens after writing 5x256+1 events. In this case the newest entry ("n"=1024)
is at address 0, not n*sizeof(entry).

(I misread your "5 sectors of an SPI flash chip" as "5 SPI flash chips"
when first replying. It makes no real difference to what I wrote, but I
might have used "chip" instead of "sector".)

You have 256 entries per flash sector, and 5 flash sectors. For the log
entry number "n" - where "n" is an abstract count that never wraps, your
index "i" into the flash array is (n % 5*256). The sector number is
then (i / 256), and the index into the sector is (i % 256). The
position in the log is determined directly by the entry number, and you
don't actually need to store it anywhere.

Think of this a different way - dispense with the log entry numbers
entirely. When you start up, scan the flash to find the next free slot.
You do this by looking at slot 0 first. If that is not empty, keep
scanning until you find a free slot - that's the next free slot. If
slot 0 is empty, scan until you have non-empty slots, then keep going
until you get a free one again, and that's the next free slot. If you
never find a used slot, or fail to find a free slot after the non-free
slots, then your first free slot is slot 0.

Any new logs are then put in this slot, moving forward. If you need to
read out old logs, move backwards. When storing new logs, as you are
nearing the end of a flash sector (how near depends on the sector erase
time and how often events can occur), start the erase of the next sector
in line.

But he problems don't stop here. What happens if an event write
failed? Should I verify the real written values, by reading back them
and comparing with original data? In a reliable system, yes, I should.

I was thinking to protect an event with a CRC, so to understand at
any time if an event is corrupted (bad written). However the
possibility to have isolated corrupted events increase the complexity
of the task a lot.

An 8-bit or 16-bit CRC is peanuts to calculate and check.

I know. Here the increased complexity wasn't related to the CRC
calculation, but to the possibility to have isolated corrupted slots in
the buffer. Taking into account these corrupted slots isn't so simple
for me.

Think how such corruption could happen, and its consequences. For most
event logs, it is simply not going to occur in the lifetime of working products - and if it does, as an isolated error in an event log, it
doesn't matter significantly. Errors in a sensibly designed SPI NOR
flash system would be an indication of serious hardware problems such as erratic power supplies, and then the log is the least of your concerns.

The only thing to consider is a reset or power failure in the middle of writing a log event.

Write the whole log entry except for a byte or two (whatever is the
minimum write size for the flashes), which must be something other
than 0xff's.  Check the entry after writing if you really want, then
write the final byte or two.  Then if there's a power-fail in the
middle, it's obvious that the log entry is bad.

I don't understand why to write in two steps. I think I can write all
the bytes of an entry, including CRC. After writing, I can read back all
the bytes and check integrity.

No, you can't. Or to be more accurate, you /can/, but it leaves you
open to the risk of undetectable errors in the event of a power fail in
the middle of writing. An 8-bit CRC has a 1 in 256 chance of passing if
data is partially written.

If you have enough capacitance on the board to ensure that any write completes, and detection of failure of external power so that you never
start a log write after power has failed, then it is probably safe to do
the write as a single action.

For SPI NOR flash, the risk of a bad write - other than through
unexpected resets or power fails - is negligible for most purposes.
Don't over-complicate things worrying about something that will never
happen.

Indeed unexpected resets, but mainly power failes, are possible. Rare,
but possible.

Yes.

Suppose to write the 4th event with ID=3 at position 4 in the sector.
Now suppose the write failed. I can try to re-write the same event at
poisition 5. Should I use ID=4 or ID=5? At first, I think it's better
to use ID=5. The CRC should detect that at position 4 is stored a
corrupted event.
After that two events are written as well, ID=6 and 7.

Now the higher application wants to read the 4th event of the log
(starting from the newest). We know that the newest is ID=7, so the
4th event is ID=7-4+1=4. However the event with ID=4 is corrupted, so
we should check the previous ID=3... and so on.
Now we can't have a mathematical function that returns the position
of an event starting from it's ordinal number from the newest event.


Eventually I think it's better to use a public domain library, if it
exists.

In my experience, these things are simple enough to implement, and
specific enough to the application and target, that I don't see the need
of a library.


--- Synchronet 3.20a-Linux NewsLink 1.114

From Newsgroup: comp.arch.embedded

Il 19/04/2024 10:28, David Brown ha scritto:

On 18/04/2024 22:36, pozz wrote:

Il 18/04/2024 21:30, David Brown ha scritto:

On 18/04/2024 16:38, pozz wrote:

The request is very common: when some interested events occur in the
system, they, with the related timestamp, must be saved in a log.
The log must be saved in an external SPI Flash connected to a MCU.
The log has a maximum number of events. After the log is completely
filled, the new event overwrite the oldest event.

I tried to implement such library, but I eventually found it's not a
simple task, mostly if you want a reliable library that works even
when errors occur (for example, when one writing fails).

I started by assigning 5 sectors of SPI Flash to the log. There are
256 events in a sector (the event is a fixed struct).
In this scenario, the maximum log size is 1024 events, because the
5th sector can be erased when the pointer reaches the end of a sector. >>>>
The first challenge is how the system can understand what is the
first (newest) event in the log at startup. I solved saving a
16-bits counter ID to each event. The initialization routine starts
reading all the IDs and taks the greatest as the last event.
However initially the log is empty, so all the IDs are 0xFFFF, the
maximum. One solution is to stop reading events when 0xFFFF is read
and wrap-around ID at 0xFFFE and not 0xFFFF.

Start at 0xfffe, and count down.

And what to do when the counter reaches zero? It can wrap-around up to
0xfffe (that is very similar to an increasing counter from
0x0000-0xFFFE).

How big are your log entries?  How many entries are realistic in the lifetime of the system?

16 bytes each entry. It's difficult to reach 0xFFFF events in the log,
but why limit our fantasy? :-D

With 20 events per day, the log will be filled after 9 years. It's a
very long life, but I think I have solved the problem to understand if
the ID was wrapped-around.

Or xor with 0xffff for storage.  Or
wrap at 0xfffe, as you suggest.  Or use 32-bit values.  Or have
another way to indicate that the log entry is valid.

I will add a CRC for each entry and that can be used to validate the
event. An empty/erased slot filled with 0xFF will not pass CRC
validation.

That's usually fine.

Or, since you have a timestamp, there's no need to track an ID - the
timestamp will be increasing monotonically.

I don't want to use timestamps for two reasons:

- the system wall clock can be changed (the system is isolated)
- the library I'm writing doesn't know the content of "events", for
   it the event is an opaque sequence of bytes.

OK.

However there's another problem. What happens after writing 65535
events in the log? The ID restarts from 0, so the latest event
hasn't the greatest ID anymore.

These are the saved IDs after 65536 events:

     1^ SECT    2^ SECT    3^ SECT    4^ SECT    5^SECT---------->
     0xFB00 ... 0xFC00 ... 0xFD00 ... 0xFE00 ... 0xFF00 ... 0xFFFF >>>>
The rule "newest event has greatest ID" is correct yet. Now a new
event is written:

     1^ SECT-------> 2^ SECT   3^ SECT   4^ SECT   5^SECT--------->
     0x0000 0xFB01.. 0xFC00 .. 0xFD00 .. 0xFE00 .. 0xFF00 .. 0xFFFF >>>>
Now the rule doesn't work anymore. The solution I found is to detect
discontinuity. The IDs are consecutive, so the initialization
routine continues reading while the ID(n+1)=ID(n)+1. When there's a
gap, the init function stops and found the ID and position of the
newest event.

Make your counts from 0 to 256*5 - 1, then wrap.  Log entry "n" will
be at address n * sizeof(log entry), with up to 256 log entries
blank. Then you don't need to store a number at all.

What do you mean with log entry "0"? Is it the oldest or the newest? I
think the oldest, because that formula is imho correct in this case.

However it doesn't appear correct when the log has rotated, that
happens after writing 5x256+1 events. In this case the newest entry
("n"=1024) is at address 0, not n*sizeof(entry).

(I misread your "5 sectors of an SPI flash chip" as "5 SPI flash chips"
when first replying.  It makes no real difference to what I wrote, but I might have used "chip" instead of "sector".)

You have 256 entries per flash sector, and 5 flash sectors. For the log entry number "n" - where "n" is an abstract count that never wraps, your index "i" into the flash array is (n % 5*256). The sector number is
then (i / 256), and the index into the sector is (i % 256). The
position in the log is determined directly by the entry number, and you don't actually need to store it anywhere.

Think of this a different way - dispense with the log entry numbers entirely.  When you start up, scan the flash to find the next free slot.
 You do this by looking at slot 0 first.  If that is not empty, keep scanning until you find a free slot - that's the next free slot.  If
slot 0 is empty, scan until you have non-empty slots, then keep going
until you get a free one again, and that's the next free slot.  If you never find a used slot, or fail to find a free slot after the non-free slots, then your first free slot is slot 0.

Any new logs are then put in this slot, moving forward.  If you need to read out old logs, move backwards.  When storing new logs, as you are nearing the end of a flash sector (how near depends on the sector erase
time and how often events can occur), start the erase of the next sector
in line.

Yes, it is what I already do. However I disagree on the formula.

The higher-layer application requests log entry 0. What is it? The
newest event. My eventlog library should convert 0 to the slot index in
the Flash, that is directly related to the Flash addres (I don't really
need the number of sector here).

If the log is empty, event 0 for the application is slot 0 for the
eventlog library. However, if there are three events in the log, event 0
for the application is slot 2 for the eventlog library.

The application doesn't know anything about what I named the ID of the
event. It's just a number used by the lower-layer eventlog module to
find the first free slot at startup.

But he problems don't stop here. What happens if an event write
failed? Should I verify the real written values, by reading back
them and comparing with original data? In a reliable system, yes, I
should.

I was thinking to protect an event with a CRC, so to understand at
any time if an event is corrupted (bad written). However the
possibility to have isolated corrupted events increase the
complexity of the task a lot.

An 8-bit or 16-bit CRC is peanuts to calculate and check.

I know. Here the increased complexity wasn't related to the CRC
calculation, but to the possibility to have isolated corrupted slots
in the buffer. Taking into account these corrupted slots isn't so
simple for me.

Think how such corruption could happen, and its consequences.  For most event logs, it is simply not going to occur in the lifetime of working products - and if it does, as an isolated error in an event log, it
doesn't matter significantly.  Errors in a sensibly designed SPI NOR
flash system would be an indication of serious hardware problems such as erratic power supplies, and then the log is the least of your concerns.

The only thing to consider is a reset or power failure in the middle of writing a log event.

Yes, I agree with you.

Write the whole log entry except for a byte or two (whatever is the
minimum write size for the flashes), which must be something other
than 0xff's.  Check the entry after writing if you really want, then
write the final byte or two.  Then if there's a power-fail in the
middle, it's obvious that the log entry is bad.

I don't understand why to write in two steps. I think I can write all
the bytes of an entry, including CRC. After writing, I can read back
all the bytes and check integrity.

No, you can't.  Or to be more accurate, you /can/, but it leaves you
open to the risk of undetectable errors in the event of a power fail in
the middle of writing.  An 8-bit CRC has a 1 in 256 chance of passing if data is partially written.

I can use crc16. We are talking of power-fail/reset in the middle of
event writing, this is possible, but rare. I think the case a crc16
reporting a valid slot when the writing was wrong or incomplete is
extremely rare.

If you have enough capacitance on the board to ensure that any write completes, and detection of failure of external power so that you never start a log write after power has failed, then it is probably safe to do
the write as a single action.

No I can't know in advance when the power will fail.

For SPI NOR flash, the risk of a bad write - other than through
unexpected resets or power fails - is negligible for most purposes.
Don't over-complicate things worrying about something that will never
happen.

Indeed unexpected resets, but mainly power failes, are possible. Rare,
but possible.

Yes.

Suppose to write the 4th event with ID=3 at position 4 in the
sector. Now suppose the write failed. I can try to re-write the same
event at poisition 5. Should I use ID=4 or ID=5? At first, I think
it's better to use ID=5. The CRC should detect that at position 4 is
stored a corrupted event.
After that two events are written as well, ID=6 and 7.

Now the higher application wants to read the 4th event of the log
(starting from the newest). We know that the newest is ID=7, so the
4th event is ID=7-4+1=4. However the event with ID=4 is corrupted,
so we should check the previous ID=3... and so on.
Now we can't have a mathematical function that returns the position
of an event starting from it's ordinal number from the newest event.


Eventually I think it's better to use a public domain library, if it
exists.

In my experience, these things are simple enough to implement, and
specific enough to the application and target, that I don't see the need
of a library.

--- Synchronet 3.20a-Linux NewsLink 1.114