(VRT_synth_page() still works, but is now deprecated.)

(While auditing for #2800)

In any good relationship, one should be clear about expectations. And
VCL authors want to know what VCC wants to hear, so tell them.

;)

besides the incorrect bits we want to check for.

Ref c4ec4712

Exposed by #3163

as requested by phk.

I found no way to avoid having to escape the backslashes because python,
even for uninterpolated ''' long strings, sees them as line continuations.

The generated output is identical except for two comments (in the code
and the code output by the code).

--- ./lib/libvcc/vcc_fixed_token.o.c	2021-02-02 10:40:03.725244373 +0100
+++ ./lib/libvcc/vcc_fixed_token.c	2021-02-02 10:52:04.921259718 +0100
@@ -834,10 +834,7 @@
 	    "uct vre **, const char *);\nvoid VPI_re_fini(struct vre *);\n"
 	);
 	VSB_cat(sb, "\n");
-
-	/* ../include/vcc_assert.h */
-
-	VSB_cat(sb, "/* ---===### include/vcc_assert.h ###===--- */\n\n");
+	VSB_cat(sb, "/* ---===### vgc asserts (generate.py) ###===--- */\n\n");
 	VSB_cat(sb, "#define assert(e)\t\t\t\t\t\t\t\\\ndo {\t\t\t\t"
 	    "\t\t\t\t\t\\\n\tif (!(e)) {\t\t\t\t\t\t\t\\\n\t\tVPI_Fail(__func"
 	    "__, __FILE__, __LINE__, #e);\t\t\\\n\t}\t\t\t\t\t\t\t\t\\\n"

to split the output formattings from emit_file() from reading the file

The produced output lib/libvcc/vcc_fixed_token.c is identical
before/after.

Fixes #3509

We wrap VAS_Fail() as VPI_Fail() and add an assert() macro to the
Varnish Generated C (VGC) code. This assert() definition is basically
a copy of that in vas.h, but deliberately added to a source file which
is only used by generate.py for VGC, such that this slightly different
definition be only visible to VGC.

I also pondered the option to include VCL source information in
VPI_Fail: By using the information updated by VPI_count(), we could
have that, but as the assert is, for now, intended only to ensure
correctness of VCC and core VRT code, I decided against this
complication.

With assert() in place for VGC, we arm the assertions from
75acb5cc

It seems they do not parse attributes and how I read
--macro-file-builtins they just define them to NIL.

How 16, you ask? Consider alignment...

One of the VCL 4.1 changes was to make the protocol read-only, which
we've successfully done for all variables except resp.proto that
kept the exact same definition for both 4.0 and 4.1 versions, despite
having distinct definitions.

How I added generation of the assertion code was motivated by how I
wanted to lay out more changes in that area later, but generating parts
of the function body before the header turned out to not help clarity.

Ref 75acb5cc

Bail out upon first error, as for the non-recursive case.

in the course of #3163 we want to add a preamble to VCL subs, which,
ideally, would perform as many checks based on constants as possible.

This is a first step towards such checks: We add a commented-out
assert() on ctx->method for built-in subs only. To add the same for
custom subs, we will need the okmask from #3163, for which I would first
want to coordinate with phk.

Also, I would want to coordinate with him on arming the assertion, as
there might be a good reason why assert() is not yet defined for vgc (it
is already used in unused macros, though).

Having assertions would help humans, flexelint and compilers alike. Until
armed, the assertion helps humans at least.

The resulting vgc looks like this:

void v_matchproto_(vcl_func_f)
VGC_function_vcl_backend_error(VRT_CTX)
{
  // assert(ctx->method == (VCL_MET_BACKEND_ERROR));

  /* ... from ('Builtin' Line 165 Pos 23) */

Add a VCL_BLOB for a connection preamble to struct vrt_backend (mortally offending FlexeLint in the process.)

Quote more. Do not quote too much

Ref da2409c3

We have now seen a second case after
184736fa, and a simple internal rename
from bool to boolean does the job.

Should have no user-visible impact.

In fee476ac VRB_Ignore(), VRB_Cache()
and VRB_Free() were rightly out of vmods' sight.

But VRB_Iterate() is legitemately required for vmods, and including it
in the move seems to be an accident.

Advise from shellcheck taken undogmatically.

(I always forget: is $() for `..` posix-sh or not?)

information to make them easier to clone.

Salvaged from #3503

The rationale is that if we add another type of task in the future, for
example probe tasks, we won't need to update the default mask for VMOD
functions and object methods.

Refs #3498

With a default read mask based on the type of symbol. Starting with a
default comprehensive mask should then help implement the $Restrict
stanza from VIP4, simply by AND-ing the masks.

Simply take an ssize_t and return size_t asserting that the value is
greater than or equal to zero.

The name might still not be optimimal, as well as choice of *size_t.

Addresses Coverity CID 1472401, CID 1472402

This way every failure can be tested again without selecting a specific
subdirectory.

adds a variant of PAN_dump_struct() which does not start another
indentation level on a new line.

A copy-pasta in 55dffc35 prevented the
top request from being dumped.

55dffc35 introduced a PAN_dump_struct()
call, which implies printing '{' and VSB_indent(vsb, 2). Fix missing '}'
and wrong indentation.

Also remove redundant "priv" output, which now happens in PAN_dump_struct()

Diff best viewed ignoring whitespace (-b option).

This test fails constantly now on vtest arm / Freebsd 13.0 for timeout.

As vcl compilations are probably the most expensive operations of vtcs,
we speed up this test by reducing the number of vcls loaded from 19 to
11. We do this by collapsing all of the client side failure tests into a
single vcl. Being at it, we also reduce logexpect synchronization points.

The number of vcl loads could be reduced further by also collapsing the
backend side, but as this would reduce clarity more than for the client
side alone, I first wanted to see if this first step brings down tuntime
enough.

Do not overwrite an already latched errno.

Note that it does not make any difference at the moment because we only
ever use ENOMEM

Ref #3502

This is similar to #3502, but the out-of-storage condition happens after
the object has been created successfully.

Use a logexpect to ensure we test as intended.

We set up fetch filters (VFPs) before initializing a storage object to
fetch into. If that fails, we close the filters again.

The esi_gzip VFP interacts with the storage object and was missing
error handling for the case of a teardown with no storage object.

Implementation:

We make sure that for the VFP_ERROR case, we do not interact with the
storage object (which makes no sense if that is going to be C-4'ed any
moment).

vfp_vep_callback() keeps its hands off if (struct vef_priv).error is
set, so we use that to avoid complications in the code.