Skip to contentSkip to conversation

symbols: discard stray file symbols

PATCH In review Jan Beulich 4 participants 7 messages Apr 16, 2025 – Oct 21, 2025

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 symbols: discard stray file symbols

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

---
Factor 2 may in principle still be too small: We zap what looks like real file symbols already in read_symbol(), so table_cnt doesn't really reflect the number of symbol table entries encountered. It has proven to work for me in practice though, with still some leeway left.

xen/tools/symbols.c +10 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
 	if (strstr((char *)s->sym + offset, "_compiled."))
 		return 0;
 
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)
+		return 0;
+
 	return 1;
 }
 
#1 lore raw
Quoted 24 lines · Jan Beulich

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

---
Factor 2 may in principle still be too small: We zap what looks like real file symbols already in read_symbol(), so table_cnt doesn't really reflect the number of symbol table entries encountered. It has proven to work for me in practice though, with still some leeway left.

xen/tools/symbols.c +2 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
  	if (strstr((char *)s->sym + offset, "_compiled."))
  		return 0;
 
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the

xen/tools/symbols.c static int symbol_valid(struct sym_entry · line 227

Quoted 10 lines
xen/tools/symbols.c +8 -0
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)
+		return 0;
+
  	return 1;
  }

I looked at this. It'll drop symbols, but I don't know enough to give an R-b. I can't give an actionable A-b either. Maybe someone else can chime in.

Maybe this is just showing my lack of knowledge, but could any symbol starting "?." be considered invalid? I don't think I've ever seen any like that.

Regards,
Jason

#2 lore raw
Quoted 40 lines · Jason Andryuk
Quoted 24 lines · Jan Beulich

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

---
Factor 2 may in principle still be too small: We zap what looks like real file symbols already in read_symbol(), so table_cnt doesn't really reflect the number of symbol table entries encountered. It has proven to work for me in practice though, with still some leeway left.

xen/tools/symbols.c +2 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
  	if (strstr((char *)s->sym + offset, "_compiled."))
  		return 0;
 
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the

xen/tools/symbols.c static int symbol_valid(struct sym_entry · line 227

Quoted 10 lines
xen/tools/symbols.c +8 -0
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)
+		return 0;
+
  	return 1;
  }

I looked at this. It'll drop symbols, but I don't know enough to give
an R-b. I can't give an actionable A-b either. Maybe someone else can
chime in.

Quoted 3 lines

Maybe this is just showing my lack of knowledge, but could any symbol starting "?." be considered invalid? I don't think I've ever seen any like that.

With quotation, almost any symbol name can appear in principle. I wouldn't want to judge symbol validity by its name. What's more important here, though, is that sym[0] isn't part of the name; it's the symbol's type as taken from nm's output. We're therefore heuristically looking at symbols of unknown type with a dot as the first character (as section names would conventionally have it).

Jan

#3 lore raw

Ping: [PATCH] symbols: discard stray file symbols

In reply to Jan Beulich
Quoted 10 lines · Jan Beulich

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

May I please ask for feedback here, so that hopefully we can have this sorted in 4.21?

Jan

Quoted 25 lines

---
Factor 2 may in principle still be too small: We zap what looks like real file symbols already in read_symbol(), so table_cnt doesn't really reflect the number of symbol table entries encountered. It has proven to work for me in practice though, with still some leeway left.

xen/tools/symbols.c +10 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
 	if (strstr((char *)s->sym + offset, "_compiled."))
 		return 0;
 
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)
+		return 0;
+
 	return 1;
 }
 
#4 lore raw

Re: Ping: [PATCH] symbols: discard stray file symbols

Quoted 10 lines · Jan Beulich
Quoted 9 lines · Jan Beulich

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

Quoted 3 lines
Quoted 1 line

May I please ask for feedback here, so that hopefully we can have this sorted in 4.21?

It is okay for me to have this change in 4.21:

~ Oleksii

Quoted 28 lines

Jan

Quoted 25 lines

---
Factor 2 may in principle still be too small: We zap what looks like real file symbols already in read_symbol(), so table_cnt doesn't really reflect the number of symbol table entries encountered. It has proven to work for me in practice though, with still some leeway left.

xen/tools/symbols.c +10 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
  	if (strstr((char *)s->sym + offset, "_compiled."))
  		return 0;
 
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)
+		return 0;
+
  	return 1;
  }
 
Acked-by on v1 1/1
#5 lore raw
In reply to Jan Beulich
Quoted 10 lines · Jan Beulich

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

Quoted 10 lines

---
Factor 2 may in principle still be too small: We zap what looks like real file symbols already in read_symbol(), so table_cnt doesn't really reflect the number of symbol table entries encountered. It has proven to work for me in practice though, with still some leeway left.

xen/tools/symbols.c +0 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
 	if (strstr((char *)s->sym + offset, "_compiled."))

xen/tools/symbols.c static int symbol_valid(struct sym_entry · line 223

Quoted 10 lines
xen/tools/symbols.c +8 -0
 		return 0;
 
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)

Maybe a naive question, but couldn't you drop everything below __XEN_VIRT_START, as we shouldn't have any symbols below that address?

Thanks, Roger.

#6 lore raw
Quoted 10 lines · Roger Pau Monné
Quoted 9 lines · Jan Beulich

By observation GNU ld 2.25 may emit file symbols for .data.read_mostly when linking xen.efi. Due to the nature of file symbols in COFF symbol tables (see the code comment) the symbols_offsets[] entries for such symbols would cause assembler warnings regarding value truncation. Of course the resulting entries would also be both meaningless and useless. Add a heuristic to get rid of them, really taking effect only when --all-symbols is specified (otherwise these symbols are discarded anyway).

Quoted 3 lines
Quoted 1 line

Thanks.

Quoted 6 lines
Quoted 6 lines
xen/tools/symbols.c +0 -0
@@ -213,6 +213,16 @@ static int symbol_valid(struct sym_entry
 	if (strstr((char *)s->sym + offset, "_compiled."))
 		return 0;
 
Quoted 12 lines

xen/tools/symbols.c static int symbol_valid(struct sym_entry · line 223

Quoted 8 lines
xen/tools/symbols.c +8 -0
+	/* At least GNU ld 2.25 may emit bogus file symbols referencing a
+	 * section name while linking xen.efi. In COFF symbol tables the
+	 * "value" of file symbols is a link (symbol table index) to the next
+	 * file symbol. Since file (and other) symbols (can) come with one
+	 * (or in principle more) auxiliary symbol table entries, the value in
+	 * this heuristic is bounded to twice the number of symbols we have
+	 * found. See also read_symbol() as to the '?' checked for here. */
+	if (s->sym[0] == '?' && s->sym[1] == '.' && s->addr < table_cnt * 2)

Maybe a naive question, but couldn't you drop everything below __XEN_VIRT_START, as we shouldn't have any symbols below that address?

If we assumed so, then that might be an option. Such an assumption doesn't look safe to me, though. See how e.g. hv_hcall_page is outside of the Xen image (albeit still within __XEN_VIRT_{START,END}). I wouldn't want to preclude architectures playing "interesting" games with symbols, in particular when - unlike x86 - they have the entire VA space for their use.

Jan