Add support for tracing of IRELATIVE PLT entries

- Because the IRELATIVE entries have no associated symbol name, we
  need to allow arch_elf_add_plt_entry to override the name.  This is
  done by that callback returning PLT_OK and returning the new symbol
  via libsym-chain return argument.  Filtering is postponed until we
  have that symbol, and the filter is applied to the whole returned
  chain.

- Add linux_elf_add_plt_entry_irelative to support proper naming of
  IRELATIVE PLT entries.  This needs to be called from arch backend,
  as the numbers of IRELATIVE relocations differ per-architecture.

3 files changed, 114 insertions, 12 deletions

diff --git a/ltrace-elf.c b/ltrace-elf.c
index 427c62a..5a228c5 100644
--- a/ltrace-elf.c
+++ b/ltrace-elf.c
@@ -668,6 +668,24 @@ mark_chain_latent(struct library_symbol *libsym)
 	}
 }
 
+static void
+filter_symbol_chain(struct filter *filter,
+		    struct library_symbol **libsymp, struct library *lib)
+{
+	assert(libsymp != NULL);
+	struct library_symbol **ptr = libsymp;
+	while (*ptr != NULL) {
+		if (filter_matches_symbol(filter, (*ptr)->name, lib)) {
+			ptr = &(*ptr)->next;
+		} else {
+			struct library_symbol *sym = *ptr;
+			*ptr = (*ptr)->next;
+			library_symbol_destroy(sym);
+			free(sym);
+		}
+	}
+}
+
 static int
 populate_plt(struct process *proc, const char *filename,
 	     struct ltelf *lte, struct library *lib,
@@ -693,30 +711,36 @@ populate_plt(struct process *proc, const char *filename,
 
 		char const *name = lte->dynstr + sym.st_name;
 
-		/* If the symbol wasn't matched, reject it, unless we
-		 * need to keep latent PLT breakpoints for tracing
-		 * exports.  */
 		int matched = filter_matches_symbol(options.plt_filter,
 						    name, lib);
-		if (!matched && !latent_plts)
-			continue;
 
 		struct library_symbol *libsym = NULL;
 		switch (arch_elf_add_plt_entry(proc, lte, name,
 					       &rela, i, &libsym)) {
+		case PLT_FAIL:
+				return -1;
+
 		case PLT_DEFAULT:
+			/* Add default entry to the beginning of LIBSYM.  */
 			if (default_elf_add_plt_entry(proc, lte, name,
 						      &rela, i, &libsym) < 0)
-			/* Fall through.  */
-		case PLT_FAIL:
 				return -1;
 			/* Fall through.  */
 		case PLT_OK:
+			/* If we didn't match the PLT entry up there,
+			 * filter the chain to only include the
+			 * matching symbols (but include all if we are
+			 * adding latent symbols).  This is to allow
+			 * arch_elf_add_plt_entry to override the PLT
+			 * symbol's name.  */
+			if (!matched && !latent_plts)
+				filter_symbol_chain(options.plt_filter,
+						    &libsym, lib);
 			if (libsym != NULL) {
 				/* If we are adding those symbols just
 				 * for tracing exports, mark them all
 				 * latent.  */
-				if (!matched)
+				if (!matched && latent_plts)
 					mark_chain_latent(libsym);
 				library_add_symbol(lib, libsym);
 			}
@@ -795,7 +819,6 @@ populate_this_symtab(struct process *proc, const char *filename,
 			continue;
 		}
 
-		/* XXX support IFUNC as well.  */
 		if (GELF_ST_TYPE(sym.st_info) != STT_FUNC
 		    || sym.st_value == 0
 		    || sym.st_shndx == STN_UNDEF)
diff --git a/sysdeps/linux-gnu/trace.c b/sysdeps/linux-gnu/trace.c
index 677438e..5567eb6 100644
--- a/sysdeps/linux-gnu/trace.c
+++ b/sysdeps/linux-gnu/trace.c
@@ -24,25 +24,28 @@
 #include "config.h"
 
 #include <asm/unistd.h>
-#include <sys/types.h>
-#include <sys/wait.h>
 #include <assert.h>
 #include <errno.h>
+#include <gelf.h>
+#include <stdbool.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/types.h>
+#include <sys/wait.h>
 #include <unistd.h>
 
 #ifdef HAVE_LIBSELINUX
 # include <selinux/selinux.h>
 #endif
 
-#include "linux-gnu/trace.h"
 #include "linux-gnu/trace-defs.h"
+#include "linux-gnu/trace.h"
 #include "backend.h"
 #include "breakpoint.h"
 #include "debug.h"
 #include "events.h"
+#include "ltrace-elf.h"
 #include "options.h"
 #include "proc.h"
 #include "ptrace.h"
@@ -1220,3 +1223,61 @@ umovebytes(struct process *proc, void *addr, void *laddr, size_t len)
 
 	return bytes_read;
 }
+
+struct irelative_name_data_t {
+	GElf_Addr addr;
+	const char *found_name;
+};
+
+static enum callback_status
+irelative_name_cb(GElf_Sym *symbol, const char *name, void *d)
+{
+	struct irelative_name_data_t *data = d;
+
+	if (symbol->st_value == data->addr) {
+		bool is_ifunc = false;
+#ifdef STT_GNU_IFUNC
+		is_ifunc = GELF_ST_TYPE(symbol->st_info) == STT_GNU_IFUNC;
+#endif
+		data->found_name = name;
+
+		/* Keep looking, unless we found the actual IFUNC
+		 * symbol.  What we matched may have been a symbol
+		 * denoting the resolver function, which would have
+		 * the same address.  */
+		return CBS_STOP_IF(is_ifunc);
+	}
+
+	return CBS_CONT;
+}
+
+enum plt_status
+linux_elf_add_plt_entry_irelative(struct process *proc, struct ltelf *lte,
+				  GElf_Rela *rela, size_t ndx,
+				  struct library_symbol **ret)
+
+{
+	struct irelative_name_data_t data = { rela->r_addend, NULL };
+	if (rela->r_addend != 0
+	    && elf_each_symbol(lte, 0,
+			       irelative_name_cb, &data).status < 0)
+		return -1;
+
+	const char *name;
+	if (data.found_name != NULL) {
+		name = data.found_name;
+	} else {
+#define NAME "IREL."
+		/* NAME\0 + 0x + digits.  */
+		char *tmp_name = alloca(sizeof NAME + 2 + 16);
+		sprintf(tmp_name, NAME "%#" PRIx64,
+			(uint64_t)rela->r_addend);
+		name = tmp_name;
+#undef NAME
+	}
+
+	if (default_elf_add_plt_entry(proc, lte, name, rela, ndx, ret) < 0)
+		return PLT_FAIL;
+
+	return PLT_OK;
+}
diff --git a/sysdeps/linux-gnu/trace.h b/sysdeps/linux-gnu/trace.h
index 5bb8380..bb52122 100644
--- a/sysdeps/linux-gnu/trace.h
+++ b/sysdeps/linux-gnu/trace.h
@@ -118,4 +118,22 @@ int process_install_stopping_handler
 void linux_ptrace_disable_and_singlestep(struct process_stopping_handler *self);
 void linux_ptrace_disable_and_continue(struct process_stopping_handler *self);
 
+/* When main binary needs to call an IFUNC function defined in the
+ * binary itself, a PLT entry is set up so that dynamic linker can get
+ * involved and resolve the symbol.  But unlike other PLT relocation,
+ * this one can't rely on symbol table being available.  So it doesn't
+ * reference the symbol by its name, but by its address, and
+ * correspondingly, has another type.  When arch backend wishes to
+ * support these IRELATIVE relocations, it should override
+ * arch_elf_add_plt_entry and dispatch to this function for IRELATIVE
+ * relocations.
+ *
+ * This function behaves as arch_elf_add_plt_entry, except that it
+ * doesn't take name for a parameter, but instead looks up the name in
+ * symbol tables in LTE.  */
+enum plt_status linux_elf_add_plt_entry_irelative(struct process *proc,
+						  struct ltelf *lte,
+						  GElf_Rela *rela, size_t ndx,
+						  struct library_symbol **ret);
+
 #endif /* _LTRACE_LINUX_TRACE_H_ */