| Message ID | 20170614203025.7581-1-robh@kernel.org |
|---|---|
| Headers | show |
| Series | DT printf format specifiers | expand |
On Wed, 2017-06-14 at 15:30 -0500, Rob Herring wrote: > From: Pantelis Antoniou <pantelis.antoniou@konsulko.com> I think the commit subject is wrong. It adds an "of" specific bit to vsprintf.c. The subject should be 'vsprintf: Add %p extension "%pO" for device tree' > 90% of the usage of device node's full_name is printing it out > in a kernel message. Preparing for the eventual delayed allocation > introduce a custom printk format specifier that is both more > compact and more pleasant to the eye. > > For instance typical use is: > pr_info("Frobbing node %s\n", node->full_name); > > Which can be written now as: > pr_info("Frobbing node %pOF\n", node); Somehow I think this example is poor as node->full_name is a pretty obvious to read use. %pOF requires you to look up or know what the output is going to be. > More fine-grained control of formatting includes printing the name, > flag, path-spec name, reference count and others, explained in the > documentation entry. > > Originally written by Pantelis, but pretty much rewrote the core > function using existing string/number functions. The 2 passes were > unnecessary and have been removed. Also, updated the checkpatch.pl > check. Some comments about the code. > diff --git a/lib/vsprintf.c b/lib/vsprintf.c > [] > @@ -1470,6 +1471,123 @@ char *flags_string(char *buf, char *end, void *flags_ptr, const char *fmt) > return format_flags(buf, end, flags, names); > } > > +static noinline_for_stack > +char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) > +{ > + int len, ret; > + > + if (!np || !np->parent) > + return buf; > + > + buf = device_node_gen_full_name(np->parent, buf, end); This is recursive. How many levels of parents could there be? Perhaps there should be a recursion limit. > + > + if (buf < end) > + len = end - buf; > + else > + len = 0; > + ret = snprintf(buf, len, "/%s", kbasename(np->full_name)); > + if (ret <= 0) > + return buf; > + else if (len == 0 || ret < len) > + return buf + ret; > + return buf + len; > +} Does this work with %p<len>OF for a right justified or padded length string? Perhaps widen_string should be added. > +static noinline_for_stack > +char *device_node_string(char *buf, char *end, struct device_node *dn, > + struct printf_spec spec, const char *fmt) > +{ > + char tbuf[sizeof("xxxxxxxxxx") + 1]; > + const char *fmtp, *p; > + int ret; > + char *buf_start = buf; > + struct property *prop; > + bool has_mult, pass; > + const struct printf_spec num_spec = { > + .flags = SMALL, > + .field_width = -1, > + .precision = -1, > + .base = 10, > + }; > + > + struct printf_spec str_spec = spec; > + str_spec.field_width = -1; > + > + if (!IS_ENABLED(CONFIG_OF)) > + return string(buf, end, "(!OF)", spec); > + > + if ((unsigned long)dn < PAGE_SIZE) > + return string(buf, end, "(null)", spec); > + > + /* simple case without anything any more format specifiers */ > + if (fmt[1] == '\0' || strcspn(fmt + 1,"fnpPFcCr") > 0) > + fmt = "Ff"; > + > + for (fmtp = fmt + 1, pass = false; strspn(fmtp,"fnpPFcCr"); fmtp++, pass = true) { why not while (isalpha(*++fmt)) like ip6 or isalnum like FORMAT_TYPE_PTR uses? > + if (pass && (*fmtp != 'f')) { > + if (buf < end) > + *buf = '|'; > + buf++; > + } > + > + switch (*fmtp) { > + case 'f': /* full_name */ > + if (pass) { > + if (buf < end) > + *buf = ':'; > + buf++; > + } > + buf = device_node_gen_full_name(dn, buf, end); > + break; > + case 'n': /* name */ > + buf = string(buf, end, dn->name, str_spec); > + break; > + case 'p': /* phandle */ > + buf = number(buf, end, (unsigned int)dn->phandle, num_spec); > + break; > + case 'P': /* path-spec */ > + buf = string(buf, end, kbasename(of_node_full_name(dn)), str_spec); > + break; > + case 'F': /* flags */ > + snprintf(tbuf, sizeof(tbuf), "%c%c%c%c", > + of_node_check_flag(dn, OF_DYNAMIC) ? > + 'D' : '-', > + of_node_check_flag(dn, OF_DETACHED) ? > + 'd' : '-', > + of_node_check_flag(dn, OF_POPULATED) ? > + 'P' : '-', > + of_node_check_flag(dn, > + OF_POPULATED_BUS) ? 'B' : '-'); I'd try to avoid all uses of snprintf as it's effectively another fairly large stack frame. It's probably better to avoid more recursion stack depth use and just use *buf++ as appropriate. -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, 2017-06-14 at 15:30 -0500, Rob Herring wrote: > Now that we have a custom printf format specifier, convert users of > full_name to use %pOF instead. This is preparation to remove storing > of the full path string for each node. Well, a full_name removal makes my reply in 2/4 less sensible. Maybe 2/4 should say that full_name is going away. -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Jun 14, 2017 at 3:56 PM, Joe Perches <joe@perches.com> wrote: > On Wed, 2017-06-14 at 15:30 -0500, Rob Herring wrote: >> From: Pantelis Antoniou <pantelis.antoniou@konsulko.com> > > I think the commit subject is wrong. > It adds an "of" specific bit to vsprintf.c. > The subject should be > 'vsprintf: Add %p extension "%pO" for device tree' Okay, but it was good enough for the 2-3 versions Pantelis did before... >> 90% of the usage of device node's full_name is printing it out >> in a kernel message. Preparing for the eventual delayed allocation >> introduce a custom printk format specifier that is both more >> compact and more pleasant to the eye. >> >> For instance typical use is: >> pr_info("Frobbing node %s\n", node->full_name); >> >> Which can be written now as: >> pr_info("Frobbing node %pOF\n", node); > > Somehow I think this example is poor as node->full_name > is a pretty obvious to read use. %pOF requires you to > look up or know what the output is going to be. So %pOFfullname? We've beat this one to death IMO. > >> More fine-grained control of formatting includes printing the name, >> flag, path-spec name, reference count and others, explained in the >> documentation entry. >> >> Originally written by Pantelis, but pretty much rewrote the core >> function using existing string/number functions. The 2 passes were >> unnecessary and have been removed. Also, updated the checkpatch.pl >> check. > > Some comments about the code. > >> diff --git a/lib/vsprintf.c b/lib/vsprintf.c >> [] >> @@ -1470,6 +1471,123 @@ char *flags_string(char *buf, char *end, void *flags_ptr, const char *fmt) >> return format_flags(buf, end, flags, names); >> } >> >> +static noinline_for_stack >> +char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) >> +{ >> + int len, ret; >> + >> + if (!np || !np->parent) >> + return buf; >> + >> + buf = device_node_gen_full_name(np->parent, buf, end); > > This is recursive. How many levels of parents could there be? > Perhaps there should be a recursion limit. 2-6 I'd say is typical. The FDT unflattening code limits things to 64 (which is probably way more than needed). I could re-write it to be non-recursive, but then I'll just have the max sized array of pointers on the stack. > >> + >> + if (buf < end) >> + len = end - buf; >> + else >> + len = 0; >> + ret = snprintf(buf, len, "/%s", kbasename(np->full_name)); I can replace this one too with a strcat and save some stack space. >> + if (ret <= 0) >> + return buf; >> + else if (len == 0 || ret < len) >> + return buf + ret; >> + return buf + len; >> +} > > Does this work with %p<len>OF for a right justified or padded > length string? Perhaps widen_string should be added. widen_string is called at the end of device_node_string. >> +static noinline_for_stack >> +char *device_node_string(char *buf, char *end, struct device_node *dn, >> + struct printf_spec spec, const char *fmt) >> +{ >> + char tbuf[sizeof("xxxxxxxxxx") + 1]; >> + const char *fmtp, *p; >> + int ret; >> + char *buf_start = buf; >> + struct property *prop; >> + bool has_mult, pass; >> + const struct printf_spec num_spec = { >> + .flags = SMALL, >> + .field_width = -1, >> + .precision = -1, >> + .base = 10, >> + }; >> + >> + struct printf_spec str_spec = spec; >> + str_spec.field_width = -1; >> + >> + if (!IS_ENABLED(CONFIG_OF)) >> + return string(buf, end, "(!OF)", spec); >> + >> + if ((unsigned long)dn < PAGE_SIZE) >> + return string(buf, end, "(null)", spec); >> + >> + /* simple case without anything any more format specifiers */ >> + if (fmt[1] == '\0' || strcspn(fmt + 1,"fnpPFcCr") > 0) >> + fmt = "Ff"; >> + >> + for (fmtp = fmt + 1, pass = false; strspn(fmtp,"fnpPFcCr"); fmtp++, pass = true) { > > why not > while (isalpha(*++fmt)) > like ip6 or isalnum like FORMAT_TYPE_PTR uses? Okay. > >> + if (pass && (*fmtp != 'f')) { >> + if (buf < end) >> + *buf = '|'; >> + buf++; >> + } >> + >> + switch (*fmtp) { >> + case 'f': /* full_name */ >> + if (pass) { >> + if (buf < end) >> + *buf = ':'; >> + buf++; >> + } >> + buf = device_node_gen_full_name(dn, buf, end); >> + break; >> + case 'n': /* name */ >> + buf = string(buf, end, dn->name, str_spec); >> + break; >> + case 'p': /* phandle */ >> + buf = number(buf, end, (unsigned int)dn->phandle, num_spec); >> + break; >> + case 'P': /* path-spec */ >> + buf = string(buf, end, kbasename(of_node_full_name(dn)), str_spec); >> + break; >> + case 'F': /* flags */ >> + snprintf(tbuf, sizeof(tbuf), "%c%c%c%c", >> + of_node_check_flag(dn, OF_DYNAMIC) ? >> + 'D' : '-', >> + of_node_check_flag(dn, OF_DETACHED) ? >> + 'd' : '-', >> + of_node_check_flag(dn, OF_POPULATED) ? >> + 'P' : '-', >> + of_node_check_flag(dn, >> + OF_POPULATED_BUS) ? 'B' : '-'); > > I'd try to avoid all uses of snprintf as it's effectively > another fairly > large stack frame. Okay. > It's probably better to avoid more recursion stack depth use > and just use *buf++ as appropriate. You can't use *buf++ as this code must work and increment buf even when buf is NULL. Rob -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, 2017-06-15 at 07:30 -0500, Rob Herring wrote: > On Wed, Jun 14, 2017 at 3:56 PM, Joe Perches <joe@perches.com> wrote: > > On Wed, 2017-06-14 at 15:30 -0500, Rob Herring wrote: > > > From: Pantelis Antoniou <pantelis.antoniou@konsulko.com> > > > > I think the commit subject is wrong. > > It adds an "of" specific bit to vsprintf.c. > > The subject should be > > 'vsprintf: Add %p extension "%pO" for device tree' > > Okay, but it was good enough for the 2-3 versions Pantelis did before... Which were not applied. > > > 90% of the usage of device node's full_name is printing it out > > > in a kernel message. Preparing for the eventual delayed allocation The "eventual delayed allocation" bit doesn't mean anything to me. > > > introduce a custom printk format specifier that is both more > > > compact and more pleasant to the eye. > > > > > > For instance typical use is: > > > pr_info("Frobbing node %s\n", node->full_name); > > > > > > Which can be written now as: > > > pr_info("Frobbing node %pOF\n", node); > > > > Somehow I think this example is poor as node->full_name > > is a pretty obvious to read use. %pOF requires you to > > look up or know what the output is going to be. > > So %pOFfullname? We've beat this one to death IMO. I don't doubt the utility, just the example. Just mention that full_name is going away. > > > More fine-grained control of formatting includes printing the name, > > > flag, path-spec name, reference count and others, explained in the > > > documentation entry. > > > > > > Originally written by Pantelis, but pretty much rewrote the core > > > function using existing string/number functions. The 2 passes were > > > unnecessary and have been removed. Also, updated the checkpatch.pl > > > check. > > > > Some comments about the code. > > > > > diff --git a/lib/vsprintf.c b/lib/vsprintf.c > > > [] > > > @@ -1470,6 +1471,123 @@ char *flags_string(char *buf, char *end, void *flags_ptr, const char *fmt) > > > return format_flags(buf, end, flags, names); > > > } > > > > > > +static noinline_for_stack > > > +char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) > > > +{ > > > + int len, ret; > > > + > > > + if (!np || !np->parent) > > > + return buf; > > > + > > > + buf = device_node_gen_full_name(np->parent, buf, end); > > > > This is recursive. How many levels of parents could there be? > > Perhaps there should be a recursion limit. > > 2-6 I'd say is typical. The FDT unflattening code limits things to 64 > (which is probably way more than needed). > > I could re-write it to be non-recursive, but then I'll just have the > max sized array of pointers on the stack. Which would be less stack than how many recursive calls? 5? In any case, 64 * 8 for pointers or 5+ stack frames is a fair amount of stack. Maybe too much. > > > + case 'F': /* flags */ > > > + snprintf(tbuf, sizeof(tbuf), "%c%c%c%c", > > > + of_node_check_flag(dn, OF_DYNAMIC) ? > > > + 'D' : '-', > > > + of_node_check_flag(dn, OF_DETACHED) ? > > > + 'd' : '-', > > > + of_node_check_flag(dn, OF_POPULATED) ? > > > + 'P' : '-', > > > + of_node_check_flag(dn, > > > + OF_POPULATED_BUS) ? 'B' : '-'); > > > > I'd try to avoid all uses of snprintf as it's effectively > > another fairly large stack frame. > > Okay. > > > It's probably better to avoid more recursion stack depth use > > and just use *buf++ as appropriate. > > You can't use *buf++ as this code must work and increment buf even > when buf is NULL. tbuf then. -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Jun 14, 2017 at 01:56:48PM -0700, Joe Perches wrote: > On Wed, 2017-06-14 at 15:30 -0500, Rob Herring wrote: > > From: Pantelis Antoniou <pantelis.antoniou@konsulko.com> > > I think the commit subject is wrong. > It adds an "of" specific bit to vsprintf.c. > The subject should be > 'vsprintf: Add %p extension "%pO" for device tree' > > > 90% of the usage of device node's full_name is printing it out > > in a kernel message. Preparing for the eventual delayed allocation > > introduce a custom printk format specifier that is both more > > compact and more pleasant to the eye. > > > > For instance typical use is: > > pr_info("Frobbing node %s\n", node->full_name); > > > > Which can be written now as: > > pr_info("Frobbing node %pOF\n", node); > > Somehow I think this example is poor as node->full_name > is a pretty obvious to read use. %pOF requires you to > look up or know what the output is going to be. > > > More fine-grained control of formatting includes printing the name, > > flag, path-spec name, reference count and others, explained in the > > documentation entry. > > > > Originally written by Pantelis, but pretty much rewrote the core > > function using existing string/number functions. The 2 passes were > > unnecessary and have been removed. Also, updated the checkpatch.pl > > check. > > Some comments about the code. > > > diff --git a/lib/vsprintf.c b/lib/vsprintf.c > > [] > > @@ -1470,6 +1471,123 @@ char *flags_string(char *buf, char *end, void *flags_ptr, const char *fmt) > > return format_flags(buf, end, flags, names); > > } > > > > +static noinline_for_stack > > +char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) > > +{ > > + int len, ret; > > + > > + if (!np || !np->parent) > > + return buf; > > + > > + buf = device_node_gen_full_name(np->parent, buf, end); > > This is recursive. How many levels of parents could there be? > Perhaps there should be a recursion limit. Okay, unlike unflattening code, we can easily calculate the depth and then allocate an array on the stack. So this is what I've ended up with: static int device_node_calc_depth(const struct device_node *np) { int d; for (d = 0; np; d++) np = np->parent; return d; } static noinline_for_stack char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) { int i; int depth = device_node_calc_depth(np); const struct device_node *nodes[depth]; const struct printf_spec strspec = { .field_width = -1, .precision = -1, }; if (!depth) return buf; /* special case for root node */ if (depth == 1) return string(buf, end, "/", strspec); depth--; for (i = depth - 1; i >= 0; i--) { nodes[i] = np; np = np->parent; } for (i = 0; i < depth; i++) { buf = string(buf, end, "/", strspec); buf = string(buf, end, kbasename(nodes[i]->full_name), strspec); } return buf; } > > + /* simple case without anything any more format specifiers */ > > + if (fmt[1] == '\0' || strcspn(fmt + 1,"fnpPFcCr") > 0) > > + fmt = "Ff"; > > + > > + for (fmtp = fmt + 1, pass = false; strspn(fmtp,"fnpPFcCr"); fmtp++, pass = true) { > > why not > while (isalpha(*++fmt)) > like ip6 or isalnum like FORMAT_TYPE_PTR uses? This case is more complicated with the field separators. If we have something like %pOFfxyz where xyz are not valid format specifiers, we'd end up with extra field separators. I did simplify things a bit and got rid of fmtp and just use fmt instead. Rob -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, 2017-06-15 at 16:26 -0500, Rob Herring wrote: > On Wed, Jun 14, 2017 at 01:56:48PM -0700, Joe Perches wrote: > > On Wed, 2017-06-14 at 15:30 -0500, Rob Herring wrote: > > > From: Pantelis Antoniou <pantelis.antoniou@konsulko.com> [] > > > diff --git a/lib/vsprintf.c b/lib/vsprintf.c > > > [] > > > @@ -1470,6 +1471,123 @@ char *flags_string(char *buf, char *end, void *flags_ptr, const char *fmt) > > > return format_flags(buf, end, flags, names); > > > } > > > > > > +static noinline_for_stack > > > +char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) > > > +{ > > > + int len, ret; > > > + > > > + if (!np || !np->parent) > > > + return buf; > > > + > > > + buf = device_node_gen_full_name(np->parent, buf, end); > > > > This is recursive. How many levels of parents could there be? > > Perhaps there should be a recursion limit. > > Okay, unlike unflattening code, we can easily calculate the depth and > then allocate an array on the stack. So this is what I've ended up > with: > > static int device_node_calc_depth(const struct device_node *np) > { > int d; > > for (d = 0; np; d++) > np = np->parent; > > return d; > } > > static noinline_for_stack > char *device_node_gen_full_name(const struct device_node *np, char *buf, char *end) > { > int i; > int depth = device_node_calc_depth(np); > const struct device_node *nodes[depth]; > const struct printf_spec strspec = { > .field_width = -1, > .precision = -1, > }; static const struct printf_spec strspec = { etc... and please move strspec above *nodes > > if (!depth) > return buf; > /* special case for root node */ > if (depth == 1) > return string(buf, end, "/", strspec); > > depth--; > for (i = depth - 1; i >= 0; i--) { > nodes[i] = np; > np = np->parent; > } > for (i = 0; i < depth; i++) { > buf = string(buf, end, "/", strspec); > buf = string(buf, end, kbasename(nodes[i]->full_name), strspec); > } > return buf; > } -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html