| Message ID | 1445984203-3288-1-git-send-email-linux@rasmusvillemoes.dk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 10/27/2015 05:16 PM, Rasmus Villemoes wrote: > I'm not familiar with this code, but path_info_show() (added in > 8270b86243658 "hpsa: add sysfs entry path_info to show box and > bay information") seems to be broken in multiple ways. > > First, there's > > 817 return snprintf(buf, output_len+1, "%s%s%s%s%s%s%s%s", > 818 path[0], path[1], path[2], path[3], > 819 path[4], path[5], path[6], path[7]); > > so hopefully output_len contains the combined length of the eight > strings. Otherwise, snprintf will stop copying to the output > buffer, but still end up reporting that combined length - which > in turn would result in user-space getting a bunch of useless nul > bytes (thankfully the upper sysfs layer seems to clear the output > buffer before passing it to the various ->show routines). But we have > > 767 output_len = snprintf(path[i], > 768 PATH_STRING_LEN, "[%d:%d:%d:%d] %20.20s ", > 769 h->scsi_host->host_no, > 770 hdev->bus, hdev->target, hdev->lun, > 771 scsi_device_type(hdev->devtype)); > > so output_len at best contains the length of the last string printed. > > Inside the loop, we then otherwise add to output_len. By magic, > we still have PATH_STRING_LEN available every time... This > wouldn't really be a problem if the bean-counting has been done > properly and each line actually does fit in 50 bytes, and maybe > it does, but I don't immediately see why. Suppose we end up > taking this branch: > > 802 output_len += snprintf(path[i] + output_len, > 803 PATH_STRING_LEN, > 804 "BOX: %hhu BAY: %hhu %s\n", > 805 box, bay, active); > > An optimistic estimate says this uses strlen("BOX: 1 BAY: 2 > Active\n") which is 21. Now add the 20 bytes guaranteed by the > %20.20s and then some for the rest of that format string, and > we're easily over 50 bytes. I don't think we can get over 100 > bytes even being pessimistic, so this just means we'll scribble > into the next path[i+1] and maybe get that overwritten later, > leading to some garbled output (in fact, since we'd overwrite the > previous string's 0-terminator, we could end up with one very > long string and then print various suffixes of that, leading to > much more than 400 bytes of output). Except of course when we're > filling path[7], where overrunning it means writing random stuff > to the kernel stack, which is usually a lot of fun. > > We can fix all of that and get rid of the 400 byte stack buffer by > simply writing directly to the given output buffer, which the upper > layer guarantees is at least PAGE_SIZE. s[c]nprintf doesn't care where > it is writing to, so this doesn't make the spin lock hold time any > longer. Using scnprintf ensures that output_len always represents the > number of bytes actually written to the buffer, so we'll report the > proper amount to the upper layer. > > Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk> > Thanks, I added this to my current patch set. This patch will be up with you as the author soon. -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Oct 28 2015, Don Brace <brace77070@gmail.com> wrote: > On 10/27/2015 05:16 PM, Rasmus Villemoes wrote: >> I'm not familiar with this code, but path_info_show() (added in >> 8270b86243658 "hpsa: add sysfs entry path_info to show box and >> bay information") seems to be broken in multiple ways. >> [snip] >> >> We can fix all of that and get rid of the 400 byte stack buffer by >> simply writing directly to the given output buffer, which the upper >> layer guarantees is at least PAGE_SIZE. s[c]nprintf doesn't care where >> it is writing to, so this doesn't make the spin lock hold time any >> longer. Using scnprintf ensures that output_len always represents the >> number of bytes actually written to the buffer, so we'll report the >> proper amount to the upper layer. >> >> Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk> >> > Thanks, I added this to my current patch set. This patch will be up > with you as the author soon. I see it in mainline now. May I ask why 6 out of 7 scnprintfs were changed (back) to snprintf? I don't think there's any functional difference as long as PAGE_SIZE is indeed sufficient, but mixing snprintf and scnprintf is pretty odd, and there's now a discrepancy between the commit log and the patch which wasn't in my original - I'd expect a "[use snprintf because xyz]" note added if the change was intentional. Rasmus -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
> -----Original Message----- > From: linux-scsi-owner@vger.kernel.org [mailto:linux-scsi- > owner@vger.kernel.org] On Behalf Of Rasmus Villemoes > Sent: Saturday, November 14, 2015 9:24 AM > To: Don Brace > Cc: Joe Handzik; James E.J. Bottomley; Kevin Barnett; Scott Teel; Tomas Henzl; > iss_storagedev@hp.com; dl Team ESD Storage Dev Support; linux- > scsi@vger.kernel.org; linux-kernel@vger.kernel.org > Subject: Re: [PATCH] scsi: hpsa: fix multiple issues in path_info_show > > On Wed, Oct 28 2015, Don Brace <brace77070@gmail.com> wrote: > > > On 10/27/2015 05:16 PM, Rasmus Villemoes wrote: > >> I'm not familiar with this code, but path_info_show() (added in > >> 8270b86243658 "hpsa: add sysfs entry path_info to show box and > >> bay information") seems to be broken in multiple ways. > >> > [snip] > >> > >> We can fix all of that and get rid of the 400 byte stack buffer by > >> simply writing directly to the given output buffer, which the upper > >> layer guarantees is at least PAGE_SIZE. s[c]nprintf doesn't care where > >> it is writing to, so this doesn't make the spin lock hold time any > >> longer. Using scnprintf ensures that output_len always represents the > >> number of bytes actually written to the buffer, so we'll report the > >> proper amount to the upper layer. > >> > >> Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk> > >> > > Thanks, I added this to my current patch set. This patch will be up > > with you as the author soon. > > I see it in mainline now. May I ask why 6 out of 7 scnprintfs were > changed (back) to snprintf? I don't think there's any functional > difference as long as PAGE_SIZE is indeed sufficient, but mixing > snprintf and scnprintf is pretty odd, and there's now a discrepancy > between the commit log and the patch which wasn't in my original - I'd > expect a "[use snprintf because xyz]" note added if the change was > intentional. > > Rasmus Unintentional. I'll upload a fix soon. > > -- > To unsubscribe from this list: send the line "unsubscribe linux-scsi" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/scsi/hpsa.c b/drivers/scsi/hpsa.c index 40669f8dd0df..b892ae90e292 100644 --- a/drivers/scsi/hpsa.c +++ b/drivers/scsi/hpsa.c @@ -726,7 +726,6 @@ static ssize_t host_show_hp_ssd_smart_path_enabled(struct device *dev, } #define MAX_PATHS 8 -#define PATH_STRING_LEN 50 static ssize_t path_info_show(struct device *dev, struct device_attribute *attr, char *buf) @@ -742,9 +741,7 @@ static ssize_t path_info_show(struct device *dev, u8 path_map_index = 0; char *active; unsigned char phys_connector[2]; - unsigned char path[MAX_PATHS][PATH_STRING_LEN]; - memset(path, 0, MAX_PATHS * PATH_STRING_LEN); sdev = to_scsi_device(dev); h = sdev_to_hba(sdev); spin_lock_irqsave(&h->devlock, flags); @@ -764,8 +761,9 @@ static ssize_t path_info_show(struct device *dev, else continue; - output_len = snprintf(path[i], - PATH_STRING_LEN, "[%d:%d:%d:%d] %20.20s ", + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, + "[%d:%d:%d:%d] %20.20s ", h->scsi_host->host_no, hdev->bus, hdev->target, hdev->lun, scsi_device_type(hdev->devtype)); @@ -773,9 +771,9 @@ static ssize_t path_info_show(struct device *dev, if (is_ext_target(h, hdev) || (hdev->devtype == TYPE_RAID) || is_logical_dev_addr_mode(hdev->scsi3addr)) { - output_len += snprintf(path[i] + output_len, - PATH_STRING_LEN, "%s\n", - active); + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, + "%s\n", active); continue; } @@ -787,36 +785,34 @@ static ssize_t path_info_show(struct device *dev, if (phys_connector[1] < '0') phys_connector[1] = '0'; if (hdev->phys_connector[i] > 0) - output_len += snprintf(path[i] + output_len, - PATH_STRING_LEN, + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, "PORT: %.2s ", phys_connector); if (hdev->devtype == TYPE_DISK && hdev->expose_state != HPSA_DO_NOT_EXPOSE) { if (box == 0 || box == 0xFF) { - output_len += snprintf(path[i] + output_len, - PATH_STRING_LEN, + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, "BAY: %hhu %s\n", bay, active); } else { - output_len += snprintf(path[i] + output_len, - PATH_STRING_LEN, + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, "BOX: %hhu BAY: %hhu %s\n", box, bay, active); } } else if (box != 0 && box != 0xFF) { - output_len += snprintf(path[i] + output_len, - PATH_STRING_LEN, "BOX: %hhu %s\n", + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, "BOX: %hhu %s\n", box, active); } else - output_len += snprintf(path[i] + output_len, - PATH_STRING_LEN, "%s\n", active); + output_len += scnprintf(buf + output_len, + PAGE_SIZE - output_len, "%s\n", active); } spin_unlock_irqrestore(&h->devlock, flags); - return snprintf(buf, output_len+1, "%s%s%s%s%s%s%s%s", - path[0], path[1], path[2], path[3], - path[4], path[5], path[6], path[7]); + return output_len; } static DEVICE_ATTR(raid_level, S_IRUGO, raid_level_show, NULL);
I'm not familiar with this code, but path_info_show() (added in 8270b86243658 "hpsa: add sysfs entry path_info to show box and bay information") seems to be broken in multiple ways. First, there's 817 return snprintf(buf, output_len+1, "%s%s%s%s%s%s%s%s", 818 path[0], path[1], path[2], path[3], 819 path[4], path[5], path[6], path[7]); so hopefully output_len contains the combined length of the eight strings. Otherwise, snprintf will stop copying to the output buffer, but still end up reporting that combined length - which in turn would result in user-space getting a bunch of useless nul bytes (thankfully the upper sysfs layer seems to clear the output buffer before passing it to the various ->show routines). But we have 767 output_len = snprintf(path[i], 768 PATH_STRING_LEN, "[%d:%d:%d:%d] %20.20s ", 769 h->scsi_host->host_no, 770 hdev->bus, hdev->target, hdev->lun, 771 scsi_device_type(hdev->devtype)); so output_len at best contains the length of the last string printed. Inside the loop, we then otherwise add to output_len. By magic, we still have PATH_STRING_LEN available every time... This wouldn't really be a problem if the bean-counting has been done properly and each line actually does fit in 50 bytes, and maybe it does, but I don't immediately see why. Suppose we end up taking this branch: 802 output_len += snprintf(path[i] + output_len, 803 PATH_STRING_LEN, 804 "BOX: %hhu BAY: %hhu %s\n", 805 box, bay, active); An optimistic estimate says this uses strlen("BOX: 1 BAY: 2 Active\n") which is 21. Now add the 20 bytes guaranteed by the %20.20s and then some for the rest of that format string, and we're easily over 50 bytes. I don't think we can get over 100 bytes even being pessimistic, so this just means we'll scribble into the next path[i+1] and maybe get that overwritten later, leading to some garbled output (in fact, since we'd overwrite the previous string's 0-terminator, we could end up with one very long string and then print various suffixes of that, leading to much more than 400 bytes of output). Except of course when we're filling path[7], where overrunning it means writing random stuff to the kernel stack, which is usually a lot of fun. We can fix all of that and get rid of the 400 byte stack buffer by simply writing directly to the given output buffer, which the upper layer guarantees is at least PAGE_SIZE. s[c]nprintf doesn't care where it is writing to, so this doesn't make the spin lock hold time any longer. Using scnprintf ensures that output_len always represents the number of bytes actually written to the buffer, so we'll report the proper amount to the upper layer. Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk> --- drivers/scsi/hpsa.c | 38 +++++++++++++++++--------------------- 1 file changed, 17 insertions(+), 21 deletions(-)